diff --git a/.openclaw/pmflow-routing-rules.json b/.openclaw/pmflow-routing-rules.json
new file mode 100644
index 00000000..c509e59f
--- /dev/null
+++ b/.openclaw/pmflow-routing-rules.json
@@ -0,0 +1,74 @@
+{
+  "version": "v2",
+  "updated_at": "2026-03-05T00:00:00Z",
+  "source_of_truth": "openclaw",
+  "description": "PMFlow 三引擎统一路由：所有主链路统一经 OpenClaw OpenAI-Compatible 网关。",
+  "rules": [
+    {
+      "name": "engine_low_execute_glm_flash",
+      "enabled": true,
+      "priority": 300,
+      "tier": "low",
+      "provider": "openai",
+      "model": "zai/glm-4-flash",
+      "match_any": [
+        "feishu_level_classification",
+        "feishu_signal_extraction",
+        "rag_intent_classification",
+        "classify_intent",
+        "feishu_humanized_reply",
+        "extract_context_insights"
+      ]
+    },
+    {
+      "name": "engine_high_reasoning_minimax",
+      "enabled": true,
+      "priority": 220,
+      "tier": "high",
+      "provider": "openai",
+      "model": "minimax/MiniMax-M2.5",
+      "match_any": [
+        "adaptive_rag_answer",
+        "feishu_project_prefill",
+        "feishu_task_orchestration",
+        "generate_task_breakdown",
+        "generate_charter_from_signal",
+        "decompose_tasks_from_charter",
+        "recommend_assignees_for_task",
+        "evaluate_model_requirement",
+        "calculate_alignment_score",
+        "detect_cognitive_conflicts",
+        "rag_db_project_semantic_retrieval",
+        "workload_optimize_prompt",
+        "analyze_cost_roi"
+      ]
+    },
+    {
+      "name": "engine_high_extended_reasoning",
+      "enabled": true,
+      "priority": 200,
+      "tier": "high",
+      "provider": "openai",
+      "model": "minimax/MiniMax-M2.5",
+      "match_any": [
+        "analyze_food_calories",
+        "feishu_new_project_brief",
+        "feishu_project_signal_classification",
+        "sstc_ai_breakdown",
+        "sstc_charter_localize_zh",
+        "sstc_charter_prefill",
+        "skill_score_update",
+        "test_engine_connection"
+      ]
+    },
+    {
+      "name": "engine_safety_fallback",
+      "enabled": true,
+      "priority": 50,
+      "tier": "safety",
+      "provider": "openai",
+      "model": "deepseek/deepseek-chat",
+      "match_all": true
+    }
+  ]
+}
diff --git a/ai.rtf b/ai.rtf
new file mode 100644
index 00000000..c58974db
--- /dev/null
+++ b/ai.rtf
@@ -0,0 +1,1237 @@
+{\rtf1\ansi\ansicpg1252\cocoartf2867
+\cocoatextscaling0\cocoaplatform0{\fonttbl\f0\fnil\fcharset0 HelveticaNeue;\f1\fnil\fcharset134 PingFangSC-Regular;\f2\fnil\fcharset0 HelveticaNeue-Bold;
+\f3\fnil\fcharset134 PingFangSC-Bold;\f4\fnil\fcharset0 .SFNS-Regular_wdth_opsz110000_GRAD_wght1F40000;\f5\fnil\fcharset134 PingFangSC-Medium;
+}
+{\colortbl;\red255\green255\blue255;\red26\green26\blue26;\red255\green255\blue255;\red0\green0\blue0;
+\red0\green0\blue0;\red0\green0\blue0;}
+{\*\expandedcolortbl;;\cssrgb\c13333\c13333\c13333;\cssrgb\c100000\c100000\c100000;\cssrgb\c0\c0\c0\c50196;
+\cssrgb\c0\c0\c0;\cssrgb\c0\c0\c0\c30196;}
+{\*\listtable{\list\listtemplateid1\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid1\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listname ;}\listid1}
+{\list\listtemplateid2\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid101\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{circle\}}{\leveltext\leveltemplateid102\'01\uc0\u9702 ;}{\levelnumbers;}\fi-360\li1440\lin1440 }{\listname ;}\listid2}
+{\list\listtemplateid3\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid201\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{circle\}}{\leveltext\leveltemplateid202\'01\uc0\u9702 ;}{\levelnumbers;}\fi-360\li1440\lin1440 }{\listname ;}\listid3}
+{\list\listtemplateid4\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid301\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{circle\}}{\leveltext\leveltemplateid302\'01\uc0\u9702 ;}{\levelnumbers;}\fi-360\li1440\lin1440 }{\listname ;}\listid4}
+{\list\listtemplateid5\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid401\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{circle\}}{\leveltext\leveltemplateid402\'01\uc0\u9702 ;}{\levelnumbers;}\fi-360\li1440\lin1440 }{\listname ;}\listid5}
+{\list\listtemplateid6\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid501\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{circle\}}{\leveltext\leveltemplateid502\'01\uc0\u9702 ;}{\levelnumbers;}\fi-360\li1440\lin1440 }{\listname ;}\listid6}
+{\list\listtemplateid7\listhybrid{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{disc\}}{\leveltext\leveltemplateid601\'01\uc0\u8226 ;}{\levelnumbers;}\fi-360\li720\lin720 }{\listname ;}\listid7}
+{\list\listtemplateid8\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid701\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listname ;}\listid8}
+{\list\listtemplateid9\listhybrid{\listlevel\levelnfc0\levelnfcn0\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{decimal\}}{\leveltext\leveltemplateid801\'01\'00;}{\levelnumbers\'01;}\fi-360\li720\lin720 }{\listname ;}\listid9}}
+{\*\listoverridetable{\listoverride\listid1\listoverridecount0\ls1}{\listoverride\listid2\listoverridecount0\ls2}{\listoverride\listid3\listoverridecount0\ls3}{\listoverride\listid4\listoverridecount0\ls4}{\listoverride\listid5\listoverridecount0\ls5}{\listoverride\listid6\listoverridecount0\ls6}{\listoverride\listid7\listoverridecount0\ls7}{\listoverride\listid8\listoverridecount0\ls8}{\listoverride\listid9\listoverridecount0\ls9}}
+\paperw11900\paperh16840\margl1440\margr1440\vieww11520\viewh8400\viewkind0
+\deftab720
+\pard\pardeftab720\partightenfactor0
+
+\f0\fs32 \cf2 \cb3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 AI 
+\f1 \'ca\'b1\'b4\'fa
+\f0  10 
+\f1 \'c8\'cb\'bc\'bc\'ca\'f5\'cd\'c5\'b6\'d3\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'c6\'bd\'cc\'a8\'a3\'ba
+\f0 AI 
+\f1 \'d4\'ad\'c9\'fa\'bc\'ab\'cb\'d9\'cf\'ec\'d3\'a6\'d0\'cd\'cd\'ea\'d5\'fb\'c2\'e4\'b5\'d8\'b7\'bd\'b0\'b8
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d5\'eb\'b6\'d4
+\f0  10 
+\f1 \'c8\'cb\'b1\'e2\'c6\'bd\'bb\'af\'bc\'bc\'ca\'f5\'cd\'c5\'b6\'d3\'a3\'ac
+\f0 AI 
+\f1 \'ca\'b1\'b4\'fa\'b5\'c4\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'ba\'cb\'d0\'c4\'d2\'aa\'b3\'b9\'b5\'d7\'cc\'f8\'b3\'f6\'b4\'ab\'cd\'b3\'a1\'b8\'b9\'dc\'bf\'d8\'d0\'cd\'b9\'a4\'be\'df\'a1\'b9\'b5\'c4\'c2\'df\'bc\'ad\'a3\'ac\'b4\'f2\'d4\'ec
+\f0  **
+\f1 \'a1\'b8\'cd\'c5\'b6\'d3\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb\'d6\'d0\'ca\'e0
+\f0  + AI 
+\f1 \'c8\'ab\'c1\'b4\'c2\'b7\'cc\'e1\'d0\'a7\'d2\'fd\'c7\'e6
+\f0  + 
+\f1 \'ca\'d0\'b3\'a1
+\f0  - 
+\f1 \'bc\'bc\'ca\'f5\'bc\'ab\'cb\'d9\'b1\'d5\'bb\'b7\'d4\'d8\'cc\'e5\'a1\'b9
+\f0 **\'97\'97 
+\f1 \'ba\'cb\'d0\'c4\'c4\'bf\'b1\'ea\'b2\'bb\'ca\'c7\'b9\'dc\'c8\'cb\'a1\'a2\'cc\'ee\'b1\'a8\'b1\'ed\'a3\'ac\'b6\'f8\'ca\'c7\'b3\'b9\'b5\'d7\'cf\'fb\'c3\'f0\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a2\'c1\'e3\'b6\'ee\'cd\'e2\'b8\'ba\'b5\'a3\'ca\'ca\'c5\'e4\'cd\'c5\'b6\'d3\'a1\'a2\'bc\'ab\'cb\'d9\'cf\'ec\'d3\'a6\'ca\'d0\'b3\'a1\'d3\'eb\'c4\'a3\'d0\'cd\'b5\'fc\'b4\'fa\'a1\'a2\'d3\'c3
+\f0  AI 
+\f1 \'cc\'e6\'b4\'fa
+\f0  80% 
+\f1 \'d6\'d8\'b8\'b4\'b9\'dc\'c0\'ed\'b6\'af\'d7\'f7\'a3\'ac\'c8\'c3
+\f0  10 
+\f1 \'c8\'cb\'d0\'a1\'cd\'c5\'b6\'d3\'ca\'cd\'b7\'c5\'b3\'f6\'d4\'b6\'b3\'ac\'b4\'ab\'cd\'b3\'c4\'a3\'ca\'bd\'b5\'c4\'bd\'bb\'b8\'b6\'d0\'a7\'c2\'ca\'d3\'eb\'ca\'d0\'b3\'a1\'d3\'a6\'b1\'e4\'c4\'dc\'c1\'a6\'a1\'a3
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d2\'bb\'a1\'a2\'cf\'c8\'b6\'a8\'b5\'d7\'b2\'e3\'ba\'cb\'d0\'c4\'d4\'ad\'d4\'f2\'a3\'a8
+\f0 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'ca\'ca\'c5\'e4
+\f0  AI 
+\f1 \'ca\'b1\'b4\'fa\'b5\'c4\'b8\'f9\'b1\'be\'c7\'b0\'cc\'e1\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'ba\'cb\'d0\'c4\'d3\'c5\'ca\'c6\'ca\'c7\'b1\'e2\'c6\'bd\'bb\'af\'a1\'a2\'be\'f6\'b2\'df\'c1\'b4\'c2\'b7\'b6\'cc\'a1\'a2\'ca\'d4\'b4\'ed\'c1\'e9\'bb\'ee\'a3\'ac\'be\'f8\'b6\'d4\'b2\'bb\'c4\'dc\'d5\'d5\'b0\'e1\'b4\'f3\'b3\'a7\'d6\'d8\'c1\'f7\'b3\'cc\'a1\'a2\'d6\'d8\'c9\'f3\'c5\'fa\'b5\'c4\'c4\'a3\'ca\'bd\'a3\'ac\'cb\'f9\'d3\'d0\'c6\'bd\'cc\'a8\'c9\'e8\'bc\'c6\'b1\'d8\'d0\'eb\'d1\'cf\'b8\'f1\'d7\'f1\'d1\'ad
+\f0  5 
+\f1 \'b8\'f6\'cc\'fa\'d4\'f2\'a3\'ba
+\f0 \cb1 \
+\
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls1\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b5\'a5\'d4\'b4\'d5\'e6\'c0\'ed\'d4\'ad\'d4\'f2
+\f1\b0 \'a3\'ba\'cb\'f9\'d3\'d0\'cf\'ee\'c4\'bf\'cf\'e0\'b9\'d8\'d0\'c5\'cf\'a2\'a3\'a8\'d0\'e8\'c7\'f3\'a1\'a2\'c4\'a3\'d0\'cd\'ca\'fd\'be\'dd\'a1\'a2\'be\'f6\'b2\'df\'a1\'a2\'bd\'f8\'b6\'c8\'a1\'a2\'ce\'c4\'b5\'b5\'a1\'a2\'b8\'b4\'c5\'cc\'a3\'a9\'ce\'a8\'d2\'bb\'c8\'eb\'bf\'da\'a1\'a2\'ce\'a8\'d2\'bb\'ca\'fd\'be\'dd\'d4\'b4\'a3\'ac\'b3\'b9\'b5\'d7\'b6\'c5\'be\'f8\'d0\'c5\'cf\'a2\'c9\'a2\'c2\'e4\'d4\'da\'ce\'a2\'d0\'c5\'a1\'a2\'b7\'c9\'ca\'e9\'a1\'a2\'c0\'eb\'cf\'df\'ce\'c4\'b5\'b5\'a1\'a2\'b4\'fa\'c2\'eb\'b2\'d6\'bf\'e2\'a1\'a2\'bb\'e1\'d2\'e9\'c2\'bc\'d2\'f4\'c0\'ef\'a3\'ac\'b4\'d3\'b8\'f9\'d4\'b4\'cf\'fb\'c3\'f0\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a3
+\f0 \cb1 \
+\ls1\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 AI 
+\f3 \'d4\'ad\'c9\'fa\'b6\'f8\'b7\'c7
+\f2  AI 
+\f3 \'b8\'bd\'bc\'d3\'d4\'ad\'d4\'f2
+\f1\b0 \'a3\'ba\'b2\'bb\'ca\'c7\'d4\'da\'b4\'ab\'cd\'b3\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'c0\'ef\'bc\'d3\'b8\'f6\'c1\'c4\'cc\'ec\'bf\'f2\'a3\'ac\'b6\'f8\'ca\'c7
+\f0  AI 
+\f1 \'b9\'e1\'b4\'a9\'a1\'b8\'d0\'e8\'c7\'f3\'ca\'e4\'c8\'eb
+\f0  - 
+\f1 \'c8\'ce\'ce\'f1\'b2\'f0\'bd\'e2
+\f0  - 
+\f1 \'d6\'b4\'d0\'d0\'b6\'d4\'c6\'eb
+\f0  - 
+\f1 \'b7\'e7\'cf\'d5\'d4\'a4\'be\'af
+\f0  - 
+\f1 \'b8\'b4\'c5\'cc\'b3\'c1\'b5\'ed\'a1\'b9\'c8\'ab\'c1\'b4\'c2\'b7\'a3\'ac\'ba\'cb\'d0\'c4\'ca\'c7\'bc\'f5\'c9\'d9\'cd\'c5\'b6\'d3\'b9\'a4\'d7\'f7\'c1\'bf\'a3\'ac\'b6\'f8\'b7\'c7\'d4\'f6\'bc\'d3\'cc\'ee\'b1\'a8\'b8\'ba\'b5\'a3\'a1\'a3
+\f0 \cb1 \
+\ls1\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'bc\'ab\'cb\'d9\'b1\'d5\'bb\'b7\'d4\'ad\'d4\'f2
+\f1\b0 \'a3\'ba\'b4\'f2\'cd\'a8\'a1\'b8\'ca\'d0\'b3\'a1\'b1\'e4\'bb\'af
+\f0  / 
+\f1 \'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2
+\f0 \uc0\u8594 
+\f1 \'d0\'e8\'c7\'f3\'c6\'c0\'b9\'c0
+\f0 \uc0\u8594 
+\f1 \'be\'f6\'b2\'df\'c2\'e4\'b5\'d8
+\f0 \uc0\u8594 
+\f1 \'d0\'a7\'b9\'fb\'b7\'b4\'c0\'a1\'a1\'b9\'c8\'ab\'c1\'b4\'c2\'b7\'a3\'ac\'c5\'d7\'c6\'fa\'b9\'cc\'b6\'a8\'b5\'fc\'b4\'fa\'b5\'c4\'ca\'f8\'b8\'bf\'a3\'ac\'ca\'b5\'cf\'d6\'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'b4\'d3\'c2\'bc\'c8\'eb\'b5\'bd\'c2\'e4\'b5\'d8\'b5\'c4\'d7\'ee\'b6\'cc\'c1\'b4\'c2\'b7\'a3\'ac\'ca\'ca\'c5\'e4
+\f0  AI 
+\f1 \'d0\'d0\'d2\'b5\'d6\'dc\'bc\'b6\'c9\'f5\'d6\'c1\'c8\'d5\'bc\'b6\'b5\'c4\'b5\'fc\'b4\'fa\'bd\'da\'d7\'e0\'a1\'a3
+\f0 \cb1 \
+\ls1\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b6\'d4\'c6\'eb\'d3\'c5\'cf\'c8\'a1\'a2\'b9\'dc\'bf\'d8\'ba\'f3\'d6\'c3\'d4\'ad\'d4\'f2
+\f1\b0 \'a3\'ba\'cb\'f9\'d3\'d0\'c1\'f7\'b3\'cc\'c9\'e8\'bc\'c6\'b5\'c4\'ba\'cb\'d0\'c4\'ca\'c7\'a1\'b8\'c8\'ab\'d4\'b1\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb\'a1\'b9\'a3\'ac\'b6\'f8\'b7\'c7\'a1\'b8\'bd\'f8\'b6\'c8\'b9\'dc\'bf\'d8\'a1\'b9\'a1\'a3
+\f0 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b2\'bb\'c5\'c2\'bd\'f8\'b6\'c8\'c2\'fd\'a3\'ac\'d7\'ee\'c5\'c2\'c8\'ab\'d4\'b1\'b6\'d4\'c4\'bf\'b1\'ea\'a1\'a2\'d0\'e8\'c7\'f3\'a1\'a2\'c4\'a3\'d0\'cd\'c4\'dc\'c1\'a6\'a1\'a2\'ca\'d0\'b3\'a1\'b1\'e4\'bb\'af\'b5\'c4\'c8\'cf\'d6\'aa\'b2\'bb\'d2\'bb\'d6\'c2\'a3\'ac\'b5\'bc\'d6\'c2\'b7\'b5\'b9\'a4\'a1\'a2\'c4\'da\'ba\'c4\'a1\'a2\'b7\'bd\'cf\'f2\'c5\'dc\'c6\'ab\'a1\'a3
+\f0 \cb1 \
+\ls1\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c1\'e3\'b8\'ba\'b5\'a3\'d4\'ad\'d4\'f2
+\f1\b0 \'a3\'ba\'c4\'dc\'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'b5\'c4\'be\'f8\'b2\'bb\'ca\'d6\'b6\'af\'cc\'ee\'d0\'b4\'a3\'ac\'c4\'dc\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'b5\'c4\'be\'f8\'b2\'bb\'d6\'d8\'b8\'b4\'c2\'bc\'c8\'eb\'a3\'ac\'c4\'dc\'d2\'ec\'b2\'bd\'cd\'ea\'b3\'c9\'b5\'c4\'be\'f8\'b2\'bb\'d5\'bc\'d3\'c3\'bf\'aa\'bb\'e1\'ca\'b1\'bc\'e4\'a1\'a3\'c6\'bd\'cc\'a8\'b1\'d8\'d0\'eb\'ca\'c7\'b8\'f8\'cd\'c5\'b6\'d3\'bc\'f5\'b8\'ba\'a3\'ac\'b6\'f8\'b7\'c7\'b3\'c9\'ce\'aa\'d0\'c2\'b5\'c4\'b9\'dc\'c0\'ed
+\f0  KPI
+\f1 \'a1\'a3
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b6\'fe\'a1\'a2\'c6\'bd\'cc\'a8\'ba\'cb\'d0\'c4\'bc\'dc\'b9\'b9\'d3\'eb\'c8\'ab\'c1\'b4\'c2\'b7\'b9\'a6\'c4\'dc\'c9\'e8\'bc\'c6
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d5\'fb\'cc\'e5\'bc\'dc\'b9\'b9\'b2\'c9\'d3\'c3\'a1\'b8\'d2\'bb\'b2\'e3\'c8\'eb\'bf\'da\'a1\'a2\'c1\'bd\'b4\'f3\'d2\'fd\'c7\'e6\'a1\'a2\'c8\'fd\'c0\'e0\'ca\'fd\'be\'dd\'a1\'a2
+\f0 N 
+\f1 \'b8\'f6\'bc\'af\'b3\'c9\'a1\'b9\'b5\'c4\'bc\'ab\'bc\'f2\'c9\'e8\'bc\'c6\'a3\'ac\'b1\'dc\'c3\'e2\'b8\'b4\'d4\'d3\'bc\'dc\'b9\'b9\'b4\'f8\'c0\'b4\'b5\'c4\'d4\'cb\'ce\'ac\'d3\'eb\'ca\'b9\'d3\'c3\'b8\'ba\'b5\'a3\'a3\'ac\'cd\'ea\'c3\'c0\'ca\'ca\'c5\'e4
+\f0  2 
+\f1 \'c8\'cb\'bf\'aa\'b7\'a2\'ce\'ac\'bb\'a4\'b5\'c4\'c8\'cb\'c1\'a6\'c5\'e4\'d6\'c3\'a1\'a3
+\f0 \cb1 \
+\
+
+\f1 \cb3 \'d2\'d1\'c9\'fa\'b3\'c9\'b4\'fa\'c2\'eb
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c4\'a3\'bf\'e9\'d2\'bb\'a3\'ba\'ca\'d0\'b3\'a1
+\f0  - 
+\f1 \'c4\'a3\'d0\'cd\'b1\'e4\'bb\'af\'bc\'ab\'cb\'d9\'cf\'ec\'d3\'a6\'d6\'d0\'ca\'e0\'a3\'a8\'bd\'e2\'be\'f6\'a1\'b8\'b8\'fa\'b2\'bb\'c9\'cf\'b1\'e4\'bb\'af\'a1\'a2\'cf\'ec\'d3\'a6\'c2\'fd\'a1\'b9\'b5\'c4\'ba\'cb\'d0\'c4\'cd\'b4\'b5\'e3\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \cb3 AI 
+\f1 \'ca\'b1\'b4\'fa\'cf\'ee\'c4\'bf\'b5\'c4\'d7\'ee\'b4\'f3\'b1\'e4\'c1\'bf\'a3\'ac\'d2\'bb\'ca\'c7\'ca\'d0\'b3\'a1\'d0\'e8\'c7\'f3
+\f0  / 
+\f1 \'be\'ba\'c6\'b7\'b6\'af\'d7\'f7\'b5\'c4\'bf\'ec\'cb\'d9\'b1\'e4\'bb\'af\'a3\'ac\'b6\'fe\'ca\'c7\'b4\'f3\'c4\'a3\'d0\'cd\'d6\'dc\'bc\'b6\'b5\'c4\'b0\'e6\'b1\'be\'b8\'fc\'d0\'c2\'b4\'f8\'c0\'b4\'b5\'c4\'c4\'dc\'c1\'a6\'d6\'d8\'b9\'b9\'a1\'a3\'b4\'ab\'cd\'b3\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'b4\'d3\'d0\'e8\'c7\'f3\'b7\'a2\'c6\'f0\'be\'cd\'c2\'fd\'c1\'cb\'a3\'ac\'d5\'e2\'b8\'f6\'c4\'a3\'bf\'e9\'d2\'aa\'d7\'f6\'b5\'c4\'ca\'c7\'b0\'d1\'cb\'f9\'d3\'d0\'d3\'b0\'cf\'ec\'cf\'ee\'c4\'bf\'b5\'c4\'a1\'b8\'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'a1\'b9\'a3\'ac\'b1\'e4\'b3\'c9\'bf\'c9\'c2\'e4\'b5\'d8\'b5\'c4\'d6\'b4\'d0\'d0\'b6\'af\'d7\'f7\'a3\'ac\'ca\'b5\'cf\'d6\'d7\'ee\'bf\'ec\'cf\'ec\'d3\'a6\'a1\'a3
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'ba\'cb\'d0\'c4\'b9\'a6\'c4\'dc\'c9\'e8\'bc\'c6
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls2\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'ca\'b5\'ca\'b1\'b9\'e9\'bc\'af\'b3\'d8\'a3\'a8\'c8\'ab\'d7\'d4\'b6\'af\'bb\'af\'a3\'ac\'ce\'de\'d0\'e8\'ca\'d6\'b6\'af\'ce\'ac\'bb\'a4\'a3\'a9
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b4\'ee\'bd\'a8\'ce\'a8\'d2\'bb\'b5\'c4\'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'c8\'eb\'bf\'da\'a3\'ac\'d7\'d4\'b6\'af\'d7\'a5\'c8\'a1
+\f0  + 
+\f1 \'c7\'e1\'c1\'bf\'bb\'af\'ca\'d6\'b6\'af\'c2\'bc\'c8\'eb\'a3\'ac\'b7\'d6\'c0\'e0\'b9\'dc\'c0\'ed\'cb\'f9\'d3\'d0\'d3\'b0\'cf\'ec\'cf\'ee\'c4\'bf\'b5\'c4\'b1\'e4\'c1\'bf\'a3\'ac\'c8\'ab\'d4\'b1\'bf\'c9\'bc\'fb\'a3\'ac\'b6\'c5\'be\'f8\'a1\'b8\'d6\'bb\'d3\'d0\'b2\'fa\'c6\'b7
+\f0  / 
+\f1 \'b8\'ba\'d4\'f0\'c8\'cb\'d6\'aa\'b5\'c0\'b1\'e4\'bb\'af\'a3\'ac\'d6\'b4\'d0\'d0\'b2\'e3\'d2\'bb\'ce\'de\'cb\'f9\'d6\'aa\'a1\'b9\'b5\'c4\'c8\'cf\'d6\'aa\'b2\'ee\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls2\ilvl1
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'d0\'c5\'ba\'c5
+\f1\b0 \'a3\'ba\'d7\'d4\'b6\'af\'bc\'e0\'bf\'d8\'d6\'f7\'c1\'f7\'bf\'aa\'d4\'b4
+\f0  / 
+\f1 \'b1\'d5\'d4\'b4\'b4\'f3\'c4\'a3\'d0\'cd\'a3\'a8
+\f0 Llama3
+\f1 \'a1\'a2
+\f0 Qwen2
+\f1 \'a1\'a2\'cd\'a8\'d2\'e5\'c7\'a7\'ce\'ca\'a1\'a2
+\f0 GPT 
+\f1 \'cf\'b5\'c1\'d0\'b5\'c8\'a3\'a9\'b5\'c4\'b0\'e6\'b1\'be\'b8\'fc\'d0\'c2\'a1\'a2
+\f0 API 
+\f1 \'b1\'e4\'b8\'fc\'a1\'a2\'c4\'dc\'c1\'a6\'c9\'fd\'bc\'b6
+\f0  / 
+\f1 \'bd\'b5\'bc\'b6\'a1\'a2\'bf\'aa\'d4\'b4\'d0\'c2\'c4\'a3\'d0\'cd\'b7\'a2\'b2\'bc\'a3\'ac\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'b5\'bd\'c6\'bd\'cc\'a8\'a3\'ac
+\f0 AI 
+\f1 \'cd\'ac\'b2\'bd\'c9\'fa\'b3\'c9\'a1\'b8\'b6\'d4\'cf\'d6\'d3\'d0\'cf\'ee\'c4\'bf\'b5\'c4\'d3\'b0\'cf\'ec\'b7\'d6\'ce\'f6\'a1\'b9\'a3\'a8\'c0\'fd\'a3\'ba
+\f0 Qwen2-7B 
+\f1 \'d0\'c2\'b0\'e6\'b1\'be\'d6\'a7\'b3\'d6
+\f0  128K 
+\f1 \'ce\'de\'cb\'f0\'c9\'cf\'cf\'c2\'ce\'c4\'a3\'ac\'ce\'d2\'c3\'c7\'b5\'c4\'b3\'a4\'ce\'c4\'b5\'b5\'b4\'a6\'c0\'ed\'b9\'a6\'c4\'dc\'bf\'c9\'d6\'d8\'b9\'b9\'a3\'ac\'bf\'aa\'b7\'a2\'d6\'dc\'c6\'da\'b4\'d3
+\f0  2 
+\f1 \'d6\'dc\'cb\'f5\'b6\'cc\'d6\'c1
+\f0  3 
+\f1 \'cc\'ec\'a3\'a9\'a1\'a3
+\f0 \cb1 \
+\ls2\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ca\'d0\'b3\'a1
+\f2  & 
+\f3 \'be\'ba\'c6\'b7\'d0\'c5\'ba\'c5
+\f1\b0 \'a3\'ba\'c7\'e1\'c1\'bf\'bb\'af\'c5\'c0\'b3\'e6\'d7\'a5\'c8\'a1
+\f0  + 
+\f1 \'c8\'ab\'d4\'b1\'ca\'d6\'b6\'af\'c2\'bc\'c8\'eb\'a3\'ac\'b8\'b2\'b8\'c7\'be\'ba\'c6\'b7\'d0\'c2\'b9\'a6\'c4\'dc\'a1\'a2\'bf\'cd\'bb\'a7\'ba\'cb\'d0\'c4\'b7\'b4\'c0\'a1\'a1\'a2\'d0\'d0\'d2\'b5\'d5\'fe\'b2\'df\'b1\'e4\'bb\'af\'a1\'a2\'ca\'d0\'b3\'a1\'d0\'c2\'bb\'fa\'bb\'e1\'a3\'ac\'c3\'bf\'cc\'f5\'d0\'c5\'ba\'c5\'bf\'c9\'b1\'ea\'bc\'c7\'d3\'c5\'cf\'c8\'bc\'b6\'a1\'a2\'b9\'d8\'c1\'aa\'cf\'ee\'c4\'bf\'a1\'a2\'b8\'ba\'d4\'f0\'c8\'cb\'a3\'ac\'c8\'ab\'d4\'b1\'bf\'c9\'c6\'c0\'c2\'db\'b2\'b9\'b3\'e4\'a1\'a3
+\f0 \cb1 \
+\ls2\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c4\'da\'b2\'bf\'bc\'bc\'ca\'f5\'d0\'c5\'ba\'c5
+\f1\b0 \'a3\'ba\'ba\'cd\'cb\'e3\'b7\'a8\'d4\'a4\'d1\'d0\'a1\'a2\'c4\'a3\'d0\'cd\'ce\'a2\'b5\'f7\'c6\'bd\'cc\'a8\'b4\'f2\'cd\'a8\'a3\'ac\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'cd\'c5\'b6\'d3\'c4\'da\'b2\'bf\'b5\'c4\'bc\'bc\'ca\'f5\'cd\'bb\'c6\'c6\'a1\'a2\'c4\'a3\'d0\'cd\'d3\'c5\'bb\'af\'b3\'c9\'b9\'fb\'a1\'a2\'b2\'c8\'bf\'d3\'be\'ad\'d1\'e9\'a3\'a8\'c0\'fd\'a3\'ba\'ce\'d2\'c3\'c7\'ce\'a2\'b5\'f7\'b5\'c4\'b4\'b9\'d6\'b1\'c4\'a3\'d0\'cd\'d7\'bc\'c8\'b7\'c2\'ca\'cc\'e1\'c9\'fd
+\f0  15%
+\f1 \'a3\'ac\'bf\'c9\'c2\'e4\'b5\'d8\'b5\'bd
+\f0  XX 
+\f1 \'ba\'cb\'d0\'c4\'b9\'a6\'c4\'dc\'a3\'a9\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls2\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls2\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 15 
+\f3 \'b7\'d6\'d6\'d3\'bc\'ab\'cb\'d9\'be\'f6\'b2\'df\'c1\'f7\'a3\'a8\'ca\'ca\'c5\'e4
+\f2  10 
+\f3 \'c8\'cb\'b1\'e2\'c6\'bd\'bb\'af\'cd\'c5\'b6\'d3\'a3\'ac\'ce\'de\'c8\'df\'d3\'e0\'c9\'f3\'c5\'fa\'a3\'a9
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b3\'b9\'b5\'d7\'c5\'d7\'c6\'fa\'a1\'b8\'d0\'e8\'c7\'f3\'c6\'c0\'c9\'f3
+\f0 \uc0\u8594 
+\f1 \'b2\'e3\'b2\'e3\'c9\'f3\'c5\'fa
+\f0 \uc0\u8594 
+\f1 \'c5\'c5\'c6\'da\'c2\'e4\'b5\'d8\'a1\'b9\'b5\'c4\'b3\'a4\'c1\'b4\'c2\'b7\'a3\'ac\'c9\'e8\'bc\'c6\'bc\'ab\'bc\'f2\'be\'f6\'b2\'df\'bb\'fa\'d6\'c6\'a3\'ac\'c8\'b7\'b1\'a3\'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5
+\f0  24 
+\f1 \'d0\'a1\'ca\'b1\'c4\'da\'cd\'ea\'b3\'c9\'c6\'c0\'b9\'c0\'d3\'eb\'be\'f6\'b2\'df\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls2\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'d0\'c5\'ba\'c5\'c2\'bc\'c8\'eb\'ba\'f3\'a3\'ac
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'cd\'ea\'b3\'c9\'b3\'f5\'c6\'c0\'a3\'ba\'b7\'d6\'ce\'f6\'d3\'b0\'cf\'ec\'b7\'b6\'ce\'a7\'a1\'a2\'c2\'e4\'b5\'d8\'ca\'d5\'d2\'e6\'a1\'a2\'bf\'aa\'b7\'a2\'b3\'c9\'b1\'be\'a1\'a2\'c7\'b1\'d4\'da\'b7\'e7\'cf\'d5\'a3\'ac\'c9\'fa\'b3\'c9
+\f0  3 
+\f1 \'be\'e4\'bb\'b0\'c6\'c0\'b9\'c0\'bd\'e1\'c2\'db\'a3\'ac\'b8\'f8\'b3\'f6\'d3\'c5\'cf\'c8\'bc\'b6\'bd\'a8\'d2\'e9\'a3\'ac\'ce\'de\'d0\'e8\'b2\'fa\'c6\'b7\'be\'ad\'c0\'ed\'ca\'d6\'b6\'af\'d0\'b4\'b3\'a4\'c6\'aa\'d0\'e8\'c7\'f3\'ce\'c4\'b5\'b5\'a1\'a3
+\f0 \cb1 \
+\ls2\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b8\'ba\'d4\'f0\'c8\'cb\'d2\'bb\'bc\'fc\'b7\'a2\'c6\'f0\'a1\'b8\'bc\'ab\'cb\'d9\'b6\'d4\'c6\'eb\'a1\'b9\'a3\'ac\'cf\'e0\'b9\'d8\'c8\'cb\'a3\'a8\'b2\'fa\'c6\'b7\'a1\'a2\'bc\'bc\'ca\'f5\'a1\'a2\'cb\'e3\'b7\'a8\'ba\'cb\'d0\'c4\'c8\'cb\'d4\'b1\'a3\'a9\'d6\'b1\'bd\'d3\'d4\'da\'c6\'bd\'cc\'a8\'c4\'da\'d2\'ec\'b2\'bd\'c6\'c0\'c2\'db\'a1\'a2\'cd\'b6\'c6\'b1\'a1\'a2\'c8\'b7\'c8\'cf\'be\'f6\'b2\'df\'a3\'ac\'ce\'de\'d0\'e8\'b4\'d5\'ca\'b1\'bc\'e4\'bf\'aa\'bb\'e1\'a3\'bb\'d6\'d8\'b4\'f3\'b1\'e4\'bb\'af\'d6\'b1\'bd\'d3\'b7\'a2\'c6\'f0
+\f0  15 
+\f1 \'b7\'d6\'d6\'d3\'cc\'da\'d1\'b6\'bb\'e1\'d2\'e9\'a3\'ac\'bb\'e1\'d2\'e9\'bc\'cd\'d2\'aa
+\f0  AI 
+\f1 \'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'a1\'a2\'cd\'ac\'b2\'bd\'b5\'bd\'c6\'bd\'cc\'a8\'a3\'ac\'be\'f6\'b2\'df\'bd\'e1\'b9\'fb\'d6\'b1\'bd\'d3\'d7\'aa\'ce\'aa\'d0\'e8\'c7\'f3
+\f0  / 
+\f1 \'c8\'ce\'ce\'f1\'a3\'ac\'d7\'d4\'b6\'af\'bd\'f8\'c8\'eb\'b5\'fc\'b4\'fa\'b3\'d8\'a3\'ac\'c8\'ab\'b3\'cc\'ce\'de\'bf\'e7\'b9\'a4\'be\'df\'b8\'b4\'d6\'c6\'d5\'b3\'cc\'f9\'a3\'ac\'ce\'de\'d0\'c5\'cf\'a2\'cb\'f0\'ba\'c4\'a1\'a3
+\f0 \cb1 \
+\ls2\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f6\'b2\'df\'c8\'ab\'c1\'b4\'c2\'b7\'bf\'c9\'d7\'b7\'cb\'dd\'a3\'ba\'cb\'ad\'cc\'e1\'b5\'c4\'d0\'c5\'ba\'c5\'a1\'a2\'cb\'ad\'d7\'f6\'b5\'c4\'c6\'c0\'b9\'c0\'a1\'a2\'be\'f6\'b2\'df\'c4\'da\'c8\'dd\'a1\'a2\'be\'f6\'b2\'df\'d4\'ad\'d2\'f2\'a1\'a2\'c2\'e4\'b5\'d8\'ca\'b1\'bc\'e4\'a3\'ac\'c8\'ab\'b2\'bf\'c1\'f4\'b4\'e6\'a3\'ac\'c8\'ab\'d4\'b1\'bf\'c9\'b2\'e9\'a3\'ac\'b3\'b9\'b5\'d7\'b6\'c5\'be\'f8\'a1\'b8\'b5\'b1\'ca\'b1\'cb\'ad\'cb\'b5\'b5\'c4\'a3\'bf\'a1\'b9\'a1\'b8\'ce\'aa\'ca\'b2\'c3\'b4\'d2\'aa\'d7\'f6\'d5\'e2\'b8\'f6\'d0\'e8\'c7\'f3\'a3\'bf\'a1\'b9\'b5\'c4\'b3\'b6\'c6\'a4\'d3\'eb\'c8\'cf\'d6\'aa\'c6\'ab\'b2\'ee\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls2\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls2\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b9\'f6\'b6\'af\'ca\'bd\'b6\'af\'cc\'ac\'b5\'fc\'b4\'fa\'b3\'d8\'a3\'a8\'c5\'d7\'c6\'fa\'b9\'cc\'b6\'a8\'b5\'fc\'b4\'fa\'a3\'ac\'ca\'ca\'c5\'e4\'bc\'ab\'cb\'d9\'b1\'e4\'bb\'af\'a3\'a9
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b4\'ab\'cd\'b3
+\f0  2 
+\f1 \'d6\'dc\'b9\'cc\'b6\'a8\'b5\'fc\'b4\'fa\'a3\'ac\'cd\'ea\'c8\'ab\'b8\'fa\'b2\'bb\'c9\'cf
+\f0  AI 
+\f1 \'d0\'d0\'d2\'b5\'b5\'c4\'b1\'e4\'bb\'af\'bd\'da\'d7\'e0\'a3\'ac\'b8\'c4\'ce\'aa\'a1\'b8\'d3\'c5\'cf\'c8\'bc\'b6\'c7\'fd\'b6\'af\'b5\'c4\'b9\'f6\'b6\'af\'ca\'bd\'b5\'fc\'b4\'fa\'a1\'b9\'a3\'ac\'bc\'c8\'b1\'a3\'d6\'a4\'ba\'cb\'d0\'c4\'c4\'bf\'b1\'ea\'b2\'bb\'c5\'dc\'c6\'ab\'a3\'ac\'d3\'d6\'c4\'dc\'c1\'e9\'bb\'ee\'cf\'ec\'d3\'a6\'b1\'e4\'bb\'af\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls2\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b5\'fc\'b4\'fa\'b3\'d8\'bd\'f6\'c9\'e8
+\f0  4 
+\f1 \'b8\'f6\'d3\'c5\'cf\'c8\'bc\'b6\'a3\'ac\'bc\'ab\'bc\'f2\'ce\'de\'b8\'ba\'b5\'a3\'a3\'ba
+\f0 P0
+\f1 \'a3\'a8\'b1\'be\'d6\'dc\'b1\'d8\'d0\'eb\'c9\'cf\'cf\'df\'a3\'a9\'a1\'a2
+\f0 P1
+\f1 \'a3\'a8\'b1\'be\'d6\'dc
+\f0  - 
+\f1 \'cf\'c2\'d6\'dc\'c9\'cf\'cf\'df\'a3\'a9\'a1\'a2
+\f0 P2
+\f1 \'a3\'a8
+\f0 2 
+\f1 \'d6\'dc\'c4\'da\'c2\'e4\'b5\'d8\'a3\'a9\'a1\'a2
+\f0 P3
+\f1 \'a3\'a8\'b4\'fd\'c6\'c0\'b9\'c0\'a3\'a9\'a3\'ac\'cb\'f9\'d3\'d0\'d0\'e8\'c7\'f3
+\f0  / 
+\f1 \'c8\'ce\'ce\'f1\'b1\'d8\'d0\'eb\'b9\'d8\'c1\'aa\'d3\'c5\'cf\'c8\'bc\'b6\'d3\'eb\'b1\'b1\'bc\'ab\'d0\'c7\'c4\'bf\'b1\'ea\'a1\'a3
+\f0 \cb1 \
+\ls2\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b6\'af\'cc\'ac\'c5\'c5\'c6\'da\'d7\'d4\'b6\'af\'b5\'f7\'d5\'fb\'a3\'ba
+\f0 AI 
+\f1 \'b8\'f9\'be\'dd\'d0\'e8\'c7\'f3\'d3\'c5\'cf\'c8\'bc\'b6\'a1\'a2\'cd\'c5\'b6\'d3\'c8\'cb\'c1\'a6\'b8\'ba\'d4\'d8\'a1\'a2\'c8\'ce\'ce\'f1\'d2\'c0\'c0\'b5\'b9\'d8\'cf\'b5\'a3\'ac\'d7\'d4\'b6\'af\'b5\'f7\'d5\'fb\'c5\'c5\'c6\'da\'a3\'ac\'d0\'c2\'d4\'f6\'b8\'df\'d3\'c5\'cf\'c8\'bc\'b6\'d0\'e8\'c7\'f3\'ca\'b1\'a3\'ac\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'b5\'cd\'d3\'c5\'cf\'c8\'bc\'b6\'d0\'e8\'c7\'f3\'b5\'c4\'c5\'c5\'c6\'da\'b1\'e4\'bb\'af\'a3\'ac\'c8\'ab\'d4\'b1\'ca\'b5\'ca\'b1\'bf\'c9\'bc\'fb\'a3\'ac\'ce\'de\'d0\'e8\'cf\'ee\'c4\'bf\'be\'ad\'c0\'ed\'b0\'a4\'b8\'f6\'cd\'a8\'d6\'aa\'a1\'a2\'ca\'d6\'b6\'af\'b8\'c4\'c5\'c5\'c6\'da\'a1\'a3
+\f0 \cb1 \
+\ls2\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c8\'ab\'be\'d6\'b5\'fc\'b4\'fa\'bf\'b4\'b0\'e5\'a3\'ba\'d2\'bb\'d2\'b3\'bf\'b4\'c7\'e5\'cb\'f9\'d3\'d0\'d0\'e8\'c7\'f3\'b5\'c4\'d7\'b4\'cc\'ac\'a1\'a2\'b8\'ba\'d4\'f0\'c8\'cb\'a1\'a2\'bd\'d8\'d6\'b9\'ca\'b1\'bc\'e4\'a1\'a2\'b7\'e7\'cf\'d5\'c7\'e9\'bf\'f6\'a3\'ac\'b9\'dc\'c0\'ed\'d5\'df\'d2\'bb\'d1\'db\'d5\'c6\'ce\'d5\'c8\'ab\'c3\'b2\'a3\'ac\'cd\'c5\'b6\'d3\'b3\'c9\'d4\'b1\'d6\'bb\'bf\'b4\'d7\'d4\'bc\'ba\'b8\'ba\'d4\'f0\'b5\'c4\'c4\'da\'c8\'dd\'a3\'ac\'ce\'de\'d0\'c5\'cf\'a2\'b8\'c9\'c8\'c5\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls2\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	6	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c4\'a3\'bf\'e9\'b6\'fe\'a3\'ba
+\f0 AI 
+\f1 \'c7\'fd\'b6\'af\'b5\'c4\'c8\'ab\'c1\'b4\'c2\'b7\'d6\'b4\'d0\'d0\'d3\'eb\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb\'cc\'e5\'cf\'b5\'a3\'a8\'bd\'e2\'be\'f6\'a1\'b8\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a2\'b8\'ba\'b5\'a3\'d6\'d8\'a1\'a2\'b9\'dc\'c0\'ed\'c4\'d1\'a1\'b9\'b5\'c4\'cd\'b4\'b5\'e3\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d5\'e2\'b8\'f6\'c4\'a3\'bf\'e9\'ca\'c7\'c6\'bd\'cc\'a8\'b5\'c4\'ba\'cb\'d0\'c4\'a3\'ac\'b3\'b9\'b5\'d7\'b5\'df\'b8\'b2\'b4\'ab\'cd\'b3\'a1\'b8\'d7\'d4\'c9\'cf\'b6\'f8\'cf\'c2\'b2\'f0\'bd\'e2\'c8\'ce\'ce\'f1\'a1\'a2\'d7\'d4\'cf\'c2\'b6\'f8\'c9\'cf\'bb\'e3\'b1\'a8\'bd\'f8\'b6\'c8\'a1\'b9\'b5\'c4\'c4\'a3\'ca\'bd\'a3\'ac\'d3\'c3
+\f0  AI 
+\f1 \'ca\'b5\'cf\'d6\'c8\'ab\'c1\'b4\'c2\'b7\'d7\'d4\'b6\'af\'bb\'af\'a3\'ac\'d3\'c3\'b5\'a5\'d4\'b4\'d0\'c5\'cf\'a2\'ca\'b5\'cf\'d6\'c8\'ab\'d4\'b1\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb\'a3\'ac\'c8\'c3\'cd\'c5\'b6\'d3\'b0\'d1\'be\'ab\'c1\'a6\'b4\'d3\'a1\'b8\'cc\'ee\'b1\'a8\'b1\'ed\'a1\'a2\'bf\'aa\'cd\'ac\'b2\'bd\'bb\'e1\'a1\'a2\'d5\'d2\'d0\'c5\'cf\'a2\'a1\'b9\'d7\'aa\'d2\'c6\'b5\'bd\'a1\'b8\'d0\'b4\'b4\'fa\'c2\'eb\'a1\'a2\'d7\'f6\'bd\'bb\'b8\'b6\'a1\'b9\'c9\'cf\'a1\'a3
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'ba\'cb\'d0\'c4\'b9\'a6\'c4\'dc\'c9\'e8\'bc\'c6
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls3\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b1\'b1\'bc\'ab\'d0\'c7\'c4\'bf\'b1\'ea\'b6\'d4\'c6\'eb\'c4\'a3\'bf\'e9\'a3\'a8\'b4\'d3\'b8\'f9\'d4\'b4\'cd\'b3\'d2\'bb\'c8\'ab\'d4\'b1\'c8\'cf\'d6\'aa\'a3\'a9
+\f0\b0 \cb1 \uc0\u8232 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'d7\'ee\'bc\'c9\'bb\'e4\'c4\'bf\'b1\'ea\'b7\'d6\'c9\'a2\'a1\'a2\'b8\'f7\'d7\'f6\'b8\'f7\'b5\'c4\'a3\'ac\'d3\'c3\'bc\'ab\'bc\'f2
+\f0  OKR 
+\f1 \'bb\'fa\'d6\'c6\'a3\'ac\'b0\'d1\'c4\'bf\'b1\'ea\'b6\'a4\'d4\'da\'c6\'bd\'cc\'a8\'d7\'ee\'cf\'d4\'d1\'db\'b5\'c4\'ce\'bb\'d6\'c3\'a3\'ac\'cb\'f9\'d3\'d0\'b6\'af\'d7\'f7\'b6\'bc\'ce\'a7\'c8\'c6\'c4\'bf\'b1\'ea\'d5\'b9\'bf\'aa\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls3\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c3\'bf\'b8\'f6\'cf\'ee\'c4\'bf\'bd\'f6\'c9\'e8
+\f0  1 
+\f1 \'b8\'f6\'b1\'b1\'bc\'ab\'d0\'c7\'d6\'b8\'b1\'ea\'a3\'ac
+\f0 3-5 
+\f1 \'b8\'f6\'bf\'c9\'c1\'bf\'bb\'af\'b5\'c4\'b9\'d8\'bc\'fc\'bd\'e1\'b9\'fb\'a3\'a8
+\f0 KR
+\f1 \'a3\'a9\'a3\'ac\'c8\'ab\'d4\'b1\'bf\'c9\'bc\'fb\'a1\'a2\'c8\'ab\'d4\'b1\'bf\'c9\'c6\'c0\'c2\'db\'a3\'ac\'c3\'bf\'b8\'f6\'c8\'ce\'ce\'f1
+\f0  / 
+\f1 \'d0\'e8\'c7\'f3\'b1\'d8\'d0\'eb\'b9\'d8\'c1\'aa\'b5\'bd\'b6\'d4\'d3\'a6\'b5\'c4
+\f0  KR
+\f1 \'a3\'ac\'b4\'d3\'b8\'f9\'d4\'b4\'b6\'c5\'be\'f8\'a1\'b8\'d7\'f6\'ba\'cd\'c4\'bf\'b1\'ea\'ce\'de\'b9\'d8\'b5\'c4\'ca\'c2\'a1\'b9\'a1\'a3
+\f0 \cb1 \
+\ls3\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c4\'bf\'b1\'ea\'b6\'af\'cc\'ac\'bf\'c9\'b5\'f7\'a3\'ba\'ca\'d0\'b3\'a1
+\f0  / 
+\f1 \'c4\'a3\'d0\'cd\'b7\'a2\'c9\'fa\'d6\'d8\'b4\'f3\'b1\'e4\'bb\'af\'ca\'b1\'a3\'ac\'bf\'c9\'bf\'ec\'cb\'d9\'b5\'f7\'d5\'fb\'c4\'bf\'b1\'ea\'d3\'eb
+\f0  KR
+\f1 \'a3\'ac\'b5\'f7\'d5\'fb\'ba\'f3
+\f0  AI 
+\f1 \'d7\'d4\'b6\'af\'cd\'a8\'d6\'aa\'c8\'ab\'d4\'b1\'a3\'ac\'cd\'ac\'b2\'bd\'d6\'d8\'d0\'c2\'c6\'c0\'b9\'c0\'cb\'f9\'d3\'d0\'b9\'d8\'c1\'aa\'c8\'ce\'ce\'f1\'b5\'c4\'d3\'c5\'cf\'c8\'bc\'b6\'a3\'ac\'c8\'b7\'b1\'a3\'c8\'ab\'d4\'b1\'ca\'bc\'d6\'d5\'b3\'af\'d7\'c5\'cd\'ac\'d2\'bb\'b8\'f6\'b7\'bd\'cf\'f2\'d7\'df\'a1\'a3
+\f0 \cb1 \
+\ls3\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c3\'bf\'d6\'dc
+\f0  AI 
+\f1 \'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'c4\'bf\'b1\'ea\'bd\'f8\'b6\'c8\'b1\'a8\'b8\'e6\'a3\'ba\'cd\'ac\'b2\'bd\'b1\'b1\'bc\'ab\'d0\'c7\'c4\'bf\'b1\'ea\'cd\'ea\'b3\'c9\'c7\'e9\'bf\'f6\'a1\'a2\'c4\'c4\'d0\'a9\'c8\'ce\'ce\'f1\'b9\'b1\'cf\'d7\'c1\'cb\'bd\'f8\'b6\'c8\'a1\'a2\'c4\'c4\'d0\'a9\'cd\'cf\'c1\'cb\'ba\'f3\'cd\'c8\'a1\'a2\'c4\'c4\'d0\'a9\'cd\'e2\'b2\'bf\'b1\'e4\'bb\'af\'bf\'c9\'c4\'dc\'d3\'b0\'cf\'ec\'c4\'bf\'b1\'ea\'a3\'ac\'c8\'ab\'d4\'b1\'cd\'ac\'b2\'bd\'a3\'ac\'b3\'b9\'b5\'d7\'cf\'fb\'c3\'f0\'a1\'b8\'b2\'bb\'d6\'aa\'b5\'c0\'cf\'ee\'c4\'bf\'d7\'f6\'b5\'bd\'c4\'c4\'c1\'cb\'a1\'a2\'b2\'bb\'d6\'aa\'b5\'c0\'ce\'aa\'ca\'b2\'c3\'b4\'d7\'f6\'d5\'e2\'b8\'f6\'c8\'ce\'ce\'f1\'a1\'b9\'b5\'c4\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls3\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls3\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 AI 
+\f3 \'d6\'c7\'c4\'dc\'c8\'ce\'ce\'f1\'b9\'dc\'c0\'ed\'a3\'a8\'c1\'e3\'ca\'d6\'b6\'af\'cc\'ee\'b1\'a8\'a3\'ac\'c8\'ab\'c1\'b4\'c2\'b7\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'a3\'a9
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b4\'ab\'cd\'b3\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'d7\'ee\'b4\'f3\'b5\'c4\'b8\'ba\'b5\'a3\'a3\'ac\'be\'cd\'ca\'c7\'c8\'c3\'bf\'aa\'b7\'a2\'c8\'cb\'d4\'b1\'ca\'d6\'b6\'af\'cc\'ee\'bd\'f8\'b6\'c8\'a1\'a2\'d0\'b4\'c8\'d5\'d6\'be\'a1\'a2\'bd\'a8\'c8\'ce\'ce\'f1\'a3\'ac\'d5\'e2\'b8\'f6\'c4\'a3\'bf\'e9\'b5\'c4\'ba\'cb\'d0\'c4\'ca\'c7\'a1\'b8
+\f0 AI 
+\f1 \'cc\'e6\'c4\'e3\'d7\'f6\'a3\'ac\'cf\'b5\'cd\'b3\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'a1\'b9\'a3\'ac\'cd\'ea\'c8\'ab\'b2\'bb\'d4\'f6\'bc\'d3\'cd\'c5\'b6\'d3\'b6\'ee\'cd\'e2\'b9\'a4\'d7\'f7\'c1\'bf\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls3\ilvl1
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 AI 
+\f3 \'d2\'bb\'bc\'fc\'b2\'f0\'bd\'e2\'c8\'ce\'ce\'f1
+\f1\b0 \'a3\'ba\'d0\'e8\'c7\'f3
+\f0  / 
+\f1 \'be\'f6\'b2\'df\'c8\'b7\'c8\'cf\'ba\'f3\'a3\'ac
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'bb\'f9\'d3\'da\'d6\'aa\'ca\'b6\'bf\'e2\'d6\'d0\'b5\'c4\'c0\'fa\'ca\'b7\'cf\'ee\'c4\'bf\'ca\'fd\'be\'dd\'a1\'a2\'cd\'c5\'b6\'d3\'c4\'dc\'c1\'a6\'cd\'bc\'c6\'d7\'a1\'a2\'c4\'a3\'d0\'cd\'c4\'dc\'c1\'a6\'b1\'df\'bd\'e7\'a3\'ac\'b2\'f0\'bd\'e2\'ce\'aa\'bf\'c9\'d6\'b4\'d0\'d0\'b5\'c4\'d7\'ee\'d0\'a1\'c8\'ce\'ce\'f1\'b5\'a5\'d4\'aa\'a3\'ac\'d7\'d4\'b6\'af\'b7\'d6\'c5\'e4\'b8\'ba\'d4\'f0\'c8\'cb\'a1\'a2\'d4\'a4\'b9\'c0\'b9\'a4\'ca\'b1\'a1\'a2\'c9\'e8\'d6\'c3\'bd\'d8\'d6\'b9\'ca\'b1\'bc\'e4\'a1\'a2\'b1\'ea\'bc\'c7\'d2\'c0\'c0\'b5\'b9\'d8\'cf\'b5\'a3\'ac\'b8\'ba\'d4\'f0\'c8\'cb\'bd\'f6\'d0\'e8\'d2\'bb\'bc\'fc\'c8\'b7\'c8\'cf
+\f0  / 
+\f1 \'ce\'a2\'b5\'f7\'a3\'ac\'ce\'de\'d0\'e8\'ca\'d6\'b6\'af\'bd\'a8\'c8\'ce\'ce\'f1\'a1\'a3
+\f0 \cb1 \uc0\u8232 
+\f1 \cb3 \'c0\'fd\'a3\'ba\'b2\'fa\'c6\'b7\'cc\'e1\'a1\'b8\'bb\'f9\'d3\'da\'d0\'c2\'c4\'a3\'d0\'cd\'d6\'d8\'b9\'b9\'b3\'a4\'ce\'c4\'b5\'b5\'d7\'dc\'bd\'e1\'b9\'a6\'c4\'dc\'a1\'b9\'a3\'ac
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'b2\'f0\'bd\'e2\'ce\'aa\'a3\'ba\'cb\'e3\'b7\'a8\'c4\'a3\'d0\'cd\'ca\'ca\'c5\'e4\'d3\'eb\'bc\'e6\'c8\'dd\'d0\'d4\'b2\'e2\'ca\'d4\'a1\'a2\'ba\'f3\'b6\'cb\'bd\'d3\'bf\'da\'bf\'aa\'b7\'a2\'a1\'a2\'c7\'b0\'b6\'cb\'bd\'bb\'bb\'a5\'d6\'d8\'b9\'b9\'a1\'a2\'b2\'e2\'ca\'d4\'d3\'c3\'c0\'fd\'b1\'e0\'d0\'b4\'a1\'a2\'bb\'d2\'b6\'c8\'c9\'cf\'cf\'df\'d1\'e9\'d6\'a4\'a3\'ac\'d7\'d4\'b6\'af\'b7\'d6\'c5\'e4\'b8\'f8\'b6\'d4\'d3\'a6\'bd\'c7\'c9\'ab\'c8\'cb\'d4\'b1\'a3\'ac\'cd\'ac\'ca\'b1\'b9\'d8\'c1\'aa\'d6\'aa\'ca\'b6\'bf\'e2\'d6\'d0\'b9\'fd\'cd\'f9\'b5\'c4\'c4\'a3\'d0\'cd\'ca\'ca\'c5\'e4\'b2\'c8\'bf\'d3\'bc\'c7\'c2\'bc\'a1\'a3
+\f0 \cb1 \
+\ls3\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c8\'ce\'ce\'f1\'bd\'f8\'b6\'c8\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'a3\'ac\'c1\'e3\'ca\'d6\'b6\'af\'cc\'ee\'b1\'a8
+\f1\b0 \'a3\'ba\'ba\'cd
+\f0  GitLab/Gitee 
+\f1 \'b4\'fa\'c2\'eb\'b2\'d6\'bf\'e2\'a1\'a2
+\f0 CI/CD 
+\f1 \'c6\'bd\'cc\'a8\'a1\'a2\'b2\'e2\'ca\'d4\'c6\'bd\'cc\'a8\'c9\'ee\'b6\'c8\'b4\'f2\'cd\'a8\'a3\'ac\'b4\'fa\'c2\'eb\'cc\'e1\'bd\'bb\'a1\'a2
+\f0 PR 
+\f1 \'ba\'cf\'b2\'a2\'a1\'a2\'b2\'e2\'ca\'d4\'cd\'a8\'b9\'fd\'a1\'a2\'c9\'cf\'cf\'df\'cd\'ea\'b3\'c9\'a3\'ac\'cf\'b5\'cd\'b3\'d7\'d4\'b6\'af\'b8\'fc\'d0\'c2\'c8\'ce\'ce\'f1\'bd\'f8\'b6\'c8\'a3\'ac\'ce\'de\'d0\'e8\'bf\'aa\'b7\'a2\'c8\'cb\'d4\'b1\'ca\'d6\'b6\'af\'d0\'de\'b8\'c4\'a1\'a3\'c0\'fd\'a3\'ba\'ba\'f3\'b6\'cb\'bf\'aa\'b7\'a2\'c8\'ce\'ce\'f1\'a3\'ac
+\f0 PR 
+\f1 \'ba\'cf\'b2\'a2\'b5\'bd\'b2\'e2\'ca\'d4\'b7\'d6\'d6\'a7\'a3\'ac\'d7\'d4\'b6\'af\'b8\'fc\'d0\'c2\'bd\'f8\'b6\'c8\'d6\'c1
+\f0  80%
+\f1 \'a3\'bb\'b2\'e2\'ca\'d4\'d3\'c3\'c0\'fd\'c8\'ab\'b2\'bf\'cd\'a8\'b9\'fd\'a3\'ac\'d7\'d4\'b6\'af\'b8\'fc\'d0\'c2\'ce\'aa
+\f0  100% 
+\f1 \'cd\'ea\'b3\'c9\'a1\'a3
+\f0 \cb1 \
+\ls3\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b5\'a5\'c8\'ce\'ce\'f1\'c8\'ab\'d0\'c5\'cf\'a2\'b9\'e9\'bc\'af\'a3\'ac\'cf\'fb\'c3\'f0\'d0\'c5\'cf\'a2\'b2\'ee
+\f1\b0 \'a3\'ba\'c3\'bf\'b8\'f6\'c8\'ce\'ce\'f1\'d2\'b3\'c3\'e6\'a3\'ac\'d7\'d4\'b6\'af\'be\'db\'ba\'cf\'cb\'f9\'d3\'d0\'cf\'e0\'b9\'d8\'d0\'c5\'cf\'a2\'a3\'ba\'d0\'e8\'c7\'f3\'b1\'b3\'be\'b0\'a1\'a2\'b9\'d8\'c1\'aa
+\f0  KR
+\f1 \'a1\'a2\'c4\'a3\'d0\'cd\'b2\'ce\'ca\'fd\'a1\'a2\'c9\'e8\'bc\'c6\'b8\'e5\'a1\'a2\'b4\'fa\'c2\'eb\'c1\'b4\'bd\'d3\'a1\'a2\'b2\'e2\'ca\'d4\'b1\'a8\'b8\'e6\'a1\'a2\'c6\'c0\'c2\'db\'b9\'b5\'cd\'a8\'a1\'a2\'c0\'fa\'ca\'b7\'b1\'e4\'b8\'fc\'a3\'ac\'b5\'e3\'bd\'f8\'c8\'a5\'be\'cd\'d5\'c6\'ce\'d5\'c8\'ab\'b2\'bf\'c9\'cf\'cf\'c2\'ce\'c4\'a3\'ac\'ce\'de\'d0\'e8\'b5\'bd\'b4\'a6\'b7\'ad\'c1\'c4\'cc\'ec\'bc\'c7\'c2\'bc\'a1\'a2\'d5\'d2\'c0\'eb\'cf\'df\'ce\'c4\'b5\'b5\'a3\'ac\'d0\'c2\'c8\'cb\'bd\'d3\'ca\'d6\'c8\'ce\'ce\'f1
+\f0  10 
+\f1 \'b7\'d6\'d6\'d3\'be\'cd\'c4\'dc\'d5\'c6\'ce\'d5\'c8\'ab\'c3\'b2\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls3\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls3\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c7\'e1\'c1\'bf\'bb\'af\'b6\'d4\'c6\'eb\'bb\'fa\'d6\'c6\'a3\'ac\'b0\'d1\'bb\'e1\'d2\'e9\'ca\'b1\'bc\'e4\'d1\'b9\'cb\'f5
+\f2  80%
+\f0\b0 \cb1 \uc0\u8232 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'b9\'b5\'cd\'a8\'a3\'ac\'ba\'cb\'d0\'c4\'ca\'c7\'a1\'b8\'d2\'ec\'b2\'bd\'d3\'c5\'cf\'c8\'a1\'a2\'cd\'ac\'b2\'bd\'ce\'aa\'b8\'a8\'a1\'b9\'a3\'ac\'d3\'c3\'c6\'bd\'cc\'a8\'cc\'e6\'b4\'fa\'ce\'de\'d0\'a7\'bb\'e1\'d2\'e9\'a3\'ac\'ca\'b5\'cf\'d6\'b3\'a3\'cc\'ac\'bb\'af\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls3\ilvl1
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c3\'bf\'c8\'d5
+\f2  3 
+\f3 \'be\'e4\'bb\'b0\'d5\'be\'bb\'e1\'a3\'ac\'cc\'e6\'b4\'fa
+\f2  15 
+\f3 \'b7\'d6\'d6\'d3\'b3\'bf\'bb\'e1
+\f1\b0 \'a3\'ba\'c3\'bf\'b8\'f6\'c8\'cb\'c3\'bf\'cc\'ec\'d4\'e7\'c9\'cf\'bb\'a8
+\f0  3 
+\f1 \'b7\'d6\'d6\'d3\'a3\'ac\'d4\'da\'c6\'bd\'cc\'a8\'c9\'cf\'cc\'ee\'d0\'b4
+\f0  3 
+\f1 \'be\'e4\'bb\'b0\'a3\'ba\'d7\'f2\'cc\'ec\'cd\'ea\'b3\'c9\'c1\'cb\'ca\'b2\'c3\'b4\'a1\'a2\'bd\'f1\'cc\'ec\'bc\'c6\'bb\'ae\'d7\'f6\'ca\'b2\'c3\'b4\'a1\'a2\'d3\'f6\'b5\'bd\'b5\'c4\'b7\'e7\'cf\'d5
+\f0  / 
+\f1 \'d0\'e8\'d2\'aa\'b5\'c4\'d0\'ad\'d6\'fa\'a1\'a3
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'bb\'e3\'d7\'dc\'c9\'fa\'b3\'c9\'c8\'ab\'d4\'b1\'d5\'be\'bb\'e1\'b1\'a8\'b8\'e6\'a3\'ac\'cd\'c6\'cb\'cd\'b5\'bd\'cd\'c5\'b6\'d3\'c8\'ba\'a3\'ac\'d3\'d0\'ce\'ca\'cc\'e2\'d6\'b1\'bd\'d3\'d4\'da\'b6\'d4\'d3\'a6\'c8\'ce\'ce\'f1\'cf\'c2\'c6\'c0\'c2\'db\'b9\'b5\'cd\'a8\'a3\'ac\'d2\'ec\'b2\'bd\'bd\'e2\'be\'f6\'a3\'ac\'ce\'de\'d0\'e8\'c8\'ab\'d4\'b1\'b4\'d5\'ca\'b1\'bc\'e4\'bf\'aa\'bb\'e1\'a3\'ac\'b3\'b9\'b5\'d7\'b6\'c5\'be\'f8\'a1\'b8\'d2\'bb\'c8\'cb\'bb\'e3\'b1\'a8\'a1\'a2\'be\'c5\'c8\'cb\'c3\'fe\'d3\'e3\'a1\'b9\'b5\'c4\'ce\'de\'d0\'a7\'bb\'e1\'d2\'e9\'a1\'a3
+\f0 \cb1 \
+\ls3\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b7\'e7\'cf\'d5\'ca\'b5\'ca\'b1\'d4\'a4\'be\'af\'d3\'eb\'cd\'ac\'b2\'bd
+\f1\b0 \'a3\'ba
+\f0 AI7\'d724 
+\f1 \'d0\'a1\'ca\'b1\'d7\'d4\'b6\'af\'ca\'b6\'b1\'f0\'cf\'ee\'c4\'bf\'b7\'e7\'cf\'d5\'a3\'ba\'c8\'ce\'ce\'f1\'d1\'d3\'c6\'da\'a1\'a2\'d2\'c0\'c0\'b5\'ce\'b4\'cd\'ea\'b3\'c9\'a1\'a2\'c4\'a3\'d0\'cd\'c4\'dc\'c1\'a6\'b2\'bb\'b4\'ef\'b1\'ea\'a1\'a2\'cd\'c5\'b6\'d3\'b8\'ba\'d4\'d8\'b9\'fd\'d4\'d8\'a1\'a2\'d0\'e8\'c7\'f3\'c6\'b5\'b7\'b1\'b1\'e4\'b8\'fc\'a3\'ac\'d7\'d4\'b6\'af\'b1\'ea\'bc\'c7\'b7\'e7\'cf\'d5\'b5\'c8\'bc\'b6\'a3\'ac\'cd\'c6\'cb\'cd\'b8\'f8\'b8\'ba\'d4\'f0\'c8\'cb\'d3\'eb\'cf\'e0\'b9\'d8\'c8\'cb\'d4\'b1\'a3\'ac\'cd\'ac\'ca\'b1\'d4\'da\'c8\'ab\'be\'d6\'b7\'e7\'cf\'d5\'bf\'b4\'b0\'e5\'b8\'df\'c1\'c1\'d5\'b9\'ca\'be\'a1\'a3\'b9\'dc\'c0\'ed\'d5\'df\'cc\'e1\'c7\'b0\'b7\'a2\'cf\'d6\'ce\'ca\'cc\'e2\'a1\'a2\'cc\'e1\'c7\'b0\'bd\'e2\'be\'f6\'a3\'ac\'b2\'bb\'d3\'c3\'b5\'c8\'cf\'ee\'c4\'bf\'d1\'d3\'c6\'da\'c1\'cb\'b2\'c5\'ba\'f3\'d6\'aa\'ba\'f3\'be\'f5\'a1\'a3
+\f0 \cb1 \
+\ls3\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f6\'b2\'df\'c8\'d5\'d6\'be\'c8\'ab\'c1\'f4\'b4\'e6
+\f1\b0 \'a3\'ba\'cb\'f9\'d3\'d0\'cf\'ee\'c4\'bf\'cf\'e0\'b9\'d8\'b5\'c4\'be\'f6\'b2\'df\'a3\'ac\'b2\'bb\'b9\'dc\'ca\'c7\'bb\'e1\'c9\'cf\'b6\'a8\'b5\'c4\'a1\'a2\'bb\'b9\'ca\'c7\'c8\'ba\'c0\'ef\'b9\'b5\'cd\'a8\'b6\'a8\'b5\'c4\'a3\'ac\'c8\'ab\'b2\'bf\'c7\'bf\'d6\'c6\'c2\'bc\'c8\'eb\'c6\'bd\'cc\'a8\'be\'f6\'b2\'df\'c8\'d5\'d6\'be\'a3\'ac\'b9\'d8\'c1\'aa\'b6\'d4\'d3\'a6\'cf\'ee\'c4\'bf
+\f0  / 
+\f1 \'c8\'ce\'ce\'f1\'a3\'ac\'b1\'ea\'bc\'c7\'be\'f6\'b2\'df\'c8\'cb\'a1\'a2\'be\'f6\'b2\'df\'ca\'b1\'bc\'e4\'a1\'a2\'be\'f6\'b2\'df\'c4\'da\'c8\'dd\'a1\'a2\'ba\'cb\'d0\'c4\'d4\'ad\'d2\'f2\'a3\'ac\'c8\'ab\'d4\'b1\'bf\'c9\'b2\'e9\'a1\'a2\'d3\'c0\'be\'c3\'c1\'f4\'b4\'e6\'a1\'a3\'b3\'b9\'b5\'d7\'b6\'c5\'be\'f8\'a1\'b8\'bb\'e1\'d2\'e9\'b6\'a8\'c1\'cb\'b5\'c4\'ca\'c2\'d7\'aa\'cd\'b7\'be\'cd\'cd\'fc\'a1\'b9\'a1\'b8\'b2\'bb\'cd\'ac\'c8\'cb\'b6\'d4\'be\'f6\'b2\'df\'b5\'c4\'c0\'ed\'bd\'e2\'b2\'bb\'d2\'bb\'d1\'f9\'a1\'b9\'b5\'c4\'ba\'cb\'d0\'c4\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls3\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	6	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c4\'a3\'bf\'e9\'c8\'fd\'a3\'ba\'cf\'ee\'c4\'bf\'b0\'f3\'b6\'a8\'b5\'c4
+\f0  AI 
+\f1 \'d4\'ad\'c9\'fa\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'a8\'bd\'e2\'be\'f6\'a1\'b8\'c4\'a3\'d0\'cd\'b5\'fc\'b4\'fa\'bf\'ec\'a1\'a2\'be\'ad\'d1\'e9\'b2\'bb\'b3\'c1\'b5\'ed\'a1\'a2\'c8\'cf\'d6\'aa\'b2\'bb\'cd\'b3\'d2\'bb\'a1\'b9\'b5\'c4\'cd\'b4\'b5\'e3\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d6\'aa\'ca\'b6\'bf\'e2\'b2\'bb\'ca\'c7\'b9\'c2\'c1\'a2\'b5\'c4\'ce\'c4\'b5\'b5\'bf\'e2\'a3\'ac\'b6\'f8\'ca\'c7\'cd\'c5\'b6\'d3\'b5\'c4\'a1\'b8\'b5\'da\'b6\'fe\'b4\'f3\'c4\'d4\'a1\'b9\'a3\'ac\'ca\'c7\'cf\'fb\'c3\'f0\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a2\'b8\'fa\'c9\'cf\'c4\'a3\'d0\'cd\'b5\'fc\'b4\'fa\'cb\'d9\'b6\'c8\'b5\'c4\'ba\'cb\'d0\'c4\'d4\'d8\'cc\'e5\'a3\'ac\'b1\'d8\'d0\'eb\'ba\'cd\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'c8\'ab\'c1\'b4\'c2\'b7\'c9\'ee\'b6\'c8\'b4\'f2\'cd\'a8\'a3\'ac\'b6\'f8\'b7\'c7\'b5\'a5\'b6\'c0\'b4\'e6\'d4\'da\'a1\'a3
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'ba\'cb\'d0\'c4\'b9\'a6\'c4\'dc\'c9\'e8\'bc\'c6
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls4\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'cf\'ee\'c4\'bf
+\f2  - 
+\f3 \'d6\'aa\'ca\'b6\'bf\'e2\'c7\'bf\'b0\'f3\'b6\'a8\'a3\'ac\'c8\'ab\'c1\'bf\'d0\'c5\'cf\'a2\'d7\'d4\'b6\'af\'b3\'c1\'b5\'ed
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b3\'b9\'b5\'d7\'bd\'e2\'be\'f6\'a1\'b8\'ce\'c4\'b5\'b5\'ba\'cd\'d6\'b4\'d0\'d0\'cd\'d1\'bd\'da\'a1\'a2\'be\'ad\'d1\'e9\'d0\'b4\'cd\'ea\'be\'cd\'c3\'bb\'c8\'cb\'bf\'b4\'a1\'a2\'d0\'c2\'c8\'cb\'c9\'cf\'ca\'d6\'c8\'ab\'bf\'bf\'c0\'cf\'d4\'b1\'b9\'a4\'b4\'f8\'a1\'b9\'b5\'c4\'ce\'ca\'cc\'e2\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls4\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c3\'bf\'b8\'f6\'cf\'ee\'c4\'bf\'b6\'d4\'d3\'a6\'d7\'a8\'ca\'f4\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'ac\'d4\'a4\'c9\'e8\'b9\'cc\'b6\'a8\'b7\'d6\'c0\'e0\'a3\'ba\'d0\'e8\'c7\'f3\'ce\'c4\'b5\'b5\'a1\'a2\'c4\'a3\'d0\'cd\'ca\'d6\'b2\'e1\'a1\'a2\'bc\'bc\'ca\'f5\'b7\'bd\'b0\'b8\'a1\'a2\'b2\'e2\'ca\'d4\'b1\'a8\'b8\'e6\'a1\'a2\'be\'f6\'b2\'df\'c8\'d5\'d6\'be\'a1\'a2\'b2\'c8\'bf\'d3\'bc\'c7\'c2\'bc\'a1\'a2\'b8\'b4\'c5\'cc\'d7\'dc\'bd\'e1\'a3\'ac\'ce\'de\'d0\'e8\'ca\'d6\'b6\'af\'bd\'a8\'ce\'c4\'bc\'fe\'bc\'d0\'a3\'ac\'b9\'e6\'b7\'b6\'cd\'b3\'d2\'bb\'a3\'ac\'b2\'e9\'d5\'d2\'b7\'bd\'b1\'e3\'a1\'a3
+\f0 \cb1 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls4\ilvl1
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'d7\'d4\'b6\'af\'b3\'c1\'b5\'ed\'a3\'ac\'c1\'e3\'ca\'d6\'b6\'af\'c9\'cf\'b4\'ab
+\f1\b0 \'a3\'ba\'cf\'ee\'c4\'bf\'c8\'ab\'c1\'f7\'b3\'cc\'b2\'fa\'c9\'fa\'b5\'c4\'ce\'c4\'b5\'b5\'a3\'ac\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'b5\'bd\'d6\'aa\'ca\'b6\'bf\'e2\'b6\'d4\'d3\'a6\'b7\'d6\'c0\'e0\'a1\'a3\'c0\'fd\'a3\'ba\'c8\'ce\'ce\'f1\'d6\'d0\'b5\'c4\'c4\'a3\'d0\'cd\'ca\'ca\'c5\'e4\'b2\'e2\'ca\'d4\'b1\'a8\'b8\'e6\'a3\'ac\'cd\'ea\'b3\'c9\'ba\'f3\'d7\'d4\'b6\'af\'b4\'e6\'c8\'eb\'a1\'b8\'c4\'a3\'d0\'cd\'ca\'d6\'b2\'e1\'a1\'b9\'a3\'bb\'cf\'ee\'c4\'bf\'b5\'c4\'be\'f6\'b2\'df\'c8\'d5\'d6\'be\'a3\'ac\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'b5\'bd\'b6\'d4\'d3\'a6\'b7\'d6\'c0\'e0\'a3\'bb\'b5\'fc\'b4\'fa\'b8\'b4\'c5\'cc\'d7\'dc\'bd\'e1\'a3\'ac\'d7\'d4\'b6\'af\'b9\'e9\'b5\'b5\'b5\'bd\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'ac\'ce\'de\'d0\'e8\'cd\'c5\'b6\'d3\'b3\'c9\'d4\'b1\'ca\'d6\'b6\'af\'c9\'cf\'b4\'ab\'a3\'ac\'b4\'d3\'b8\'f9\'d4\'b4\'b1\'dc\'c3\'e2\'ce\'c4\'b5\'b5\'d3\'eb\'ca\'b5\'bc\'ca\'d6\'b4\'d0\'d0\'cd\'d1\'bd\'da\'a1\'a3
+\f0 \cb1 \
+\ls4\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b0\'e6\'b1\'be\'d3\'eb\'b5\'fc\'b4\'fa
+\f2  / 
+\f3 \'c4\'a3\'d0\'cd\'c7\'bf\'b0\'f3\'b6\'a8
+\f1\b0 \'a3\'ba\'c3\'bf\'b8\'f6\'ce\'c4\'b5\'b5\'b5\'c4\'b0\'e6\'b1\'be\'a3\'ac\'b6\'bc\'b9\'d8\'c1\'aa\'b6\'d4\'d3\'a6\'b5\'c4\'cf\'ee\'c4\'bf\'b5\'fc\'b4\'fa\'a1\'a2\'c4\'a3\'d0\'cd\'b0\'e6\'b1\'be\'ba\'c5\'a3\'a8\'c0\'fd\'a3\'ba
+\f0 V2.0 
+\f1 \'b5\'fc\'b4\'fa
+\f0  - 
+\f1 \'bb\'f9\'d3\'da
+\f0  Qwen2-7B v1.5 
+\f1 \'b0\'e6\'b1\'be\'b5\'c4\'b3\'a4\'ce\'c4\'b5\'b5\'b4\'a6\'c0\'ed\'b7\'bd\'b0\'b8\'a3\'a9\'a1\'a3\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'b5\'fc\'b4\'fa\'ca\'b1\'a3\'ac\'bf\'c9\'bf\'ec\'cb\'d9\'bb\'d8\'cb\'dd\'c0\'fa\'ca\'b7\'b7\'bd\'b0\'b8\'a1\'a2\'b2\'c8\'bf\'d3\'bc\'c7\'c2\'bc\'a1\'a2\'ca\'ca\'c5\'e4\'be\'ad\'d1\'e9\'a3\'ac\'b1\'dc\'c3\'e2\'d6\'d8\'b8\'b4\'b2\'c8\'bf\'d3\'a3\'ac\'c8\'c3\'cd\'c5\'b6\'d3\'bf\'ec\'cb\'d9\'b8\'fa\'c9\'cf\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'b5\'c4\'cb\'d9\'b6\'c8\'a3\'ac\'b2\'bb\'bb\'e1\'b4\'ed\'b9\'fd\'bc\'bc\'ca\'f5\'ba\'ec\'c0\'fb\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls4\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls4\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 RAG 
+\f3 \'d4\'f6\'c7\'bf\'b5\'c4\'d6\'c7\'c4\'dc\'ce\'ca\'b4\'f0\'a3\'ac\'ca\'b5\'cf\'d6\'c8\'ab\'d4\'b1\'c8\'cf\'d6\'aa\'c0\'ad\'c6\'bd
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'bb\'f9\'d3\'da\'b1\'be\'b5\'d8\'b2\'bf\'ca\'f0\'b5\'c4\'b4\'f3\'c4\'a3\'d0\'cd\'a3\'ac\'b6\'d4\'c8\'ab\'c1\'bf\'d6\'aa\'ca\'b6\'bf\'e2\'d7\'f6\'cf\'f2\'c1\'bf\'bb\'af\'b4\'a6\'c0\'ed\'a3\'ac\'b4\'f2\'d4\'ec\'cd\'c5\'b6\'d3\'d7\'a8\'ca\'f4\'b5\'c4\'d6\'c7\'c4\'dc\'ce\'ca\'b4\'f0\'d6\'fa\'ca\'d6\'a3\'ac\'b3\'b9\'b5\'d7\'bd\'e2\'be\'f6\'a1\'b8\'c0\'cf\'d4\'b1\'b9\'a4\'be\'ad\'d1\'e9\'b4\'f8\'b2\'bb\'d7\'df\'a1\'a2\'d0\'c2\'c8\'cb\'c9\'cf\'ca\'d6\'c2\'fd\'a1\'a2\'d0\'c5\'cf\'a2\'d5\'d2\'b2\'bb\'b5\'bd\'a1\'b9\'b5\'c4\'ce\'ca\'cc\'e2\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls4\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c8\'ab\'c6\'bd\'cc\'a8\'d2\'bb\'bc\'fc\'bb\'bd\'c6\'f0\'ce\'ca\'b4\'f0\'a3\'ac\'d4\'da\'c8\'ce\'ce\'f1\'d2\'b3\'c3\'e6\'a1\'a2\'d0\'e8\'c7\'f3\'d2\'b3\'c3\'e6\'a1\'a2\'d6\'aa\'ca\'b6\'bf\'e2\'d2\'b3\'c3\'e6\'a3\'ac\'b6\'bc\'bf\'c9\'d2\'d4\'d6\'b1\'bd\'d3\'cc\'e1\'ce\'ca\'a3\'ac
+\f0 AI 
+\f1 \'bb\'f9\'d3\'da\'d6\'aa\'ca\'b6\'bf\'e2\'c4\'da\'c8\'dd\'b8\'f8\'b3\'f6\'be\'ab\'d7\'bc\'b4\'f0\'b0\'b8\'a3\'ac\'b6\'c5\'be\'f8\'bb\'c3\'be\'f5\'a1\'a3\'c0\'fd\'a3\'ba\'d4\'da\'c4\'a3\'d0\'cd\'ca\'ca\'c5\'e4\'c8\'ce\'ce\'f1\'d6\'d0\'a3\'ac\'d6\'b1\'bd\'d3\'ce\'ca\'a1\'b8\'d5\'e2\'b8\'f6\'c4\'a3\'d0\'cd\'d6\'ae\'c7\'b0\'b5\'c4\'b0\'e6\'b1\'be\'ca\'ca\'c5\'e4\'d3\'d0\'ca\'b2\'c3\'b4\'bf\'d3\'a3\'bf\'a1\'b9\'a3\'ac
+\f0 AI 
+\f1 \'d6\'b1\'bd\'d3\'bc\'ec\'cb\'f7\'d6\'aa\'ca\'b6\'bf\'e2\'d6\'d0\'b5\'c4\'c0\'fa\'ca\'b7\'bc\'c7\'c2\'bc\'a3\'ac\'b8\'f8\'b3\'f6\'cd\'ea\'d5\'fb\'b5\'c4\'b1\'dc\'bf\'d3\'b7\'bd\'b0\'b8\'a3\'ac\'ce\'de\'d0\'e8\'b5\'bd\'b4\'a6\'d5\'d2\'c0\'cf\'d4\'b1\'b9\'a4\'d1\'af\'ce\'ca\'a1\'a3
+\f0 \cb1 \
+\ls4\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'d3\'b0\'cf\'ec\'b7\'d6\'ce\'f6\'a3\'ba\'d0\'c2\'c4\'a3\'d0\'cd\'b7\'a2\'b2\'bc\'ba\'f3\'a3\'ac
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'b6\'d4\'b1\'c8\'d0\'c2\'be\'c9\'c4\'a3\'d0\'cd\'b5\'c4\'c4\'dc\'c1\'a6\'b2\'ee\'d2\'ec\'a3\'ac\'bc\'ec\'cb\'f7\'d6\'aa\'ca\'b6\'bf\'e2\'d6\'d0\'cb\'f9\'d3\'d0\'d3\'c3\'b5\'bd\'b8\'c3\'c4\'a3\'d0\'cd\'b5\'c4\'cf\'ee\'c4\'bf\'a1\'a2\'b9\'a6\'c4\'dc\'a1\'a2\'bc\'bc\'ca\'f5\'b7\'bd\'b0\'b8\'a3\'ac\'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'a1\'b6\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'d3\'b0\'cf\'ec\'b7\'d6\'ce\'f6\'b1\'a8\'b8\'e6\'a1\'b7\'a3\'ac\'b8\'e6\'cb\'df\'cd\'c5\'b6\'d3\'c4\'c4\'d0\'a9\'b9\'a6\'c4\'dc\'bf\'c9\'d2\'d4\'d3\'c5\'bb\'af\'a1\'a2\'c4\'c4\'d0\'a9\'b7\'bd\'b0\'b8\'d0\'e8\'d2\'aa\'d6\'d8\'b9\'b9\'a1\'a2\'c4\'c4\'d0\'a9\'b7\'e7\'cf\'d5\'d0\'e8\'d2\'aa\'cc\'e1\'c7\'b0\'b9\'e6\'b1\'dc\'a3\'ac\'c8\'c3\'cd\'c5\'b6\'d3\'d4\'da
+\f0  24 
+\f1 \'d0\'a1\'ca\'b1\'c4\'da\'cd\'ea\'b3\'c9\'b6\'d4\'d0\'c2\'c4\'a3\'d0\'cd\'b5\'c4\'c6\'c0\'b9\'c0\'d3\'eb\'c2\'e4\'b5\'d8\'b9\'e6\'bb\'ae\'a3\'ac\'ca\'bc\'d6\'d5\'b8\'fa\'c9\'cf\'bc\'bc\'ca\'f5\'b5\'fc\'b4\'fa\'cb\'d9\'b6\'c8\'a1\'a3
+\f0 \cb1 \
+\ls4\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c3\'bf\'d6\'dc\'c8\'cf\'d6\'aa\'cd\'ac\'b2\'bd\'a3\'ba
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'a1\'b6\'d6\'aa\'ca\'b6\'bf\'e2\'b8\'fc\'d0\'c2\'d6\'dc\'b1\'a8\'a1\'b7\'a3\'ac\'cd\'ac\'b2\'bd\'b1\'be\'d6\'dc\'d0\'c2\'d4\'f6\'b5\'c4\'ce\'c4\'b5\'b5\'a1\'a2\'c4\'a3\'d0\'cd\'be\'ad\'d1\'e9\'a1\'a2\'b2\'c8\'bf\'d3\'bc\'c7\'c2\'bc\'a1\'a2\'be\'f6\'b2\'df\'b1\'e4\'b8\'fc\'a3\'ac\'c8\'ab\'d4\'b1\'cd\'c6\'cb\'cd\'a3\'ac\'c8\'b7\'b1\'a3\'cb\'f9\'d3\'d0\'c8\'cb\'b6\'bc\'d5\'c6\'ce\'d5\'d7\'ee\'d0\'c2\'b5\'c4\'d0\'c5\'cf\'a2\'a3\'ac\'b1\'dc\'c3\'e2\'a1\'b8\'d6\'bb\'d3\'d0\'c9\'d9\'ca\'fd\'c8\'cb\'d6\'aa\'b5\'c0\'b8\'fc\'d0\'c2\'a3\'ac\'c6\'e4\'cb\'fb\'c8\'cb\'bb\'b9\'d4\'da\'d3\'c3\'be\'c9\'b7\'bd\'b0\'b8\'a1\'b9\'b5\'c4\'c8\'cf\'d6\'aa\'b2\'ee\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls4\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c4\'a3\'bf\'e9\'cb\'c4\'a3\'ba
+\f0 AI 
+\f1 \'b8\'a8\'d6\'fa\'b5\'c4\'b9\'dc\'c0\'ed\'d3\'eb\'b8\'b4\'c5\'cc\'b1\'d5\'bb\'b7\'a3\'a8\'bd\'e2\'be\'f6\'a1\'b8\'b9\'dc\'c0\'ed\'ce\'de\'d7\'a5\'ca\'d6\'a1\'a2\'d3\'c5\'bb\'af\'ce\'de\'b7\'bd\'cf\'f2\'a1\'a2\'be\'ad\'d1\'e9\'b2\'bb\'b3\'c1\'b5\'ed\'a1\'b9\'b5\'c4\'cd\'b4\'b5\'e3\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'b9\'dc\'c0\'ed\'a3\'ac\'ba\'cb\'d0\'c4\'b2\'bb\'ca\'c7\'b6\'a2\'c8\'cb\'a3\'ac\'b6\'f8\'ca\'c7\'b6\'a2\'c4\'bf\'b1\'ea\'a1\'a2\'bf\'d8\'b7\'e7\'cf\'d5\'a1\'a2\'cc\'e1\'c8\'cb\'d0\'a7\'a1\'a2\'b3\'d6\'d0\'f8\'d3\'c5\'bb\'af\'a1\'a3\'c6\'bd\'cc\'a8\'d2\'aa\'d7\'f6\'b5\'c4\'ca\'c7\'b8\'f8\'b9\'dc\'c0\'ed\'d5\'df\'cc\'e1\'b9\'a9\'c8\'ab\'be\'d6\'ca\'d3\'bd\'c7\'a3\'ac\'d3\'c3
+\f0  AI 
+\f1 \'cc\'e6\'b4\'fa\'c8\'cb\'b9\'a4\'cd\'b3\'bc\'c6\'d3\'eb\'b7\'d6\'ce\'f6\'a3\'ac\'ca\'b5\'cf\'d6\'a1\'b8\'ca\'fd\'be\'dd\'c7\'fd\'b6\'af\'b9\'dc\'c0\'ed\'a1\'b9\'a3\'ac\'b6\'f8\'b7\'c7\'a1\'b8\'be\'ad\'d1\'e9\'c7\'fd\'b6\'af\'b9\'dc\'c0\'ed\'a1\'b9\'a1\'a3
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'ba\'cb\'d0\'c4\'b9\'a6\'c4\'dc\'c9\'e8\'bc\'c6
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls5\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b9\'dc\'c0\'ed\'d5\'df\'c8\'ab\'be\'d6\'bc\'dd\'ca\'bb\'b2\'d5\'a3\'ac\'d2\'bb\'d2\'b3\'d5\'c6\'ce\'d5\'c8\'ab\'c3\'b2
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'bc\'ab\'bc\'f2\'d2\'c7\'b1\'ed\'c5\'cc\'a3\'ac\'ce\'de\'c8\'df\'d3\'e0\'ca\'fd\'be\'dd\'a3\'ac\'d6\'bb\'d5\'b9\'ca\'be\'b9\'dc\'c0\'ed\'d5\'df\'d7\'ee\'b9\'d8\'d0\'c4\'b5\'c4\'ba\'cb\'d0\'c4\'d0\'c5\'cf\'a2\'a3\'ac\'d2\'bb\'d1\'db\'bf\'b4\'c7\'e5\'cf\'ee\'c4\'bf\'c8\'ab\'be\'d6\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls5\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ba\'cb\'d0\'c4\'ca\'fd\'be\'dd\'bf\'b4\'b0\'e5\'a3\'ba\'b1\'b1\'bc\'ab\'d0\'c7\'c4\'bf\'b1\'ea\'bd\'f8\'b6\'c8\'a1\'a2\'b5\'fc\'b4\'fa\'b3\'d8\'bd\'bb\'b8\'b6\'c7\'e9\'bf\'f6\'a1\'a2\'b7\'e7\'cf\'d5\'d4\'a4\'be\'af\'c7\'e5\'b5\'a5\'a1\'a2\'cd\'c5\'b6\'d3\'c8\'cb\'c1\'a6\'b8\'ba\'d4\'d8\'a1\'a2\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'d3\'b0\'cf\'ec\'a1\'a2\'ca\'d0\'b3\'a1\'b7\'b4\'c0\'a1\'bb\'e3\'d7\'dc\'a3\'ac\'ce\'de\'d0\'e8\'b7\'ad\'b6\'e0\'b8\'f6\'d2\'b3\'c3\'e6\'a1\'a2\'ce\'de\'d0\'e8\'b0\'a4\'b8\'f6\'d5\'d2\'c8\'cb\'ce\'ca\'bd\'f8\'b6\'c8\'a1\'a3
+\f0 \cb1 \
+\ls5\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 AI 
+\f1 \'c3\'bf\'c8\'d5
+\f0  / 
+\f1 \'c3\'bf\'d6\'dc\'b9\'dc\'c0\'ed\'bc\'f2\'b1\'a8\'a3\'ba\'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'b9\'dc\'c0\'ed\'bd\'e1\'c2\'db\'a3\'ac\'b6\'f8\'b7\'c7\'b5\'a5\'b4\'bf\'b5\'c4\'ca\'fd\'be\'dd\'c2\'de\'c1\'d0\'a1\'a3\'c0\'fd\'a3\'ba\'a1\'b8
+\f0 XX 
+\f1 \'cd\'ac\'d1\'a7\'b8\'ba\'d4\'d8\'b3\'ac\'b9\'fd
+\f0  120%
+\f1 \'a3\'ac\'bd\'a8\'d2\'e9\'bd\'ab
+\f0  XX 
+\f1 \'b5\'cd\'d3\'c5\'cf\'c8\'bc\'b6\'c8\'ce\'ce\'f1\'b5\'f7\'d5\'fb\'b8\'f8
+\f0  XX 
+\f1 \'cd\'ac\'d1\'a7\'a3\'ac\'b1\'dc\'c3\'e2\'cf\'ee\'c4\'bf\'d1\'d3\'c6\'da\'a1\'b9\'a1\'b8\'b1\'be\'d6\'dc\'d0\'e8\'c7\'f3\'b1\'e4\'b8\'fc\'c2\'ca\'b4\'ef
+\f0  35%
+\f1 \'a3\'ac\'bd\'a8\'d2\'e9\'d3\'c5\'bb\'af\'d0\'e8\'c7\'f3\'c6\'c0\'b9\'c0\'c1\'f7\'b3\'cc\'a3\'ac\'bc\'f5\'c9\'d9\'c6\'b5\'b7\'b1\'b1\'e4\'b8\'fc\'b6\'d4\'cd\'c5\'b6\'d3\'bd\'da\'d7\'e0\'b5\'c4\'d3\'b0\'cf\'ec\'a1\'b9\'a3\'ac\'d6\'b1\'bd\'d3\'b8\'f8\'b3\'f6\'bf\'c9\'c2\'e4\'b5\'d8\'b5\'c4\'b9\'dc\'c0\'ed\'bd\'a8\'d2\'e9\'a3\'ac\'b8\'a8\'d6\'fa\'be\'f6\'b2\'df\'a1\'a3
+\f0 \cb1 \
+\ls5\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'cd\'c5\'b6\'d3\'bd\'a1\'bf\'b5\'b6\'c8\'b7\'d6\'ce\'f6\'a3\'ba
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'b7\'d6\'ce\'f6\'cd\'c5\'b6\'d3\'b5\'c4\'d0\'ad\'d7\'f7\'d0\'a7\'c2\'ca\'a1\'a2\'b5\'fc\'b4\'fa\'bd\'bb\'b8\'b6\'c2\'ca\'a1\'a2\'d0\'e8\'c7\'f3\'cf\'ec\'d3\'a6\'cb\'d9\'b6\'c8\'a1\'a2
+\f0 bug 
+\f1 \'c2\'ca\'a1\'a2\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb\'c7\'e9\'bf\'f6\'a3\'ac\'cc\'e1\'c7\'b0\'b7\'a2\'cf\'d6\'cd\'c5\'b6\'d3\'b5\'c4\'c7\'b1\'d4\'da\'ce\'ca\'cc\'e2\'a3\'ac\'b1\'c8\'c8\'e7\'b9\'b5\'cd\'a8\'c4\'da\'ba\'c4\'a1\'a2\'d0\'e8\'c7\'f3\'d6\'ca\'c1\'bf\'b2\'ee\'a1\'a2\'c8\'cb\'c1\'a6\'b7\'d6\'c5\'e4\'b2\'bb\'be\'f9\'a3\'ac\'b6\'f8\'b7\'c7\'b5\'c8\'b3\'f6\'c1\'cb\'ce\'ca\'cc\'e2\'d4\'d9\'b2\'b9\'be\'c8\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls5\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls5\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'bc\'ab\'cb\'d9\'b8\'b4\'c5\'cc\'b1\'d5\'bb\'b7\'a3\'ac\'ca\'b5\'cf\'d6\'b3\'d6\'d0\'f8\'b5\'fc\'b4\'fa\'d3\'c5\'bb\'af
+\f0\b0 \cb1 \uc0\u8232 \cb3 AI 
+\f1 \'ca\'b1\'b4\'fa\'b5\'c4\'b8\'b4\'c5\'cc\'a3\'ac\'b2\'bb\'c4\'dc\'b5\'c8\'d2\'bb\'b8\'f6\'bc\'be\'b6\'c8\'a1\'a2\'c9\'f5\'d6\'c1\'d2\'bb\'b8\'f6\'b5\'fc\'b4\'fa\'a3\'ac\'d2\'aa\'d7\'f6\'b5\'bd\'a1\'b8\'c9\'cf\'cf\'df\'bc\'b4\'b8\'b4\'c5\'cc\'a1\'a2\'b1\'e4\'bb\'af\'bc\'b4\'b8\'b4\'c5\'cc\'a1\'b9\'a3\'ac\'d3\'c3\'c6\'bd\'cc\'a8\'ca\'b5\'cf\'d6\'a1\'b8\'b8\'b4\'c5\'cc
+\f0  - 
+\f1 \'b8\'c4\'bd\'f8
+\f0  - 
+\f1 \'c2\'e4\'b5\'d8
+\f0  - 
+\f1 \'d1\'e9\'d6\'a4\'a1\'b9\'b5\'c4\'cd\'ea\'d5\'fb\'b1\'d5\'bb\'b7\'a3\'ac\'c8\'c3\'cd\'c5\'b6\'d3\'d4\'bd\'d7\'f6\'d4\'bd\'bf\'ec\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls5\ilvl1
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 AI 
+\f3 \'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'b8\'b4\'c5\'cc\'b1\'a8\'b8\'e6
+\f1\b0 \'a3\'ba\'c3\'bf\'b8\'f6\'b5\'fc\'b4\'fa
+\f0  / 
+\f1 \'b9\'a6\'c4\'dc\'c9\'cf\'cf\'df\'ba\'f3\'a3\'ac
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'c0\'ad\'c8\'a1\'c8\'ce\'ce\'f1\'ca\'fd\'be\'dd\'a1\'a2\'b4\'fa\'c2\'eb\'ca\'fd\'be\'dd\'a1\'a2\'b2\'e2\'ca\'d4\'ca\'fd\'be\'dd\'a1\'a2\'c9\'cf\'cf\'df\'d0\'a7\'b9\'fb\'a1\'a2\'ca\'d0\'b3\'a1\'b7\'b4\'c0\'a1\'a3\'ac\'c9\'fa\'b3\'c9\'b8\'b4\'c5\'cc\'b3\'f5\'b8\'e5\'a3\'ac\'d7\'dc\'bd\'e1\'d7\'f6\'b5\'c3\'ba\'c3\'b5\'c4\'b5\'d8\'b7\'bd\'a1\'a2\'ba\'cb\'d0\'c4\'ce\'ca\'cc\'e2\'a1\'a2\'b8\'f9\'d2\'f2\'b7\'d6\'ce\'f6\'a1\'a2\'b8\'c4\'bd\'f8\'bd\'a8\'d2\'e9\'a3\'ac\'cd\'c5\'b6\'d3\'bd\'f6\'d0\'e8\'d4\'da\'b3\'f5\'b8\'e5\'c9\'cf\'b2\'b9\'b3\'e4\'b5\'f7\'d5\'fb\'a3\'ac\'ce\'de\'d0\'e8\'ca\'d6\'b6\'af\'d5\'fb\'c0\'ed\'ca\'fd\'be\'dd\'a3\'ac\'b0\'d1\'b8\'b4\'c5\'cc\'ca\'b1\'bc\'e4\'b4\'d3
+\f0  2 
+\f1 \'d0\'a1\'ca\'b1\'d1\'b9\'cb\'f5\'b5\'bd
+\f0  30 
+\f1 \'b7\'d6\'d6\'d3\'a1\'a3
+\f0 \cb1 \
+\ls5\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b8\'b4\'c5\'cc\'bd\'e1\'c2\'db\'d7\'d4\'b6\'af\'c2\'e4\'b5\'d8
+\f1\b0 \'a3\'ba\'b8\'b4\'c5\'cc\'b3\'f6\'c0\'b4\'b5\'c4\'ce\'ca\'cc\'e2\'a1\'a2\'b8\'c4\'bd\'f8\'b4\'eb\'ca\'a9\'a3\'ac\'d6\'b1\'bd\'d3\'c9\'fa\'b3\'c9\'c8\'ce\'ce\'f1\'a3\'ac\'bd\'f8\'c8\'eb\'b5\'fc\'b4\'fa\'b3\'d8\'a3\'ac\'b9\'d8\'c1\'aa\'b6\'d4\'d3\'a6\'b5\'c4\'cf\'ee\'c4\'bf\'d3\'eb\'b8\'ba\'d4\'f0\'c8\'cb\'a3\'ac\'cf\'b5\'cd\'b3\'d7\'d4\'b6\'af\'b8\'fa\'d7\'d9\'bd\'f8\'b6\'c8\'a3\'ac\'cd\'ea\'b3\'c9\'ba\'f3\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'b5\'bd\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'ac\'b3\'b9\'b5\'d7\'b6\'c5\'be\'f8\'a1\'b8\'b8\'b4\'c5\'cc\'c1\'cb\'be\'cd\'cd\'fc\'c1\'cb\'a3\'ac\'ce\'ca\'cc\'e2\'cf\'c2\'b4\'ce\'bb\'b9\'b7\'b8\'a1\'b9\'b5\'c4\'cd\'e7\'bc\'b2\'a1\'a3
+\f0 \cb1 \
+\ls5\ilvl1
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'d7\'a8\'cf\'ee\'b8\'b4\'c5\'cc
+\f1\b0 \'a3\'ba\'c3\'bf\'b4\'ce\'b4\'f3\'b5\'c4\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'c2\'e4\'b5\'d8\'ba\'f3\'a3\'ac
+\f0 AI 
+\f1 \'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'d7\'a8\'cf\'ee\'b8\'b4\'c5\'cc\'b1\'a8\'b8\'e6\'a3\'ac\'d7\'dc\'bd\'e1\'b1\'be\'b4\'ce\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'b4\'f8\'c0\'b4\'b5\'c4\'ca\'d5\'d2\'e6\'a1\'a2\'d3\'f6\'b5\'bd\'b5\'c4\'bf\'d3\'a1\'a2\'d3\'c5\'bb\'af\'bf\'d5\'bc\'e4\'a1\'a2\'bc\'bc\'ca\'f5\'b3\'c1\'b5\'ed\'a3\'ac\'b4\'e6\'c8\'eb\'cd\'c5\'b6\'d3\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'ac\'c8\'c3\'cd\'c5\'b6\'d3\'b5\'c4\'c4\'a3\'d0\'cd\'ca\'ca\'c5\'e4\'be\'ad\'d1\'e9\'b3\'d6\'d0\'f8\'bb\'fd\'c0\'db\'a3\'ac\'ba\'f3\'d0\'f8\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'b5\'c4\'c2\'e4\'b5\'d8\'cb\'d9\'b6\'c8\'d4\'bd\'c0\'b4\'d4\'bd\'bf\'ec\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls5\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls5\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'cd\'c5\'b6\'d3\'c4\'dc\'c1\'a6\'d3\'eb\'b8\'ba\'d4\'d8\'b9\'dc\'c0\'ed
+\f0\b0 \cb1 \uc0\u8232 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'d2\'bb\'c8\'cb\'b6\'e0\'d4\'f0\'a3\'ac\'d7\'ee\'c8\'dd\'d2\'d7\'b3\'f6\'cf\'d6\'b8\'ba\'d4\'d8\'b2\'bb\'be\'f9\'a1\'a2\'c4\'dc\'c1\'a6\'b4\'ed\'c5\'e4\'b5\'c4\'ce\'ca\'cc\'e2\'a3\'ac\'c6\'bd\'cc\'a8\'d7\'d4\'b6\'af\'ca\'b5\'cf\'d6\'be\'ab\'cf\'b8\'bb\'af\'b9\'dc\'c0\'ed\'a3\'ba
+\f0 \cb1 \uc0\u8232 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls5\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'cd\'c5\'b6\'d3\'b8\'ba\'d4\'d8\'ca\'b5\'ca\'b1\'bc\'e0\'bf\'d8\'a3\'ba\'d7\'d4\'b6\'af\'cd\'b3\'bc\'c6\'c3\'bf\'b8\'f6\'c8\'cb\'b5\'c4\'c8\'ce\'ce\'f1\'c1\'bf\'a1\'a2\'b9\'a4\'ca\'b1\'a1\'a2\'b8\'ba\'d4\'d8\'c7\'e9\'bf\'f6\'a3\'ac\'b8\'ba\'d4\'d8\'b9\'fd\'b8\'df
+\f0  / 
+\f1 \'b9\'fd\'b5\'cd\'d7\'d4\'b6\'af\'b8\'df\'c1\'c1\'a3\'ac\'b9\'dc\'c0\'ed\'d5\'df\'bf\'c9\'bf\'ec\'cb\'d9\'b5\'f7\'d5\'fb\'c8\'ce\'ce\'f1\'b7\'d6\'c5\'e4\'a3\'ac\'b1\'dc\'c3\'e2\'a1\'b8\'c3\'a6\'b5\'c4\'c3\'a6\'cb\'c0\'a1\'a2\'cf\'d0\'b5\'c4\'cf\'d0\'cb\'c0\'a1\'b9\'a3\'ac\'d7\'ee\'b4\'f3\'bb\'af
+\f0  10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'c8\'cb\'d0\'a7\'a1\'a3
+\f0 \cb1 \
+\ls5\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'cd\'c5\'b6\'d3\'c4\'dc\'c1\'a6\'cd\'bc\'c6\'d7\'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'a3\'ba
+\f0 AI 
+\f1 \'b8\'f9\'be\'dd\'c3\'bf\'b8\'f6\'c8\'cb\'cd\'ea\'b3\'c9\'b5\'c4\'c8\'ce\'ce\'f1\'a1\'a2\'b3\'c1\'b5\'ed\'b5\'c4\'ce\'c4\'b5\'b5\'a1\'a2\'bd\'e2\'be\'f6\'b5\'c4\'ce\'ca\'cc\'e2\'a3\'ac\'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'cd\'c5\'b6\'d3\'c4\'dc\'c1\'a6\'cd\'bc\'c6\'d7\'a3\'ac\'b1\'c8\'c8\'e7\'cb\'ad\'c9\'c3\'b3\'a4\'c4\'a3\'d0\'cd\'ce\'a2\'b5\'f7\'a1\'a2\'cb\'ad\'c9\'c3\'b3\'a4\'b8\'df\'b2\'a2\'b7\'a2\'bf\'aa\'b7\'a2\'a1\'a2\'cb\'ad\'c9\'c3\'b3\'a4\'c7\'b0\'b6\'cb\'bd\'bb\'bb\'a5\'a3\'ac\'d3\'f6\'b5\'bd\'d0\'c2\'b5\'c4\'d0\'e8\'c7\'f3
+\f0  / 
+\f1 \'ca\'d0\'b3\'a1\'b1\'e4\'bb\'af\'a3\'ac\'bf\'c9\'bf\'ec\'cb\'d9\'b7\'d6\'c5\'e4\'b8\'f8\'d7\'ee\'ba\'cf\'ca\'ca\'b5\'c4\'c8\'cb\'a3\'ac\'b4\'f3\'b7\'f9\'cc\'e1\'c9\'fd\'cf\'ec\'d3\'a6\'cb\'d9\'b6\'c8\'d3\'eb\'bd\'bb\'b8\'b6\'d6\'ca\'c1\'bf\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls5\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	6	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c8\'fd\'a1\'a2\'ca\'ca\'c5\'e4
+\f0  10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'c2\'e4\'b5\'d8\'c2\'b7\'be\'b6\'d3\'eb\'bc\'bc\'ca\'f5\'d1\'a1\'d0\'cd
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d3\'c5\'cf\'c8\'b7\'bd\'b0\'b8\'a3\'ba\'bf\'aa\'d4\'b4\'b9\'a4\'be\'df\'d7\'e9\'ba\'cf
+\f0  + 
+\f1 \'c7\'e1\'c1\'bf\'b6\'fe\'b4\'ce\'bf\'aa\'b7\'a2\'a3\'a8
+\f0 2 
+\f1 \'c8\'cb\'bf\'aa\'b7\'a2\'a3\'ac
+\f0 2 
+\f1 \'b8\'f6\'d4\'c2\'c8\'ab\'c1\'bf\'c9\'cf\'cf\'df\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'ce\'de\'d0\'e8\'b4\'d3\'c1\'e3\'d7\'d4\'d1\'d0\'a3\'ac\'d3\'c5\'cf\'c8\'d1\'a1\'d4\'f1\'b3\'c9\'ca\'ec\'bf\'aa\'d4\'b4\'b9\'a4\'be\'df\'d7\'f6\'d7\'e9\'ba\'cf\'d3\'eb\'b6\'fe\'b4\'ce\'bf\'aa\'b7\'a2\'a3\'ac\'be\'db\'bd\'b9\'ba\'cb\'d0\'c4\'b5\'c4
+\f0  AI 
+\f1 \'c4\'dc\'c1\'a6\'a1\'a2\'c1\'f7\'b3\'cc\'b1\'d5\'bb\'b7\'a1\'a2\'ca\'fd\'be\'dd\'b4\'f2\'cd\'a8\'a3\'ac\'d7\'ee\'bf\'ec\'cb\'d9\'b6\'c8\'c2\'e4\'b5\'d8\'a3\'ac\'cf\'c8\'c5\'dc\'cd\'a8\'c1\'f7\'b3\'cc\'d4\'d9\'d3\'c5\'bb\'af\'b6\'a8\'d6\'c6\'a1\'a3
+\f0 \cb1 \
+\
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'cd\'c6\'bc\'f6\'bf\'aa\'d4\'b4\'d7\'e9\'ba\'cf\'a3\'a8\'bf\'aa\'cf\'e4\'bc\'b4\'d3\'c3\'a1\'a2
+\f0 AI 
+\f1 \'d3\'d1\'ba\'c3\'a1\'a2\'b2\'bf\'ca\'f0\'bc\'f2\'b5\'a5\'a1\'a2\'b6\'fe\'b4\'ce\'bf\'aa\'b7\'a2\'b3\'c9\'b1\'be\'b5\'cd\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b1\'ed\'b8\'f1
+\f0 \cb1 \
+\
+\
+\
+
+\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrt\brdrnil \trbrdrl\brdrnil \trbrdrr\brdrnil 
+\clvertalc \clshdrawnil \clwWidth1568\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx2880
+\clvertalc \clshdrawnil \clwWidth1306\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx5760
+\clvertalc \clshdrawnil \clwWidth10725\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx8640
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f3\b \cf2 \cb3 \'ba\'cb\'d0\'c4\'c4\'a3\'bf\'e9
+\f2 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f3 \cf2 \cb3 \'d1\'a1\'d0\'cd\'cd\'c6\'bc\'f6
+\f2 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f3 \cf2 \cb3 \'ba\'cb\'d0\'c4\'c0\'ed\'d3\'c9
+\f2 \cb1 \cell \row
+
+\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil 
+\clvertalc \clshdrawnil \clwWidth1568\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx2880
+\clvertalc \clshdrawnil \clwWidth1306\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx5760
+\clvertalc \clshdrawnil \clwWidth10725\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx8640
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1\b0 \cf2 \cb3 \'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed
+\f0  + 
+\f1 \'d6\'aa\'ca\'b6\'bf\'e2\'b5\'d7\'d7\'f9
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+\cf2 \cb3 Plane\cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'bf\'aa\'d4\'b4\'cc\'e6\'b4\'fa
+\f0  Jira+Notion
+\f1 \'a3\'ac\'d7\'d4\'b4\'f8\'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'a1\'a2\'bf\'b4\'b0\'e5\'a1\'a2
+\f0 OKR
+\f1 \'a1\'a2\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'ac
+\f0 API 
+\f1 \'cd\'ea\'c9\'c6\'a3\'ac\'c7\'e1\'c1\'bf\'d2\'d7\'b2\'bf\'ca\'f0\'a3\'ac\'cd\'ea\'c3\'c0\'ca\'ca\'c5\'e4
+\f0  10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'a3\'ac\'d6\'a7\'b3\'d6\'d7\'d4\'b6\'a8\'d2\'e5\'d7\'d6\'b6\'ce\'d3\'eb\'c1\'f7\'b3\'cc\'a3\'ac\'b6\'fe\'b4\'ce\'bf\'aa\'b7\'a2\'d3\'d1\'ba\'c3
+\f0 \cb1 \cell \row
+
+\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil 
+\clvertalc \clshdrawnil \clwWidth1568\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx2880
+\clvertalc \clshdrawnil \clwWidth1306\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx5760
+\clvertalc \clshdrawnil \clwWidth10725\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx8640
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b1\'be\'b5\'d8
+\f0  AI 
+\f1 \'d2\'fd\'c7\'e6
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+\cf2 \cb3 Ollama + Spring AI\cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d2\'bb\'bc\'fc\'b2\'bf\'ca\'f0\'b1\'be\'b5\'d8\'b4\'f3\'c4\'a3\'d0\'cd\'a3\'ac\'d7\'d4\'b4\'f8
+\f0  OpenAI 
+\f1 \'bc\'e6\'c8\'dd
+\f0  API
+\f1 \'a3\'ac\'cd\'ea\'c3\'c0\'ca\'ca\'c5\'e4\'b9\'fa\'c4\'da\'bf\'aa\'d4\'b4\'c4\'a3\'d0\'cd\'a3\'ac
+\f0 Spring AI 
+\f1 \'ce\'de\'b7\'ec\'bc\'af\'b3\'c9\'a3\'ac\'bf\'ec\'cb\'d9\'ca\'b5\'cf\'d6\'c8\'ce\'ce\'f1\'b2\'f0\'bd\'e2\'a1\'a2
+\f0 RAG 
+\f1 \'ce\'ca\'b4\'f0\'a1\'a2\'d6\'c7\'c4\'dc\'b7\'d6\'ce\'f6\'b5\'c8\'c4\'dc\'c1\'a6\'a3\'ac\'d4\'cb\'ce\'ac\'b3\'c9\'b1\'be\'bc\'ab\'b5\'cd
+\f0 \cb1 \cell \row
+
+\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil 
+\clvertalc \clshdrawnil \clwWidth1568\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx2880
+\clvertalc \clshdrawnil \clwWidth1306\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx5760
+\clvertalc \clshdrawnil \clwWidth10725\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx8640
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d6\'aa\'ca\'b6\'bf\'e2
+\f0  RAG 
+\f1 \'d4\'f6\'c7\'bf
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+\cf2 \cb3 MaxKB\cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b9\'fa\'b2\'fa\'bf\'aa\'d4\'b4\'d6\'aa\'ca\'b6\'bf\'e2
+\f0  RAG 
+\f1 \'c6\'bd\'cc\'a8\'a3\'ac\'cd\'ea\'c3\'c0\'bc\'e6\'c8\'dd
+\f0  Ollama
+\f1 \'a3\'ac\'d2\'bb\'bc\'fc\'b2\'bf\'ca\'f0\'a3\'ac\'ba\'cd
+\f0  Plane 
+\f1 \'cb\'ab\'cf\'f2\'b4\'f2\'cd\'a8\'a3\'ac\'ca\'b5\'cf\'d6\'ce\'c4\'b5\'b5\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'a1\'a2\'cf\'f2\'c1\'bf\'bb\'af\'a1\'a2\'d6\'c7\'c4\'dc\'ce\'ca\'b4\'f0\'a3\'ac\'ce\'de\'d0\'e8\'b4\'d3\'c1\'e3\'bf\'aa\'b7\'a2
+\f0  RAG 
+\f1 \'c1\'f7\'b3\'cc
+\f0 \cb1 \cell \row
+
+\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil 
+\clvertalc \clshdrawnil \clwWidth1568\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx2880
+\clvertalc \clshdrawnil \clwWidth1306\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx5760
+\clvertalc \clshdrawnil \clwWidth10725\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx8640
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'d7\'a5\'c8\'a1
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+\cf2 \cb3 Python + 
+\f1 \'b6\'a8\'ca\'b1\'c8\'ce\'ce\'f1
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c7\'e1\'c1\'bf\'c5\'c0\'b3\'e6\'ca\'b5\'cf\'d6\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'a1\'a2\'be\'ba\'c6\'b7\'d0\'c5\'cf\'a2\'d7\'d4\'b6\'af\'d7\'a5\'c8\'a1\'a3\'ac\'cd\'a8\'b9\'fd
+\f0  API 
+\f1 \'cd\'ac\'b2\'bd\'b5\'bd
+\f0  Plane 
+\f1 \'b5\'c4\'d0\'e8\'c7\'f3\'b3\'d8\'a3\'ac\'bf\'aa\'b7\'a2\'c1\'bf\'bc\'ab\'d0\'a1
+\f0 \cb1 \cell \row
+
+\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrt\brdrnil \trbrdrr\brdrnil 
+\clvertalc \clshdrawnil \clwWidth1568\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx2880
+\clvertalc \clshdrawnil \clwWidth1306\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx5760
+\clvertalc \clshdrawnil \clwWidth10725\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadl0 \clpadr0 \gaph\cellx8640
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b5\'da\'c8\'fd\'b7\'bd\'bc\'af\'b3\'c9
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+\cf2 \cb3 Plane 
+\f1 \'d4\'ad\'c9\'fa\'bc\'af\'b3\'c9
+\f0 \cb1 \cell 
+\pard\intbl\itap1\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'d7\'d4\'b4\'f8
+\f0  GitLab/Gitee
+\f1 \'a1\'a2\'b7\'c9\'ca\'e9
+\f0  / 
+\f1 \'c6\'f3\'d2\'b5\'ce\'a2\'d0\'c5\'a1\'a2
+\f0 Jenkins 
+\f1 \'b4\'f2\'cd\'a8\'a3\'ac\'ca\'b5\'cf\'d6\'b4\'fa\'c2\'eb\'cc\'e1\'bd\'bb\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'c8\'ce\'ce\'f1\'bd\'f8\'b6\'c8\'a1\'a2\'cf\'fb\'cf\'a2\'d7\'d4\'b6\'af\'cd\'c6\'cb\'cd\'a3\'ac\'ce\'de\'d0\'e8\'b6\'ee\'cd\'e2\'bf\'aa\'b7\'a2
+\f0 \cb1 \cell \lastrow\row
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b7\'d6\'bd\'d7\'b6\'ce\'c2\'e4\'b5\'d8\'c2\'b7\'be\'b6\'a3\'a8
+\f0 2 
+\f1 \'c8\'cb\'b7\'d6\'b9\'a4\'a3\'ac
+\f0 8 
+\f1 \'d6\'dc\'c9\'cf\'cf\'df\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls6\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'bd\'d7\'b6\'ce
+\f2  1
+\f3 \'a3\'ba\'bb\'f9\'b4\'a1\'b5\'d7\'d7\'f9\'b4\'ee\'bd\'a8\'d3\'eb\'c1\'f7\'b3\'cc\'c5\'dc\'cd\'a8\'a3\'a8
+\f2 2 
+\f3 \'d6\'dc\'a3\'a9
+\f0\b0 \cb1 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls6\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b7\'d6\'b9\'a4\'a3\'ba\'c7\'b0\'ba\'f3\'b6\'cb\'c5\'e4\'ba\'cf\'a3\'ac\'cd\'ea\'b3\'c9
+\f0  Plane
+\f1 \'a1\'a2
+\f0 Ollama
+\f1 \'a1\'a2
+\f0 MaxKB 
+\f1 \'b5\'c4\'b2\'bf\'ca\'f0\'a3\'ac\'b4\'f2\'cd\'a8\'b7\'c9\'ca\'e9
+\f0  / 
+\f1 \'c6\'f3\'d2\'b5\'ce\'a2\'d0\'c5\'a1\'a2\'b4\'fa\'c2\'eb\'b2\'d6\'bf\'e2\'a3\'ac\'c5\'e4\'d6\'c3\'ba\'c3\'bb\'f9\'b4\'a1\'b5\'c4\'cf\'ee\'c4\'bf\'a1\'a2\'c8\'ce\'ce\'f1\'a1\'a2\'d6\'aa\'ca\'b6\'bf\'e2\'a1\'a2
+\f0 OKR 
+\f1 \'c4\'a3\'bf\'e9\'a1\'a3
+\f0 \cb1 \
+\ls6\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ba\'cb\'d0\'c4\'b2\'fa\'b3\'f6\'a3\'ba\'c6\'bd\'cc\'a8\'bb\'f9\'b4\'a1\'bb\'b7\'be\'b3\'bf\'c9\'d3\'c3\'a3\'ac\'cd\'c5\'b6\'d3\'c5\'dc\'cd\'a8\'a1\'b8\'d0\'e8\'c7\'f3
+\f0  - 
+\f1 \'c8\'ce\'ce\'f1
+\f0  - 
+\f1 \'bd\'bb\'b8\'b6\'a1\'b9\'b5\'c4\'bb\'f9\'b4\'a1\'c1\'f7\'b3\'cc\'a3\'ac\'c8\'b7\'c8\'cf\'ba\'cb\'d0\'c4\'c1\'f7\'b3\'cc\'d3\'eb\'d0\'e8\'c7\'f3\'b1\'df\'bd\'e7\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls6\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls6\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'bd\'d7\'b6\'ce
+\f2  2
+\f3 \'a3\'ba
+\f2 AI 
+\f3 \'ba\'cb\'d0\'c4\'c4\'dc\'c1\'a6\'bf\'aa\'b7\'a2\'d3\'eb\'bc\'af\'b3\'c9\'a3\'a8
+\f2 4 
+\f3 \'d6\'dc\'a3\'a9
+\f0\b0 \cb1 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls6\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b7\'d6\'b9\'a4\'a3\'ba\'ba\'f3\'b6\'cb\'b8\'ba\'d4\'f0\'c8\'cb\'d6\'f7\'b9\'a5
+\f0  AI 
+\f1 \'d2\'fd\'c7\'e6\'a1\'a2\'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'d7\'a5\'c8\'a1\'a1\'a2\'b5\'da\'c8\'fd\'b7\'bd\'bc\'af\'b3\'c9\'a1\'a2\'d7\'d4\'b6\'af\'bb\'af\'c1\'f7\'b3\'cc\'bf\'aa\'b7\'a2\'a3\'bb\'c7\'b0\'b6\'cb\'b8\'ba\'d4\'f0\'c8\'cb\'d6\'f7\'b9\'a5\'bd\'bb\'bb\'a5\'d3\'c5\'bb\'af\'a1\'a2
+\f0 AI 
+\f1 \'b6\'d4\'bb\'b0\'c8\'eb\'bf\'da\'a1\'a2\'bc\'dd\'ca\'bb\'b2\'d5\'d2\'b3\'c3\'e6\'bf\'aa\'b7\'a2\'a1\'a3
+\f0 \cb1 \
+\ls6\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ba\'cb\'d0\'c4\'b2\'fa\'b3\'f6\'a3\'ba\'cd\'ea\'b3\'c9\'b1\'e4\'bb\'af\'d0\'c5\'ba\'c5\'b9\'e9\'bc\'af\'a1\'a2
+\f0 AI 
+\f1 \'c8\'ce\'ce\'f1\'b2\'f0\'bd\'e2\'a1\'a2\'bd\'f8\'b6\'c8\'d7\'d4\'b6\'af\'cd\'ac\'b2\'bd\'a1\'a2\'b7\'e7\'cf\'d5\'d4\'a4\'be\'af\'a1\'a2
+\f0 RAG 
+\f1 \'d6\'c7\'c4\'dc\'ce\'ca\'b4\'f0\'a1\'a2\'b9\'dc\'c0\'ed\'bc\'dd\'ca\'bb\'b2\'d5\'b5\'c4\'ba\'cb\'d0\'c4\'b9\'a6\'c4\'dc\'a3\'ac\'c8\'ab\'c1\'b4\'c2\'b7\'b4\'f2\'cd\'a8\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls6\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls6\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'bd\'d7\'b6\'ce
+\f2  3
+\f3 \'a3\'ba\'d3\'c5\'bb\'af\'b5\'fc\'b4\'fa\'d3\'eb\'c8\'ab\'c1\'bf\'c9\'cf\'cf\'df\'a3\'a8
+\f2 2 
+\f3 \'d6\'dc\'a3\'a9
+\f0\b0 \cb1 \
+\pard\tx940\tx1440\pardeftab720\li1440\fi-1440\partightenfactor0
+\ls6\ilvl1\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b7\'d6\'b9\'a4\'a3\'ba\'c1\'bd\'c8\'cb\'c5\'e4\'ba\'cf\'cd\'ea\'b3\'c9\'c8\'ab\'c1\'f7\'b3\'cc\'b2\'e2\'ca\'d4\'a1\'a2
+\f0 AI 
+\f1 \'c4\'dc\'c1\'a6\'b5\'f7\'d3\'c5\'a1\'a2
+\f0 Prompt 
+\f1 \'d3\'c5\'bb\'af\'a3\'ac\'cd\'c5\'b6\'d3\'d0\'a1\'b7\'b6\'ce\'a7\'ca\'d4\'d3\'c3\'a3\'ac\'ca\'d5\'bc\'af\'b7\'b4\'c0\'a1\'bf\'ec\'cb\'d9\'b5\'fc\'b4\'fa\'a3\'ac\'cd\'ea\'b3\'c9\'d7\'ee\'d6\'d5\'d3\'c5\'bb\'af\'a1\'a3
+\f0 \cb1 \
+\ls6\ilvl1\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u9702 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ba\'cb\'d0\'c4\'b2\'fa\'b3\'f6\'a3\'ba\'c6\'bd\'cc\'a8\'c8\'ab\'c1\'bf\'c9\'cf\'cf\'df\'a3\'ac\'cd\'c5\'b6\'d3\'c8\'ab\'d4\'b1\'ca\'b9\'d3\'c3\'a3\'ac\'cd\'ea\'b3\'c9\'cd\'ea\'d5\'fb\'b5\'c4\'a1\'b8\'b1\'e4\'bb\'af
+\f0  - 
+\f1 \'be\'f6\'b2\'df
+\f0  - 
+\f1 \'d6\'b4\'d0\'d0
+\f0  - 
+\f1 \'b8\'b4\'c5\'cc
+\f0  - 
+\f1 \'b3\'c1\'b5\'ed\'a1\'b9\'b1\'d5\'bb\'b7\'a3\'ac\'c5\'e4\'cc\'d7\'ca\'b9\'d3\'c3\'b9\'e6\'b7\'b6\'c2\'e4\'b5\'d8\'a1\'a3
+\f0 \cb1 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls6\ilvl0\cf2 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	6	}\expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b1\'b8\'d1\'a1\'b7\'bd\'b0\'b8\'a3\'ba\'c8\'ab\'d7\'d4\'d1\'d0\'a3\'a8
+\f0 4-6 
+\f1 \'b8\'f6\'d4\'c2\'c9\'cf\'cf\'df\'a3\'ac\'ca\'ca\'c5\'e4\'c7\'bf\'b6\'a8\'d6\'c6\'bb\'af\'d0\'e8\'c7\'f3\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'c8\'e7\'b9\'fb\'cd\'c5\'b6\'d3\'d3\'d0\'c7\'bf\'b6\'a8\'d6\'c6\'bb\'af\'d0\'e8\'c7\'f3\'a1\'a2\'ca\'fd\'be\'dd\'b0\'b2\'c8\'ab\'d2\'aa\'c7\'f3\'bc\'ab\'b8\'df\'a3\'ac\'bf\'c9\'bb\'f9\'d3\'da\'d6\'ae\'c7\'b0\'b5\'c4\'d7\'d4\'d1\'d0\'bc\'dc\'b9\'b9\'c9\'fd\'bc\'b6\'a3\'ac\'bc\'bc\'ca\'f5\'d5\'bb\'b1\'a3\'b3\'d6\'d2\'bb\'d6\'c2\'a3\'ac\'be\'db\'bd\'b9\'ba\'cb\'d0\'c4\'c4\'a3\'bf\'e9\'bf\'aa\'b7\'a2\'a3\'ba
+\f0 \cb1 \
+\
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls7\ilvl0\cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u8226 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'c7\'b0\'b6\'cb\'a3\'ba
+\f0 Vue3 + Vite + Element Plus
+\f1 \'a3\'ac\'b1\'a3\'b3\'d6\'cd\'b3\'d2\'bb\'c9\'fa\'cc\'ac\'a3\'ac\'bd\'b5\'b5\'cd\'bf\'aa\'b7\'a2\'b3\'c9\'b1\'be
+\f0 \cb1 \
+\ls7\ilvl0\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u8226 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ba\'f3\'b6\'cb\'a3\'ba
+\f0 Spring Boot 3.x + MyBatis-Plus + Spring AI
+\f1 \'a3\'ac\'b3\'c9\'ca\'ec\'ce\'c8\'b6\'a8\'a3\'ac
+\f0 AI 
+\f1 \'bc\'af\'b3\'c9\'d3\'d1\'ba\'c3
+\f0 \cb1 \
+\ls7\ilvl0\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u8226 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ca\'fd\'be\'dd\'bf\'e2\'a3\'ba
+\f0 PostgreSQL 16+ pgvector
+\f1 \'a3\'ac\'d2\'bb\'cc\'d7\'ca\'fd\'be\'dd\'bf\'e2\'b8\'e3\'b6\'a8\'bd\'e1\'b9\'b9\'bb\'af\'ca\'fd\'be\'dd\'a1\'a2\'c8\'ab\'ce\'c4\'bc\'ec\'cb\'f7\'a1\'a2\'cf\'f2\'c1\'bf\'b4\'e6\'b4\'a2\'a3\'ac\'bc\'ab\'d6\'c2\'bd\'b5\'b5\'cd\'d4\'cb\'ce\'ac\'b3\'c9\'b1\'be
+\f0 \cb1 \
+\ls7\ilvl0\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u8226 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b1\'be\'b5\'d8\'c4\'a3\'d0\'cd\'a3\'ba
+\f0 Ollama + Qwen2-7B/14B 
+\f1 \'d6\'d0\'ce\'c4\'d4\'f6\'c7\'bf\'b0\'e6
+\f0  + bge-small-zh-v1.5 Embedding 
+\f1 \'c4\'a3\'d0\'cd\'a3\'ac\'ca\'ca\'c5\'e4
+\f0  10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'d3\'b2\'bc\'fe\'c5\'e4\'d6\'c3
+\f0 \cb1 \
+\ls7\ilvl0\cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	\uc0\u8226 	}
+\f1 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'ba\'cb\'d0\'c4\'b7\'d6\'b9\'a4\'a3\'ba\'c7\'b0\'b6\'cb\'b8\'ba\'d4\'f0\'c8\'cb\'b8\'ba\'d4\'f0\'c8\'ab\'c1\'bf\'d2\'b3\'c3\'e6\'d3\'eb\'bd\'bb\'bb\'a5\'bf\'aa\'b7\'a2\'a3\'ac\'ba\'f3\'b6\'cb\'b8\'ba\'d4\'f0\'c8\'cb\'b8\'ba\'d4\'f0\'bc\'dc\'b9\'b9\'a1\'a2\'d2\'b5\'ce\'f1\'c2\'df\'bc\'ad\'a1\'a2
+\f0 AI 
+\f1 \'c4\'dc\'c1\'a6\'a1\'a2\'d4\'cb\'ce\'ac\'b2\'bf\'ca\'f0\'a3\'ac\'c1\'bd\'c8\'cb\'bd\'bb\'b2\'e6\'c6\'c0\'c9\'f3\'a3\'ac\'d1\'cf\'bf\'d8\'d0\'e8\'c7\'f3\'b1\'df\'bd\'e7\'a1\'a3
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'cb\'c4\'a1\'a2\'ba\'cb\'d0\'c4\'b1\'dc\'bf\'d3\'d6\'b8\'c4\'cf\'a3\'a8
+\f0 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3
+\f0  AI 
+\f1 \'cf\'ee\'c4\'bf\'b9\'dc\'c0\'ed\'b5\'c4\'c9\'fa\'cb\'c0\'cf\'df\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls8\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f8\'b6\'d4\'b2\'bb\'d2\'aa\'ce\'aa\'c1\'cb
+\f2  AI 
+\f3 \'b6\'f8
+\f2  AI
+\f3 \'a3\'ac\'ba\'cb\'d0\'c4\'ca\'c7\'bc\'f5\'b8\'ba\'b6\'f8\'b7\'c7\'d4\'f6\'b8\'ba
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'d7\'ee\'b3\'a3\'bc\'fb\'b5\'c4\'bf\'d3\'a3\'ba\'b6\'d1\'c6\'f6\'d2\'bb\'b6\'d1
+\f0  AI 
+\f1 \'b9\'a6\'c4\'dc\'a3\'ac\'c8\'b4\'b2\'bb\'bd\'e2\'be\'f6\'ca\'b5\'bc\'ca\'ce\'ca\'cc\'e2\'a3\'ac\'b7\'b4\'b6\'f8\'d4\'f6\'bc\'d3\'cd\'c5\'b6\'d3\'b9\'a4\'d7\'f7\'c1\'bf\'a1\'a3\'b1\'c8\'c8\'e7
+\f0  AI 
+\f1 \'c9\'fa\'b3\'c9\'b5\'c4\'c8\'ce\'ce\'f1\'b2\'f0\'bd\'e2\'b2\'bb\'b7\'fb\'ba\'cf\'ca\'b5\'bc\'ca\'a3\'ac\'bb\'b9\'d2\'aa\'ca\'d6\'b6\'af\'b8\'c4\'b0\'eb\'cc\'ec\'a3\'bb\'c7\'bf\'d6\'c6\'d2\'aa\'c7\'f3\'cd\'c5\'b6\'d3\'ba\'cd
+\f0  AI 
+\f1 \'b6\'d4\'bb\'b0\'a3\'ac\'b7\'b4\'b6\'f8\'b6\'e0\'c1\'cb\'d2\'bb\'b2\'bd\'b2\'d9\'d7\'f7\'a1\'a3\'bc\'c7\'d7\'a1\'a3\'ba\'cb\'f9\'d3\'d0
+\f0  AI 
+\f1 \'b9\'a6\'c4\'dc\'a3\'ac\'b1\'d8\'d0\'eb\'c4\'dc\'bc\'f5\'c9\'d9\'cd\'c5\'b6\'d3\'b5\'c4\'d6\'d8\'b8\'b4\'c0\'cd\'b6\'af\'a3\'ac\'b7\'f1\'d4\'f2\'d2\'bb\'c2\'c9\'bf\'b3\'b5\'f4\'a1\'a3
+\f0 \cb1 \
+\ls8\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f8\'b6\'d4\'b2\'bb\'d2\'aa\'b8\'e3\'d6\'d8\'c1\'f7\'b3\'cc\'a1\'a2\'d6\'d8\'c9\'f3\'c5\'fa\'a3\'ac\'b1\'a3\'b3\'d6\'b1\'e2\'c6\'bd\'bb\'af\'bc\'ab\'cb\'d9\'be\'f6\'b2\'df
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b2\'bb\'d2\'aa\'d5\'d5\'b0\'e1\'b4\'f3\'b3\'a7\'b5\'c4\'a1\'b8\'d0\'e8\'c7\'f3\'c8\'fd\'bc\'b6\'c6\'c0\'c9\'f3\'a1\'a2\'b2\'e3\'b2\'e3\'c9\'f3\'c5\'fa\'a1\'b9\'c4\'a3\'ca\'bd\'a3\'ac
+\f0 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3\'b5\'c4\'ba\'cb\'d0\'c4\'d3\'c5\'ca\'c6\'be\'cd\'ca\'c7\'c1\'e9\'bb\'ee\'a3\'ac\'d2\'bb\'b8\'f6\'d0\'e8\'c7\'f3\'c9\'f3\'c5\'fa
+\f0  3 
+\f1 \'cc\'ec\'a3\'ac\'ca\'d0\'b3\'a1\'d4\'e7\'be\'cd\'b1\'e4\'c1\'cb\'a1\'a3\'bc\'e1\'b3\'d6\'a1\'b8\'cb\'ad\'b8\'ba\'d4\'f0\'cb\'ad\'be\'f6\'b2\'df\'a3\'ac
+\f0 15 
+\f1 \'b7\'d6\'d6\'d3\'c4\'da\'cd\'ea\'b3\'c9\'b6\'d4\'c6\'eb\'a3\'ac
+\f0 24 
+\f1 \'d0\'a1\'ca\'b1\'c4\'da\'c2\'e4\'b5\'d8\'d6\'b4\'d0\'d0\'a1\'b9\'a3\'ac\'c1\'f7\'b3\'cc\'d4\'bd\'bc\'f2\'b5\'a5\'a3\'ac\'cf\'ec\'d3\'a6\'cb\'d9\'b6\'c8\'d4\'bd\'bf\'ec\'a1\'a3
+\f0 \cb1 \
+\ls8\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f8\'b6\'d4\'b2\'bb\'d2\'aa\'b4\'f2\'c6\'c6\'b5\'a5\'d4\'b4\'d5\'e6\'c0\'ed\'d4\'ad\'d4\'f2\'a3\'ac\'b6\'c5\'be\'f8\'d0\'c5\'cf\'a2\'b7\'d6\'c9\'a2
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'c8\'cf\'d6\'aa\'b2\'ee\'b5\'c4\'b8\'f9\'d4\'b4\'a3\'ac\'be\'cd\'ca\'c7\'d0\'c5\'cf\'a2\'b7\'d6\'c9\'a2\'d4\'da\'b6\'e0\'b8\'f6\'b9\'a4\'be\'df\'c0\'ef\'a1\'a3\'b1\'d8\'d0\'eb\'c7\'bf\'d6\'c6\'cd\'c5\'b6\'d3\'a3\'ba\'cb\'f9\'d3\'d0\'cf\'ee\'c4\'bf\'cf\'e0\'b9\'d8\'b5\'c4\'d0\'c5\'cf\'a2\'a3\'ac\'c8\'ab\'b2\'bf\'c2\'bc\'c8\'eb\'c6\'bd\'cc\'a8\'a3\'ac\'b2\'bb\'d4\'ca\'d0\'ed\'d4\'da\'ce\'a2\'d0\'c5
+\f0  / 
+\f1 \'b7\'c9\'ca\'e9\'c0\'ef\'b6\'a8\'be\'f6\'b2\'df\'a1\'a2\'b2\'bb\'d4\'ca\'d0\'ed\'d3\'c3\'c0\'eb\'cf\'df\'ce\'c4\'b5\'b5\'d0\'b4\'b7\'bd\'b0\'b8\'a1\'a2\'b2\'bb\'d4\'ca\'d0\'ed\'cb\'bd\'cf\'c2\'cd\'ac\'b2\'bd\'bd\'f8\'b6\'c8\'a3\'ac\'c8\'b7\'b1\'a3\'c6\'bd\'cc\'a8\'ca\'c7\'ce\'a8\'d2\'bb\'b5\'c4\'d0\'c5\'cf\'a2\'c8\'eb\'bf\'da\'a3\'ac\'b4\'d3\'b8\'f9\'d4\'b4\'cf\'fb\'c3\'f0\'d0\'c5\'cf\'a2\'b2\'ee\'a1\'a3
+\f0 \cb1 \
+\ls8\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f8\'b6\'d4\'b2\'bb\'d2\'aa\'d3\'c3\'b9\'cc\'b6\'a8\'b5\'fc\'b4\'fa\'ca\'f8\'b8\'bf\'cd\'c5\'b6\'d3\'a3\'ac\'ca\'ca\'c5\'e4
+\f2  AI 
+\f3 \'d0\'d0\'d2\'b5\'b5\'c4\'bf\'ec\'bd\'da\'d7\'e0
+\f0\b0 \cb1 \uc0\u8232 
+\f1 \cb3 \'b4\'ab\'cd\'b3
+\f0  2 
+\f1 \'d6\'dc\'b9\'cc\'b6\'a8\'b5\'fc\'b4\'fa\'a3\'ac\'d4\'da
+\f0  AI 
+\f1 \'ca\'b1\'b4\'fa\'d2\'d1\'be\'ad\'b9\'fd\'ca\'b1\'a1\'a3\'b2\'bb\'d2\'aa\'ce\'aa\'c1\'cb\'a1\'b8\'b5\'fc\'b4\'fa\'ce\'c8\'b6\'a8\'a1\'b9\'a3\'ac\'be\'dc\'be\'f8\'b8\'df\'d3\'c5\'cf\'c8\'bc\'b6\'b5\'c4\'ca\'d0\'b3\'a1
+\f0  / 
+\f1 \'c4\'a3\'d0\'cd\'b1\'e4\'bb\'af\'a3\'ac\'d3\'c3\'b9\'f6\'b6\'af\'ca\'bd\'b6\'af\'cc\'ac\'b5\'fc\'b4\'fa\'a3\'ac\'d3\'c5\'cf\'c8\'bc\'b6\'b8\'df\'b5\'c4\'d0\'e8\'c7\'f3\'cb\'e6\'ca\'b1\'bd\'f8\'c8\'eb\'a1\'a2\'cb\'e6\'ca\'b1\'c9\'cf\'cf\'df\'a3\'ac\'b5\'cd\'d3\'c5\'cf\'c8\'bc\'b6\'d0\'e8\'c7\'f3\'d7\'d4\'b6\'af\'cb\'b3\'d1\'d3\'a3\'ac\'bc\'c8\'b1\'a3\'d6\'a4\'ba\'cb\'d0\'c4\'bd\'bb\'b8\'b6\'a3\'ac\'d3\'d6\'c4\'dc\'c1\'e9\'bb\'ee\'cf\'ec\'d3\'a6\'b1\'e4\'bb\'af\'a1\'a3
+\f0 \cb1 \
+\ls8\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f8\'b6\'d4\'b2\'bb\'d2\'aa\'d6\'bb\'b6\'a2\'bd\'f8\'b6\'c8\'a3\'ac\'b2\'bb\'b6\'a2\'c8\'cf\'d6\'aa\'b6\'d4\'c6\'eb
+\f0\b0 \cb1 \uc0\u8232 \cb3 10 
+\f1 \'c8\'cb\'cd\'c5\'b6\'d3
+\f0  80% 
+\f1 \'b5\'c4\'b7\'b5\'b9\'a4\'a3\'ac\'b6\'bc\'c0\'b4\'d7\'d4\'c8\'cf\'d6\'aa\'b2\'bb\'d2\'bb\'d6\'c2\'a3\'ba\'b2\'fa\'c6\'b7\'cf\'eb\'b5\'c4\'ba\'cd\'bc\'bc\'ca\'f5\'d7\'f6\'b5\'c4\'b2\'bb\'d2\'bb\'d1\'f9\'a3\'ac\'cb\'e3\'b7\'a8\'b6\'d4\'c4\'a3\'d0\'cd\'c4\'dc\'c1\'a6\'b5\'c4\'c5\'d0\'b6\'cf\'ba\'cd\'b2\'fa\'c6\'b7\'d4\'a4\'c6\'da\'b2\'bb\'d2\'bb\'d1\'f9\'a3\'ac\'cd\'c5\'b6\'d3\'b6\'d4\'c4\'bf\'b1\'ea\'b5\'c4\'c0\'ed\'bd\'e2\'b2\'bb\'d2\'bb\'d1\'f9\'a1\'a3\'c6\'bd\'cc\'a8\'b5\'c4\'ba\'cb\'d0\'c4\'ca\'c7\'cf\'c8\'b6\'d4\'c6\'eb\'c8\'cf\'d6\'aa\'a3\'ac\'d4\'d9\'b9\'dc\'bf\'d8\'bd\'f8\'b6\'c8\'a3\'ac\'b7\'f1\'d4\'f2\'d7\'f6\'b5\'c3\'d4\'bd\'bf\'ec\'a3\'ac\'b4\'ed\'b5\'c3\'d4\'bd\'d4\'b6\'a1\'a3
+\f0 \cb1 \
+\ls8\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	6	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'be\'f8\'b6\'d4\'b2\'bb\'d2\'aa\'ba\'f6\'ca\'d3\'c4\'a3\'d0\'cd\'be\'ad\'d1\'e9\'b5\'c4\'b3\'c1\'b5\'ed\'a3\'ac\'b1\'dc\'c3\'e2\'d6\'d8\'b8\'b4\'b2\'c8\'bf\'d3
+\f0\b0 \cb1 \uc0\u8232 \cb3 AI 
+\f1 \'d0\'d0\'d2\'b5\'c4\'a3\'d0\'cd\'b5\'fc\'b4\'fa\'bc\'ab\'bf\'ec\'a3\'ac\'b2\'bb\'b3\'c1\'b5\'ed\'be\'ad\'d1\'e9\'a3\'ac\'cd\'c5\'b6\'d3\'be\'cd\'bb\'e1\'d2\'bb\'d6\'b1\'d4\'da\'d4\'ad\'b5\'d8\'b2\'c8\'bf\'d3\'a1\'a3\'b1\'d8\'d0\'eb\'c7\'bf\'d6\'c6\'cb\'f9\'d3\'d0\'c4\'a3\'d0\'cd\'ca\'ca\'c5\'e4\'a1\'a2\'d4\'a4\'d1\'d0\'a1\'a2\'b2\'e2\'ca\'d4\'b5\'c4\'be\'ad\'d1\'e9\'a3\'ac\'c8\'ab\'b2\'bf\'b3\'c1\'b5\'ed\'b5\'bd\'d6\'aa\'ca\'b6\'bf\'e2\'a3\'ac\'ba\'cd\'cf\'ee\'c4\'bf\'a1\'a2\'c4\'a3\'d0\'cd\'b0\'e6\'b1\'be\'c7\'bf\'b0\'f3\'b6\'a8\'a3\'ac\'c8\'c3\'cd\'c5\'b6\'d3\'b5\'c4\'be\'ad\'d1\'e9\'cb\'e6\'d7\'c5\'c4\'a3\'d0\'cd\'b5\'fc\'b4\'fa\'b3\'d6\'d0\'f8\'bb\'fd\'c0\'db\'a3\'ac\'d4\'bd\'d7\'f6\'d4\'bd\'bf\'ec\'a3\'ac\'b6\'f8\'b2\'bb\'ca\'c7\'c3\'bf\'b4\'ce\'c4\'a3\'d0\'cd\'b8\'fc\'d0\'c2\'b6\'bc\'b4\'d3\'c1\'e3\'bf\'aa\'ca\'bc\'a1\'a3
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'ce\'e5\'a1\'a2\'c5\'e4\'cc\'d7\'cd\'c5\'b6\'d3\'d0\'ad\'d7\'f7\'b9\'e6\'b7\'b6\'a3\'a8\'b9\'a4\'be\'df
+\f0  + 
+\f1 \'b9\'e6\'b7\'b6\'a3\'ac\'b2\'c5\'c4\'dc\'b7\'a2\'bb\'d3\'d7\'ee\'b4\'f3\'bc\'db\'d6\'b5\'a3\'a9
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\partightenfactor0
+
+\f1 \cf2 \cb3 \'b9\'a4\'be\'df\'d6\'bb\'ca\'c7\'d4\'d8\'cc\'e5\'a3\'ac\'b1\'d8\'d0\'eb\'c5\'e4\'cc\'d7\'bc\'ab\'bc\'f2\'b5\'c4\'d0\'ad\'d7\'f7\'b9\'e6\'b7\'b6\'a3\'ac\'b2\'c5\'c4\'dc\'b3\'b9\'b5\'d7\'c2\'e4\'b5\'d8\'a3\'ba
+\f0 \cb1 \
+\
+\pard\tx220\tx720\pardeftab720\li720\fi-720\partightenfactor0
+\ls9\ilvl0
+\f2\b \cf2 \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	1	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'d0\'c5\'cf\'a2\'c2\'bc\'c8\'eb\'b9\'e6\'b7\'b6
+\f1\b0 \'a3\'ba\'cb\'f9\'d3\'d0\'be\'f6\'b2\'df\'a1\'a2\'d0\'e8\'c7\'f3\'a1\'a2\'b7\'bd\'b0\'b8\'a1\'a2\'ce\'ca\'cc\'e2\'a3\'ac\'b1\'d8\'d0\'eb\'b5\'da\'d2\'bb\'ca\'b1\'bc\'e4\'c2\'bc\'c8\'eb\'c6\'bd\'cc\'a8\'a3\'ac\'b2\'bb\'d4\'ca\'d0\'ed\'cb\'bd\'cf\'c2\'b7\'a2\'c9\'a2\'b9\'b5\'cd\'a8\'a1\'a3
+\f0 \cb1 \
+\ls9\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	2	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'bb\'e1\'d2\'e9\'b9\'e6\'b7\'b6
+\f1\b0 \'a3\'ba\'c4\'dc\'d2\'ec\'b2\'bd\'b9\'b5\'cd\'a8\'b5\'c4\'be\'f8\'b2\'bb\'bf\'aa\'bb\'e1\'a3\'ac\'bf\'aa\'bb\'e1\'b1\'d8\'d0\'eb\'d3\'d0\'c3\'f7\'c8\'b7\'d6\'f7\'cc\'e2\'a3\'ac\'bb\'e1\'ba\'f3
+\f0  AI 
+\f1 \'d7\'d4\'b6\'af\'c9\'fa\'b3\'c9\'bc\'cd\'d2\'aa\'a3\'ac\'cd\'ac\'b2\'bd\'b5\'bd\'c6\'bd\'cc\'a8\'a3\'ac\'be\'f6\'b2\'df\'c4\'da\'c8\'dd\'d6\'b1\'bd\'d3\'d7\'aa\'ce\'aa\'c8\'ce\'ce\'f1\'a1\'a3
+\f0 \cb1 \
+\ls9\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	3	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'d3\'c5\'cf\'c8\'bc\'b6\'b9\'e6\'b7\'b6
+\f1\b0 \'a3\'ba
+\f0 P0 
+\f1 \'d0\'e8\'c7\'f3\'b2\'bb\'b3\'ac\'b9\'fd
+\f0  2 
+\f1 \'b8\'f6\'a3\'ac
+\f0 P1 
+\f1 \'d0\'e8\'c7\'f3\'b2\'bb\'b3\'ac\'b9\'fd
+\f0  5 
+\f1 \'b8\'f6\'a3\'ac\'b6\'c5\'be\'f8\'c8\'ab\'d4\'b1\'cd\'ac\'ca\'b1\'d7\'f6\'d2\'bb\'b6\'d1\'ca\'c2\'a3\'ac\'c8\'b7\'b1\'a3\'ba\'cb\'d0\'c4\'c4\'bf\'b1\'ea\'bf\'ec\'cb\'d9\'c2\'e4\'b5\'d8\'a1\'a3
+\f0 \cb1 \
+\ls9\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	4	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'b8\'b4\'c5\'cc\'b9\'e6\'b7\'b6
+\f1\b0 \'a3\'ba\'c3\'bf\'d6\'dc\'b9\'cc\'b6\'a8
+\f0  30 
+\f1 \'b7\'d6\'d6\'d3\'b8\'b4\'c5\'cc\'a3\'ac\'d6\'bb\'cc\'b8\'ce\'ca\'cc\'e2\'ba\'cd\'b8\'c4\'bd\'f8\'b4\'eb\'ca\'a9\'a3\'ac\'b2\'bb\'cc\'b8\'bf\'d5\'bb\'b0\'cc\'d7\'bb\'b0\'a3\'ac\'b8\'c4\'bd\'f8\'b4\'eb\'ca\'a9\'b1\'d8\'d0\'eb\'c9\'fa\'b3\'c9\'bf\'c9\'c2\'e4\'b5\'d8\'b5\'c4\'c8\'ce\'ce\'f1\'a1\'a3
+\f0 \cb1 \
+\ls9\ilvl0
+\f2\b \cb3 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext	5	}
+\f3 \expnd0\expndtw0\kerning0
+\outl0\strokewidth0 \strokec2 \'d6\'aa\'ca\'b6\'bf\'e2\'b9\'e6\'b7\'b6
+\f1\b0 \'a3\'ba\'cb\'ad\'d6\'b4\'d0\'d0\'a1\'a2\'cb\'ad\'b3\'c1\'b5\'ed\'a3\'ac\'c8\'ce\'ce\'f1\'cd\'ea\'b3\'c9\'ba\'f3\'a3\'ac\'b6\'d4\'d3\'a6\'b5\'c4\'be\'ad\'d1\'e9\'a1\'a2\'b7\'bd\'b0\'b8\'a1\'a2\'b2\'c8\'bf\'d3\'bc\'c7\'c2\'bc\'a3\'ac\'b1\'d8\'d0\'eb\'d7\'d4\'b6\'af
+\f0  / 
+\f1 \'ca\'d6\'b6\'af\'cd\'ac\'b2\'bd\'b5\'bd\'d6\'aa\'ca\'b6\'bf\'e2\'a1\'a3
+\f0 \cb1 \
+\pard\pardeftab720\partightenfactor0
+\cf2 \
+\pard\pardeftab720\qc\partightenfactor0
+\cf4 \strokec4 \
+\pard\pardeftab720\partightenfactor0
+\cf0 \strokec5 \
+\pard\pardeftab720\qc\partightenfactor0
+
+\f4 \cf3 \strokec3 \
+
+\f5 \'bf\'aa\'ca\'bc\'ca\'d4\'d3\'c3\'b6\'b9\'b0\'fc
+\f4 \
+\pard\pardeftab720\qc\partightenfactor0
+
+\f0 \cf6 \strokec6 \
+}
\ No newline at end of file
diff --git a/backend/.env.example b/backend/.env.example
new file mode 100644
index 00000000..d9dd0e2b
--- /dev/null
+++ b/backend/.env.example
@@ -0,0 +1,96 @@
+# Database Configuration
+DATABASE_URL=postgresql+asyncpg://pmflow:pmflow_pass@db:5432/pmflow
+# 可选：启动方言守卫（postgresql / sqlite）。配置后若不一致将拒绝启动。
+PMFLOW_EXPECTED_DB_DIALECT=postgresql
+# For PostgreSQL with pgvector:
+# DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5435/neural_sync
+# For local SQLite quick-start:
+# DATABASE_URL=sqlite+aiosqlite:///./neural_sync.db
+
+# Redis Configuration
+REDIS_URL=redis://:pmflow_redis_pass@redis:6379/0
+
+# Knowledge Graph Incremental Scheduler
+KG_INCREMENTAL_INTERVAL_MINUTES=5
+KG_INCREMENTAL_DELTA_LIMIT=1000
+KG_INCREMENTAL_EVENT_BATCH=200
+KG_CODE_SCAN_INTERVAL_HOURS=6
+
+# ========================================
+# LLM / OpenClaw Control Plane
+# ========================================
+# 单一控制面：PMFlow 仅调用 OpenClaw 网关（OpenAI-Compatible）
+LLM_PROVIDER=openai
+OPENAI_BASE_URL=http://openclaw:18789/v1
+OPENAI_API_KEY=<OPENCLAW_GATEWAY_TOKEN>
+
+# OpenClaw 函数路由
+OPENCLAW_ROUTING_ENABLED=true
+OPENCLAW_ROUTING_RULES_PATH=/app/backend/.openclaw/pmflow-routing-rules.json
+# OpenClaw agent 默认超时（秒）：用于注入 chat/completions -> agents.defaults.timeoutSeconds
+OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS=120
+
+# 关闭 PMFlow 内部 mixed routing，优先以 OpenClaw 规则为准
+LLM_MIXED_ROUTING_ENABLED=false
+LLM_ROUTING_DEFAULT_TIER=active
+LLM_LOW_TO_HIGH_ESCALATION_ENABLED=true
+
+# low/high/safety tier 默认模型（由网关解释）
+LLM_ROUTING_LOW_PROVIDER=openai
+LLM_ROUTING_LOW_MODEL=zai/glm-4-flash
+LLM_ROUTING_HIGH_PROVIDER=openai
+LLM_ROUTING_HIGH_MODEL=minimax/MiniMax-M2.5
+LLM_ROUTING_SAFETY_PROVIDER=openai
+LLM_ROUTING_SAFETY_MODEL=deepseek/deepseek-chat
+
+# 应急开关：仅在网关不可用时允许直连 provider（默认关闭）
+LLM_DIRECT_EMERGENCY_ENABLED=false
+LLM_FALLBACK_PROVIDER=ollama
+
+# Adaptive RAG 深入思考模式（前端开关=deep 时生效）
+RAG_DEEP_RETRIEVAL_STEPS=5
+RAG_DEEP_VECTOR_TOP_K=12
+RAG_DEEP_CONTEXT_BLOCKS=16
+RAG_DEEP_ANSWER_MAX_TOKENS=1400
+
+# --- Emergency Direct Provider Keys (optional) ---
+GLM_API_KEY=
+GLM_BASE_URL=https://open.bigmodel.cn/api/paas/v4
+GLM_MODEL=glm-4-flash
+
+DEEPSEEK_API_KEY=
+DEEPSEEK_BASE_URL=https://api.deepseek.com/v1
+
+# --- Ollama 本地配置（应急本地兜底） ---
+OLLAMA_HOST=http://127.0.0.1:11434
+OLLAMA_CHAT_MODEL=llama3:8b
+OLLAMA_EMBEDDING_MODEL=nomic-embed-text
+OLLAMA_BASE_URL=http://127.0.0.1:11434
+OLLAMA_MODEL=llama3:latest
+
+# ========================================
+
+# Multi-Model Voting Configuration (仅 Ollama 支持)
+# For high-impact signals (impact_score >= VOTING_THRESHOLD), multiple models will vote
+# Available models: llama3, llama3.2, qwen2.5, mistral, deepseek-coder, etc.
+VOTING_MODELS=llama3,qwen2.5,mistral
+VOTING_THRESHOLD=0.8
+
+# Business Context for AI Strategic Analysis
+# Customize this to match your actual product/team
+# This context is used by the AI to evaluate external signals
+BUSINESS_CONTEXT=Pixelle-Video: An AI-powered automated short video generation platform. Core features: topic-to-video automation (script generation → AI images/videos → TTS → BGM → final video). Tech stack: Python, Streamlit, FastAPI, ComfyUI workflows. AI models: FLUX/Qwen/SDXL for images, WAN 2.1/2.2 for videos, Edge TTS/Index TTS for voice. Supports multiple pipelines: standard generation, custom assets, digital human, image-to-video, action transfer.
+
+# Plane PM Tool Integration (optional, simulation mode if not set)
+PLANE_API_URL=https://api.plane.so/api/v1
+PLANE_API_KEY=
+PLANE_WORKSPACE_SLUG=
+PLANE_PROJECT_ID=
+
+# Lark/Feishu Bot Integration (optional)
+LARK_APP_ID=
+LARK_APP_SECRET=
+LARK_VERIFICATION_TOKEN=
+# 飞书长连接重试策略（指数退避 1s -> 2s -> 4s，最多 3 次）
+FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS=1
+FEISHU_LONG_CONNECTION_RETRY_MAX_ATTEMPTS=3
diff --git a/backend/.openclaw/pmflow-routing-rules.json b/backend/.openclaw/pmflow-routing-rules.json
new file mode 100644
index 00000000..c509e59f
--- /dev/null
+++ b/backend/.openclaw/pmflow-routing-rules.json
@@ -0,0 +1,74 @@
+{
+  "version": "v2",
+  "updated_at": "2026-03-05T00:00:00Z",
+  "source_of_truth": "openclaw",
+  "description": "PMFlow 三引擎统一路由：所有主链路统一经 OpenClaw OpenAI-Compatible 网关。",
+  "rules": [
+    {
+      "name": "engine_low_execute_glm_flash",
+      "enabled": true,
+      "priority": 300,
+      "tier": "low",
+      "provider": "openai",
+      "model": "zai/glm-4-flash",
+      "match_any": [
+        "feishu_level_classification",
+        "feishu_signal_extraction",
+        "rag_intent_classification",
+        "classify_intent",
+        "feishu_humanized_reply",
+        "extract_context_insights"
+      ]
+    },
+    {
+      "name": "engine_high_reasoning_minimax",
+      "enabled": true,
+      "priority": 220,
+      "tier": "high",
+      "provider": "openai",
+      "model": "minimax/MiniMax-M2.5",
+      "match_any": [
+        "adaptive_rag_answer",
+        "feishu_project_prefill",
+        "feishu_task_orchestration",
+        "generate_task_breakdown",
+        "generate_charter_from_signal",
+        "decompose_tasks_from_charter",
+        "recommend_assignees_for_task",
+        "evaluate_model_requirement",
+        "calculate_alignment_score",
+        "detect_cognitive_conflicts",
+        "rag_db_project_semantic_retrieval",
+        "workload_optimize_prompt",
+        "analyze_cost_roi"
+      ]
+    },
+    {
+      "name": "engine_high_extended_reasoning",
+      "enabled": true,
+      "priority": 200,
+      "tier": "high",
+      "provider": "openai",
+      "model": "minimax/MiniMax-M2.5",
+      "match_any": [
+        "analyze_food_calories",
+        "feishu_new_project_brief",
+        "feishu_project_signal_classification",
+        "sstc_ai_breakdown",
+        "sstc_charter_localize_zh",
+        "sstc_charter_prefill",
+        "skill_score_update",
+        "test_engine_connection"
+      ]
+    },
+    {
+      "name": "engine_safety_fallback",
+      "enabled": true,
+      "priority": 50,
+      "tier": "safety",
+      "provider": "openai",
+      "model": "deepseek/deepseek-chat",
+      "match_all": true
+    }
+  ]
+}
diff --git a/backend/app/api/ai_logs.py b/backend/app/api/ai_logs.py
new file mode 100644
index 00000000..bfe7f8bc
--- /dev/null
+++ b/backend/app/api/ai_logs.py
@@ -0,0 +1,341 @@
+"""AI 处理可视化日志 API。"""
+
+from __future__ import annotations
+
+import json
+from datetime import date, datetime
+from typing import Any, Optional
+
+from fastapi import APIRouter, Depends, HTTPException, Query
+from sqlalchemy import and_, case, func, or_, select, union_all
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.response import success_response
+from ..database import get_db
+from ..dependencies.auth import get_current_user
+from ..models import (
+    AIProcessingLog,
+    FeishuItem,
+    ProjectSignalLink,
+    ProjectSuggestion,
+    Signal,
+    User,
+    WorkItem,
+)
+
+
+router = APIRouter(prefix="/ai-logs", tags=["AI Logs"])
+
+
+def _normalize_bullets(value: Any) -> list[str]:
+    if isinstance(value, list):
+        return [str(x).strip() for x in value if str(x).strip()][:3]
+    if isinstance(value, str):
+        raw = value.strip()
+        if not raw:
+            return []
+        try:
+            parsed = json.loads(raw)
+            if isinstance(parsed, list):
+                return [str(x).strip() for x in parsed if str(x).strip()][:3]
+        except Exception:
+            pass
+        return [line.strip(" -•\t") for line in raw.splitlines() if line.strip()][:3]
+    return []
+
+
+def _safe_json(value: Optional[str]) -> Any:
+    text = (value or "").strip()
+    if not text:
+        return None
+    try:
+        return json.loads(text)
+    except Exception:
+        return None
+
+
+def _serialize_log_row(row: AIProcessingLog) -> dict[str, Any]:
+    return {
+        "id": str(row.id),
+        "feishu_message_id": row.feishu_message_id,
+        "feishu_item_id": str(row.feishu_item_id) if row.feishu_item_id is not None else None,
+        "raw_message": row.raw_message,
+        "project_score": row.project_score,
+        "signal_confidence": row.signal_confidence,
+        "level": row.level,
+        "routing_decision": row.routing_decision,
+        "routing_reason": row.routing_reason,
+        "reasoning_bullets": _normalize_bullets(row.reasoning_bullets),
+        "processing_status": row.processing_status,
+        "error_message": row.error_message,
+        "duration_ms": row.duration_ms,
+        "created_at": row.created_at.isoformat() if row.created_at else None,
+    }
+
+
+def _serialize_log_detail(row: AIProcessingLog) -> dict[str, Any]:
+    payload = _serialize_log_row(row)
+    payload["ai_raw_output"] = row.ai_raw_output
+    payload["ai_raw_output_json"] = _safe_json(row.ai_raw_output)
+    return payload
+
+
+def _apply_project_filter(stmt, *, tenant_id: int, project_id: Optional[int]):
+    if project_id is None:
+        return stmt
+
+    work_item_message_ids = select(func.distinct(WorkItem.source_message_id)).where(
+        and_(
+            WorkItem.tenant_id == tenant_id,
+            WorkItem.project_id == int(project_id),
+            WorkItem.source_system == "feishu",
+            WorkItem.source_message_id.is_not(None),
+        )
+    )
+    direct_items = select(FeishuItem.id.label("item_id")).where(
+        and_(
+            FeishuItem.tenant_id == tenant_id,
+            FeishuItem.matched_project_id == int(project_id),
+        )
+    )
+    signal_matched_items = select(Signal.feishu_item_id.label("item_id")).where(
+        and_(
+            Signal.tenant_id == tenant_id,
+            Signal.matched_project_id == int(project_id),
+            Signal.feishu_item_id.is_not(None),
+        )
+    )
+    linked_items = (
+        select(Signal.feishu_item_id.label("item_id"))
+        .select_from(Signal)
+        .join(
+            ProjectSignalLink,
+            and_(
+                ProjectSignalLink.tenant_id == tenant_id,
+                ProjectSignalLink.project_id == int(project_id),
+                ProjectSignalLink.signal_id == Signal.id,
+            ),
+        )
+        .where(
+            and_(
+                Signal.tenant_id == tenant_id,
+                Signal.feishu_item_id.is_not(None),
+            )
+        )
+    )
+    auto_created_items = select(ProjectSuggestion.feishu_item_id.label("item_id")).where(
+        and_(
+            ProjectSuggestion.tenant_id == tenant_id,
+            ProjectSuggestion.created_project_id == int(project_id),
+            ProjectSuggestion.feishu_item_id.is_not(None),
+        )
+    )
+
+    all_item_ids = union_all(
+        direct_items,
+        signal_matched_items,
+        linked_items,
+        auto_created_items,
+    ).subquery("project_item_ids")
+    item_ids_stmt = select(func.distinct(all_item_ids.c.item_id)).where(all_item_ids.c.item_id.is_not(None))
+    message_ids_stmt = select(func.distinct(FeishuItem.feishu_message_id)).where(
+        and_(
+            FeishuItem.tenant_id == tenant_id,
+            FeishuItem.id.in_(item_ids_stmt),
+            FeishuItem.feishu_message_id.is_not(None),
+        )
+    )
+    return stmt.where(
+        or_(
+            AIProcessingLog.feishu_item_id.in_(item_ids_stmt),
+            AIProcessingLog.feishu_message_id.in_(message_ids_stmt),
+            AIProcessingLog.feishu_message_id.in_(work_item_message_ids),
+        )
+    )
+
+
+@router.get("")
+async def list_ai_logs(
+    project_id: Optional[int] = Query(default=None, ge=1),
+    limit: int = Query(default=20, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    total_stmt = select(func.count(AIProcessingLog.id)).where(AIProcessingLog.tenant_id == tenant_id)
+    data_stmt = (
+        select(AIProcessingLog)
+        .where(AIProcessingLog.tenant_id == tenant_id)
+        .order_by(AIProcessingLog.created_at.desc(), AIProcessingLog.id.desc())
+        .offset(offset)
+        .limit(limit)
+    )
+    total_stmt = _apply_project_filter(total_stmt, tenant_id=tenant_id, project_id=project_id)
+    data_stmt = _apply_project_filter(data_stmt, tenant_id=tenant_id, project_id=project_id)
+
+    total = int((await db.execute(total_stmt)).scalar() or 0)
+    rows = (await db.execute(data_stmt)).scalars().all()
+    items = [_serialize_log_row(row) for row in rows]
+    return success_response(
+        {
+            "items": items,
+            "total": total,
+            "limit": limit,
+            "offset": offset,
+            "has_more": (offset + len(items)) < total,
+        }
+    )
+
+
+@router.get("/stats")
+async def get_ai_log_stats(
+    project_id: Optional[int] = Query(default=None, ge=1),
+    date_value: str = Query(default="today", alias="date"),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    if date_value == "today":
+        target_date = datetime.now().date()
+    else:
+        try:
+            target_date = date.fromisoformat(date_value)
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=f"invalid_date:{date_value}") from exc
+
+    stats_stmt = (
+        select(
+            func.count(AIProcessingLog.id).label("received_count"),
+            func.sum(case((AIProcessingLog.processing_status == "success", 1), else_=0)).label("success_count"),
+            func.sum(case((AIProcessingLog.processing_status == "failed", 1), else_=0)).label("failed_count"),
+            func.sum(
+                case((AIProcessingLog.routing_decision == "duplicate_filtered", 1), else_=0)
+            ).label("duplicate_filtered_count"),
+            func.sum(
+                case((AIProcessingLog.routing_decision == "execute_reply_guard_triggered", 1), else_=0)
+            ).label("execute_reply_guard_count"),
+            func.sum(
+                case(
+                    (
+                        AIProcessingLog.routing_decision.in_(
+                            ["auto_create_project", "auto_enter_signal"]
+                        ),
+                        1,
+                    ),
+                    else_=0,
+                )
+            ).label("auto_entered_count"),
+            func.sum(
+                case(
+                    (
+                        AIProcessingLog.routing_decision.in_(
+                            [
+                                "pending_low_confidence",
+                                "pending_cold_start",
+                                "pending_uncertain_project",
+                            ]
+                        ),
+                        1,
+                    ),
+                    else_=0,
+                )
+            ).label("pending_count"),
+            func.sum(case((AIProcessingLog.processing_status == "processing", 1), else_=0)).label(
+                "processing_count"
+            ),
+            func.avg(AIProcessingLog.signal_confidence).label("avg_signal_confidence"),
+            func.avg(AIProcessingLog.project_score).label("avg_project_score"),
+        )
+        .where(
+            and_(
+                AIProcessingLog.tenant_id == tenant_id,
+                func.date(AIProcessingLog.created_at) == target_date.isoformat(),
+            )
+        )
+    )
+    stats_stmt = _apply_project_filter(stats_stmt, tenant_id=tenant_id, project_id=project_id)
+    row = (await db.execute(stats_stmt)).one()
+    return success_response(
+        {
+            "date": target_date.isoformat(),
+            "received_count": int(row.received_count or 0),
+            "success_count": int(row.success_count or 0),
+            "failed_count": int(row.failed_count or 0),
+            "duplicate_filtered_count": int(row.duplicate_filtered_count or 0),
+            "execute_reply_guard_count": int(row.execute_reply_guard_count or 0),
+            "auto_entered_count": int(row.auto_entered_count or 0),
+            "pending_count": int(row.pending_count or 0),
+            "processing_count": int(row.processing_count or 0),
+            "avg_signal_confidence": round(float(row.avg_signal_confidence or 0.0), 4),
+            "avg_project_score": round(float(row.avg_project_score or 0.0), 4),
+        }
+    )
+
+
+@router.get("/alerts/execute-reply-guard")
+async def list_execute_reply_guard_alerts(
+    limit: int = Query(default=20, ge=1, le=100),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    rows = (
+        await db.execute(
+            select(AIProcessingLog)
+            .where(
+                and_(
+                    AIProcessingLog.tenant_id == tenant_id,
+                    AIProcessingLog.routing_decision == "execute_reply_guard_triggered",
+                )
+            )
+            .order_by(AIProcessingLog.created_at.desc(), AIProcessingLog.id.desc())
+            .limit(limit)
+        )
+    ).scalars().all()
+    return success_response(
+        {
+            "count": len(rows),
+            "items": [_serialize_log_row(row) for row in rows],
+        }
+    )
+
+
+@router.get("/latest")
+async def get_latest_ai_log(
+    project_id: Optional[int] = Query(default=None, ge=1),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    stmt = (
+        select(AIProcessingLog)
+        .where(AIProcessingLog.tenant_id == tenant_id)
+        .order_by(AIProcessingLog.created_at.desc(), AIProcessingLog.id.desc())
+        .limit(1)
+    )
+    stmt = _apply_project_filter(stmt, tenant_id=tenant_id, project_id=project_id)
+    row = (await db.execute(stmt)).scalar_one_or_none()
+    return success_response(_serialize_log_detail(row) if row else None)
+
+
+@router.get("/{log_id}")
+async def get_ai_log_detail(
+    log_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    row = (
+        await db.execute(
+            select(AIProcessingLog).where(
+                and_(
+                    AIProcessingLog.tenant_id == tenant_id,
+                    AIProcessingLog.id == int(log_id),
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if row is None:
+        raise HTTPException(status_code=404, detail="ai_log_not_found")
+    return success_response(_serialize_log_detail(row))
diff --git a/backend/app/api/ai_pmo.py b/backend/app/api/ai_pmo.py
new file mode 100644
index 00000000..d18fedd1
--- /dev/null
+++ b/backend/app/api/ai_pmo.py
@@ -0,0 +1,417 @@
+"""AI PMO 决策中枢 API。"""
+
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Any, Optional
+
+from fastapi import APIRouter, Depends, HTTPException
+from pydantic import BaseModel, Field
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.response import success_response
+from ..database import get_db
+from ..dependencies.auth import get_current_user
+from ..dependencies.tenant import get_tenant_id
+from ..models import DecisionRecord, IngestEvent, PlanBundle, PortfolioSnapshot, Project, ReplyReceipt, User, WorkItem
+from ..services.ai_pmo_cadence_service import AIPMOCadenceService
+from ..services.ai_pmo_kernel_service import AIPMOKernelService
+
+router = APIRouter(prefix="/ai-pmo", tags=["AI PMO"])
+
+
+class PMOProcessTextRequest(BaseModel):
+    message_text: str = Field(..., min_length=1, max_length=4000)
+    sender: Optional[str] = Field(default=None, max_length=120)
+    source_message_id: Optional[str] = Field(default=None, max_length=128)
+    project_id: Optional[int] = Field(default=None)
+
+
+class PMOCadenceRunRequest(BaseModel):
+    cadence_type: str = Field(..., min_length=1, max_length=32)
+    project_id: Optional[int] = Field(default=None)
+
+
+class PMOResumeWorkItemRequest(BaseModel):
+    project_id: Optional[int] = Field(default=None)
+    source_message_id: Optional[str] = Field(default=None, max_length=128)
+
+
+@router.post("/process-text")
+async def process_text(
+    request: PMOProcessTextRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AIPMOKernelService(db, tenant_id=tenant_id)
+    result = await service.process_ingress(
+        source_system="pmflow_manual",
+        source_message_id=request.source_message_id,
+        message_text=request.message_text,
+        sender=str(request.sender or current_user.name or current_user.username or "PMFlow用户"),
+        channel_ref="pmflow_manual",
+        source_sender_open_id=None,
+        captured_at=None,
+        chat_type="manual",
+        project_hint_id=request.project_id,
+    )
+    if not result.get("handled"):
+        raise HTTPException(status_code=400, detail=str(result.get("reason") or "AI PMO kernel 未处理该输入"))
+    return success_response(result)
+
+
+@router.get("/work-items")
+async def list_work_items(
+    limit: int = 20,
+    state: Optional[str] = None,
+    project_id: Optional[int] = None,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    stmt = (
+        select(WorkItem, IngestEvent, Project)
+        .join(IngestEvent, IngestEvent.id == WorkItem.ingest_event_id)
+        .outerjoin(Project, and_(Project.tenant_id == tenant_id, Project.id == WorkItem.project_id))
+        .where(WorkItem.tenant_id == tenant_id)
+        .order_by(WorkItem.updated_at.desc(), WorkItem.id.desc())
+    )
+    if state:
+        stmt = stmt.where(WorkItem.state == state)
+    if project_id is not None:
+        stmt = stmt.where(WorkItem.project_id == int(project_id))
+    rows = (await db.execute(stmt.limit(max(1, min(limit, 100))))).all()
+    items = [
+        _serialize_work_item_summary(row=row, ingest=ingest, project=project)
+        for row, ingest, project in rows
+    ]
+    return success_response({"items": items})
+
+
+@router.get("/work-items/{work_item_id}")
+async def get_work_item(
+    work_item_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    row = (
+        await db.execute(
+            select(WorkItem).where(
+                and_(
+                    WorkItem.tenant_id == tenant_id,
+                    WorkItem.id == int(work_item_id),
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if row is None:
+        raise HTTPException(status_code=404, detail="work_item_not_found")
+    decision_rows = (
+        await db.execute(
+            select(DecisionRecord).where(
+                and_(
+                    DecisionRecord.tenant_id == tenant_id,
+                    DecisionRecord.work_item_id == int(work_item_id),
+                )
+            ).order_by(DecisionRecord.id.asc())
+        )
+    ).scalars().all()
+    plan_rows = (
+        await db.execute(
+            select(PlanBundle).where(
+                and_(
+                    PlanBundle.tenant_id == tenant_id,
+                    PlanBundle.work_item_id == int(work_item_id),
+                )
+            ).order_by(PlanBundle.id.asc())
+        )
+    ).scalars().all()
+    reply_rows = (
+        await db.execute(
+            select(ReplyReceipt).where(
+                and_(
+                    ReplyReceipt.tenant_id == tenant_id,
+                    ReplyReceipt.work_item_id == int(work_item_id),
+                )
+            ).order_by(ReplyReceipt.id.asc())
+        )
+    ).scalars().all()
+    payload: dict[str, Any] = {
+        "id": int(row.id),
+        "source_system": row.source_system,
+        "source_message_id": row.source_message_id,
+        "state": row.state,
+        "intent": row.intent,
+        "decision_level": row.decision_level,
+        "project_id": _safe_int(row.project_id),
+        "signal_id": _safe_int(row.signal_id),
+        "context_snapshot": row.context_snapshot,
+        "action_bundle": row.action_bundle,
+        "execution_result": row.execution_result,
+        "review_summary": row.review_summary,
+        "reply_template_key": row.reply_template_key,
+        "decision_records": [
+            {
+                "id": int(item.id),
+                "decision_level": item.decision_level,
+                "decision_object_type": item.decision_object_type,
+                "decision_object_id": _safe_int(item.decision_object_id),
+                "rationale": item.rationale,
+                "risk_score": item.risk_score,
+                "budget_impact": item.budget_impact,
+                "resource_impact": item.resource_impact,
+                "simulation_result": item.simulation_result,
+                "policy_gate_hit": bool(item.policy_gate_hit),
+                "policy_gate_code": item.policy_gate_code,
+                "approval_requirement": item.approval_requirement,
+                "execution_result": item.execution_result,
+                "final_outcome": item.final_outcome,
+            }
+            for item in decision_rows
+        ],
+        "plan_bundles": [
+            {
+                "id": int(item.id),
+                "plan_type": item.plan_type,
+                "summary": item.summary,
+                "payload": item.payload,
+            }
+            for item in plan_rows
+        ],
+        "reply_receipts": [
+            {
+                "id": int(item.id),
+                "receipt_phase": item.receipt_phase,
+                "template_key": item.template_key,
+                "payload": item.payload,
+                "status": item.status,
+            }
+            for item in reply_rows
+        ],
+    }
+    return success_response(payload)
+
+
+@router.post("/work-items/{work_item_id}/resume")
+async def resume_work_item(
+    work_item_id: int,
+    request: PMOResumeWorkItemRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    target = (
+        await db.execute(
+            select(WorkItem).where(
+                and_(
+                    WorkItem.tenant_id == tenant_id,
+                    WorkItem.id == int(work_item_id),
+                )
+            ).limit(1)
+        )
+    ).scalar_one_or_none()
+    if target is None:
+        raise HTTPException(status_code=404, detail="work_item_not_found")
+    service = AIPMOKernelService(db, tenant_id=tenant_id)
+    result = await service.process_ingress(
+        source_system="ai_pmo_ui",
+        source_message_id=request.source_message_id or f"resume-{int(work_item_id)}-{int(datetime.now().timestamp() * 1000)}",
+        message_text=f"继续处理 work_item #{int(work_item_id)}",
+        sender=str(current_user.name or current_user.username or "PMFlow用户"),
+        channel_ref="ai_pmo_ui",
+        source_sender_open_id=None,
+        captured_at=None,
+        chat_type="manual",
+        project_hint_id=request.project_id or _safe_int(target.project_id),
+    )
+    if not result.get("handled"):
+        raise HTTPException(status_code=400, detail=str(result.get("reason") or "resume_failed"))
+    return success_response(result)
+
+
+@router.post("/cadence/run")
+async def run_cadence(
+    request: PMOCadenceRunRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    service = AIPMOCadenceService(db, tenant_id=tenant_id)
+    try:
+        snapshot = await service.generate_snapshot(
+            cadence_type=request.cadence_type,
+            project_id=request.project_id,
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+    await db.commit()
+    return success_response(snapshot)
+
+
+@router.get("/cadence/latest")
+async def get_latest_cadence_snapshot(
+    cadence_type: str,
+    project_id: Optional[int] = None,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    service = AIPMOCadenceService(db, tenant_id=tenant_id)
+    snapshot = await service.get_latest_snapshot(
+        cadence_type=cadence_type,
+        project_id=project_id,
+    )
+    return success_response(snapshot)
+
+
+@router.get("/portfolio/program-view")
+async def get_program_view(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    service = AIPMOCadenceService(db, tenant_id=tenant_id)
+    return success_response(await service.build_program_view())
+
+
+@router.get("/portfolio/latest")
+async def get_latest_portfolio_snapshot(
+    scope_type: str = "tenant",
+    project_id: Optional[int] = None,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    scope_ref = f"project:{int(project_id)}" if project_id is not None else f"tenant:{tenant_id}"
+    row = (
+        await db.execute(
+            select(PortfolioSnapshot).where(
+                and_(
+                    PortfolioSnapshot.tenant_id == tenant_id,
+                    PortfolioSnapshot.scope_type == scope_type,
+                    PortfolioSnapshot.scope_ref == scope_ref,
+                )
+            ).order_by(PortfolioSnapshot.id.desc()).limit(1)
+        )
+    ).scalar_one_or_none()
+    if row is None:
+        return success_response(None)
+    return success_response(
+        {
+            "id": int(row.id),
+            "scope_type": row.scope_type,
+            "scope_ref": row.scope_ref,
+            "summary": row.summary,
+            "payload": row.payload,
+            "created_at": row.created_at.isoformat() if row.created_at else None,
+        }
+    )
+
+
+def _safe_int(value: Any) -> Optional[int]:
+    try:
+        if value is None or value == "":
+            return None
+        return int(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def _compact_text(value: Any, limit: int) -> str:
+    text = " ".join(str(value or "").split()).strip()
+    if len(text) <= limit:
+        return text
+    return f"{text[: max(limit - 1, 1)].rstrip()}…"
+
+
+def _priority_from_review(review_summary: Any) -> tuple[str, int]:
+    payload = review_summary if isinstance(review_summary, dict) else {}
+    risk_score = float(payload.get("risk_score_after") or 0.0)
+    if risk_score >= 0.75:
+        return "high", 3
+    if risk_score >= 0.45:
+        return "medium", 2
+    return "low", 1
+
+
+def _blocking_reason(row: WorkItem) -> Optional[str]:
+    execution_result = row.execution_result if isinstance(row.execution_result, dict) else {}
+    review_summary = row.review_summary if isinstance(row.review_summary, dict) else {}
+    reason = str(execution_result.get("reason") or execution_result.get("error_message") or "").strip()
+    if reason:
+        return reason
+    divergences = review_summary.get("divergences") if isinstance(review_summary.get("divergences"), list) else []
+    if divergences and isinstance(divergences[0], dict):
+        return str(divergences[0].get("detail") or "").strip() or None
+    return None
+
+
+def _next_step_text(row: WorkItem) -> Optional[str]:
+    review_summary = row.review_summary if isinstance(row.review_summary, dict) else {}
+    next_actions = review_summary.get("next_actions") if isinstance(review_summary.get("next_actions"), list) else []
+    if next_actions:
+        return str(next_actions[0]).strip()
+    action_bundle = row.action_bundle if isinstance(row.action_bundle, dict) else {}
+    summary = str(action_bundle.get("summary") or "").strip()
+    return summary or None
+
+
+def _stage_text(state: str) -> str:
+    mapping = {
+        "received": "已接收",
+        "modeled": "已建模",
+        "decided": "已决策",
+        "planned": "已规划",
+        "executing": "执行中",
+        "waiting_manual": "待人工确认",
+        "succeeded": "执行完成",
+        "failed": "执行失败",
+        "ignored": "已忽略",
+    }
+    return mapping.get(str(state or "").strip(), str(state or "").strip() or "-")
+
+
+def _serialize_work_item_summary(*, row: WorkItem, ingest: IngestEvent, project: Optional[Project]) -> dict[str, Any]:
+    payload = ingest.payload if isinstance(ingest.payload, dict) else {}
+    message_text = str(payload.get("message_text") or "").strip()
+    summary = _compact_text(message_text or _next_step_text(row) or row.intent or "AI PMO 工作项", 180)
+    title = _compact_text(message_text or f"AI PMO · {row.intent or 'work_item'}", 64)
+    priority, priority_rank = _priority_from_review(row.review_summary)
+    project_name = None
+    if project is not None:
+        project_name = str(project.title or project.name or "").strip() or None
+    project_id = _safe_int(row.project_id)
+    return {
+        "id": int(row.id),
+        "source_system": row.source_system,
+        "source_message_id": row.source_message_id,
+        "state": row.state,
+        "intent": row.intent,
+        "decision_level": row.decision_level,
+        "project_id": project_id,
+        "project_name": project_name,
+        "signal_id": _safe_int(row.signal_id),
+        "reply_template_key": row.reply_template_key,
+        "title": title or f"工作项#{int(row.id)}",
+        "summary": summary,
+        "ai_plan": _compact_text(((row.action_bundle or {}) if isinstance(row.action_bundle, dict) else {}).get("summary") or "", 180) or None,
+        "blocking_reason": _blocking_reason(row),
+        "next_step_text": _next_step_text(row),
+        "stage_text": _stage_text(str(row.state or "")),
+        "priority": priority,
+        "priority_rank": priority_rank,
+        "target_url": f"/projects/{project_id}/board" if project_id is not None else None,
+        "review_summary": row.review_summary,
+        "execution_result": row.execution_result,
+        "created_at": row.created_at.isoformat() if row.created_at else None,
+        "updated_at": row.updated_at.isoformat() if row.updated_at else None,
+    }
diff --git a/backend/app/api/ai_steward.py b/backend/app/api/ai_steward.py
new file mode 100644
index 00000000..8378a1f4
--- /dev/null
+++ b/backend/app/api/ai_steward.py
@@ -0,0 +1,1430 @@
+"""
+AI Steward API - 主动式 AI 管家接口
+
+核心功能：
+1. 零配置项目创建
+2. 主动信号捕获
+3. 认知冲突检测
+4. 阻塞自动解决
+5. 个性化进度简报
+"""
+from fastapi import APIRouter, Depends, HTTPException, BackgroundTasks
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+from pydantic import BaseModel, Field
+from typing import List, Optional, Dict, Any, Literal
+from datetime import datetime, timezone, timedelta
+import json
+
+from ..core.response import success_response
+from ..database import get_db
+from ..dependencies.auth import get_current_user
+from ..dependencies.permissions import require_manager_user
+from ..dependencies.tenant import get_tenant_id
+from ..models import ProjectSignalLink, Signal, User
+from ..services.ai_steward import get_ai_steward_service, AIStewardService
+from ..services.signal_capture import ProactiveSignalCaptureService
+from ..services.cognitive_conflict import get_cognitive_conflict_service
+from ..services.blocker_detection import get_blocker_detection_service
+from ..services.daily_briefing import get_daily_briefing_service
+from ..services.event_workflow import get_event_workflow_engine, WorkflowEventType
+from ..services.adaptive_rag_service import AdaptiveRAGService
+from ..services.ai_orchestration_run_service import AIOrchestrationRunService
+from ..services.openclaw_commander_service import OpenClawCommanderService
+from ..services.signal_source_orchestration_service import SignalSourceOrchestrationService
+from ..schemas.rag import RAGAskRequest
+
+
+router = APIRouter(prefix="/ai-steward", tags=["AI Steward"])
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+async def _ensure_project_signal_link(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    signal_id: int,
+    link_type: str = "related",
+) -> None:
+    existing = (
+        await db.execute(
+            select(ProjectSignalLink.id).where(
+                and_(
+                    ProjectSignalLink.tenant_id == int(tenant_id),
+                    ProjectSignalLink.project_id == int(project_id),
+                    ProjectSignalLink.signal_id == int(signal_id),
+                )
+            ).limit(1)
+        )
+    ).scalar_one_or_none()
+    if existing is not None:
+        return
+    db.add(
+        ProjectSignalLink(
+            tenant_id=int(tenant_id),
+            project_id=int(project_id),
+            signal_id=int(signal_id),
+            link_type=link_type,
+            relevance_score=1.0,
+        )
+    )
+
+
+# ==================== Request/Response Models ====================
+
+class ProjectFromTextRequest(BaseModel):
+    """零配置项目创建请求"""
+    input_text: str = Field(
+        ...,
+        description="自然语言项目描述，例如：我们要做一个 AI 客服工具，基于 GPT-4o mini，对标 XX 竞品，3 周上线",
+        min_length=10,
+        max_length=500
+    )
+
+
+class ProjectFromTextResponse(BaseModel):
+    """零配置项目创建响应"""
+    project: Dict[str, Any]
+    signal: Dict[str, Any]
+    charter: Dict[str, Any]
+    tasks: List[Dict[str, Any]]
+    risks: List[Dict[str, Any]]
+    analysis: Dict[str, Any]
+
+
+class CognitiveConflictResponse(BaseModel):
+    """认知冲突响应"""
+    id: str
+    type: str
+    description: str
+    severity: str
+    resolution_suggestion: Optional[str] = None
+    related_users: List[str]
+
+
+class DailyBriefingResponse(BaseModel):
+    """每日简报响应"""
+    date: str
+    total_updates: int
+    your_updates: int
+    relevant_changes: List[Dict[str, Any]]
+    risk_alerts: List[Dict[str, Any]]
+    alignment_needed: List[Dict[str, Any]]
+
+
+class FeishuCaptureRequest(BaseModel):
+    """飞书信号捕获请求"""
+    group_chat_id: str
+    message_content: str
+    sender: str
+    feishu_message_id: Optional[str] = None
+
+
+class BrowserCaptureRequest(BaseModel):
+    """浏览器信号捕获请求"""
+    user_id: int
+    url: str
+    title: str
+    content_preview: str
+
+
+class GitHubCaptureRequest(BaseModel):
+    """GitHub信号捕获请求"""
+    repo_url: str
+    event_type: str
+    event_data: Dict[str, Any]
+
+
+class ConflictDetectRequest(BaseModel):
+    """认知冲突检测请求"""
+    project_id: int
+    detect_type: str = "all"  # all, charter, definition
+
+
+class OpenClawCommandDispatchRequest(BaseModel):
+    """OpenClaw 指挥层任务分配请求"""
+    project_id: int = Field(..., ge=1)
+    instruction: str = Field(..., min_length=4, max_length=2000)
+    dry_run: bool = False
+    max_operations: int = Field(default=8, ge=1, le=20)
+    response_mode: Literal["fast", "complete"] = "fast"
+
+
+class SignalSourceOnboardRequest(BaseModel):
+    project_id: int = Field(..., ge=1)
+    instruction: str = Field(..., min_length=6, max_length=2000)
+    dry_run: bool = False
+    response_mode: Literal["fast", "complete"] = "fast"
+
+
+class SignalSourceListResponse(BaseModel):
+    items: List[Dict[str, Any]]
+
+
+def _parse_related_users(raw_value: Any) -> List[str]:
+    """兼容 involved_user_ids 的 list / json-string / 单值字符串。"""
+    if raw_value is None:
+        return []
+    if isinstance(raw_value, list):
+        return [str(item) for item in raw_value if item is not None]
+    if isinstance(raw_value, (bytes, bytearray)):
+        raw_value = raw_value.decode("utf-8", errors="ignore")
+    if isinstance(raw_value, str):
+        text = raw_value.strip()
+        if not text:
+            return []
+        try:
+            parsed = json.loads(text)
+        except (TypeError, ValueError, json.JSONDecodeError):
+            if "," in text:
+                return [v.strip() for v in text.split(",") if v.strip()]
+            return [text]
+        if isinstance(parsed, list):
+            return [str(item) for item in parsed if item is not None]
+        if parsed is None:
+            return []
+        return [str(parsed)]
+    return [str(raw_value)]
+
+
+# ==================== Zero-Config Project Creation ====================
+
+@router.post("/rag/ask")
+async def ask_ai_steward_rag(
+    request: RAGAskRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    effective_tenant_id = int(getattr(current_user, "tenant_id", tenant_id) or tenant_id)
+    service = AdaptiveRAGService(db, tenant_id=effective_tenant_id)
+    result = await service.answer_question(
+        question=request.question,
+        asked_by=int(current_user.id),
+        asked_via="pmflow",
+        project_id=request.project_id,
+        source_chat_id=request.source_chat_id,
+        session_id=request.session_id,
+        force_new_session=bool(request.new_session),
+        thinking_mode=request.thinking_mode,
+        action_mode=request.action_mode,
+        sync_requirement=bool(request.sync_requirement),
+    )
+    return success_response(result)
+
+
+@router.post("/command/dispatch")
+async def dispatch_openclaw_command(
+    request: OpenClawCommandDispatchRequest,
+    current_user: User = Depends(require_manager_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    effective_tenant_id = int(getattr(current_user, "tenant_id", tenant_id) or tenant_id)
+    service = OpenClawCommanderService(db)
+    try:
+        result = await service.dispatch_command(
+            tenant_id=effective_tenant_id,
+            operator_user=current_user,
+            project_id=request.project_id,
+            instruction=request.instruction,
+            dry_run=request.dry_run,
+            max_operations=request.max_operations,
+            response_mode=request.response_mode,
+        )
+    except ValueError as exc:
+        detail = str(exc)
+        if detail == "project not found":
+            raise HTTPException(status_code=404, detail=detail) from exc
+        raise HTTPException(status_code=400, detail=detail) from exc
+    return success_response(result)
+
+
+@router.post("/signal-sources/onboard")
+async def onboard_signal_source(
+    request: SignalSourceOnboardRequest,
+    current_user: User = Depends(require_manager_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    effective_tenant_id = int(getattr(current_user, "tenant_id", tenant_id) or tenant_id)
+    signal = Signal(
+        tenant_id=effective_tenant_id,
+        source="PMFlow",
+        content=f"[PMFlow用户#{int(current_user.id)}]: {request.instruction[:1800]}",
+        impact_score=0.7,
+        priority="medium",
+        processed=False,
+        strategic_analysis=json.dumps(
+            {
+                "source": "pmflow_manual_signal_source_onboard",
+                "instruction": request.instruction[:2000],
+                "created_at": datetime.now(timezone.utc).isoformat(),
+            },
+            ensure_ascii=False,
+        ),
+        tasks_generated=json.dumps([], ensure_ascii=False),
+        matched_project_id=int(request.project_id),
+        title=(request.instruction[:80] or "新增信号源"),
+        signal_type="新需求",
+        summary=request.instruction[:300],
+        reasoning=json.dumps(["来自 AI Steward 手动信号源接入请求。"], ensure_ascii=False),
+        confidence=0.82,
+        recommended_action="enter_only",
+        status="待处理",
+    )
+    if not request.dry_run:
+        db.add(signal)
+        await db.flush()
+        await _ensure_project_signal_link(
+            db,
+            tenant_id=effective_tenant_id,
+            project_id=int(request.project_id),
+            signal_id=int(signal.id),
+        )
+        await db.commit()
+        await db.refresh(signal)
+    signal_id_value = int(signal.id) if not request.dry_run and getattr(signal, "id", None) is not None else None
+
+    service = SignalSourceOrchestrationService(db)
+    result = await service.orchestrate_from_signal(
+        tenant_id=effective_tenant_id,
+        project_id=int(request.project_id),
+        signal_id=signal_id_value,
+        initiating_text=request.instruction,
+        requested_via="pmflow",
+        operator_user_id=int(current_user.id),
+        dry_run=bool(request.dry_run),
+        response_mode=request.response_mode,
+    )
+    return success_response(
+        {
+            "signal_id": signal_id_value,
+            **result,
+        }
+    )
+
+
+@router.get("/orchestration-runs/{trace_id}")
+async def get_orchestration_run(
+    trace_id: str,
+    current_user: User = Depends(require_manager_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    effective_tenant_id = int(getattr(current_user, "tenant_id", tenant_id) or tenant_id)
+    row = await AIOrchestrationRunService(db).get_run(
+        tenant_id=effective_tenant_id,
+        trace_id=str(trace_id or "").strip(),
+    )
+    if row is None:
+        raise HTTPException(status_code=404, detail="orchestration run not found")
+    return success_response(AIOrchestrationRunService.serialize_run(row))
+
+
+@router.get("/projects/{project_id}/signal-sources")
+async def list_project_signal_sources(
+    project_id: int,
+    current_user: User = Depends(require_manager_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    effective_tenant_id = int(getattr(current_user, "tenant_id", tenant_id) or tenant_id)
+    items = await SignalSourceOrchestrationService(db).list_project_sources(
+        tenant_id=effective_tenant_id,
+        project_id=int(project_id),
+    )
+    return success_response(SignalSourceListResponse(items=items).model_dump())
+
+@router.post("/project-from-text", response_model=ProjectFromTextResponse)
+async def create_project_from_text(
+    request: ProjectFromTextRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    零配置项目创建 - 从一句话生成完整项目
+
+    示例输入：
+    ```
+    我们要做一个 AI 客服工具，基于 GPT-4o mini，对标 XX 竞品，3 周上线
+    ```
+
+    AI 将自动生成：
+    - 完整的项目配置
+    - 竞品分析信号
+    - 项目宪章
+    - 任务链（含依赖关系）
+    - 风险预警规则
+    """
+    try:
+        service = get_ai_steward_service(db)
+
+        result = await service.create_project_from_natural_language(
+            input_text=request.input_text,
+            creator_id=int(current_user.id),
+            tenant_id=tenant_id
+        )
+
+        return ProjectFromTextResponse(**result)
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"项目创建失败: {str(e)}"
+        )
+
+
+# ==================== Signal Capture Endpoints ====================
+
+@router.post("/signal/capture-feishu")
+async def capture_feishu_signal(
+    request: FeishuCaptureRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    从飞书群聊捕获信号
+
+    AI 会自动判断消息是否包含有价值的项目或技术信息，
+    如果是，则创建信号并自动关联到相关项目
+    """
+    try:
+        service = ProactiveSignalCaptureService(db)
+        result = await service.capture_from_feishu(
+            group_chat_id=request.group_chat_id,
+            message_content=request.message_content,
+            sender=request.sender,
+            tenant_id=tenant_id,
+            feishu_message_id=request.feishu_message_id,
+        )
+        return result
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"飞书信号捕获失败: {str(e)}")
+
+
+@router.post("/signal/capture-browser")
+async def capture_browser_signal(
+    request: BrowserCaptureRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    从浏览器标签页捕获信号
+
+    监控用户访问的技术文档、竞品页面、GitHub issues等
+    """
+    try:
+        service = ProactiveSignalCaptureService(db)
+        result = await service.capture_from_browser(
+            user_id=request.user_id,
+            url=request.url,
+            title=request.title,
+            content_preview=request.content_preview,
+            tenant_id=tenant_id
+        )
+        return result
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"浏览器信号捕获失败: {str(e)}")
+
+
+@router.post("/signal/capture-github")
+async def capture_github_signal(
+    request: GitHubCaptureRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    从GitHub事件捕获信号
+
+    监控竞品仓库更新、上游依赖更新等
+    """
+    try:
+        service = ProactiveSignalCaptureService(db)
+        result = await service.capture_from_github(
+            repo_url=request.repo_url,
+            event_type=request.event_type,
+            event_data=request.event_data,
+            tenant_id=tenant_id
+        )
+        return result
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"GitHub信号捕获失败: {str(e)}")
+
+
+@router.post("/signal/start-hn-monitor")
+async def start_hn_monitor(
+    background_tasks: BackgroundTasks,
+    interval_minutes: int = 30,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    启动 HackerNews 定时监控（后台任务）
+
+    定期抓取 HackerNews 首页，分析技术趋势
+    """
+    async def run_monitor():
+        service = ProactiveSignalCaptureService(db)
+        await service.start_hackernews_monitor(interval_minutes=interval_minutes, tenant_id=tenant_id)
+
+    background_tasks.add_task(run_monitor)
+    return {
+        "status": "started",
+        "message": f"HackerNews 监控已启动，间隔: {interval_minutes} 分钟"
+    }
+
+
+# ==================== Cognitive Conflict Endpoints ====================
+
+@router.get("/cognitive-conflicts", response_model=List[CognitiveConflictResponse])
+async def get_cognitive_conflicts(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+    project_id: Optional[int] = None
+):
+    """
+    获取认知冲突列表
+
+    AI 会自动识别团队中的隐性认知差异，例如：
+    - 产品说"用户留存提升10%"，开发实现的是"DAU提升10%"
+    - 双方没有争论，但目标不一致，AI 会主动识别并推送对齐提醒
+    """
+    from ..models import CognitiveConflict, TruthNode
+    from sqlalchemy.future import select
+
+    query = select(CognitiveConflict).where(
+        CognitiveConflict.tenant_id == tenant_id,
+        CognitiveConflict.status == 'active'
+    )
+
+    if project_id:
+        # 通过 truth_node 关联到项目
+        query = query.join(
+            TruthNode, CognitiveConflict.truth_node_id == TruthNode.id
+        ).where(
+            TruthNode.project_id == project_id
+        )
+
+    result = await db.execute(query)
+    conflicts = result.scalars().all()
+
+    return [
+        CognitiveConflictResponse(
+            id=str(c.id),
+            type=c.conflict_type,
+            description=c.description,
+            severity="high" if "critical" in (c.description or "").lower() else "medium",
+            related_users=_parse_related_users(c.involved_user_ids),
+        )
+        for c in conflicts
+    ]
+
+
+@router.post("/conflicts/detect")
+async def detect_conflicts(
+    request: ConflictDetectRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    检测项目中的认知冲突
+
+    支持检测类型：
+    - all: 全部检测（宪章冲突 + 定义冲突）
+    - charter: 仅检测宪章中的潜在冲突
+    - definition: 仅检测用户定义冲突
+    """
+    try:
+        service = get_cognitive_conflict_service(db)
+
+        detected_conflicts = []
+
+        if request.detect_type in ["all", "charter"]:
+            charter_conflicts = await service.detect_charter_conflicts(
+                project_id=request.project_id,
+                tenant_id=tenant_id
+            )
+            detected_conflicts.extend(charter_conflicts)
+
+        if request.detect_type in ["all", "definition"]:
+            definition_conflicts = await service.detect_user_definition_conflicts(
+                project_id=request.project_id,
+                tenant_id=tenant_id
+            )
+            detected_conflicts.extend(definition_conflicts)
+
+        return {
+            "project_id": request.project_id,
+            "detected_count": len(detected_conflicts),
+            "conflicts": detected_conflicts
+        }
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"冲突检测失败: {str(e)}")
+
+
+@router.post("/conflicts/{conflict_id}/propose-resolution")
+async def propose_conflict_resolution(
+    conflict_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    AI 生成认知冲突解决方案
+
+    为冲突双方生成对齐方案：
+    1. 分析冲突根因
+    2. 提出妥协方案
+    3. 生成可执行的对齐计划
+    """
+    try:
+        service = get_cognitive_conflict_service(db)
+        resolution = await service.propose_alignment_resolution(
+            conflict_id=conflict_id,
+            tenant_id=tenant_id
+        )
+        return resolution
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"方案生成失败: {str(e)}")
+
+
+@router.post("/conflicts/{conflict_id}/align")
+async def trigger_async_alignment(
+    conflict_id: int,
+    user1_id: int,
+    user2_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    触发异步 1v1 认知对齐
+
+    为冲突双方创建独立的对齐任务：
+    1. 发送对齐邀请
+    2. 提供背景信息
+    3. 收集双方观点
+    4. AI 生成共识方案
+    """
+    try:
+        service = get_cognitive_conflict_service(db)
+        alignment = await service.trigger_async_alignment(
+            conflict_id=conflict_id,
+            user1_id=user1_id,
+            user2_id=user2_id,
+            tenant_id=tenant_id
+        )
+        return alignment
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"对齐触发失败: {str(e)}")
+
+
+@router.post("/cognitive-conflicts/{conflict_id}/resolve")
+async def resolve_cognitive_conflict(
+    conflict_id: int,
+    resolution_note: str,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    解决认知冲突
+
+    AI 会生成解决方案建议，双方确认后标记为已解决
+    """
+    from ..models import CognitiveConflict
+    from sqlalchemy.future import select
+
+    result = await db.execute(
+        select(CognitiveConflict).where(
+            CognitiveConflict.id == conflict_id,
+            CognitiveConflict.tenant_id == tenant_id
+        )
+    )
+    conflict = result.scalar_one_or_none()
+
+    if not conflict:
+        raise HTTPException(status_code=404, detail="冲突记录不存在")
+
+    conflict.status = 'resolved'
+    conflict.resolved_at = _utc_now_naive()
+    conflict.description += f"\n\n解决方案（{current_user.username}）: {resolution_note}"
+
+    await db.commit()
+
+    return {"message": "冲突已标记为已解决"}
+
+
+@router.get("/conflicts/task-mismatch/{task_id}")
+async def check_task_charter_mismatch(
+    task_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    检测任务与宪章的一致性
+
+    检查：
+    1. 任务描述是否与宪章目标一致
+    2. 技术选型是否符合约束
+    3. 是否偏离项目方向
+    """
+    from ..models import Task, TruthNode
+    from sqlalchemy.future import select
+
+    try:
+        # 获取任务
+        result = await db.execute(
+            select(Task).where(
+                Task.id == task_id,
+                Task.tenant_id == tenant_id
+            )
+        )
+        task = result.scalar_one_or_none()
+
+        if not task:
+            raise HTTPException(status_code=404, detail="任务不存在")
+
+        # 获取宪章
+        charter_result = await db.execute(
+            select(TruthNode).where(
+                TruthNode.id == task.truth_node_id,
+                TruthNode.tenant_id == tenant_id
+            )
+        )
+        charter = charter_result.scalar_one_or_none()
+
+        if not charter:
+            raise HTTPException(status_code=404, detail="宪章不存在")
+
+        # 检测一致性
+        service = get_cognitive_conflict_service(db)
+        mismatch = await service.detect_task_charter_mismatch(
+            task=task,
+            charter_content=charter.content,
+            tenant_id=tenant_id
+        )
+
+        if mismatch:
+            return {
+                "is_aligned": False,
+                "mismatch": mismatch
+            }
+
+        return {
+            "is_aligned": True,
+            "message": "任务与宪章一致"
+        }
+
+    except HTTPException:
+        raise
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"一致性检测失败: {str(e)}")
+
+
+# ==================== Other AI Steward Endpoints ====================
+
+@router.get("/daily-briefing")
+async def get_daily_briefing(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    获取个性化每日简报
+
+    AI 为每个角色生成定制化简报，只包含：
+    - 与你相关的任务变更
+    - 风险提醒
+    - 需要对齐的事项
+    - 高影响信号
+    - 团队更新
+
+    彻底过滤无关信息，替代传统日报/周报
+    """
+    try:
+        service = get_daily_briefing_service(db)
+        briefing = await service.generate_personalized_briefing(
+            user_id=int(current_user.id),
+            tenant_id=tenant_id
+        )
+        return briefing
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"生成简报失败: {str(e)}"
+        )
+
+
+@router.get("/daily-briefing/team/{project_id}")
+async def get_team_briefing(
+    project_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    获取团队简报（给管理者）
+
+    包含：
+    - 整体进度
+    - 每个成员的工作量
+    - 风险和阻塞
+    - 需要决策的事项
+    """
+    try:
+        service = get_daily_briefing_service(db)
+        briefing = await service.generate_team_briefing(
+            project_id=project_id,
+            tenant_id=tenant_id
+        )
+        return briefing
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"生成团队简报失败: {str(e)}"
+        )
+
+
+@router.post("/analyze-blocker")
+async def analyze_blocker(
+    code_context: str,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    分析代码阻塞
+
+    当开发在代码注释中提到「这个接口达不到效果」时，
+    AI 自动触发分析，推荐替代方案或生成代码片段
+    """
+    from ..services.llm import LLMService
+
+    service = get_ai_steward_service(db)
+
+    prompt = f"""分析以下代码/注释中提到的问题，提供解决方案：
+
+代码上下文：
+{code_context}
+
+请提供：
+1. 问题分析
+2. 推荐的替代方案
+3. 示例代码（如适用）
+
+以 JSON 格式返回。"""
+
+    try:
+        llm = LLMService()
+        response = await llm.generate(
+            prompt,
+            db=db,
+            tenant_id=tenant_id,
+        )
+
+        return {
+            'analysis': response,
+            'suggestions': [],
+            'auto_generated': True
+        }
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"分析失败: {str(e)}"
+        )
+
+
+# ==================== Blocker Detection Endpoints ====================
+
+class CodeCommentScanRequest(BaseModel):
+    """代码注释扫描请求"""
+    code_content: str
+    file_path: str
+    line_number: int
+    project_id: int
+
+
+class BlockerResolveRequest(BaseModel):
+    """阻塞解决请求"""
+    blocker_id: int
+    resolution_note: str
+
+
+@router.post("/blockers/scan-comment")
+async def scan_code_comment_for_blockers(
+    request: CodeCommentScanRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    扫描代码注释，检测阻塞
+
+    自动识别阻塞模式：
+    - "等待 X 接口完成"
+    - "这个 API 达不到效果"
+    - "由于 Y 的限制，暂时跳过"
+
+    如果检测到阻塞，AI 会自动：
+    1. 分析严重程度
+    2. 生成解决方案建议
+    3. 创建追踪任务（如需要）
+    """
+    try:
+        service = get_blocker_detection_service(db)
+        result = await service.scan_code_comment(
+            code_content=request.code_content,
+            file_path=request.file_path,
+            line_number=request.line_number,
+            project_id=request.project_id,
+            user_id=int(current_user.id),
+            tenant_id=tenant_id
+        )
+
+        if result:
+            return {
+                'status': 'blocker_detected',
+                'blocker': result
+            }
+
+        return {
+            'status': 'no_blocker',
+            'message': '未检测到阻塞'
+        }
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"扫描失败: {str(e)}"
+        )
+
+
+@router.post("/blockers/detect-tasks")
+async def detect_task_blockers(
+    background_tasks: BackgroundTasks,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    检测所有任务中的阻塞
+
+    扫描所有超过3天未更新的进行中任务，
+    使用 AI 判断是否存在阻塞
+    """
+    try:
+        service = get_blocker_detection_service(db)
+        blockers = await service.detect_task_blockers(tenant_id=tenant_id)
+
+        return {
+            'total_detected': len(blockers),
+            'blockers': blockers
+        }
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"检测失败: {str(e)}"
+        )
+
+
+@router.get("/blockers/active")
+async def get_active_blockers(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    获取所有活跃的阻塞
+
+    返回所有未解决的阻塞，按严重程度排序
+    """
+    try:
+        service = get_blocker_detection_service(db)
+        blockers = await service.get_active_blockers(tenant_id=tenant_id)
+
+        # 按严重程度排序
+        severity_order = {'high': 0, 'medium': 1, 'low': 2}
+        blockers.sort(key=lambda b: severity_order.get(b['severity'], 1))
+
+        return {
+            'total': len(blockers),
+            'blockers': blockers
+        }
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"获取失败: {str(e)}"
+        )
+
+
+@router.post("/blockers/resolve")
+async def resolve_blocker(
+    request: BlockerResolveRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    标记阻塞为已解决
+
+    记录解决方案并关闭阻塞记录
+    """
+    try:
+        service = get_blocker_detection_service(db)
+        result = await service.resolve_blocker(
+            blocker_id=request.blocker_id,
+            resolution_note=request.resolution_note,
+            tenant_id=tenant_id
+        )
+
+        return result
+
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"解决失败: {str(e)}"
+        )
+
+
+@router.get("/stats")
+async def get_ai_steward_stats(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    获取 AI Steward 统计数据
+
+    展示 AI 为团队节省的工作量：
+    - 自动捕获的信号数
+    - 避免的认知冲突数
+    - 自动解决的阻塞数
+    - 节省的会议时间（小时）
+    """
+    from ..models import Signal, CognitiveConflict
+    from sqlalchemy import func, text
+
+    # 统计信号数量
+    week_ago = _utc_now_naive() - timedelta(days=7)
+    month_ago = _utc_now_naive() - timedelta(days=30)
+
+    signal_result = await db.execute(
+        select(func.count(Signal.id)).where(
+            Signal.tenant_id == tenant_id,
+            Signal.created_at >= week_ago
+        )
+    )
+    signals_captured = signal_result.scalar() or 0
+
+    # 统计高影响信号
+    high_impact_result = await db.execute(
+        select(func.count(Signal.id)).where(
+            Signal.tenant_id == tenant_id,
+            Signal.impact_score >= 0.8,
+            Signal.created_at >= week_ago
+        )
+    )
+    high_impact_count = high_impact_result.scalar() or 0
+
+    # 统计认知冲突
+    conflict_result = await db.execute(
+        select(func.count(CognitiveConflict.id)).where(
+            CognitiveConflict.tenant_id == tenant_id,
+            CognitiveConflict.status.in_(['detected', 'active'])
+        )
+    )
+    active_conflicts = conflict_result.scalar() or 0
+
+    # 统计已解决的冲突
+    resolved_conflict_result = await db.execute(
+        select(func.count(CognitiveConflict.id)).where(
+            CognitiveConflict.tenant_id == tenant_id,
+            CognitiveConflict.status == 'resolved',
+            CognitiveConflict.resolved_at >= month_ago
+        )
+    )
+    conflicts_resolved = resolved_conflict_result.scalar() or 0
+
+    # 统计阻塞
+    blocker_result = await db.execute(
+        text("""
+            SELECT COUNT(*) FROM cognitive_conflicts
+            WHERE tenant_id = :tenant_id
+            AND conflict_type IN ('code_blocker', 'task_blocker')
+            AND status = 'detected'
+        """),
+        {"tenant_id": tenant_id}
+    )
+    active_blockers = blocker_result.scalar() or 0
+
+    # 活跃项目数
+    from ..models import Project
+    project_result = await db.execute(
+        select(func.count(Project.id)).where(
+            Project.tenant_id == tenant_id,
+            Project.status == 'active'
+        )
+    )
+    active_projects = project_result.scalar() or 0
+
+    # 估算节省时间（每个冲突/阻塞避免约2小时会议）
+    hours_saved = conflicts_resolved * 2
+
+    stats = {
+        'signals_captured': signals_captured,
+        'high_impact_signals': high_impact_count,
+        'active_conflicts': active_conflicts,
+        'conflicts_resolved': conflicts_resolved,
+        'active_blockers': active_blockers,
+        'active_projects': active_projects,
+        'hours_saved': hours_saved,
+        'last_updated': datetime.now(timezone.utc).isoformat()
+    }
+
+    return stats
+
+
+@router.post("/sync-signal")
+async def sync_external_signal(
+    source: str,
+    content: str,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    手动同步外部信号
+
+    支持从以下来源同步：
+    - 飞书群聊
+    - GitHub commits/PRs
+    - 浏览器标签页
+    """
+    from ..models import Signal
+    from ..services.ai_steward import get_ai_steward_service
+
+    service = get_ai_steward_service(db)
+
+    # AI 分析信号重要性
+    prompt = f"""分析以下信号的重要性：
+
+来源：{source}
+内容：{content}
+
+请评估：
+1. impact_score (0-1)
+2. priority (low/medium/high/critical)
+3. 是否需要立即行动
+
+以 JSON 格式返回。"""
+
+    try:
+        from ..services.llm import LLMService
+        llm = LLMService()
+        response = await llm.generate(
+            prompt,
+            db=db,
+            tenant_id=tenant_id,
+        )
+
+        # 创建信号
+        signal = Signal(
+            source=source,
+            content=content,
+            impact_score=0.7,
+            priority='medium',
+            processed=False,
+            tenant_id=tenant_id
+        )
+        db.add(signal)
+        await db.commit()
+
+        return {
+            'signal_id': str(signal.id),
+            'analysis': response,
+            'status': 'captured'
+        }
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"信号同步失败: {str(e)}"
+        )
+
+
+# ==================== Event-Driven Workflow Endpoints ====================
+
+class WorkflowTriggerRequest(BaseModel):
+    """工作流触发请求"""
+    event_type: str
+    event_data: Dict[str, Any]
+
+
+@router.post("/workflow/trigger")
+async def trigger_workflow(
+    request: WorkflowTriggerRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    触发事件驱动工作流
+
+    支持的事件类型：
+    - high_impact_signal: 高影响信号到达
+    - blocker_detected: 阻塞检测
+    - cognitive_conflict: 认知冲突
+    - tech_stack_change: 技术栈变更
+    - competitor_update: 竞品动态
+    - task_completed: 任务完成
+    - task_blocked: 任务阻塞
+    - project_risk: 项目风险
+
+    工作流会自动：
+    1. 调整项目优先级
+    2. 创建响应任务
+    3. 通知相关角色
+    4. 更新依赖关系
+    """
+    try:
+        engine = get_event_workflow_engine(db)
+        result = await engine.trigger_workflow(
+            event_type=request.event_type,
+            event_data=request.event_data,
+            tenant_id=tenant_id
+        )
+        return result
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"工作流触发失败: {str(e)}"
+        )
+
+
+@router.post("/workflow/signal/{signal_id}")
+async def trigger_signal_workflow(
+    signal_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    当接收到高影响信号时自动触发工作流
+
+    检查信号的 impact_score，如果超过阈值（0.85），
+    自动触发工作流来调整相关项目的优先级
+    """
+    from ..models import Signal
+    from sqlalchemy.future import select
+
+    signal_result = await db.execute(
+        select(Signal).where(
+            Signal.id == signal_id,
+            Signal.tenant_id == tenant_id
+        )
+    )
+    signal = signal_result.scalar_one_or_none()
+
+    if not signal:
+        raise HTTPException(status_code=404, detail="信号不存在")
+
+    # 检查是否达到阈值
+    if signal.impact_score < 0.85:
+        return {
+            'status': 'ignored',
+            'reason': f'信号影响分数 {signal.impact_score} 低于阈值 0.85'
+        }
+
+    # 触发高影响信号工作流
+    engine = get_event_workflow_engine(db)
+    result = await engine.trigger_workflow(
+        event_type=WorkflowEventType.HIGH_IMPACT_SIGNAL,
+        event_data={'signal_id': signal_id},
+        tenant_id=tenant_id
+    )
+
+    return result
+
+
+@router.get("/workflow/status")
+async def get_workflow_status(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    获取工作流状态
+
+    显示最近触发的工作流及其结果
+    """
+    from ..models import AICallLog
+    from sqlalchemy.future import select
+
+    # 获取最近的工作流记录
+    # 这里简化处理，实际应该有专门的工作流记录表
+    return {
+        'active_workflows': [],
+        'recent_triggers': [],
+        'message': '工作流状态追踪'
+    }
+
+
+@router.post("/workflow/auto-enable")
+async def enable_auto_workflow(
+    project_id: int,
+    enable: bool = True,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    启用/禁用项目的自动工作流
+
+    启用后，当以下情况发生时会自动触发工作流：
+    - 高影响信号到达（>0.85）
+    - 任务阻塞检测
+    - 认知冲突检测
+    """
+    from ..models import Project
+    from sqlalchemy.future import select
+
+    project_result = await db.execute(
+        select(Project).where(
+            Project.id == project_id,
+            Project.tenant_id == tenant_id
+        )
+    )
+    project = project_result.scalar_one_or_none()
+
+    if not project:
+        raise HTTPException(status_code=404, detail="项目不存在")
+
+    # 这里应该设置项目的 auto_workflow 标志
+    # 由于 Project 模型可能没有这个字段，暂时返回成功
+    return {
+        'project_id': project_id,
+        'auto_workflow': enable,
+        'message': f'自动工作流已{"启用" if enable else "禁用"}'
+    }
+
+
+# ==================== AI Config Suggestion ====================
+
+@router.get("/config/suggest")
+async def get_ai_config_suggestion(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    AI 建议配置方案
+
+    根据团队现状，智能推荐应该优先配置哪些集成源
+    """
+    from ..models import Integration, Project, Signal
+    from sqlalchemy.future import select
+    from sqlalchemy import func
+
+    # 检查已配置的集成
+    configured_sources = []
+
+    # 检查飞书配置
+    feishu_result = await db.execute(
+        select(Integration).where(
+            Integration.tenant_id == tenant_id,
+            Integration.type == 'feishu',
+            Integration.is_active == True
+        )
+    )
+    if feishu_result.scalar_one_or_none():
+        configured_sources.append('feishu')
+
+    # 检查活跃项目数
+    project_result = await db.execute(
+        select(func.count(Project.id)).where(
+            Project.tenant_id == tenant_id,
+            Project.status == 'active'
+        )
+    )
+    active_projects = project_result.scalar() or 0
+
+    # 检查近期信号数量
+    signal_result = await db.execute(
+        select(func.count(Signal.id)).where(
+            Signal.tenant_id == tenant_id,
+            Signal.created_at >= (_utc_now_naive() - timedelta(days=7))
+        )
+    )
+    recent_signals = signal_result.scalar() or 0
+
+    # AI 分析并推荐配置
+    recommended = []
+    reasoning_parts = []
+
+    # 根据项目数量推荐
+    if active_projects > 0:
+        if 'feishu' not in configured_sources:
+            recommended.append('feishu')
+            reasoning_parts.append(f"您有 {active_projects} 个活跃项目，建议配置飞书群聊集成以自动捕获需求讨论")
+
+    # 根据信号量推荐
+    if recent_signals < 10:
+        if 'hn' not in [s.split('_')[0] for s in configured_sources]:
+            recommended.append('hn')
+            reasoning_parts.append("近期信号较少，建议启用 HackerNews 监控获取更多技术趋势")
+
+    # 默认推荐
+    if not recommended:
+        if 'feishu' not in configured_sources:
+            recommended.append('feishu')
+        if 'hn' not in [s.split('_')[0] for s in configured_sources]:
+            recommended.append('hn')
+        reasoning_parts.append("建议配置基础集成源以启动 AI 主动捕获能力")
+
+    reasoning = "；".join(reasoning_parts) if reasoning_parts else "所有基础集成已配置，AI 主动捕获能力已全面启用"
+
+    # 估算影响
+    impact_levels = {
+        0: "暂无明显影响",
+        1: "预计每周节省 2 小时手动录入时间",
+        2: "预计每周节省 4 小时手动录入时间，捕获 10+ 个需求信号",
+        3: "预计每周节省 6 小时会议时间，自动对齐 5+ 个认知冲突",
+        4: "预计每周节省 8 小时会议时间，主动捕获 20+ 个需求信号"
+    }
+    impact = impact_levels.get(len(recommended), impact_levels[4])
+
+    return {
+        'recommended_sources': recommended,
+        'reasoning': reasoning,
+        'estimated_impact': impact,
+        'current_config': {
+            'feishu_configured': 'feishu' in configured_sources,
+            'hn_configured': any('hn' in s for s in configured_sources),
+            'github_configured': any('github' in s for s in configured_sources),
+            'browser_installed': False  # 需要前端检测
+        }
+    }
diff --git a/backend/app/api/feishu.py b/backend/app/api/feishu.py
new file mode 100644
index 00000000..0eb0ec0a
--- /dev/null
+++ b/backend/app/api/feishu.py
@@ -0,0 +1,376 @@
+"""Feishu Integration API - 飞书集成接口。"""
+from fastapi import APIRouter, Depends, HTTPException
+from sqlalchemy.ext.asyncio import AsyncSession
+from pydantic import BaseModel, Field
+from typing import List, Optional, Dict, Any
+from ..database import get_db
+from ..dependencies.auth import get_current_user
+from ..dependencies.tenant import get_tenant_id
+from ..models import User
+from ..services.auth import has_any_role
+from ..services.feishu_integration import get_feishu_service
+from ..services.feishu_long_connection import get_feishu_long_connection_manager
+
+
+router = APIRouter(prefix="/feishu", tags=["Feishu Integration"])
+
+
+def _build_long_connection_failure(long_status: Dict[str, Any]) -> tuple[str, Optional[str]]:
+    code = str(long_status.get("last_error_code") or "").strip().lower()
+    last_error = str(long_status.get("last_error") or "").strip()
+    retryable = bool(long_status.get("last_error_retryable"))
+    if code in {"network_error", "thread_exited"} and not retryable:
+        return (
+            "长连接自动重连次数已耗尽，请检查网络/凭证后手动重启飞书长连接",
+            last_error or None,
+        )
+    if code == "auth_failed":
+        return (
+            "长连接鉴权失败，请检查 App ID / App Secret、应用凭证是否已生效，以及飞书应用权限是否完整",
+            last_error or None,
+        )
+    if code == "network_error":
+        return (
+            "长连接预检失败，后端到飞书开放平台的网络或 TLS 证书链异常",
+            last_error or None,
+        )
+    if code == "sdk_missing":
+        return ("长连接 SDK 未安装（缺少 lark-oapi）", None)
+    if code == "thread_exited":
+        return ("长连接已退出，系统正在等待下一次自动重连窗口", last_error or None)
+    if last_error:
+        return ("长连接未建立，请检查配置与后端日志", last_error)
+    return ("长连接未建立，请检查配置后重试", None)
+
+
+# ==================== Request/Response Models ====================
+
+class FeishuConfigRequest(BaseModel):
+    """飞书配置请求"""
+    app_id: str = Field(..., description="飞书 App ID")
+    app_secret: str = Field(..., description="飞书 App Secret")
+    encrypt_key: Optional[str] = Field(None, description="已废弃字段，保留仅为兼容旧客户端")
+    verification_token: Optional[str] = Field(None, description="已废弃字段，保留仅为兼容旧客户端")
+    event_mode: str = Field("long_connection", description="事件接收模式，现仅支持 long_connection")
+
+
+class FeishuSendMessageRequest(BaseModel):
+    """发送消息请求"""
+    chat_id: str = Field(..., description="群聊 ID")
+    message: str = Field(..., description="消息内容")
+
+
+class FeishuSendCardRequest(BaseModel):
+    """发送卡片请求"""
+    chat_id: str = Field(..., description="群聊 ID")
+    title: str = Field(..., description="卡片标题")
+    content: List[Dict[str, Any]] = Field(..., description="卡片内容")
+
+
+# ==================== Configuration Endpoints ====================
+
+@router.post("/config")
+async def save_feishu_config(
+    request: FeishuConfigRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    保存飞书配置
+
+    配置信息说明：
+    - app_id: 飞书应用的 App ID，从飞书开放平台获取
+    - app_secret: 飞书应用的 App Secret
+
+    获取步骤：
+    1. 访问 https://open.feishu.cn/
+    2. 创建企业自建应用
+    3. 在「凭证与基础信息」页面获取 App ID 和 App Secret
+    4. 添加权限：im:message, im:message:group_at_msg, im:chat
+    5. 订阅事件：接收消息 v1 (message.receive_v1)
+    """
+    # 检查用户权限
+    if not has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"]):
+        raise HTTPException(status_code=403, detail="只有管理员可以配置飞书集成")
+
+    try:
+        event_mode = "long_connection"
+
+        service = get_feishu_service(db)
+        await service.save_config(
+            app_id=request.app_id,
+            app_secret=request.app_secret,
+            encrypt_key="",
+            verification_token="",
+            event_mode=event_mode,
+            tenant_id=tenant_id
+        )
+        manager = get_feishu_long_connection_manager()
+        if manager:
+            await manager.restart()
+
+        next_steps = [
+            "1. 飞书应用事件订阅至少包含：接收消息 v1 (message.receive_v1)",
+            "2. 当前模式：long_connection",
+            "3. 后端已自动重启飞书长连接",
+        ]
+
+        return {
+            'status': 'success',
+            'message': '飞书配置已保存，并已切换到长连接模式',
+            'next_steps': next_steps
+        }
+    except HTTPException:
+        raise
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"保存配置失败: {str(e)}"
+        )
+
+
+@router.get("/config")
+async def get_feishu_config(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """获取飞书配置状态。管理员可获取完整配置用于回填修改表单。"""
+    try:
+        service = get_feishu_service(db)
+        await service.load_config(tenant_id)
+        manager = get_feishu_long_connection_manager()
+        can_view_secrets = has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"])
+
+        return {
+            'is_configured': service.is_configured(),
+            'app_id': service.app_id,
+            'app_secret': service.app_secret if can_view_secrets else None,
+            'event_mode': "long_connection",
+            'long_connection': manager.status() if manager else None,
+        }
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"获取配置失败: {str(e)}"
+        )
+
+
+@router.delete("/config")
+async def delete_feishu_config(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """删除飞书配置"""
+    from ..models import Integration
+    from sqlalchemy.future import select
+
+    if not has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"]):
+        raise HTTPException(status_code=403, detail="只有管理员可以删除配置")
+
+    try:
+        result = await db.execute(
+            select(Integration).where(
+                Integration.tenant_id == tenant_id,
+                Integration.type == 'feishu'
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration:
+            integration.is_active = False
+            await db.commit()
+        manager = get_feishu_long_connection_manager()
+        if manager:
+            await manager.restart()
+
+        return {'status': 'success', 'message': '飞书配置已删除'}
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"删除配置失败: {str(e)}"
+        )
+
+
+@router.get("/long-connection/status")
+async def feishu_long_connection_status(
+    current_user: User = Depends(get_current_user),
+):
+    """查询飞书长连接状态。"""
+    manager = get_feishu_long_connection_manager()
+    if not manager:
+        return {
+            "status": "error",
+            "message": "Feishu 长连接管理器未初始化",
+            "mode": "long_connection",
+            "data": None,
+        }
+    return {
+        "status": "success",
+        "mode": "long_connection",
+        "data": manager.status(),
+    }
+
+
+@router.post("/long-connection/restart")
+async def restart_feishu_long_connection(
+    current_user: User = Depends(get_current_user),
+):
+    """手动重启飞书长连接。"""
+    if not has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"]):
+        raise HTTPException(status_code=403, detail="只有管理员可以重启飞书长连接")
+
+    manager = get_feishu_long_connection_manager()
+    if not manager:
+        raise HTTPException(status_code=500, detail="Feishu 长连接管理器未初始化")
+
+    await manager.restart()
+    return {
+        "status": "success",
+        "message": "Feishu 长连接已重启",
+        "mode": "long_connection",
+        "data": manager.status(),
+    }
+
+# ==================== Message Endpoints ====================
+
+@router.post("/send-message")
+async def send_feishu_message(
+    request: FeishuSendMessageRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    发送文本消息到飞书群
+
+    用于发送简报、通知等
+    """
+    try:
+        service = get_feishu_service(db)
+        await service.load_config(tenant_id)
+
+        result = await service.send_message_to_group(
+            chat_id=request.chat_id,
+            message=request.message
+        )
+
+        return {
+            'status': 'success',
+            'result': result
+        }
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"发送消息失败: {str(e)}"
+        )
+
+
+@router.post("/send-card")
+async def send_feishu_card(
+    request: FeishuSendCardRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    发送富文本卡片到飞书群
+
+    用于发送格式化的每日简报、风险预警等
+    """
+    try:
+        service = get_feishu_service(db)
+        await service.load_config(tenant_id)
+
+        result = await service.send_rich_card(
+            chat_id=request.chat_id,
+            title=request.title,
+            content=request.content
+        )
+
+        return {
+            'status': 'success',
+            'result': result
+        }
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"发送卡片失败: {str(e)}"
+        )
+
+
+# ==================== Test Endpoints ====================
+
+@router.post("/test-connection")
+async def test_feishu_connection(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """测试飞书连接是否正常"""
+    try:
+        service = get_feishu_service(db)
+        await service.load_config(tenant_id)
+        manager = get_feishu_long_connection_manager()
+
+        if not service.is_configured():
+            return {
+                'status': 'error',
+                'message': '飞书未配置'
+            }
+
+        # 尝试获取 tenant_access_token 测试连接
+        try:
+            access_token = await service._get_tenant_access_token()
+            event_mode = "long_connection"
+            long_status = manager.status() if manager else None
+
+            if not long_status:
+                return {
+                    'status': 'error',
+                    'message': '长连接管理器未初始化',
+                    'event_mode': event_mode,
+                }
+            if not long_status.get("sdk_available"):
+                return {
+                    'status': 'error',
+                    'message': '长连接 SDK 未安装（缺少 lark-oapi）',
+                    'event_mode': event_mode,
+                    'long_connection': long_status,
+                }
+            if not long_status.get("thread_alive"):
+                message, detail = _build_long_connection_failure(long_status)
+                return {
+                    'status': 'error',
+                    'message': message,
+                    'event_mode': event_mode,
+                    'long_connection': long_status,
+                    'detail': detail,
+                }
+
+            return {
+                'status': 'success',
+                'message': '飞书连接正常',
+                'has_token': bool(access_token),
+                'mode': 'long_connection',
+                'event_mode': event_mode,
+                'long_connection': long_status,
+            }
+        except Exception as token_error:
+            return {
+                'status': 'error',
+                'message': f'获取 Token 失败: {str(token_error)}',
+                'hint': '请检查 App ID 和 App Secret 是否正确'
+            }
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"测试连接失败: {str(e)}"
+        )
diff --git a/backend/app/api/model_capability.py b/backend/app/api/model_capability.py
new file mode 100644
index 00000000..7c788788
--- /dev/null
+++ b/backend/app/api/model_capability.py
@@ -0,0 +1,491 @@
+"""模型能力矩阵 API"""
+
+from typing import Optional
+
+from fastapi import APIRouter, Depends, Query
+from fastapi.responses import JSONResponse
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.response import ErrorCode, error_response, success_response
+from ..database import get_db
+from ..dependencies import get_current_user, get_tenant_id
+from ..models import Experiment, TruthNode, User
+from ..schemas.model_capability import (
+    CostOptimizationAnalyzeRequest,
+    ExperimentSyncRequest,
+    ExperimentSyncReviewRequest,
+    ExperimentSyncRollbackRequest,
+    GenerateMigrationTasksRequest,
+    LifecycleImpactRequest,
+    ValidateMatchRequest,
+)
+from ..services.ai_service import get_ai_service
+from ..services.cost_optimization_service import CostOptimizationService
+from ..services.experiment_sync_service import ExperimentSyncService
+from ..services.model_capability_service import ModelCapabilityService
+from ..services.model_lifecycle_service import ModelLifecycleService
+
+router = APIRouter(prefix="/model-capability", tags=["model-capability"])
+
+
+def _build_service(db: AsyncSession) -> ModelCapabilityService:
+    return ModelCapabilityService(db=db, ai_service=get_ai_service())
+
+
+@router.post("/validate-match")
+async def validate_model_requirement_match(
+    request: ValidateMatchRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """需求-模型匹配校验"""
+    _ = current_user
+    service = _build_service(db)
+    try:
+        data = await service.validate_model_requirement_match(
+            truth_node_id=request.truth_node_id,
+            model_version_id=request.model_version_id,
+            tenant_id=tenant_id,
+            project_id=request.project_id,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=404,
+            content=error_response(ErrorCode.NOT_FOUND, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"匹配校验失败: {exc}"),
+        )
+
+
+@router.get("/capability-gaps")
+async def list_capability_gaps(
+    status: Optional[str] = Query(default=None),
+    severity: Optional[str] = Query(default=None),
+    limit: int = Query(default=100, ge=1, le=500),
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """列出能力差距预警"""
+    _ = current_user
+    service = _build_service(db)
+    try:
+        data = await service.list_capability_gaps(
+            tenant_id=tenant_id,
+            status=status,
+            severity=severity,
+            limit=limit,
+        )
+        return success_response(data=data)
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"查询能力差距失败: {exc}"),
+        )
+
+
+@router.get("/cost-tracking/{truth_node_id}")
+async def get_cost_tracking(
+    truth_node_id: int,
+    period: str = Query(default="weekly", pattern="^(daily|weekly|monthly|project)$"),
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """查询成本跟踪与 ROI"""
+    _ = current_user
+    service = _build_service(db)
+    try:
+        data = await service.get_cost_tracking(
+            truth_node_id=truth_node_id,
+            period=period,
+            tenant_id=tenant_id,
+        )
+        return success_response(data=data)
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"查询成本跟踪失败: {exc}"),
+        )
+
+
+@router.get("/cost-optimization/{truth_node_id}")
+async def get_cost_optimization(
+    truth_node_id: int,
+    period: str = Query(default="weekly", pattern="^(daily|weekly|monthly|project)$"),
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """查询成本优化建议（必要时自动触发一次计算）"""
+    _ = current_user
+    service = _build_service(db)
+    try:
+        tracking = await service.get_cost_tracking(
+            truth_node_id=truth_node_id,
+            period=period,
+            tenant_id=tenant_id,
+        )
+
+        optimization = tracking.get("optimization_suggestions") or {}
+        if not optimization:
+            computed = await service.track_cost_and_roi(
+                truth_node_id=truth_node_id,
+                period=period,
+                tenant_id=tenant_id,
+            )
+            optimization = computed.get("optimization_suggestions", {})
+            tracking = await service.get_cost_tracking(
+                truth_node_id=truth_node_id,
+                period=period,
+                tenant_id=tenant_id,
+            )
+
+        return success_response(
+            data={
+                "truth_node_id": truth_node_id,
+                "period": period,
+                "optimization_suggestions": optimization,
+                "roi_score": tracking.get("roi_score"),
+                "actual_cost_usd": tracking.get("actual_cost_usd", 0.0),
+            }
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"查询成本优化失败: {exc}"),
+        )
+
+
+@router.post("/lifecycle/assess-impact")
+async def assess_model_update_impact(
+    request: LifecycleImpactRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """评估模型升级对现有需求集合的影响。"""
+    _ = current_user
+    service = ModelLifecycleService(db)
+    try:
+        data = await service.assess_model_update_impact(
+            old_model_version_id=request.old_model_version_id,
+            new_model_version_id=request.new_model_version_id,
+            tenant_id=tenant_id,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=404,
+            content=error_response(ErrorCode.NOT_FOUND, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"影响评估失败: {exc}"),
+        )
+
+
+@router.get("/lifecycle/impacts")
+async def list_lifecycle_impacts(
+    limit: int = Query(default=50, ge=1, le=200),
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """列出模型生命周期影响评估记录。"""
+    _ = current_user
+    service = ModelLifecycleService(db)
+    try:
+        data = await service.list_impacts(tenant_id=tenant_id, limit=limit)
+        return success_response(data=data)
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"查询影响评估失败: {exc}"),
+        )
+
+
+@router.post("/lifecycle/{impact_id}/migration-tasks")
+async def create_migration_tasks(
+    impact_id: int,
+    request: GenerateMigrationTasksRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """基于影响评估自动生成迁移任务建议。"""
+    _ = current_user
+    service = ModelLifecycleService(db)
+    try:
+        data = await service.generate_migration_tasks(
+            impact_id=impact_id,
+            tenant_id=tenant_id,
+            max_tasks=request.max_tasks,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=404,
+            content=error_response(ErrorCode.NOT_FOUND, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"生成迁移任务失败: {exc}"),
+        )
+
+
+@router.post("/experiment-sync")
+async def create_experiment_sync(
+    request: ExperimentSyncRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """创建实验结果同步记录，满足阈值时自动同步。"""
+    _ = current_user
+    service = ExperimentSyncService(db)
+    try:
+        data = await service.auto_sync_experiment_result(
+            experiment_id=request.experiment_id,
+            truth_node_id=request.truth_node_id,
+            tenant_id=tenant_id,
+            project_id=request.project_id,
+            sync_type=request.sync_type,
+            auto_apply_threshold=request.auto_apply_threshold,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=404,
+            content=error_response(ErrorCode.NOT_FOUND, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"创建同步失败: {exc}"),
+        )
+
+
+@router.get("/experiment-sync")
+async def list_experiment_sync_records(
+    truth_node_id: Optional[int] = Query(default=None),
+    limit: int = Query(default=50, ge=1, le=200),
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """查询实验同步记录。"""
+    _ = current_user
+    service = ExperimentSyncService(db)
+    try:
+        data = await service.list_sync_records(
+            tenant_id=tenant_id,
+            truth_node_id=truth_node_id,
+            limit=limit,
+        )
+        return success_response(data=data)
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"查询同步记录失败: {exc}"),
+        )
+
+
+@router.post("/experiment-sync/{sync_id}/review")
+async def review_experiment_sync(
+    sync_id: int,
+    request: ExperimentSyncReviewRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """人工审核实验同步记录。"""
+    service = ExperimentSyncService(db)
+    try:
+        data = await service.review_sync(
+            sync_id=sync_id,
+            tenant_id=tenant_id,
+            reviewer_id=current_user.id,
+            action=request.action,
+            review_notes=request.review_notes,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=400,
+            content=error_response(ErrorCode.BAD_REQUEST, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"审核同步失败: {exc}"),
+        )
+
+
+@router.post("/experiment-sync/{sync_id}/rollback")
+async def rollback_experiment_sync(
+    sync_id: int,
+    request: ExperimentSyncRollbackRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """回滚已应用的实验同步。"""
+    service = ExperimentSyncService(db)
+    try:
+        data = await service.rollback_sync(
+            sync_id=sync_id,
+            tenant_id=tenant_id,
+            reviewer_id=current_user.id,
+            rollback_notes=request.rollback_notes,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=400,
+            content=error_response(ErrorCode.BAD_REQUEST, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"回滚同步失败: {exc}"),
+        )
+
+
+@router.post("/cost-optimization/analyze")
+async def analyze_cost_optimization(
+    request: CostOptimizationAnalyzeRequest,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """生成并持久化成本优化建议。"""
+    _ = current_user
+    service = CostOptimizationService(db)
+    try:
+        data = await service.analyze_cost_optimization(
+            truth_node_id=request.truth_node_id,
+            tenant_id=tenant_id,
+            period=request.period,
+            project_id=request.project_id,
+        )
+        return success_response(data=data)
+    except ValueError as exc:
+        return JSONResponse(
+            status_code=400,
+            content=error_response(ErrorCode.BAD_REQUEST, str(exc)),
+        )
+    except Exception as exc:
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, f"成本优化分析失败: {exc}"),
+        )
+
+
+@router.get("/context/projects/{project_id}/truth-nodes")
+async def list_project_truth_nodes(
+    project_id: int,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """获取项目上下文中的宪章列表，用于前端选择器。"""
+    _ = current_user
+    result = await db.execute(
+        select(TruthNode).where(
+            and_(
+                TruthNode.tenant_id == tenant_id,
+                TruthNode.project_id == project_id,
+            )
+        ).order_by(TruthNode.created_at.desc())
+    )
+    nodes = result.scalars().all()
+    return success_response(
+        data={
+            "items": [
+                {
+                    "id": n.id,
+                    "title": n.title,
+                    "status": n.status,
+                    "node_type": n.node_type,
+                }
+                for n in nodes
+            ]
+        }
+    )
+
+
+@router.get("/context/projects/{project_id}/experiments")
+async def list_project_experiments(
+    project_id: int,
+    current_user: User = Depends(get_current_user),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    """获取项目上下文中的实验列表（基于 linked_context_nodes 关联）。"""
+    _ = current_user
+    truth_ids_result = await db.execute(
+        select(TruthNode.id).where(
+            and_(
+                TruthNode.tenant_id == tenant_id,
+                TruthNode.project_id == project_id,
+            )
+        )
+    )
+    truth_ids = {int(row[0]) for row in truth_ids_result.all()}
+    if not truth_ids:
+        return success_response(data={"items": []})
+
+    exp_result = await db.execute(
+        select(Experiment)
+        .where(Experiment.tenant_id == tenant_id)
+        .order_by(Experiment.created_at.desc())
+    )
+    experiments = exp_result.scalars().all()
+
+    def _linked_ids(raw) -> set[int]:
+        ids = set()
+        if isinstance(raw, list):
+            for x in raw:
+                try:
+                    ids.add(int(x))
+                except (TypeError, ValueError):
+                    continue
+        return ids
+
+    items = []
+    for e in experiments:
+        if e.project_id is not None:
+            if int(e.project_id) != int(project_id):
+                continue
+            items.append(
+                {
+                    "id": e.id,
+                    "title": e.title or f"Experiment {e.id}",
+                    "status": e.status,
+                    "target_metric": e.target_metric,
+                }
+            )
+            continue
+
+        # 迁移兼容：project_id 为空时，回退到 linked_context_nodes 交叉判断。
+        if _linked_ids(e.linked_context_nodes).isdisjoint(truth_ids):
+            continue
+        items.append(
+            {
+                "id": e.id,
+                "title": e.title or f"Experiment {e.id}",
+                "status": e.status,
+                "target_metric": e.target_metric,
+            }
+        )
+    return success_response(data={"items": items})
+
diff --git a/backend/app/api/projects.py b/backend/app/api/projects.py
new file mode 100644
index 00000000..adc542e4
--- /dev/null
+++ b/backend/app/api/projects.py
@@ -0,0 +1,386 @@
+"""项目管理 API。"""
+
+from __future__ import annotations
+
+import json
+from typing import Any, Optional
+
+from fastapi import APIRouter, Depends, HTTPException, Query
+from pydantic import BaseModel
+from sqlalchemy import and_, delete, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..database import get_db
+from ..dependencies import get_current_user, get_tenant_id, require_roles
+from ..models import Project, ProjectSignalLink, Signal, Task, User
+from ..services.context_cache import context_cache_service
+from ..utils.project_members import ensure_project_default_members
+from ..utils.sequence import ensure_project_id_sequence
+from ..utils.title import compact_title
+
+router = APIRouter(prefix="/projects", tags=["projects"])
+
+DONE_STATUSES = {"done", "completed", "closed", "resolved", "archived"}
+
+
+def _as_dict(value: object) -> dict[str, Any]:
+    if isinstance(value, dict):
+        return value
+    if isinstance(value, str) and value.strip():
+        try:
+            parsed = json.loads(value)
+            return parsed if isinstance(parsed, dict) else {"text": value}
+        except (TypeError, ValueError):
+            return {"text": value}
+    return {}
+
+
+class ProjectCreateRequest(BaseModel):
+    title: str
+    description: Optional[str] = None
+    status: str = "active"
+    domain_id: Optional[int] = None
+    assigned_role: Optional[str] = None
+
+
+class ProjectUpdateRequest(BaseModel):
+    title: Optional[str] = None
+    description: Optional[str] = None
+    status: Optional[str] = None
+    domain_id: Optional[int] = None
+    assigned_role: Optional[str] = None
+
+
+class LinkSignalRequest(BaseModel):
+    link_type: Optional[str] = "related"
+    relevance_score: Optional[float] = 1.0
+
+
+def _serialize_project_row(project: Project) -> dict:
+    return {
+        "id": str(project.id),
+        "title": project.title or project.name,
+        "description": project.description,
+        "status": project.status,
+        "domain_id": str(project.domain_id) if project.domain_id is not None else None,
+        "assigned_role": project.assigned_role,
+        "signal_count": int(project.signal_count or 0),
+        "created_at": project.created_at.isoformat() if project.created_at else None,
+        "updated_at": project.updated_at.isoformat() if project.updated_at else None,
+    }
+
+
+async def _build_project_summary(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project: Project,
+) -> dict:
+    base = _serialize_project_row(project)
+
+    signal_count_result = await db.execute(
+        select(func.count(ProjectSignalLink.id)).where(
+            and_(
+                ProjectSignalLink.tenant_id == tenant_id,
+                ProjectSignalLink.project_id == int(project.id),
+            )
+        )
+    )
+    signal_count = int(signal_count_result.scalar() or 0)
+    base["signal_count"] = signal_count
+
+    task_rows = (
+        await db.execute(
+            select(Task).where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.project_id == int(project.id),
+                )
+            )
+        )
+    ).scalars().all()
+
+    total_tasks = len(task_rows)
+    completed_tasks = 0
+    in_progress_tasks = 0
+    pending_tasks = 0
+    task_items: list[dict] = []
+    for task in task_rows:
+        status_value = (task.status or "").lower()
+        if status_value in DONE_STATUSES:
+            completed_tasks += 1
+        elif status_value in {"in_progress", "review", "blocked"}:
+            in_progress_tasks += 1
+        else:
+            pending_tasks += 1
+
+        task_items.append(
+            {
+                "task_name": task.title,
+                "type": "task",
+                "status": task.status,
+                "signal_id": str(task.truth_node_id) if task.truth_node_id is not None else "",
+            }
+        )
+
+    base["progress"] = round((completed_tasks / total_tasks) * 100, 2) if total_tasks else 0.0
+    base["total_tasks"] = total_tasks
+    base["completed_tasks"] = completed_tasks
+    base["in_progress_tasks"] = in_progress_tasks
+    base["pending_tasks"] = pending_tasks
+    base["tasks"] = task_items
+    return base
+
+
+async def _get_project_or_404(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+) -> Project:
+    project = (
+        await db.execute(
+            select(Project).where(
+                and_(
+                    Project.tenant_id == tenant_id,
+                    Project.id == project_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if not project:
+        raise HTTPException(status_code=404, detail="Project not found")
+    return project
+
+
+@router.get("/")
+async def list_projects(
+    summary: Optional[int] = Query(default=0),
+    status: Optional[str] = Query(default=None),
+    assigned_role: Optional[str] = Query(default=None),
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    query = select(Project).where(Project.tenant_id == tenant_id)
+    if status:
+        query = query.where(Project.status == status)
+    if assigned_role:
+        query = query.where(Project.assigned_role == assigned_role)
+    query = query.order_by(Project.updated_at.desc(), Project.id.desc())
+    projects = (await db.execute(query)).scalars().all()
+
+    if int(summary or 0) == 1:
+        rows = [
+            await _build_project_summary(
+                db,
+                tenant_id=tenant_id,
+                project=project,
+            )
+            for project in projects
+        ]
+        return {"projects": rows}
+
+    return [_serialize_project_row(project) for project in projects]
+
+
+@router.get("/{project_id}")
+async def get_project(
+    project_id: int,
+    tenant_id: int = Depends(get_tenant_id),
+    db: AsyncSession = Depends(get_db),
+):
+    project = await _get_project_or_404(db, tenant_id=tenant_id, project_id=project_id)
+    return _serialize_project_row(project)
+
+
+@router.post("/")
+async def create_project(
+    body: ProjectCreateRequest,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+):
+    tenant_id = int(current_user.tenant_id)
+    title = compact_title(body.title, max_len=24)
+    if not title:
+        raise HTTPException(status_code=400, detail="title required")
+    await ensure_project_id_sequence(db)
+    project = Project(
+        tenant_id=tenant_id,
+        name=title,
+        title=title,
+        description=body.description,
+        status=body.status or "active",
+        domain_id=body.domain_id,
+        assigned_role=body.assigned_role,
+        owner_id=int(current_user.id),
+        signal_count=0,
+    )
+    db.add(project)
+    await db.flush()
+    await ensure_project_default_members(
+        db,
+        tenant_id=tenant_id,
+        project_id=int(project.id),
+        owner_id=int(current_user.id),
+        assigned_role=project.assigned_role,
+    )
+    await db.commit()
+    await db.refresh(project)
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    return _serialize_project_row(project)
+
+
+@router.patch("/{project_id}")
+async def update_project(
+    project_id: int,
+    body: ProjectUpdateRequest,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+):
+    tenant_id = int(current_user.tenant_id)
+    project = await _get_project_or_404(db, tenant_id=tenant_id, project_id=project_id)
+    if body.title is not None:
+        title = compact_title(body.title, max_len=24)
+        if not title:
+            raise HTTPException(status_code=400, detail="title cannot be empty")
+        project.title = title
+        project.name = title
+    if body.description is not None:
+        project.description = body.description
+    if body.status is not None:
+        project.status = body.status
+    if body.domain_id is not None:
+        project.domain_id = body.domain_id
+    if body.assigned_role is not None:
+        project.assigned_role = body.assigned_role
+    await db.commit()
+    await db.refresh(project)
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    return _serialize_project_row(project)
+
+
+@router.delete("/{project_id}")
+async def delete_project(
+    project_id: int,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+):
+    tenant_id = int(current_user.tenant_id)
+    _ = current_user
+    await _get_project_or_404(db, tenant_id=tenant_id, project_id=project_id)
+    await db.execute(
+        delete(ProjectSignalLink).where(
+            and_(
+                ProjectSignalLink.tenant_id == tenant_id,
+                ProjectSignalLink.project_id == project_id,
+            )
+        )
+    )
+    await db.execute(
+        delete(Task).where(
+            and_(
+                Task.tenant_id == tenant_id,
+                Task.project_id == project_id,
+            )
+        )
+    )
+    await db.execute(
+        delete(Project).where(
+            and_(
+                Project.tenant_id == tenant_id,
+                Project.id == project_id,
+            )
+        )
+    )
+    await db.commit()
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    return {"message": "Project deleted"}
+
+
+@router.post("/{project_id}/signals/{signal_id}")
+async def link_signal_to_project(
+    project_id: int,
+    signal_id: int,
+    body: LinkSignalRequest,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+):
+    tenant_id = int(current_user.tenant_id)
+    await _get_project_or_404(db, tenant_id=tenant_id, project_id=project_id)
+
+    signal = (
+        await db.execute(
+            select(Signal).where(
+                and_(
+                    Signal.tenant_id == tenant_id,
+                    Signal.id == signal_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if not signal:
+        raise HTTPException(status_code=404, detail="Signal not found")
+
+    existing = (
+        await db.execute(
+            select(ProjectSignalLink).where(
+                and_(
+                    ProjectSignalLink.tenant_id == tenant_id,
+                    ProjectSignalLink.project_id == project_id,
+                    ProjectSignalLink.signal_id == signal_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if existing:
+        return {
+            "id": str(existing.id),
+            "project_id": str(existing.project_id),
+            "signal_id": str(existing.signal_id),
+            "link_type": existing.link_type,
+            "relevance_score": float(existing.relevance_score or 0.0),
+            "created_at": existing.created_at.isoformat() if existing.created_at else None,
+        }
+
+    link = ProjectSignalLink(
+        tenant_id=tenant_id,
+        project_id=project_id,
+        signal_id=signal_id,
+        link_type=body.link_type or "related",
+        relevance_score=float(body.relevance_score or 1.0),
+    )
+    db.add(link)
+    await db.commit()
+    await db.refresh(link)
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    return {
+        "id": str(link.id),
+        "project_id": str(link.project_id),
+        "signal_id": str(link.signal_id),
+        "link_type": link.link_type,
+        "relevance_score": float(link.relevance_score or 0.0),
+        "created_at": link.created_at.isoformat() if link.created_at else None,
+    }
+
+
+@router.delete("/{project_id}/signals/{signal_id}")
+async def unlink_signal_from_project(
+    project_id: int,
+    signal_id: int,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+):
+    tenant_id = int(current_user.tenant_id)
+    _ = current_user
+    await db.execute(
+        delete(ProjectSignalLink).where(
+            and_(
+                ProjectSignalLink.tenant_id == tenant_id,
+                ProjectSignalLink.project_id == project_id,
+                ProjectSignalLink.signal_id == signal_id,
+            )
+        )
+    )
+    await db.commit()
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    return {"message": "Signal unlinked"}
diff --git a/backend/app/api/rag.py b/backend/app/api/rag.py
new file mode 100644
index 00000000..11fdc8c4
--- /dev/null
+++ b/backend/app/api/rag.py
@@ -0,0 +1,570 @@
+"""Adaptive RAG API。"""
+
+from __future__ import annotations
+
+import asyncio
+import io
+import re
+from datetime import datetime
+
+from fastapi import APIRouter, Depends, File, Form, HTTPException, Query, Response, UploadFile
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.response import success_response
+from ..database import get_db
+from ..dependencies.auth import get_current_user, require_roles
+from ..dependencies.tenant import get_tenant_id
+from ..models import User
+from ..schemas.rag import (
+    EntityResolutionRequest,
+    KnowledgeDocCreateRequest,
+    KnowledgeURLCreateRequest,
+    MultiStepSearchRequest,
+    QAExportQuery,
+    QAFeedbackCreateRequest,
+    QAListQuery,
+    QASessionListQuery,
+    QASessionRenameRequest,
+    RAGAskRequest,
+    SymbolSearchRequest,
+    SymbolSyncRequest,
+)
+from ..services.adaptive_rag_service import AdaptiveRAGService
+from ..services.symbol_index_service import SymbolIndexService
+
+router = APIRouter(prefix="/rag", tags=["Adaptive RAG"])
+
+
+def _resolve_tenant_id(user: User, fallback_tenant_id: int) -> int:
+    try:
+        return int(getattr(user, "tenant_id") or fallback_tenant_id or 1)
+    except Exception:
+        return int(fallback_tenant_id or 1)
+
+
+def _safe_decode_bytes(raw: bytes) -> str:
+    for encoding in ("utf-8", "utf-8-sig", "gb18030", "latin-1"):
+        try:
+            return raw.decode(encoding)
+        except Exception:
+            continue
+    return raw.decode("utf-8", errors="ignore")
+
+
+def _extract_text_from_upload(filename: str, raw: bytes) -> tuple[str, str]:
+    name = str(filename or "uploaded.txt")
+    suffix = name.rsplit(".", 1)[-1].lower() if "." in name else "txt"
+
+    if suffix in {"txt", "md", "markdown", "csv", "json", "py", "ts", "tsx", "js", "jsx"}:
+        return ("md" if suffix in {"md", "markdown"} else "txt", _safe_decode_bytes(raw))
+
+    if suffix == "pdf":
+        try:
+            from pypdf import PdfReader  # type: ignore
+        except Exception as exc:
+            raise HTTPException(status_code=400, detail=f"PDF 解析依赖缺失: {exc}") from exc
+        reader = PdfReader(io.BytesIO(raw))
+        parts: list[str] = []
+        for idx, page in enumerate(reader.pages, start=1):
+            page_text = (page.extract_text() or "").strip()
+            if not page_text:
+                continue
+            # 写入页码标记，向量分块时会保留 page_number 元数据，便于引用来源页码。
+            parts.append(f"[[PAGE:{idx}]]\n{page_text}")
+        text = "\n\n".join(parts)
+        return ("pdf", text)
+
+    if suffix in {"docx", "doc"}:
+        try:
+            from docx import Document  # type: ignore
+        except Exception as exc:
+            raise HTTPException(status_code=400, detail=f"Word 解析依赖缺失: {exc}") from exc
+        doc = Document(io.BytesIO(raw))
+        text = "\n".join([p.text for p in doc.paragraphs if p.text])
+        return ("word", text)
+
+    raise HTTPException(status_code=400, detail=f"不支持的文件类型: .{suffix}")
+
+
+@router.post("/ask")
+async def rag_ask(
+    request: RAGAskRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    result = await service.answer_question(
+        question=request.question,
+        asked_by=int(current_user.id),
+        asked_via=request.asked_via,
+        project_id=request.project_id,
+        source_chat_id=request.source_chat_id,
+        session_id=request.session_id,
+        force_new_session=bool(request.new_session),
+        thinking_mode=request.thinking_mode,
+        action_mode=request.action_mode,
+        sync_requirement=bool(request.sync_requirement),
+    )
+    return success_response(result)
+
+
+@router.get("/manifest")
+async def rag_manifest(
+    project_id: int | None = Query(default=None),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.get_project_manifest(project_id=project_id)
+    return success_response(data)
+
+
+@router.get("/features/catalog")
+async def rag_feature_catalog(
+    query: str | None = Query(default=None),
+    limit: int = Query(default=20, ge=1, le=200),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.get_feature_catalog(query=query, limit=limit)
+    return success_response(data)
+
+
+@router.post("/entity-resolution")
+async def rag_entity_resolution(
+    request: EntityResolutionRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.entity_resolution(
+        question=request.question,
+        project_id=request.project_id,
+        limit=request.limit,
+    )
+    return success_response(data)
+
+
+@router.post("/multi-step-search")
+async def rag_multi_step_search(
+    request: MultiStepSearchRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.multi_step_search(
+        question=request.question,
+        project_id=request.project_id,
+        source_chat_id=request.source_chat_id,
+        session_state=None,
+    )
+    return success_response(data)
+
+
+@router.post("/symbols/sync")
+async def rag_symbols_sync(
+    request: SymbolSyncRequest,
+    current_user: User = Depends(require_roles("ADMIN", "BOSS")),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = SymbolIndexService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.sync_codebase_symbols(
+        base_dir=request.base_dir,
+        max_files=request.max_files,
+        force_reindex=request.force_reindex,
+    )
+    return success_response(data)
+
+
+@router.post("/symbols/search")
+async def rag_symbols_search(
+    request: SymbolSearchRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = SymbolIndexService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.search_symbols(
+        query=request.query,
+        language=request.language,
+        symbol_type=request.symbol_type,
+        limit=request.limit,
+    )
+    return success_response(data)
+
+
+@router.get("/symbols/overview")
+async def rag_symbols_overview(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = SymbolIndexService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    data = await service.get_overview(top_n_files=10, auto_sync_if_empty=False)
+    return success_response(data)
+
+
+@router.get("/symbols/{symbol_id}")
+async def rag_symbol_detail(
+    symbol_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = SymbolIndexService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    try:
+        data = await service.get_symbol_detail(symbol_id=symbol_id)
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    return success_response(data)
+
+
+@router.get("/knowledge-base/docs")
+async def list_knowledge_docs(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    return success_response(await service.list_knowledge_docs())
+
+
+@router.post("/knowledge-base/docs")
+async def create_knowledge_doc(
+    request: KnowledgeDocCreateRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    result = await service.create_knowledge_doc(
+        title=request.title,
+        file_type=request.file_type,
+        raw_content=request.content,
+        uploaded_by=int(current_user.id),
+        source_type=request.source_type,
+        file_url=request.file_url,
+        project_id=request.project_id,
+    )
+    return success_response(result)
+
+
+@router.post("/knowledge-base/upload")
+async def upload_knowledge_doc(
+    file: UploadFile = File(...),
+    project_id: int | None = Form(default=None),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    raw = await file.read()
+    if not raw:
+        raise HTTPException(status_code=400, detail="上传文件为空")
+    if len(raw) > 15 * 1024 * 1024:
+        raise HTTPException(status_code=400, detail="文件过大，限制 15MB")
+
+    file_type, text = _extract_text_from_upload(file.filename or "uploaded.txt", raw)
+    if not str(text or "").strip():
+        raise HTTPException(status_code=400, detail="无法提取有效文本内容")
+
+    effective_tenant_id = _resolve_tenant_id(current_user, tenant_id)
+    service = AdaptiveRAGService(db, tenant_id=effective_tenant_id)
+    doc = await service.create_knowledge_doc_placeholder(
+        title=(file.filename or "uploaded.txt")[:255],
+        file_type=file_type,
+        uploaded_by=int(current_user.id),
+        source_type="file",
+        file_url=None,
+    )
+    asyncio.create_task(
+        AdaptiveRAGService.process_knowledge_doc_content_async(
+            tenant_id=effective_tenant_id,
+            doc_id=int(doc.id),
+            raw_content=text,
+            project_id=project_id,
+        )
+    )
+    return success_response(service._serialize_doc(doc))
+
+
+@router.post("/knowledge-base/urls")
+async def create_knowledge_url(
+    request: KnowledgeURLCreateRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    url = str(request.url or "").strip()
+    if not re.match(r"^https?://", url, flags=re.IGNORECASE):
+        raise HTTPException(status_code=400, detail="URL 必须以 http:// 或 https:// 开头")
+
+    effective_tenant_id = _resolve_tenant_id(current_user, tenant_id)
+    service = AdaptiveRAGService(db, tenant_id=effective_tenant_id)
+    doc = await service.create_knowledge_doc_placeholder(
+        title=(request.title or url)[:255],
+        file_type="url",
+        uploaded_by=int(current_user.id),
+        source_type="url",
+        file_url=url,
+    )
+    asyncio.create_task(
+        AdaptiveRAGService.refresh_url_doc_async(
+            tenant_id=effective_tenant_id,
+            doc_id=int(doc.id),
+            project_id=request.project_id,
+        )
+    )
+    return success_response(service._serialize_doc(doc))
+
+
+@router.post("/knowledge-base/docs/{doc_id}/refresh")
+async def refresh_knowledge_doc(
+    doc_id: int,
+    project_id: int | None = Query(default=None),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    effective_tenant_id = _resolve_tenant_id(current_user, tenant_id)
+    service = AdaptiveRAGService(db, tenant_id=effective_tenant_id)
+    try:
+        doc = await service.get_knowledge_doc(doc_id=doc_id)
+        if str(doc.get("source_type") or "").lower() != "url":
+            raise HTTPException(status_code=400, detail="仅 URL 类型条目支持 refresh")
+        marked = await service.mark_doc_processing(doc_id=doc_id)
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+    asyncio.create_task(
+        AdaptiveRAGService.refresh_url_doc_async(
+            tenant_id=effective_tenant_id,
+            doc_id=doc_id,
+            project_id=project_id,
+        )
+    )
+    return success_response(marked)
+
+
+@router.delete("/knowledge-base/docs/{doc_id}")
+async def delete_knowledge_doc(
+    doc_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    ok = await service.delete_knowledge_doc(doc_id=doc_id)
+    if not ok:
+        raise HTTPException(status_code=404, detail="文档不存在")
+    return success_response({"deleted": True, "doc_id": doc_id})
+
+
+@router.post("/qa-records/{qa_id}/feedback")
+async def submit_qa_feedback(
+    qa_id: int,
+    request: QAFeedbackCreateRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    try:
+        result = await service.submit_feedback(
+            qa_id=qa_id,
+            rating=request.rating,
+            helpfulness=request.helpfulness,
+            error_types=request.error_types,
+            correct_answer=request.correct_answer,
+            submitted_by=int(current_user.id),
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    return success_response(result)
+
+
+@router.get("/qa-records")
+async def list_qa_records(
+    project_id: int | None = Query(default=None),
+    date_from: datetime | None = Query(default=None),
+    date_to: datetime | None = Query(default=None),
+    rating: str | None = Query(default=None),
+    page: int = Query(default=1, ge=1),
+    page_size: int = Query(default=20, ge=1, le=100),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    query = QAListQuery(
+        project_id=project_id,
+        date_from=date_from,
+        date_to=date_to,
+        rating=rating,
+        page=page,
+        page_size=page_size,
+    )
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    result = await service.list_qa_records(
+        project_id=query.project_id,
+        date_from=query.date_from,
+        date_to=query.date_to,
+        rating=query.rating,
+        page=query.page,
+        page_size=query.page_size,
+    )
+    return success_response(result)
+
+
+@router.get("/qa-sessions")
+async def list_qa_sessions(
+    date_from: datetime | None = Query(default=None),
+    date_to: datetime | None = Query(default=None),
+    page: int = Query(default=1, ge=1),
+    page_size: int = Query(default=20, ge=1, le=100),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    query = QASessionListQuery(
+        date_from=date_from,
+        date_to=date_to,
+        page=page,
+        page_size=page_size,
+    )
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    result = await service.list_qa_sessions(
+        user_id=int(current_user.id),
+        date_from=query.date_from,
+        date_to=query.date_to,
+        page=query.page,
+        page_size=query.page_size,
+    )
+    return success_response(result)
+
+
+@router.get("/qa-sessions/{session_id}")
+async def qa_session_detail(
+    session_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    try:
+        result = await service.get_qa_session_detail(session_id=session_id, user_id=int(current_user.id))
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    return success_response(result)
+
+
+@router.get("/qa-sessions/{session_id}/state")
+async def qa_session_state_replay(
+    session_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    try:
+        result = await service.get_session_state_replay(
+            session_id=session_id,
+            user_id=int(current_user.id),
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    return success_response(result)
+
+
+@router.patch("/qa-sessions/{session_id}/title")
+async def rename_qa_session(
+    session_id: int,
+    request: QASessionRenameRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    try:
+        result = await service.rename_qa_session(
+            session_id=session_id,
+            user_id=int(current_user.id),
+            title=request.title,
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    return success_response(result)
+
+
+@router.delete("/qa-sessions/{session_id}")
+async def delete_qa_session(
+    session_id: int,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    deleted = await service.delete_qa_session(session_id=session_id, user_id=int(current_user.id))
+    if not deleted:
+        raise HTTPException(status_code=404, detail="qa session 不存在")
+    return success_response({"deleted": True, "session_id": session_id})
+
+
+@router.get("/qa-records/stats")
+async def qa_records_stats(
+    project_id: int | None = Query(default=None),
+    date_from: datetime | None = Query(default=None),
+    date_to: datetime | None = Query(default=None),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    result = await service.get_qa_stats(
+        project_id=project_id,
+        date_from=date_from,
+        date_to=date_to,
+    )
+    return success_response(result)
+
+
+@router.get("/qa-records/export")
+async def export_qa_records_jsonl(
+    include_unrated: bool = Query(default=False),
+    project_id: int | None = Query(default=None),
+    rating: str | None = Query(default=None),
+    domain: str | None = Query(default=None),
+    date_from: datetime | None = Query(default=None),
+    date_to: datetime | None = Query(default=None),
+    current_user: User = Depends(require_roles("ADMIN", "BOSS")),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    _ = current_user
+    query = QAExportQuery(
+        include_unrated=include_unrated,
+        project_id=project_id,
+        rating=rating,
+        domain=domain,
+        date_from=date_from,
+        date_to=date_to,
+    )
+
+    service = AdaptiveRAGService(db, tenant_id=_resolve_tenant_id(current_user, tenant_id))
+    payload = await service.export_qa_jsonl(
+        include_unrated=query.include_unrated,
+        project_id=query.project_id,
+        rating=query.rating,
+        domain=query.domain,
+        date_from=query.date_from,
+        date_to=query.date_to,
+    )
+    return Response(content=payload, media_type="application/x-ndjson")
diff --git a/backend/app/api/signal_sources.py b/backend/app/api/signal_sources.py
new file mode 100644
index 00000000..0ab97679
--- /dev/null
+++ b/backend/app/api/signal_sources.py
@@ -0,0 +1,401 @@
+"""食谱系统信号源 API
+
+功能：
+1. 接收食谱系统的 webhook 请求
+2. 将食谱识别需求转换为 PMFlow 信号
+3. 自动关联到指定项目（食谱系统项目）
+"""
+from fastapi import APIRouter, Depends, HTTPException, Query
+from sqlalchemy.ext.asyncio import AsyncSession
+from pydantic import BaseModel, Field
+from typing import Optional, Dict, Any
+import json
+
+from ..database import get_db
+from ..dependencies.tenant import get_tenant_id
+from ..core.logging import get_logger
+from ..models import Signal, ProjectSignalLink, Project
+
+router = APIRouter(prefix="/signal-sources", tags=["Signal Sources"])
+logger = get_logger(__name__)
+
+RECIPE_PROJECT_KEYWORDS = ("食谱", "recipe", "sstc")
+
+
+async def _resolve_target_project_id(
+    *,
+    db: AsyncSession,
+    tenant_id: int,
+    project_id: Optional[int],
+) -> int:
+    """解析目标项目：优先显式 project_id，其次按关键字自动匹配。"""
+    from sqlalchemy import select, or_
+
+    if project_id is not None:
+        result = await db.execute(
+            select(Project.id).where(
+                Project.tenant_id == tenant_id,
+                Project.id == int(project_id),
+            ).limit(1)
+        )
+        found = result.scalar_one_or_none()
+        if found is None:
+            raise HTTPException(status_code=404, detail=f"project_id={project_id} 不存在或无权限")
+        return int(found)
+
+    keyword_conditions = []
+    for keyword in RECIPE_PROJECT_KEYWORDS:
+        keyword_conditions.extend([
+            Project.title.ilike(f"%{keyword}%"),
+            Project.name.ilike(f"%{keyword}%"),
+        ])
+
+    match_result = await db.execute(
+        select(Project.id).where(
+            Project.tenant_id == tenant_id,
+            or_(*keyword_conditions),
+        ).limit(1)
+    )
+    matched = match_result.scalar_one_or_none()
+    if matched is not None:
+        return int(matched)
+
+    # 兜底：若租户仅存在一个活跃项目，自动绑定，避免“已调用但未入库”。
+    active_result = await db.execute(
+        select(Project.id).where(
+            Project.tenant_id == tenant_id,
+            Project.status == "active",
+        ).limit(2)
+    )
+    active_ids = list(active_result.scalars().all())
+    if len(active_ids) == 1:
+        return int(active_ids[0])
+
+    raise HTTPException(
+        status_code=422,
+        detail="未找到目标项目，请在请求中显式传入 project_id（或确保项目名包含 食谱/recipe/sstc）",
+    )
+
+
+# ============================================================================
+# 食谱系统信号源 Webhook
+# ============================================================================
+
+class RecipeRecognitionRequest(BaseModel):
+    """食谱识别请求"""
+    # 食谱信息
+    food_name: str = Field(..., description="食物名称")
+    food_description: Optional[str] = Field(None, description="食物描述")
+    image_url: Optional[str] = Field(None, description="图片URL")
+    category: Optional[str] = Field(None, description="分类")
+    cuisine: Optional[str] = Field(None, description="菜系")
+    
+    # 营养信息（如果来源系统已分析）
+    calories: Optional[float] = Field(None, description="热量（千卡）")
+    protein: Optional[float] = Field(None, description="蛋白质（克）")
+    fat: Optional[float] = Field(None, description="脂肪（克）")
+    carbs: Optional[float] = Field(None, description="碳水化合物（克）")
+    fiber: Optional[float] = Field(None, description="膳食纤维（克）")
+    
+    # 意图
+    intent: str = Field(
+        default="analyze",
+        description="意图：analyze-热量分析, save-保存食谱, recommend-推荐菜谱"
+    )
+    
+    # 用户信息（可选）
+    user_name: Optional[str] = Field(None, description="用户名")
+
+
+class RecipeRecognitionResponse(BaseModel):
+    """响应"""
+    success: bool
+    signal_id: Optional[int] = None
+    project_id: Optional[int] = None
+    message: str
+    details: Optional[Dict[str, Any]] = None
+
+
+@router.post("/recipe/webhook", response_model=RecipeRecognitionResponse, summary="食谱系统 Webhook")
+async def recipe_webhook(
+    request: RecipeRecognitionRequest,
+    project_id: Optional[int] = Query(None, description="项目ID，不传则自动查找食谱系统项目"),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    """
+    接收食谱系统的识别请求
+    
+    用法：
+    - 食谱系统完成拍照识别后，调用此 API 将需求推送到 PMFlow
+    - PMFlow 会自动创建信号并关联到食谱系统项目
+    
+    请求示例：
+    ```json
+    {
+        "food_name": "红烧肉",
+        "food_description": "一块红烧肉，约200克，肥瘦相间",
+        "image_url": "https://example.com/image.jpg",
+        "category": "肉类",
+        "cuisine": "中式",
+        "calories": 350,
+        "protein": 20,
+        "fat": 25,
+        "carbs": 10,
+        "fiber": 0,
+        "intent": "analyze",
+        "user_name": "张三"
+    }
+    ```
+    """
+    try:
+        # 1. 查找目标项目
+        target_project_id = await _resolve_target_project_id(
+            db=db,
+            tenant_id=tenant_id,
+            project_id=project_id,
+        )
+        
+        # 2. 构建信号内容
+        content_parts = []
+        content_parts.append(f"食物：{request.food_name}")
+        
+        if request.food_description:
+            content_parts.append(f"描述：{request.food_description}")
+        
+        # 营养信息
+        if request.calories or request.protein or request.fat or request.carbs or request.fiber:
+            nutrition = []
+            if request.calories:
+                nutrition.append(f"热量 {request.calories} kcal")
+            if request.protein:
+                nutrition.append(f"蛋白 {request.protein}g")
+            if request.fat:
+                nutrition.append(f"脂肪 {request.fat}g")
+            if request.carbs:
+                nutrition.append(f"碳水 {request.carbs}g")
+            if request.fiber:
+                nutrition.append(f"纤维 {request.fiber}g")
+            content_parts.append("营养：" + ", ".join(nutrition))
+        
+        if request.category:
+            content_parts.append(f"分类：{request.category}")
+        if request.cuisine:
+            content_parts.append(f"菜系：{request.cuisine}")
+        content_parts.append(f"意图：{request.intent}")
+        
+        content = "\n".join(content_parts)
+        
+        # 3. 构建标题
+        intent_text = {
+            "analyze": "热量分析",
+            "save": "保存食谱",
+            "recommend": "推荐菜谱"
+        }.get(request.intent, request.intent)
+        
+        title = f"🍽️ 食谱{intent_text}：{request.food_name}"
+        
+        # 4. 构建摘要
+        summary = f"食物：{request.food_name}"
+        if request.calories:
+            summary += f"\n热量：{request.calories} kcal"
+        if request.category:
+            summary += f"\n分类：{request.category}"
+        summary += f"\n意图：{intent_text}"
+        
+        # 5. 构建信号元数据
+        signal_metadata = {
+            "source": "recipe_system",
+            "intent": request.intent,
+            "category": request.category,
+            "cuisine": request.cuisine,
+            "nutrition": {
+                "calories": request.calories,
+                "protein": request.protein,
+                "fat": request.fat,
+                "carbs": request.carbs,
+                "fiber": request.fiber,
+            },
+            "image_url": request.image_url,
+            "user_name": request.user_name,
+        }
+        
+        # 6. 创建信号
+        signal = Signal(
+            tenant_id=tenant_id,
+            source="recipe_system",
+            title=title,
+            content=content,
+            summary=summary,
+            signal_type="recipe_recognition",
+            priority="medium",
+            impact_score=0.5,
+            status="待处理",
+            processed=False,
+            strategic_analysis=json.dumps(signal_metadata, ensure_ascii=False),
+        )
+        
+        db.add(signal)
+        await db.flush()  # 获取 signal.id
+        
+        # 7. 关联到项目
+        link = ProjectSignalLink(
+            tenant_id=tenant_id,
+            project_id=target_project_id,
+            signal_id=signal.id,
+        )
+        db.add(link)
+        
+        await db.commit()
+        await db.refresh(signal)
+        
+        return RecipeRecognitionResponse(
+            success=True,
+            signal_id=signal.id,
+            project_id=target_project_id,
+            message="信号创建成功并已关联到食谱系统项目",
+            details={
+                "signal_id": signal.id,
+                "project_id": target_project_id,
+                "title": title,
+                "status": signal.status,
+            }
+        )
+        
+    except HTTPException:
+        await db.rollback()
+        raise
+    except Exception as e:
+        await db.rollback()
+        logger.exception(
+            "recipe_webhook_create_signal_failed",
+            tenant_id=tenant_id,
+            project_id=project_id,
+            food_name=request.food_name,
+            error=str(e),
+        )
+        raise HTTPException(status_code=500, detail=f"创建信号失败: {str(e)}")
+
+
+@router.get("/recipe/status/{signal_id}", summary="查询食谱信号状态")
+async def get_recipe_signal_status(
+    signal_id: int,
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    """查询食谱识别信号的处理状态"""
+    from sqlalchemy import select
+    
+    result = await db.execute(
+        select(Signal).where(
+            Signal.id == signal_id,
+            Signal.tenant_id == tenant_id,
+            Signal.source == "recipe_system",
+        )
+    )
+    signal = result.scalar_one_or_none()
+    
+    if not signal:
+        raise HTTPException(status_code=404, detail="信号不存在")
+    
+    return {
+        "signal_id": signal.id,
+        "title": signal.title,
+        "status": signal.status,
+        "processed": signal.processed,
+        "summary": signal.summary,
+    }
+
+
+# ============================================================================
+# 批量导入食谱记录
+# ============================================================================
+
+class RecipeBatchRequest(BaseModel):
+    """批量导入请求"""
+    recipes: list[RecipeRecognitionRequest] = Field(..., description="食谱列表")
+
+
+@router.post("/recipe/batch", response_model=RecipeRecognitionResponse, summary="批量导入食谱")
+async def recipe_batch_import(
+    request: RecipeBatchRequest,
+    project_id: Optional[int] = Query(None, description="项目ID"),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    """批量导入食谱记录，每个都会创建信号"""
+    try:
+        target_project_id = await _resolve_target_project_id(
+            db=db,
+            tenant_id=tenant_id,
+            project_id=project_id,
+        )
+        
+        signal_ids = []
+        for recipe in request.recipes:
+            # 构建内容
+            content = f"食物：{recipe.food_name}"
+            if recipe.calories:
+                content += f"\n热量：{recipe.calories} kcal"
+            
+            title = f"🍽️ 食谱记录：{recipe.food_name}"
+            
+            metadata = {
+                "source": "recipe_system",
+                "intent": recipe.intent,
+                "nutrition": {
+                    "calories": recipe.calories,
+                    "protein": recipe.protein,
+                    "fat": recipe.fat,
+                    "carbs": recipe.carbs,
+                }
+            }
+            
+            signal = Signal(
+                tenant_id=tenant_id,
+                source="recipe_system",
+                title=title,
+                content=content,
+                summary=f"食物：{recipe.food_name}",
+                signal_type="recipe_recognition",
+                priority="low",
+                impact_score=0.3,
+                status="待处理",
+                processed=True,  # 批量导入的视为已处理
+                strategic_analysis=json.dumps(metadata),
+            )
+            db.add(signal)
+            await db.flush()
+            
+            # 关联项目
+            link = ProjectSignalLink(
+                tenant_id=tenant_id,
+                project_id=target_project_id,
+                signal_id=signal.id,
+            )
+            db.add(link)
+            signal_ids.append(signal.id)
+        
+        await db.commit()
+        
+        return RecipeRecognitionResponse(
+            success=True,
+            signal_id=len(signal_ids),  # 返回数量
+            project_id=target_project_id,
+            message=f"成功导入 {len(signal_ids)} 条食谱记录",
+            details={"signal_count": len(signal_ids)}
+        )
+        
+    except HTTPException:
+        await db.rollback()
+        raise
+    except Exception as e:
+        await db.rollback()
+        logger.exception(
+            "recipe_batch_import_failed",
+            tenant_id=tenant_id,
+            project_id=project_id,
+            recipe_count=len(request.recipes),
+            error=str(e),
+        )
+        raise HTTPException(status_code=500, detail=f"批量导入失败: {str(e)}")
diff --git a/backend/app/api/signals.py b/backend/app/api/signals.py
new file mode 100644
index 00000000..28b9ae64
--- /dev/null
+++ b/backend/app/api/signals.py
@@ -0,0 +1,462 @@
+"""信号相关 API 路由"""
+
+import json
+from typing import List, Optional
+
+from fastapi import APIRouter, Depends, HTTPException, BackgroundTasks, Query
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+
+from ..database import get_db
+from ..dependencies.tenant import get_tenant_id
+from ..models import Signal, SignalFeedback, Project, ProjectSignalLink
+from ..schemas import (
+    SignalCreate,
+    SignalPriority,
+    classify_signal_priority,
+    SignalFeedbackCreate,
+    SignalFeedbackResponse,
+)
+from ..services.fetcher import fetch_hn_stories
+from ..services.analyzer import analyze_impact
+from ..services.llm import llm_service
+from ..services.plane import plane_service
+from ..services.lark import LarkService
+
+router = APIRouter(prefix="/signals", tags=["signals"])
+lark_service = LarkService()
+
+
+@router.get("/")
+async def list_signals(
+    roles: Optional[str] = Query(None, description="逗号分隔的角色列表，如 'frontend,backend'"),
+    db: AsyncSession = Depends(get_db),
+):
+    """获取信号列表，支持按角色过滤"""
+    result = await db.execute(select(Signal).order_by(Signal.impact_score.desc()))
+    signals = result.scalars().all()
+
+    if roles:
+        role_list = [r.strip() for r in roles.split(",")]
+        filtered_signals = []
+        for signal in signals:
+            tasks = []
+            if signal.tasks_generated:
+                try:
+                    tasks = json.loads(signal.tasks_generated)
+                except Exception:
+                    tasks = []
+
+            filtered_tasks = [t for t in tasks if t.get("type") in role_list]
+
+            if filtered_tasks:
+                signal_dict = {
+                    "id": str(signal.id),
+                    "source": signal.source,
+                    "content": signal.content,
+                    "impact_score": signal.impact_score,
+                    "priority": signal.priority,
+                    "processed": signal.processed,
+                    "strategic_analysis": signal.strategic_analysis,
+                    "tasks_generated": signal.tasks_generated,
+                    "tasks_filtered": filtered_tasks,
+                    "created_at": signal.created_at.isoformat() if signal.created_at else None,
+                }
+                filtered_signals.append(signal_dict)
+
+        return filtered_signals
+
+    return signals
+
+
+@router.post("/", response_model=dict)
+async def create_signal(signal: SignalCreate, db: AsyncSession = Depends(get_db)):
+    new_signal = Signal(
+        source=signal.source,
+        content=signal.content,
+        impact_score=signal.impact_score,
+    )
+    db.add(new_signal)
+    await db.commit()
+    await db.refresh(new_signal)
+    return {"id": str(new_signal.id), "status": "received"}
+
+
+@router.post("/ingest", response_model=dict)
+async def ingest_signal(
+    signal: SignalCreate,
+    background_tasks: BackgroundTasks,
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    """接入外部信号，按优先级处理"""
+    initial_priority = signal.priority or classify_signal_priority(signal.impact_score)
+
+    new_signal = Signal(
+        source=signal.source,
+        content=signal.content,
+        impact_score=signal.impact_score,
+        priority=initial_priority.value,
+    )
+    db.add(new_signal)
+    await db.commit()
+    await db.refresh(new_signal)
+
+    if initial_priority == SignalPriority.LOW:
+        return {
+            "id": str(new_signal.id),
+            "priority": initial_priority.value,
+            "status": "recorded",
+            "message": "Low priority signal recorded without processing",
+        }
+
+    async def process_signal_internal():
+        try:
+            relevant_context = await llm_service.find_relevant_context(
+                signal_content=signal.content,
+                db_session=db,
+                top_k=3,
+                min_similarity=0.3,
+                tenant_id=tenant_id,
+            )
+
+            strategy_result = await llm_service.evaluate_with_voting_async(
+                signal_content=signal.content,
+                relevant_context=relevant_context,
+                impact_score=signal.impact_score,
+                db=db,
+                tenant_id=tenant_id,
+            )
+        except RuntimeError:
+            if initial_priority == SignalPriority.CRITICAL:
+                raise HTTPException(status_code=403, detail="预算超限，AI 调用被阻止（hard limit）")
+            return {
+                "id": str(new_signal.id),
+                "priority": initial_priority.value,
+                "status": "blocked",
+                "processing_mode": "background",
+                "message": "预算超限，AI 调用被阻止（hard limit）",
+            }
+
+        is_strategic = strategy_result.get("is_strategic", False)
+        decision_proposal = strategy_result.get("decision_proposal", {})
+        tasks = strategy_result.get("tasks", [])
+        confidence = strategy_result.get("confidence", 0.0)
+
+        final_priority = classify_signal_priority(signal.impact_score, confidence)
+        new_signal.priority = final_priority.value
+
+        sync_result = None
+        notification_sent = False
+
+        if final_priority in [SignalPriority.CRITICAL, SignalPriority.HIGH]:
+            if is_strategic and tasks:
+                sync_result = await plane_service.sync_tasks(
+                    tasks, is_urgent=(final_priority == SignalPriority.CRITICAL)
+                )
+                default_chat_id = "oc_1234567890"
+                await lark_service.send_strategic_alert(
+                    default_chat_id, signal.content, decision_proposal, tasks
+                )
+                notification_sent = True
+                new_signal.impact_score = 1.0
+
+        analysis_data = {
+            "decision_proposal": decision_proposal,
+            "competitiveness_radar": strategy_result.get("competitiveness_radar"),
+            "detected_tech_stacks": strategy_result.get("detected_tech_stacks", []),
+            "related_knowledge_nodes": strategy_result.get("related_knowledge_nodes", []),
+            "rag_context_used": len(relevant_context) if relevant_context else 0,
+            "rag_context_sources": [ctx["id"] for ctx in (relevant_context or [])],
+        }
+
+        if "_voting" in strategy_result:
+            voting_info = strategy_result["_voting"]
+            analysis_data["voting"] = {
+                "enabled": voting_info.get("enabled"),
+                "models_used": voting_info.get("models_used"),
+                "strategic_votes": voting_info.get("strategic_votes"),
+                "total_votes": voting_info.get("total_votes"),
+                "unanimous": voting_info.get("unanimous"),
+            }
+
+        new_signal.strategic_analysis = json.dumps(analysis_data)
+        new_signal.tasks_generated = json.dumps(tasks)
+        new_signal.processed = True
+
+        await db.commit()
+        await db.refresh(new_signal)
+
+        return {
+            "id": str(new_signal.id),
+            "priority": final_priority.value,
+            "is_strategic": is_strategic,
+            "confidence": confidence,
+            "decision_proposal": decision_proposal,
+            "rag_context_used": len(relevant_context) if relevant_context else 0,
+            "voting": strategy_result.get("_voting"),
+            "tasks_generated": len(tasks),
+            "sync_result": sync_result,
+            "notification_sent": notification_sent,
+        }
+
+    if initial_priority == SignalPriority.CRITICAL:
+        result = await process_signal_internal()
+        result["processing_mode"] = "synchronous"
+        return result
+    else:
+        background_tasks.add_task(process_signal_internal)
+        return {
+            "id": str(new_signal.id),
+            "priority": initial_priority.value,
+            "status": "queued",
+            "processing_mode": "background",
+            "message": f"{initial_priority.value.upper()} priority signal queued for processing",
+        }
+
+
+@router.post("/fetch", response_model=dict)
+async def trigger_signal_fetch(
+    background_tasks: BackgroundTasks,
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id),
+):
+    """手动触发抓取并分析信号"""
+
+    def _stop_ollama():
+        import subprocess
+        import signal
+        import os
+
+        killed = False
+        try:
+            result = subprocess.run(
+                ["pgrep", "-f", "ollama"],
+                capture_output=True,
+                text=True,
+            )
+            if result.returncode == 0:
+                for pid in result.stdout.strip().split("\n"):
+                    if pid:
+                        try:
+                            os.kill(int(pid), signal.SIGTERM)
+                            killed = True
+                        except ProcessLookupError:
+                            pass
+        except Exception as e:
+            print(f"[AI] pgrep failed: {e}")
+
+        if not killed:
+            try:
+                result = subprocess.run(
+                    ["pkill", "-9", "-f", "ollama"],
+                    capture_output=True,
+                    timeout=5,
+                )
+                killed = result.returncode == 0
+            except Exception as e:
+                print(f"[AI] pkill failed: {e}")
+
+        if killed:
+            print("[AI] Ollama stopped - analysis complete")
+
+    raw_stories = await fetch_hn_stories(limit=30)
+    new_signals_count = 0
+    processed_count = 0
+
+    for story in raw_stories:
+        result = await db.execute(select(Signal).where(Signal.content == story["content"]))
+        if result.scalars().first():
+            continue
+
+        impact = analyze_impact(story["content"])
+
+        if impact > 0.4:
+            initial_priority = classify_signal_priority(impact)
+
+            new_signal = Signal(
+                source=story["source"],
+                content=story["content"],
+                impact_score=impact,
+                priority=initial_priority.value,
+            )
+            db.add(new_signal)
+            await db.commit()
+            await db.refresh(new_signal)
+            new_signals_count += 1
+
+            if initial_priority == SignalPriority.LOW:
+                continue
+
+            async def process_fetched():
+                nonlocal processed_count
+                try:
+                    relevant_context = await llm_service.find_relevant_context(
+                        signal_content=story["content"],
+                        db_session=db,
+                        top_k=3,
+                        min_similarity=0.3,
+                        tenant_id=tenant_id,
+                    )
+                    strategy_result = await llm_service.evaluate_with_voting_async(
+                        signal_content=story["content"],
+                        relevant_context=relevant_context,
+                        impact_score=impact,
+                        db=db,
+                        tenant_id=tenant_id,
+                    )
+                except RuntimeError:
+                    return
+                is_strategic = strategy_result.get("is_strategic", False)
+                decision_proposal = strategy_result.get("decision_proposal", {})
+                confidence = strategy_result.get("confidence", 0.0)
+                final_priority = classify_signal_priority(impact, confidence)
+                new_signal.priority = final_priority.value
+
+                analysis_data = {
+                    "decision_proposal": decision_proposal,
+                    "competitiveness_radar": strategy_result.get("competitiveness_radar"),
+                    "detected_tech_stacks": strategy_result.get("detected_tech_stacks", []),
+                    "related_knowledge_nodes": strategy_result.get(
+                        "related_knowledge_nodes", []
+                    ),
+                    "rag_context_used": len(relevant_context) if relevant_context else 0,
+                    "rag_context_sources": [ctx["id"] for ctx in (relevant_context or [])],
+                    "confidence": confidence,
+                    "is_strategic": is_strategic,
+                }
+                if "_voting" in strategy_result:
+                    vi = strategy_result["_voting"]
+                    analysis_data["voting"] = {
+                        "enabled": vi.get("enabled"),
+                        "models_used": vi.get("models_used"),
+                        "strategic_votes": vi.get("strategic_votes"),
+                        "total_votes": vi.get("total_votes"),
+                        "unanimous": vi.get("unanimous"),
+                    }
+                new_signal.strategic_analysis = json.dumps(analysis_data)
+                new_signal.tasks_generated = json.dumps(strategy_result.get("tasks", []))
+                new_signal.processed = True
+                await db.commit()
+                processed_count += 1
+
+            if initial_priority in [SignalPriority.MEDIUM, SignalPriority.HIGH]:
+                background_tasks.add_task(process_fetched)
+            else:
+                await process_fetched()
+
+    _stop_ollama()
+
+    return {
+        "status": "success",
+        "new_signals_count": new_signals_count,
+        "processed_count": processed_count,
+        "message": f"Fetched {new_signals_count} signals, {processed_count} analyzed",
+    }
+
+
+@router.get("/{signal_id}/projects")
+async def get_signal_projects(
+    signal_id: str,
+    db: AsyncSession = Depends(get_db),
+):
+    """获取信号关联的项目"""
+    try:
+        sid = int(signal_id)
+    except (ValueError, TypeError):
+        raise HTTPException(status_code=400, detail="Invalid signal_id")
+
+    signal_result = await db.execute(select(Signal).where(Signal.id == sid))
+    if not signal_result.scalar_one_or_none():
+        raise HTTPException(status_code=404, detail="Signal not found")
+
+    result = await db.execute(
+        select(Project, ProjectSignalLink)
+        .join(ProjectSignalLink, Project.id == ProjectSignalLink.project_id)
+        .where(ProjectSignalLink.signal_id == sid)
+    )
+    rows = result.all()
+
+    return [
+        {
+            "id": str(p.id),
+            "title": p.title,
+            "status": p.status,
+            "assigned_role": p.assigned_role,
+            "link_type": link.link_type,
+            "relevance_score": link.relevance_score,
+        }
+        for p, link in rows
+    ]
+
+
+@router.post("/{signal_id}/feedback", response_model=SignalFeedbackResponse)
+async def submit_signal_feedback(
+    signal_id: int,
+    feedback: SignalFeedbackCreate,
+    db: AsyncSession = Depends(get_db),
+):
+    """提交信号反馈"""
+    result = await db.execute(select(Signal).where(Signal.id == signal_id))
+    signal = result.scalar_one_or_none()
+    if not signal:
+        raise HTTPException(status_code=404, detail="Signal not found")
+
+    ai_predicted_strategic = False
+    ai_confidence = 0.0
+    if signal.strategic_analysis:
+        try:
+            analysis = json.loads(signal.strategic_analysis)
+            ai_predicted_strategic = analysis.get("is_strategic", signal.impact_score >= 0.9)
+            ai_confidence = analysis.get("confidence", 0.0)
+        except Exception:
+            ai_predicted_strategic = signal.impact_score >= 0.9
+
+    new_feedback = SignalFeedback(
+        signal_id=signal_id,
+        was_strategic=feedback.was_strategic,
+        reason=feedback.reason,
+        user_id=feedback.user_id,
+        ai_predicted_strategic=ai_predicted_strategic,
+        ai_confidence=ai_confidence,
+    )
+    db.add(new_feedback)
+    await db.commit()
+    await db.refresh(new_feedback)
+
+    is_correct = ai_predicted_strategic == feedback.was_strategic
+
+    return SignalFeedbackResponse(
+        id=new_feedback.id,
+        signal_id=new_feedback.signal_id,
+        was_strategic=new_feedback.was_strategic,
+        ai_predicted_strategic=new_feedback.ai_predicted_strategic,
+        ai_confidence=new_feedback.ai_confidence,
+        is_correct=is_correct,
+    )
+
+
+@router.get("/{signal_id}/feedbacks", response_model=List[SignalFeedbackResponse])
+async def get_signal_feedbacks(
+    signal_id: int,
+    db: AsyncSession = Depends(get_db),
+):
+    """获取某信号的全部反馈"""
+    result = await db.execute(
+        select(SignalFeedback)
+        .where(SignalFeedback.signal_id == signal_id)
+        .order_by(SignalFeedback.created_at.desc())
+    )
+    feedbacks = result.scalars().all()
+
+    return [
+        SignalFeedbackResponse(
+            id=f.id,
+            signal_id=f.signal_id,
+            was_strategic=f.was_strategic,
+            ai_predicted_strategic=f.ai_predicted_strategic,
+            ai_confidence=f.ai_confidence or 0.0,
+            is_correct=f.ai_predicted_strategic == f.was_strategic,
+        )
+        for f in feedbacks
+    ]
diff --git a/backend/app/api/skill_tags.py b/backend/app/api/skill_tags.py
new file mode 100644
index 00000000..8a156c20
--- /dev/null
+++ b/backend/app/api/skill_tags.py
@@ -0,0 +1,320 @@
+"""技能标签与用户技能绑定标准路由。"""
+
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Optional
+
+from fastapi import APIRouter, Depends, HTTPException, Query, status
+from pydantic import BaseModel, Field
+from sqlalchemy import and_, delete, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..database import get_db
+from ..dependencies import get_current_user, get_tenant_id, require_roles
+from ..models import SkillTag, User, UserSkillBinding
+from ..services import skill_service
+from ..services.auth import has_any_role
+
+router = APIRouter(tags=["skill-tags"])
+
+
+class SkillTagCreateRequest(BaseModel):
+    name: str
+    slug: str
+    category: Optional[str] = None
+    description: Optional[str] = None
+    sort_order: int = 0
+
+
+class SkillTagUpdateRequest(BaseModel):
+    name: Optional[str] = None
+    category: Optional[str] = None
+    description: Optional[str] = None
+    sort_order: Optional[int] = None
+
+
+class UserSkillCreateRequest(BaseModel):
+    skill_tag_id: int
+    level: str = "beginner"
+    proficiency: int = Field(default=50, ge=0, le=100)
+    quality_score: float = 0.0
+    efficiency_score: float = 0.0
+
+
+class UserSkillUpdateRequest(BaseModel):
+    level: Optional[str] = None
+    proficiency: Optional[int] = Field(default=None, ge=0, le=100)
+    quality_score: Optional[float] = None
+    efficiency_score: Optional[float] = None
+
+
+def _to_tag_payload(tag: SkillTag) -> dict:
+    return {
+        "id": int(tag.id),
+        "tenant_id": int(tag.tenant_id),
+        "name": tag.name,
+        "slug": tag.slug,
+        "category": tag.category,
+        "description": tag.description,
+        "is_system": bool(tag.is_system),
+        "sort_order": int(tag.sort_order or 0),
+        "created_at": tag.created_at.isoformat() if tag.created_at else None,
+        "updated_at": tag.updated_at.isoformat() if tag.updated_at else None,
+    }
+
+
+@router.get("/skill-tags")
+async def list_skill_tags(
+    category: Optional[str] = Query(None),
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    tags = await skill_service.list_skill_tags(
+        db,
+        tenant_id=tenant_id,
+        category=category,
+        include_system=True,
+    )
+    count_rows = await db.execute(
+        select(UserSkillBinding.skill_tag_id, func.count(UserSkillBinding.id))
+        .where(UserSkillBinding.tenant_id == tenant_id)
+        .group_by(UserSkillBinding.skill_tag_id)
+    )
+    count_map = {int(skill_id): int(total) for skill_id, total in count_rows.all()}
+    items = []
+    for t in tags:
+        payload = _to_tag_payload(t)
+        payload["linked_user_count"] = count_map.get(int(t.id), 0)
+        items.append(payload)
+    return {"items": items, "total": len(items)}
+
+
+@router.post("/skill-tags")
+async def create_skill_tag(
+    body: SkillTagCreateRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    try:
+        tag = await skill_service.create_skill_tag(
+            db,
+            tenant_id=tenant_id,
+            name=body.name,
+            slug=body.slug,
+            category=body.category,
+            description=body.description,
+            sort_order=body.sort_order,
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(exc))
+    return _to_tag_payload(tag)
+
+
+@router.put("/skill-tags/{tag_id}")
+async def update_skill_tag(
+    tag_id: int,
+    body: SkillTagUpdateRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    tag = await skill_service.get_skill_tag(db, tag_id, tenant_id)
+    if not tag:
+        raise HTTPException(status_code=404, detail="Skill tag not found")
+
+    # 系统标签仅允许改分类，禁止改名。
+    if tag.is_system and body.name is not None and body.name != tag.name:
+        raise HTTPException(status_code=400, detail="System tag name cannot be changed")
+
+    if body.name is not None and not tag.is_system:
+        tag.name = body.name
+    if body.category is not None:
+        tag.category = body.category
+    if body.description is not None and not tag.is_system:
+        tag.description = body.description
+    if body.sort_order is not None:
+        tag.sort_order = body.sort_order
+    tag.updated_at = datetime.utcnow()
+    await db.commit()
+    await db.refresh(tag)
+    return _to_tag_payload(tag)
+
+
+@router.delete("/skill-tags/{tag_id}")
+async def delete_skill_tag(
+    tag_id: int,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    tag = await skill_service.get_skill_tag(db, tag_id, tenant_id)
+    if not tag:
+        raise HTTPException(status_code=404, detail="Skill tag not found")
+    if tag.is_system:
+        raise HTTPException(status_code=400, detail="System tag cannot be deleted")
+
+    await db.execute(
+        delete(UserSkillBinding).where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.skill_tag_id == tag_id,
+            )
+        )
+    )
+    await db.delete(tag)
+    await db.commit()
+    return {"message": "Skill tag deleted"}
+
+
+@router.get("/users/{user_id}/skills")
+async def get_user_skills(
+    user_id: int,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    bindings = await skill_service.get_user_skill_bindings(db, tenant_id=tenant_id, user_id=user_id)
+    items = []
+    for row in bindings:
+        items.append(
+            {
+                "skill_tag_id": int(row.skill_tag_id),
+                "skill_name": row.skill_tag.name if row.skill_tag else "",
+                "category": row.skill_tag.category if row.skill_tag else None,
+                "level": row.level or row.proficiency_level,
+                "proficiency": int(row.proficiency or 0),
+                "quality_score": float(row.quality_score or 0.0),
+                "efficiency_score": float(row.efficiency_score or 0.0),
+                "last_used_at": row.last_used_at.isoformat() if row.last_used_at else None,
+                "is_system": bool(row.skill_tag.is_system) if row.skill_tag else False,
+            }
+        )
+    return {"items": items, "total": len(items)}
+
+
+@router.post("/users/{user_id}/skills")
+async def bind_user_skill(
+    user_id: int,
+    body: UserSkillCreateRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    exists = (
+        await db.execute(
+            select(UserSkillBinding).where(
+                and_(
+                    UserSkillBinding.tenant_id == tenant_id,
+                    UserSkillBinding.user_id == user_id,
+                    UserSkillBinding.skill_tag_id == body.skill_tag_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if exists:
+        raise HTTPException(status_code=400, detail="Skill already bound")
+
+    row = UserSkillBinding(
+        tenant_id=tenant_id,
+        user_id=user_id,
+        skill_tag_id=body.skill_tag_id,
+        level=body.level,
+        proficiency_level=body.level,
+        proficiency=body.proficiency,
+        quality_score=body.quality_score,
+        efficiency_score=body.efficiency_score,
+    )
+    db.add(row)
+    await db.commit()
+    return {"message": "Skill bound"}
+
+
+@router.put("/users/{user_id}/skills/{skill_id}")
+async def update_user_skill(
+    user_id: int,
+    skill_id: int,
+    body: UserSkillUpdateRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    current_role = str(current_user.role or "")
+    is_manager = has_any_role(current_role, ["BOSS", "ADMIN"])
+    is_self = int(current_user.id) == user_id
+    if not is_manager and not is_self:
+        raise HTTPException(status_code=403, detail="No permission")
+
+    if is_self and not is_manager:
+        if body.quality_score is not None or body.efficiency_score is not None:
+            raise HTTPException(
+                status_code=403,
+                detail="Members can only update own proficiency",
+            )
+
+    row = (
+        await db.execute(
+            select(UserSkillBinding).where(
+                and_(
+                    UserSkillBinding.tenant_id == tenant_id,
+                    UserSkillBinding.user_id == user_id,
+                    UserSkillBinding.skill_tag_id == skill_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if not row:
+        raise HTTPException(status_code=404, detail="Binding not found")
+
+    if body.level is not None:
+        row.level = body.level
+        row.proficiency_level = body.level
+    if body.proficiency is not None:
+        row.proficiency = body.proficiency
+    if body.quality_score is not None:
+        row.quality_score = body.quality_score
+    if body.efficiency_score is not None:
+        row.efficiency_score = body.efficiency_score
+    row.updated_at = datetime.utcnow()
+    await db.commit()
+    return {"message": "Skill updated"}
+
+
+@router.delete("/users/{user_id}/skills/{skill_id}")
+async def unbind_user_skill(
+    user_id: int,
+    skill_id: int,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    _ = current_user
+    row = (
+        await db.execute(
+            select(UserSkillBinding)
+            .join(SkillTag, SkillTag.id == UserSkillBinding.skill_tag_id)
+            .where(
+                and_(
+                    UserSkillBinding.tenant_id == tenant_id,
+                    UserSkillBinding.user_id == user_id,
+                    UserSkillBinding.skill_tag_id == skill_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if not row:
+        raise HTTPException(status_code=404, detail="Binding not found")
+    if row.skill_tag and row.skill_tag.is_system:
+        raise HTTPException(status_code=400, detail="System bound skill cannot be removed")
+
+    await db.delete(row)
+    await db.commit()
+    return {"message": "Skill unbound"}
diff --git a/backend/app/api/sstc.py b/backend/app/api/sstc.py
new file mode 100644
index 00000000..79e233e2
--- /dev/null
+++ b/backend/app/api/sstc.py
@@ -0,0 +1,3934 @@
+"""
+SSTC 唯一事实链路 API
+
+覆盖模块：
+- 项目概览/链路导航/风险预警
+- 信号源
+- 需求宪章（含对齐）
+- 任务拆解与执行
+- 发布验证
+- 复盘沉淀
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import re
+from datetime import datetime, timedelta, timezone
+from io import BytesIO
+from pathlib import Path
+from typing import Any, Optional
+from uuid import uuid4
+
+from fastapi import APIRouter, BackgroundTasks, Body, Depends, File, HTTPException, Query, UploadFile
+from fastapi.responses import JSONResponse, StreamingResponse
+from pydantic import BaseModel
+from sqlalchemy import and_, delete, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.response import error_response, success_response
+from ..database import get_db
+from ..models import (
+    AICallLog,
+    AlignmentRecord,
+    Integration,
+    ModelRequirementBinding,
+    ModelVersion,
+    ProjectSignalLink,
+    Signal,
+    Task,
+    TruthNode,
+    User,
+    UserProject,
+)
+from ..schemas.sstc import (
+    AlignmentConfirmRequest,
+    AlignmentStartRequest,
+    CharterUpdateRequest,
+    ReleaseVerificationItemPatchRequest,
+    ReleaseCreateRequest,
+    ReleaseStatusPatchRequest,
+    ReviewArchiveRequest,
+    ReviewContentPatchRequest,
+    ReviewCreateRequest,
+    ReviewImprovementToTaskRequest,
+    WorkloadOptimizeRequest,
+    TaskStatusPatchRequest,
+)
+from ..services.llm import llm_service
+from ..services.ai_service import AIService
+from ..services.operation_log_service import log_operation
+from ..services.auth import has_any_role
+from ..services.skill_service import recommend_personnel
+from ..services import skill_score_service
+from ..services.sstc_service import (
+    ALIGNED_STATUSES,
+    DONE_STATUSES,
+    RELEASE_VERIFIED_STATUSES,
+    SSTCService,
+)
+from ..services import workload_service
+from ..services.budget_enforcement_service import BudgetEnforcementService
+from ..services.feishu_integration import get_feishu_service
+from ..services.adaptive_rag_service import AdaptiveRAGService
+from ..services.context_cache import context_cache_service
+from ..services.task_persistence import build_task_create_draft, persist_task
+from ..dependencies import get_current_user, require_project_member_or_manager
+from ..utils.title import compact_title
+
+
+router = APIRouter(tags=["sstc"])
+_FEISHU_PLACEHOLDER_RE = re.compile(r"^飞书用户\(([^)]+)\)$")
+_SIGNAL_PREFIX_RE = re.compile(r"^\[(.*?)\]\s*:\s*")
+
+DEFAULT_SIGNAL_AUTO_ENTER_THRESHOLD = 0.80
+DEFAULT_PROJECT_AUTO_CREATE_THRESHOLD = 0.85
+DEFAULT_PROJECT_UNCERTAIN_THRESHOLD = 0.60
+
+_SIGNAL_STATUS_CODE_PENDING = "pending"
+_SIGNAL_STATUS_CODE_LINKED = "linked"
+_SIGNAL_STATUS_CODE_CLOSED = "closed"
+_SIGNAL_STATUS_LABEL_BY_CODE = {
+    _SIGNAL_STATUS_CODE_PENDING: "待处理",
+    _SIGNAL_STATUS_CODE_LINKED: "已关联宪章",
+    _SIGNAL_STATUS_CODE_CLOSED: "已关闭",
+}
+_SIGNAL_STATUS_FILTER_ALIAS = {
+    "pending": _SIGNAL_STATUS_CODE_PENDING,
+    "待处理": _SIGNAL_STATUS_CODE_PENDING,
+    "linked": _SIGNAL_STATUS_CODE_LINKED,
+    "已关联宪章": _SIGNAL_STATUS_CODE_LINKED,
+    "closed": _SIGNAL_STATUS_CODE_CLOSED,
+    "archived": _SIGNAL_STATUS_CODE_CLOSED,
+    "已关闭": _SIGNAL_STATUS_CODE_CLOSED,
+    "已归档": _SIGNAL_STATUS_CODE_CLOSED,
+}
+_SIGNAL_LOCK_REASON = "信号已关联宪章，禁止重复新建"
+_BEIJING_UTC_OFFSET = timedelta(hours=8)
+_ALLOWED_ATTACHMENT_TYPES = {"image", "file", "media", "document"}
+_SIGNAL_ATTACHMENT_MAX_BYTES = 20 * 1024 * 1024
+_SIGNAL_ATTACHMENT_ALLOWED_EXTENSIONS = {
+    ".jpg": "image",
+    ".jpeg": "image",
+    ".png": "image",
+    ".gif": "image",
+    ".pdf": "document",
+    ".doc": "document",
+    ".docx": "document",
+    ".zip": "file",
+}
+_SIGNAL_ATTACHMENT_ALLOWED_MIME_TYPES = {
+    "image/jpeg",
+    "image/png",
+    "image/gif",
+    "application/pdf",
+    "application/msword",
+    "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
+    "application/zip",
+    "application/x-zip-compressed",
+    "multipart/x-zip",
+    "application/octet-stream",
+}
+
+_CHARTER_REQUIRED_FIELDS: dict[str, tuple[str, ...]] = {
+    "module_1": ("problem_statement", "goal_statement", "kpis"),
+    "module_2": ("in_scope", "out_of_scope", "alignment_qa"),
+    "module_3": ("primary_model", "fallback_model", "capability_table", "dependency_map"),
+    "module_4": ("milestones", "role_effort_estimates", "capability_owner_assignment"),
+    "module_5": ("risk_register", "trigger_rules", "execution_rule"),
+    "module_6": ("attachment_metadata", "attachment_summary", "search_tags"),
+}
+
+RELEASE_STATUS_TRANSITIONS: dict[str, set[str]] = {
+    "pending": {"releasing", "online"},
+    "releasing": {"online", "rolled_back"},
+    "online": {"rolled_back"},
+    "rolled_back": {"releasing"},
+    "verified": set(),
+}
+
+
+def _sync_pmflow_rag_async(
+    *,
+    tenant_id: int,
+    record_type: str,
+    record_id: int,
+    project_id: int,
+    title: str,
+    content: dict[str, Any] | str | None,
+) -> None:
+    try:
+        if isinstance(content, str):
+            payload = content
+        else:
+            payload = json.dumps(content or {}, ensure_ascii=False)
+        asyncio.create_task(
+            AdaptiveRAGService.upsert_pmflow_record_async(
+                tenant_id=tenant_id,
+                record_type=record_type,
+                record_id=record_id,
+                project_id=project_id,
+                title=title,
+                content=payload,
+            )
+        )
+    except Exception:
+        return
+
+
+def _parse_project_id(project_id: str) -> int:
+    try:
+        return int(project_id)
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail="Invalid project_id") from exc
+
+
+def _parse_node_id(node_id: str, label: str) -> int:
+    try:
+        return int(node_id)
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=f"Invalid {label}") from exc
+
+
+def _as_dict(value: Any) -> dict:
+    if isinstance(value, dict):
+        return value
+    if isinstance(value, str):
+        import json
+        try:
+            parsed = json.loads(value)
+            return parsed if isinstance(parsed, dict) else {"text": value}
+        except (TypeError, ValueError):
+            return {"text": value}
+    return {}
+
+
+def _signal_ai_analysis(signal: Signal) -> dict:
+    payload = _as_dict(signal.strategic_analysis)
+    ai_payload = payload.get("ai_analysis")
+    return ai_payload if isinstance(ai_payload, dict) else {}
+
+
+def _signal_project_scope(signal: Signal) -> Optional[str]:
+    scope = str(_signal_ai_analysis(signal).get("requirement_project_scope") or "").strip().lower()
+    if scope in {"new_project", "existing_project", "pending_classification", "non_requirement"}:
+        return scope
+    return None
+
+
+def _signal_display_type(signal: Signal) -> str:
+    scope = _signal_project_scope(signal)
+    if scope:
+        return scope
+    return "pending_classification"
+
+
+def _signal_source_context(signal: Signal) -> dict:
+    payload = _as_dict(signal.strategic_analysis)
+    source_context = payload.get("source_context")
+    return source_context if isinstance(source_context, dict) else {}
+
+
+def _safe_int(value: Any) -> Optional[int]:
+    try:
+        return int(str(value).strip())
+    except (TypeError, ValueError, AttributeError):
+        return None
+
+
+def _signal_attachment_storage_root() -> Path:
+    root = Path(__file__).resolve().parents[2] / "storage" / "signal_attachments"
+    root.mkdir(parents=True, exist_ok=True)
+    return root
+
+
+def _build_signal_attachment_download_url(project_id: int, signal_id: int, attachment_id: str) -> str:
+    return f"/projects/{project_id}/signals/{signal_id}/attachments/{attachment_id}/download"
+
+
+def _resolve_local_attachment_path(file_key: str) -> Optional[Path]:
+    raw = str(file_key or "").strip()
+    if not raw.startswith("local://"):
+        return None
+    relative = raw.removeprefix("local://").strip().lstrip("/")
+    if not relative:
+        return None
+    storage_root = _signal_attachment_storage_root().resolve()
+    candidate = (storage_root / relative).resolve()
+    try:
+        candidate.relative_to(storage_root)
+    except ValueError:
+        return None
+    return candidate
+
+
+def _error_json(status_code: int, message: str) -> JSONResponse:
+    return JSONResponse(status_code=status_code, content=error_response(status_code, message))
+
+
+def _normalize_attachment_type(value: Any) -> str:
+    raw = str(value or "").strip().lower()
+    alias = {
+        "img": "image",
+        "picture": "image",
+        "attachment": "file",
+        "doc": "document",
+        "docs": "document",
+    }
+    mapped = alias.get(raw, raw)
+    return mapped if mapped in _ALLOWED_ATTACHMENT_TYPES else "file"
+
+
+def _normalize_signal_attachment(item: Any, *, index: int) -> Optional[dict[str, Any]]:
+    if not isinstance(item, dict):
+        return None
+    attachment_type = _normalize_attachment_type(item.get("attachment_type") or item.get("type"))
+    attachment_id = (
+        str(item.get("attachment_id") or item.get("id") or item.get("file_key") or item.get("image_key") or item.get("media_id") or "")
+        .strip()
+    )
+    if not attachment_id:
+        attachment_id = f"attachment-{index + 1}"
+    file_name = (
+        str(item.get("file_name") or item.get("name") or item.get("title") or "").strip()
+        or f"{attachment_type}-{attachment_id}"
+    )
+    file_size = _safe_int(item.get("file_size") or item.get("size"))
+    mime_type = str(item.get("mime_type") or item.get("mime") or "").strip() or None
+    occurred_at = str(item.get("occurred_at") or "").strip() or None
+    download_url = str(item.get("download_url") or item.get("url") or "").strip() or None
+    file_key = str(item.get("file_key") or "").strip() or None
+    image_key = str(item.get("image_key") or "").strip() or None
+    media_id = str(item.get("media_id") or "").strip() or None
+    source = str(item.get("source") or "").strip() or None
+    if not any([file_key, image_key, media_id, download_url]):
+        return None
+    return {
+        "id": attachment_id,
+        "attachment_id": attachment_id,
+        "attachment_type": attachment_type,
+        "file_name": file_name,
+        "mime_type": mime_type,
+        "file_size": file_size,
+        "file_key": file_key,
+        "image_key": image_key,
+        "media_id": media_id,
+        "source": source,
+        "occurred_at": occurred_at,
+        "download_url": download_url,
+    }
+
+
+def _extract_source_attachments(source_context: dict[str, Any]) -> list[dict[str, Any]]:
+    raw_items = source_context.get("attachments")
+    if not isinstance(raw_items, list):
+        return []
+    items: list[dict[str, Any]] = []
+    seen_ids: set[str] = set()
+    for idx, raw in enumerate(raw_items):
+        normalized = _normalize_signal_attachment(raw, index=idx)
+        if normalized is None:
+            continue
+        aid = str(normalized.get("attachment_id") or "").strip()
+        if not aid or aid in seen_ids:
+            continue
+        seen_ids.add(aid)
+        items.append(normalized)
+    return items
+
+
+def _extract_signal_attachments(signal: Signal) -> list[dict[str, Any]]:
+    source_context = _signal_source_context(signal)
+    return _extract_source_attachments(source_context)
+
+
+def _normalize_signal_status_code(
+    *,
+    linked_state: Optional[str],
+    signal_status: Optional[str],
+    processed: bool,
+) -> str:
+    if linked_state in {_SIGNAL_STATUS_CODE_LINKED, _SIGNAL_STATUS_CODE_CLOSED}:
+        return linked_state
+    raw = str(signal_status or "").strip().lower()
+    if "关联" in raw or raw == _SIGNAL_STATUS_CODE_LINKED:
+        return _SIGNAL_STATUS_CODE_LINKED
+    if "关闭" in raw or "归档" in raw or raw in {"closed", "archived"}:
+        return _SIGNAL_STATUS_CODE_CLOSED
+    if processed:
+        return _SIGNAL_STATUS_CODE_CLOSED
+    return _SIGNAL_STATUS_CODE_PENDING
+
+
+def _signal_status_label_from_code(status_code: str, fallback_status: Optional[str] = None) -> str:
+    fallback = str(fallback_status or "").strip()
+    normalized_code = str(status_code or "").strip().lower()
+    if (
+        normalized_code == _SIGNAL_STATUS_CODE_PENDING
+        and fallback
+        and all(token not in fallback for token in ["关联", "关闭", "归档"])
+    ):
+        return fallback
+    label = _SIGNAL_STATUS_LABEL_BY_CODE.get(normalized_code)
+    if label:
+        return label
+    return fallback or _SIGNAL_STATUS_LABEL_BY_CODE[_SIGNAL_STATUS_CODE_PENDING]
+
+
+def _extract_charter_related_signal_ids(charter: TruthNode) -> list[int]:
+    payload = _as_dict(charter.content)
+    signal_ids: list[int] = []
+    raw_ids = payload.get("signal_ids")
+    if isinstance(raw_ids, list):
+        for raw_id in raw_ids:
+            sid = _safe_int(raw_id)
+            if sid is not None:
+                signal_ids.append(sid)
+    charter_signal_id = _safe_int(charter.signal_id)
+    if charter_signal_id is not None:
+        signal_ids.append(charter_signal_id)
+    return sorted(set(signal_ids))
+
+
+def _collect_locked_signal_conflicts(
+    *,
+    charters: list[TruthNode],
+    candidate_signal_ids: list[int],
+) -> tuple[list[int], dict[int, list[dict[str, Any]]]]:
+    candidate_set = set(int(sid) for sid in candidate_signal_ids)
+    conflicts: dict[int, list[dict[str, Any]]] = {}
+    for charter in charters:
+        charter_signal_ids = _extract_charter_related_signal_ids(charter)
+        if not charter_signal_ids:
+            continue
+        for sid in charter_signal_ids:
+            if sid not in candidate_set:
+                continue
+            conflicts.setdefault(sid, []).append(
+                {
+                    "charter_id": int(charter.id),
+                    "charter_title": str(charter.title or "").strip() or f"宪章#{int(charter.id)}",
+                }
+            )
+    conflict_ids = sorted(conflicts.keys())
+    return conflict_ids, conflicts
+
+
+def _parse_beijing_day_start(raw: str, *, label: str) -> datetime:
+    text = str(raw or "").strip()
+    if not text:
+        raise HTTPException(status_code=400, detail=f"{label} 不能为空")
+    if "T" in text:
+        text = text.split("T")[0]
+    try:
+        local_day = datetime.strptime(text, "%Y-%m-%d")
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=f"Invalid {label}") from exc
+    return local_day - _BEIJING_UTC_OFFSET
+
+
+def _build_signal_created_at_range(
+    *,
+    created_from: Optional[str],
+    created_to: Optional[str],
+) -> tuple[Optional[datetime], Optional[datetime]]:
+    start_utc = _parse_beijing_day_start(created_from, label="created_from") if created_from else None
+    end_utc = None
+    if created_to:
+        end_day_start_utc = _parse_beijing_day_start(created_to, label="created_to")
+        end_utc = end_day_start_utc + timedelta(days=1)
+    if start_utc and end_utc and start_utc >= end_utc:
+        raise HTTPException(status_code=400, detail="created_from 不能晚于 created_to")
+    return start_utc, end_utc
+
+
+def _normalize_signal_status_filter(value: Optional[str]) -> Optional[str]:
+    raw = str(value or "").strip()
+    if not raw:
+        return None
+    return _SIGNAL_STATUS_FILTER_ALIAS.get(raw.lower(), _SIGNAL_STATUS_FILTER_ALIAS.get(raw, raw.lower()))
+
+
+def _match_signal_filters(
+    *,
+    row: dict[str, Any],
+    status_filter: Optional[str],
+    keyword: Optional[str],
+    source: Optional[str],
+    priority: Optional[str],
+    created_from_utc: Optional[datetime],
+    created_to_utc: Optional[datetime],
+    created_at_dt: Optional[datetime],
+) -> bool:
+    if status_filter:
+        if str(row.get("status_code") or "").strip().lower() != str(status_filter).strip().lower():
+            return False
+
+    keyword_text = str(keyword or "").strip().lower()
+    if keyword_text:
+        haystack = " ".join(
+            [
+                str(row.get("id") or ""),
+                str(row.get("title") or ""),
+                str(row.get("content") or ""),
+                str(row.get("proposer_name") or ""),
+                str(row.get("charter_name") or ""),
+            ]
+        ).lower()
+        if keyword_text not in haystack:
+            return False
+
+    source_text = str(source or "").strip().lower()
+    if source_text and source_text != "all":
+        row_source = str(row.get("source") or "").strip().lower()
+        if row_source != source_text:
+            return False
+
+    priority_text = str(priority or "").strip().lower()
+    if priority_text and priority_text != "all":
+        row_priority = str(row.get("priority") or "").strip().lower()
+        if row_priority != priority_text:
+            return False
+
+    if created_from_utc or created_to_utc:
+        if created_at_dt is None:
+            return False
+        if created_from_utc and created_at_dt < created_from_utc:
+            return False
+        if created_to_utc and created_at_dt >= created_to_utc:
+            return False
+
+    return True
+
+
+def _strip_signal_sender_prefix(content: str) -> str:
+    text = str(content or "")
+    return _SIGNAL_PREFIX_RE.sub("", text, count=1).strip() or text
+
+
+def _extract_sender_from_content(content: str) -> Optional[str]:
+    text = str(content or "")
+    matched = _SIGNAL_PREFIX_RE.match(text)
+    if not matched:
+        return None
+    sender = (matched.group(1) or "").strip()
+    return sender or None
+
+
+def _signal_title_text(signal: Signal) -> str:
+    return _strip_signal_sender_prefix(str(signal.content or ""))
+
+
+def _signal_proposer_name(signal: Signal) -> str:
+    source_context = _signal_source_context(signal)
+    sender = str(source_context.get("sender") or "").strip()
+    if sender:
+        return sender
+    sender_from_content = _extract_sender_from_content(str(signal.content or ""))
+    return sender_from_content or "未知用户"
+
+
+def _replace_signal_sender_in_content(content: str, sender_name: str) -> str:
+    text = str(content or "")
+    if not text:
+        return text
+    if _SIGNAL_PREFIX_RE.match(text):
+        return _SIGNAL_PREFIX_RE.sub(f"[{sender_name}]: ", text, count=1)
+    return f"[{sender_name}]: {text}"
+
+
+def _safe_threshold(value: Any, default: float) -> float:
+    try:
+        number = float(value)
+    except (TypeError, ValueError):
+        number = float(default)
+    return min(1.0, max(0.0, number))
+
+
+def _parse_json_object_text(raw: str) -> dict[str, Any]:
+    text = str(raw or "").strip()
+    if not text:
+        return {}
+    if text.startswith("```"):
+        text = text.strip("`")
+        if text.lower().startswith("json"):
+            text = text[4:].strip()
+    try:
+        parsed = json.loads(text)
+        return parsed if isinstance(parsed, dict) else {}
+    except Exception:
+        return {}
+
+
+def _seed_text_snapshot(seed: dict[str, Any]) -> str:
+    pieces: list[str] = []
+    for value in seed.values():
+        if isinstance(value, str):
+            pieces.append(value)
+        elif isinstance(value, (list, tuple)):
+            pieces.extend([str(item) for item in value if item is not None])
+    return "\n".join([part for part in pieces if str(part).strip()]).strip()
+
+
+def _needs_seed_localize_zh(seed: dict[str, Any]) -> bool:
+    text = _seed_text_snapshot(seed)
+    if not text:
+        return False
+    zh_chars = len(re.findall(r"[\u4e00-\u9fff]", text))
+    latin_chars = len(re.findall(r"[A-Za-z]", text))
+    return zh_chars < 20 and latin_chars > 40
+
+
+async def _localize_seed_to_zh(
+    *,
+    db: AsyncSession,
+    tenant_id: int,
+    project_id: int,
+    seed: dict[str, Any],
+) -> dict[str, Any]:
+    if not seed:
+        return seed
+    prompt = (
+        "将下列 JSON 草稿的自然语言内容翻译为简体中文。"
+        "要求：仅返回 JSON；键名与结构保持不变；模型名/专有技术名词可保留英文。\n\n"
+        f"{json.dumps(seed, ensure_ascii=False)}"
+    )
+    try:
+        raw = await llm_service.generate(
+            prompt,
+            function_name="sstc_charter_localize_zh",
+            db=db,
+            tenant_id=tenant_id,
+            project_id=project_id,
+        )
+        localized = _parse_json_object_text(raw)
+        return localized if localized else seed
+    except Exception:
+        return seed
+
+
+def _to_string_list(value: Any) -> list[str]:
+    if value is None:
+        return []
+    if isinstance(value, str):
+        lines = [line.strip(" -\t") for line in value.replace("\r", "").split("\n")]
+        return [line for line in lines if line]
+    if isinstance(value, (list, tuple)):
+        output: list[str] = []
+        for item in value:
+            text = str(item or "").strip(" -\t")
+            if text:
+                output.append(text)
+        return output
+    text = str(value).strip()
+    return [text] if text else []
+
+
+def _to_bullet_text(items: list[str]) -> str:
+    cleaned = [str(item or "").strip() for item in items if str(item or "").strip()]
+    if not cleaned:
+        return ""
+    return "\n".join([f"- {item}" for item in cleaned])
+
+
+def _extract_charter_signal_ids(content: dict[str, Any], charter_signal_id: Optional[int] = None) -> list[int]:
+    signal_ids: list[int] = []
+    raw_ids = content.get("signal_ids")
+    if isinstance(raw_ids, list):
+        for sid in raw_ids:
+            try:
+                signal_ids.append(int(sid))
+            except Exception:
+                continue
+    if charter_signal_id:
+        try:
+            cid = int(charter_signal_id)
+            if cid not in signal_ids:
+                signal_ids.append(cid)
+        except Exception:
+            pass
+    return sorted(set(signal_ids))
+
+
+def _build_signal_context_text(
+    signal_rows: list[Signal],
+    extra_text: str = "",
+    raw_text: str = "",
+) -> str:
+    parts: list[str] = []
+    for idx, signal in enumerate(signal_rows, start=1):
+        title = _signal_title_text(signal) or f"信号{idx}"
+        parts.append(
+            "\n".join(
+                [
+                    f"[信号{idx}] 标题：{title}",
+                    f"来源：{signal.source or '-'}",
+                    f"优先级：{signal.priority or '-'}",
+                    f"内容：{str(signal.content or '').strip()}",
+                ]
+            ).strip()
+        )
+    extra_blocks: list[str] = []
+    for block in (extra_text, raw_text):
+        text = str(block or "").strip()
+        if text and text not in extra_blocks:
+            extra_blocks.append(text)
+    for text in extra_blocks:
+        parts.append(f"[补充上下文]\n{text}")
+    return "\n\n".join([part for part in parts if part]).strip()
+
+
+def _has_non_empty_module_data(content: dict[str, Any]) -> bool:
+    modules = content.get("modules")
+    if not isinstance(modules, dict):
+        return False
+    for module in modules.values():
+        if not isinstance(module, dict):
+            continue
+        for value in module.values():
+            if str(value or "").strip():
+                return True
+    return False
+
+
+def _has_complete_module_data(content: dict[str, Any]) -> bool:
+    modules = content.get("modules")
+    if not isinstance(modules, dict):
+        return False
+    for module_key, required_fields in _CHARTER_REQUIRED_FIELDS.items():
+        module_payload = modules.get(module_key)
+        if not isinstance(module_payload, dict):
+            return False
+        for field_key in required_fields:
+            if not str(module_payload.get(field_key) or "").strip():
+                return False
+    return True
+
+
+def _merge_prefill_content(
+    existing: dict[str, Any], incoming: dict[str, Any], *, overwrite_existing: bool
+) -> dict[str, Any]:
+    merged: dict[str, Any] = dict(existing)
+    for key, value in incoming.items():
+        if overwrite_existing or key not in merged:
+            merged[key] = value
+            continue
+        current = merged.get(key)
+        if isinstance(current, dict) and isinstance(value, dict):
+            merged[key] = _merge_prefill_content(
+                current, value, overwrite_existing=overwrite_existing
+            )
+            continue
+        if isinstance(current, list):
+            if not current and value:
+                merged[key] = value
+            continue
+        if current is None or not str(current).strip():
+            merged[key] = value
+    return merged
+
+
+def _default_ai_validations(module_name: str) -> str:
+    return _to_bullet_text(
+        [
+            f"校验「{module_name}」与信号源证据链一致",
+            "校验字段可验证、可执行、可追踪",
+            "发现冲突时自动标注并等待人工否决/确认",
+        ]
+    )
+
+
+def _default_escalation_trigger() -> str:
+    return _to_bullet_text(
+        [
+            "置信度低于 0.75",
+            "关键字段缺失或来源证据冲突",
+            "超过模型能力边界或预算/时限约束",
+        ]
+    )
+
+
+async def _generate_charter_seed(
+    *,
+    db: AsyncSession,
+    tenant_id: int,
+    project_id: int,
+    signal_context: str,
+) -> dict[str, Any]:
+    if not signal_context.strip():
+        return {}
+
+    prompt = (
+        "你是 AI 项目经理，请基于输入信号生成可落地的宪章草稿。"
+        "所有自然语言字段必须使用简体中文，禁止英文句子；模型名称或技术名词可保留英文。"
+        "只输出 JSON，不要 markdown。\n"
+        "字段：\n"
+        "{\n"
+        '  "summary": "一句话问题摘要",\n'
+        '  "background": "背景与动机",\n'
+        '  "objectives": ["目标1","目标2"],\n'
+        '  "scope": "范围说明",\n'
+        '  "in_scope": ["范围内事项1","范围内事项2"],\n'
+        '  "out_of_scope": ["范围外事项1","范围外事项2"],\n'
+        '  "success_metrics": ["指标1","指标2"],\n'
+        '  "risks": ["风险1","风险2"],\n'
+        '  "dependencies": ["依赖1","依赖2"],\n'
+        '  "primary_model": "模型名",\n'
+        '  "fallback_model": "备用模型名"\n'
+        "}\n\n"
+        f"信号输入：\n{signal_context[:6000]}"
+    )
+    try:
+        raw = await llm_service.generate(
+            prompt,
+            function_name="sstc_charter_prefill",
+            db=db,
+            tenant_id=tenant_id,
+            project_id=project_id,
+        )
+        parsed = _parse_json_object_text(raw)
+        if parsed:
+            if _needs_seed_localize_zh(parsed):
+                return await _localize_seed_to_zh(
+                    db=db, tenant_id=tenant_id, project_id=project_id, seed=parsed
+                )
+            return parsed
+    except Exception:
+        pass
+
+    try:
+        ai_seed = await AIService().generate_charter_from_signal(
+            signal_context,
+            tenant_id=tenant_id,
+            db=db,
+        )
+        if isinstance(ai_seed, dict):
+            if _needs_seed_localize_zh(ai_seed):
+                return await _localize_seed_to_zh(
+                    db=db, tenant_id=tenant_id, project_id=project_id, seed=ai_seed
+                )
+            return ai_seed
+        return {}
+    except Exception:
+        return {}
+
+
+async def _build_ai_first_charter_content(
+    *,
+    db: AsyncSession,
+    tenant_id: int,
+    project_id: int,
+    signal_rows: list[Signal],
+    signal_ids: list[int],
+    raw_text: str = "",
+) -> dict[str, Any]:
+    signal_context = _build_signal_context_text(signal_rows, raw_text=raw_text)
+    seed = await _generate_charter_seed(
+        db=db,
+        tenant_id=tenant_id,
+        project_id=project_id,
+        signal_context=signal_context,
+    )
+
+    summary = str(seed.get("summary") or "").strip()
+    background = str(seed.get("background") or "").strip()
+    objectives = _to_string_list(seed.get("objectives"))
+    scope = str(seed.get("scope") or "").strip()
+    in_scope_items = _to_string_list(seed.get("in_scope"))
+    out_of_scope_items = _to_string_list(seed.get("out_of_scope"))
+    success_metrics = _to_string_list(seed.get("success_metrics"))
+    risks = _to_string_list(seed.get("risks"))
+    dependencies = _to_string_list(seed.get("dependencies"))
+    primary_model = str(seed.get("primary_model") or "").strip()
+    fallback_model = str(seed.get("fallback_model") or "").strip()
+
+    if not summary:
+        summary = "目标用户在当前流程中存在效率损耗，需要通过 AI 驱动的流程优化提升交付质量与速度。"
+    if not background:
+        background = signal_context[:500] or "基于已绑定信号源自动生成。"
+    if not objectives:
+        objectives = ["缩短端到端交付周期", "提升关键路径质量稳定性", "降低跨角色协同成本"]
+    if not success_metrics:
+        success_metrics = ["需求到上线周期缩短 >= 30%", "关键任务按期交付率 >= 90%", "线上回滚率 <= 3%"]
+    if not in_scope_items:
+        in_scope_items = ["需求澄清与宪章生成自动化", "任务拆解与负责人推荐", "对齐状态自动追踪与留痕"]
+    if not out_of_scope_items:
+        out_of_scope_items = ["替代团队组织架构决策", "绕过人工否决机制自动上线", "非项目相关流程改造"]
+    if not risks:
+        risks = ["需求波动导致范围蔓延", "模型能力边界与输入复杂度不匹配", "关键成员负载过高导致延期"]
+    if not dependencies:
+        dependencies = ["FastAPI", "React", "LLM 推理服务", "项目知识库/日志系统"]
+    if not primary_model:
+        primary_model = "待 AI 评估后绑定"
+    if not fallback_model:
+        fallback_model = "同供应商低成本模型"
+
+    source_refs = [f"signal:{sid}" for sid in signal_ids]
+    now_iso = datetime.now(timezone.utc).isoformat()
+    validations_m1 = _default_ai_validations("模块1：触发与目标")
+    validations_m2 = _default_ai_validations("模块2：范围与对齐")
+    validations_m3 = _default_ai_validations("模块3：模型与依赖")
+    validations_m4 = _default_ai_validations("模块4：计划与资源")
+    validations_m5 = _default_ai_validations("模块5：风险与兜底")
+    validations_m6 = _default_ai_validations("模块6：附件与治理")
+    escalation = _default_escalation_trigger()
+
+    alignment_qa = _to_bullet_text(
+        [
+            f"核心用户与痛点：{summary}",
+            "上线时间与硬约束：由里程碑计划自动回填并持续校验",
+            f"明确不做项：{'; '.join(out_of_scope_items[:3])}",
+            f"关键依赖：{'; '.join(dependencies[:4])}",
+            f"核心风险与缓解：{'; '.join(risks[:3])}",
+        ]
+    )
+
+    modules = {
+        "module_1": {
+            "problem_statement": summary,
+            "goal_statement": _to_bullet_text(objectives),
+            "kpis": _to_bullet_text(success_metrics),
+            "ai_validations": validations_m1,
+            "escalation_trigger": escalation,
+        },
+        "module_2": {
+            "in_scope": _to_bullet_text(in_scope_items),
+            "out_of_scope": _to_bullet_text(out_of_scope_items),
+            "alignment_qa": alignment_qa,
+            "ai_validations": validations_m2,
+            "escalation_trigger": escalation,
+        },
+        "module_3": {
+            "primary_model": primary_model,
+            "fallback_model": fallback_model,
+            "capability_table": _to_bullet_text(
+                [
+                    "上下文窗口/并发/延迟由模型能力中心实时读取",
+                    "按 P50/P90 工作负载自动估算 token 与成本",
+                    "超边界输入自动触发替代模型建议",
+                ]
+            ),
+            "dependency_map": _to_bullet_text(dependencies),
+            "business_view": "以业务术语展示能力上限、成本区间与上线风险。",
+            "engineering_view": "以工程参数展示 context window、QPS、成本与错误预算。",
+            "ai_validations": validations_m3,
+            "escalation_trigger": escalation,
+        },
+        "module_4": {
+            "milestones": _to_bullet_text(
+                [
+                    "需求冻结 -> 技术评审 -> 开发完成 -> QA 验收 -> 发布上线",
+                    "关键依赖阻塞超过阈值自动重排计划",
+                ]
+            ),
+            "role_effort_estimates": _to_bullet_text(
+                [
+                    "产品/PM：需求澄清与验收口径",
+                    "前后端工程：核心功能开发与联调",
+                    "AI 工程：模型评估、提示词与推理链路优化",
+                ]
+            ),
+            "capability_owner_assignment": "负责人由能力图谱自动推荐并可人工否决。",
+            "buffer_policy": "默认 15% 进度缓冲，关键路径超阈值触发自动预警。",
+            "ai_validations": validations_m4,
+            "escalation_trigger": escalation,
+        },
+        "module_5": {
+            "risk_register": _to_bullet_text(risks),
+            "trigger_rules": _to_bullet_text(
+                [
+                    "进度偏差 > 15%",
+                    "关键接口错误率连续超过阈值",
+                    "模型调用成本超预算预警线",
+                ]
+            ),
+            "execution_rule": _to_bullet_text(
+                [
+                    "触发后优先执行预设 fallback",
+                    "自动通知责任人并写入审计日志",
+                    "连续失败时升级到人工决策",
+                ]
+            ),
+            "ai_validations": validations_m5,
+            "escalation_trigger": escalation,
+        },
+        "module_6": {
+            "attachment_metadata": _to_bullet_text(
+                [
+                    "原始信号记录",
+                    "需求讨论记录",
+                    "设计与评审文档",
+                    "发布与复盘材料",
+                ]
+            ),
+            "attachment_summary": "AI 自动提炼附件决策结论并关联到对应模块字段。",
+            "search_tags": _to_bullet_text(["需求宪章", "AI First", "SSTC", "对齐记录"]),
+            "ai_validations": validations_m6,
+            "governance_rule": "附件仅作证据引用，核心事实字段以模块1-5为准。",
+        },
+    }
+
+    return {
+        "schema_version": "charter_ai_first_v1",
+        "auto_generated_by_ai": True,
+        "metadata": {
+            "status": "draft",
+            "confidence_score": 0.82 if seed else 0.72,
+            "last_validated_at": now_iso,
+            "evidence_refs": source_refs,
+            "escalation_reason": None,
+        },
+        "modules": modules,
+        "signal_ids": signal_ids,
+        "summary": summary,
+        "background": background,
+        "objectives": objectives,
+        "scope": scope,
+        "success_metrics": success_metrics,
+        "core_metrics": success_metrics,
+        "rationale": summary,
+        "decision": _to_bullet_text(objectives),
+        "in_scope": in_scope_items,
+        "out_of_scope": out_of_scope_items,
+        "risks": risks,
+        "raw_source_text": raw_text.strip() or None,
+    }
+
+
+def _extract_feishu_routing_config(integration: Optional[Integration]) -> dict[str, Any]:
+    config: dict[str, Any] = {}
+    if integration is not None:
+        if isinstance(integration.config, dict):
+            config = integration.config
+        elif isinstance(integration.config, str):
+            try:
+                parsed = json.loads(integration.config)
+                if isinstance(parsed, dict):
+                    config = parsed
+            except (TypeError, ValueError):
+                config = {}
+
+    cold_start_enabled = bool(config.get("cold_start_enabled"))
+    signal_threshold = _safe_threshold(
+        config.get("signal_auto_enter_threshold"),
+        DEFAULT_SIGNAL_AUTO_ENTER_THRESHOLD,
+    )
+    project_threshold = _safe_threshold(
+        config.get("project_auto_create_threshold"),
+        DEFAULT_PROJECT_AUTO_CREATE_THRESHOLD,
+    )
+    uncertain_threshold = min(
+        project_threshold,
+        _safe_threshold(
+            config.get("project_uncertain_threshold"),
+            DEFAULT_PROJECT_UNCERTAIN_THRESHOLD,
+        ),
+    )
+    return {
+        "raw": config,
+        "cold_start_enabled": cold_start_enabled,
+        "signal_auto_enter_threshold": signal_threshold,
+        "project_auto_create_threshold": project_threshold,
+        "project_uncertain_threshold": uncertain_threshold,
+    }
+
+
+async def _backfill_feishu_sender_names(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    signals: list[Signal],
+) -> None:
+    pending: dict[int, dict[str, str]] = {}
+    for signal in signals:
+        if str(signal.source or "").lower() != "feishu":
+            continue
+        source_context = _signal_source_context(signal)
+        open_id = str(source_context.get("sender_open_id") or "").strip()
+        sender = _signal_proposer_name(signal).strip()
+        if not open_id or not sender:
+            continue
+        if _FEISHU_PLACEHOLDER_RE.match(sender):
+            pending[int(signal.id)] = {
+                "open_id": open_id,
+                "current_sender": sender,
+            }
+    if not pending:
+        return
+
+    service = get_feishu_service(db)
+    await service.load_config(tenant_id=tenant_id)
+    if not service.app_id or not service.app_secret:
+        return
+
+    resolved_name_by_open_id: dict[str, str] = {}
+    for open_id in sorted({meta["open_id"] for meta in pending.values()})[:20]:
+        try:
+            remote_name = await service._fetch_user_name_by_open_id(open_id)  # noqa: SLF001
+        except Exception:
+            remote_name = None
+        normalized = (remote_name or "").strip()
+        if not normalized or _FEISHU_PLACEHOLDER_RE.match(normalized):
+            continue
+        resolved_name_by_open_id[open_id] = normalized
+
+    if not resolved_name_by_open_id:
+        return
+
+    changed = False
+    for signal in signals:
+        meta = pending.get(int(signal.id))
+        if not meta:
+            continue
+        sender_name = resolved_name_by_open_id.get(meta["open_id"])
+        if not sender_name:
+            continue
+        payload = _as_dict(signal.strategic_analysis)
+        source_context = payload.get("source_context")
+        if not isinstance(source_context, dict):
+            source_context = {}
+        source_context["sender"] = sender_name
+        source_context["sender_open_id"] = meta["open_id"]
+        payload["source_context"] = source_context
+        signal.strategic_analysis = json.dumps(payload, ensure_ascii=False)
+        signal.content = _replace_signal_sender_in_content(str(signal.content or ""), sender_name)
+        changed = True
+
+    if changed:
+        await db.commit()
+
+
+def _parse_due_date(value: Optional[str]) -> Optional[datetime]:
+    if not value:
+        return None
+    try:
+        return datetime.fromisoformat(value.replace("Z", "+00:00")).replace(tzinfo=None)
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail="Invalid due_date") from exc
+
+
+async def _ensure_project(
+    svc: SSTCService,
+    project_id: int,
+):
+    project = await svc.get_project(project_id)
+    if not project:
+        raise HTTPException(status_code=404, detail="Project not found")
+    return project
+
+
+async def _ensure_charter(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    charter_id: int,
+) -> TruthNode:
+    result = await db.execute(
+        select(TruthNode).where(
+            and_(
+                TruthNode.id == charter_id,
+                TruthNode.project_id == project_id,
+                TruthNode.tenant_id == tenant_id,
+                TruthNode.node_type == "charter",
+            )
+        )
+    )
+    charter = result.scalar_one_or_none()
+    if not charter:
+        raise HTTPException(status_code=404, detail="Charter not found")
+    return charter
+
+
+class AIRoutingConfigUpdateRequest(BaseModel):
+    cold_start_enabled: Optional[bool] = None
+    signal_auto_enter_threshold: Optional[float] = None
+    project_auto_create_threshold: Optional[float] = None
+    project_uncertain_threshold: Optional[float] = None
+
+
+def _routing_payload(data: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "cold_start_enabled": bool(data["cold_start_enabled"]),
+        "signal_auto_enter_threshold": round(float(data["signal_auto_enter_threshold"]), 2),
+        "project_auto_create_threshold": round(float(data["project_auto_create_threshold"]), 2),
+        "project_uncertain_threshold": round(float(data["project_uncertain_threshold"]), 2),
+        "tips": {
+            "cold_start_enabled": "开启后，所有飞书消息进入待处理，不会自动入库或自动建项。",
+            "signal_auto_enter_threshold": "建议范围 0.75-0.90，数值越高越保守。",
+            "project_auto_create_threshold": "建议范围 0.80-0.95，数值越高越严格。",
+            "project_uncertain_threshold": "低于该值视为低置信度，进入人工待处理。",
+        },
+    }
+
+
+@router.get("/projects/{project_id}/ai-routing-config")
+async def get_ai_routing_config(
+    project_id: str,
+    current_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+
+    integration = (
+        await db.execute(
+            select(Integration).where(
+                and_(
+                    Integration.tenant_id == tenant_id,
+                    Integration.type == "feishu",
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    return success_response(_routing_payload(_extract_feishu_routing_config(integration)))
+
+
+@router.put("/projects/{project_id}/ai-routing-config")
+async def update_ai_routing_config(
+    project_id: str,
+    request: AIRoutingConfigUpdateRequest,
+    current_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+):
+    tenant_id = int(current_user.tenant_id)
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+
+    integration = (
+        await db.execute(
+            select(Integration).where(
+                and_(
+                    Integration.tenant_id == tenant_id,
+                    Integration.type == "feishu",
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    if integration is None:
+        integration = Integration(
+            tenant_id=tenant_id,
+            type="feishu",
+            config={},
+            is_active=True,
+        )
+        db.add(integration)
+        await db.flush()
+
+    current = _extract_feishu_routing_config(integration)
+    config = dict(current["raw"])
+
+    if request.cold_start_enabled is not None:
+        enabled = bool(request.cold_start_enabled)
+        config["cold_start_enabled"] = enabled
+
+    if request.signal_auto_enter_threshold is not None:
+        config["signal_auto_enter_threshold"] = _safe_threshold(
+            request.signal_auto_enter_threshold,
+            current["signal_auto_enter_threshold"],
+        )
+
+    if request.project_auto_create_threshold is not None:
+        config["project_auto_create_threshold"] = _safe_threshold(
+            request.project_auto_create_threshold,
+            current["project_auto_create_threshold"],
+        )
+
+    if request.project_uncertain_threshold is not None:
+        config["project_uncertain_threshold"] = _safe_threshold(
+            request.project_uncertain_threshold,
+            current["project_uncertain_threshold"],
+        )
+
+    project_threshold = _safe_threshold(
+        config.get("project_auto_create_threshold"),
+        current["project_auto_create_threshold"],
+    )
+    uncertain_threshold = min(
+        project_threshold,
+        _safe_threshold(
+            config.get("project_uncertain_threshold"),
+            current["project_uncertain_threshold"],
+        ),
+    )
+    config["project_auto_create_threshold"] = project_threshold
+    config["project_uncertain_threshold"] = uncertain_threshold
+
+    integration.config = config
+    await db.commit()
+    await db.refresh(integration)
+    return success_response(_routing_payload(_extract_feishu_routing_config(integration)))
+
+
+async def _ensure_task(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    task_id: int,
+) -> Task:
+    result = await db.execute(
+        select(Task).where(
+            and_(
+                Task.id == task_id,
+                Task.project_id == project_id,
+                Task.tenant_id == tenant_id,
+            )
+        )
+    )
+    task = result.scalar_one_or_none()
+    if not task:
+        raise HTTPException(status_code=404, detail="Task not found")
+    return task
+
+
+async def _ensure_release(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    release_node_id: int,
+) -> TruthNode:
+    result = await db.execute(
+        select(TruthNode).where(
+            and_(
+                TruthNode.id == release_node_id,
+                TruthNode.project_id == project_id,
+                TruthNode.tenant_id == tenant_id,
+                TruthNode.node_type == "release",
+            )
+        )
+    )
+    release_node = result.scalar_one_or_none()
+    if not release_node:
+        raise HTTPException(status_code=404, detail="Release plan not found")
+    return release_node
+
+
+async def _ensure_review(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    review_node_id: int,
+) -> TruthNode:
+    result = await db.execute(
+        select(TruthNode).where(
+            and_(
+                TruthNode.id == review_node_id,
+                TruthNode.project_id == project_id,
+                TruthNode.tenant_id == tenant_id,
+                TruthNode.node_type == "review",
+            )
+        )
+    )
+    review_node = result.scalar_one_or_none()
+    if not review_node:
+        raise HTTPException(status_code=404, detail="Review not found")
+    return review_node
+
+
+def _normalize_review_content(raw: dict[str, Any] | None) -> dict[str, Any]:
+    payload = dict(raw or {})
+    summary = str(payload.get("summary") or "").strip()
+    goals = payload.get("goals")
+    if not isinstance(goals, list):
+        goals = []
+    what_went_well = payload.get("what_went_well")
+    if not isinstance(what_went_well, list):
+        what_went_well = []
+    root_causes = payload.get("root_causes")
+    if not isinstance(root_causes, list):
+        root_causes = []
+    improvements = payload.get("improvements")
+    if not isinstance(improvements, list):
+        improvements = []
+
+    normalized_improvements: list[dict[str, Any]] = []
+    def _optional_int(value: Any) -> int | None:
+        if value in (None, ""):
+            return None
+        try:
+            return int(value)
+        except (TypeError, ValueError):
+            return None
+
+    for idx, item in enumerate(improvements):
+        if isinstance(item, str):
+            title = item.strip()
+            if not title:
+                continue
+            normalized_improvements.append(
+                {
+                    "id": f"imp-{idx+1}",
+                    "title": title,
+                    "status": "pending",
+                    "owner_id": None,
+                    "due_date": None,
+                    "task_id": None,
+                }
+            )
+            continue
+        if not isinstance(item, dict):
+            continue
+        title = str(item.get("title") or "").strip()
+        if not title:
+            continue
+        normalized_improvements.append(
+            {
+                "id": str(item.get("id") or f"imp-{idx+1}"),
+                "title": title,
+                "status": str(item.get("status") or "pending"),
+                "owner_id": _optional_int(item.get("owner_id")),
+                "due_date": item.get("due_date"),
+                "task_id": _optional_int(item.get("task_id")),
+                "description": str(item.get("description") or "").strip() or None,
+            }
+        )
+
+    return {
+        "summary": summary,
+        "goals": [str(x).strip() for x in goals if str(x).strip()],
+        "what_went_well": [str(x).strip() for x in what_went_well if str(x).strip()],
+        "root_causes": [str(x).strip() for x in root_causes if str(x).strip()],
+        "improvements": normalized_improvements,
+        "last_updated_at": datetime.now(timezone.utc).isoformat(),
+    }
+
+
+async def _project_task_metrics(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+) -> dict[str, int]:
+    result = await db.execute(
+        select(Task.status).where(
+            and_(
+                Task.project_id == project_id,
+                Task.tenant_id == tenant_id,
+            )
+        )
+    )
+    statuses = [str(row[0] or "").lower() for row in result.all()]
+    total = len(statuses)
+    done = len([s for s in statuses if s in DONE_STATUSES])
+    return {"total": total, "done": done}
+
+
+async def _build_governance_snapshot(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    charter_id: int | None = None,
+) -> dict[str, Any]:
+    charter: TruthNode | None = None
+    if charter_id:
+        charter = (
+            await db.execute(
+                select(TruthNode).where(
+                    and_(
+                        TruthNode.id == charter_id,
+                        TruthNode.project_id == project_id,
+                        TruthNode.tenant_id == tenant_id,
+                        TruthNode.node_type == "charter",
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+    if charter is None:
+        charter = (
+            await db.execute(
+                select(TruthNode)
+                .where(
+                    and_(
+                        TruthNode.project_id == project_id,
+                        TruthNode.tenant_id == tenant_id,
+                        TruthNode.node_type == "charter",
+                    )
+                )
+                .order_by(TruthNode.created_at.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+
+    confidence_score = None
+    confidence_risk = False
+    charter_id_value = int(charter.id) if charter else None
+    if charter:
+        charter_content = _as_dict(charter.content)
+        metadata = charter_content.get("metadata")
+        if isinstance(metadata, dict):
+            raw_score = metadata.get("confidence_score")
+            try:
+                confidence_score = float(raw_score)
+            except (TypeError, ValueError):
+                confidence_score = None
+        confidence_risk = confidence_score is not None and confidence_score < 0.8
+
+    binding = None
+    if charter_id_value:
+        binding = (
+            await db.execute(
+                select(ModelRequirementBinding, ModelVersion)
+                .join(
+                    ModelVersion,
+                    ModelVersion.id == ModelRequirementBinding.model_version_id,
+                    isouter=True,
+                )
+                .where(
+                    and_(
+                        ModelRequirementBinding.tenant_id == tenant_id,
+                        ModelRequirementBinding.truth_node_id == charter_id_value,
+                    )
+                )
+                .order_by(ModelRequirementBinding.created_at.desc())
+                .limit(1)
+            )
+        ).first()
+
+    model_block = False
+    model_info: dict[str, Any] = {
+        "bound": False,
+        "compatibility_score": None,
+        "model_version_id": None,
+        "model_name": None,
+        "provider": None,
+    }
+    est_cost = 0.01
+    if binding:
+        binding_row, model_row = binding
+        score = int(binding_row.compatibility_score or 0)
+        model_block = score < 70
+        model_info = {
+            "bound": True,
+            "compatibility_score": score,
+            "model_version_id": int(binding_row.model_version_id),
+            "model_name": getattr(model_row, "name", None),
+            "provider": getattr(model_row, "provider", None),
+        }
+        try:
+            est_cost = max(float(binding_row.estimated_cost or 0.01), 0.01)
+        except (TypeError, ValueError):
+            est_cost = 0.01
+
+    budget_service = BudgetEnforcementService(db)
+    budget_check = await budget_service.check_budget_before_ai_call(
+        tenant_id=tenant_id,
+        estimated_cost_usd=est_cost,
+        project_id=project_id,
+    )
+
+    budget_block = not bool(budget_check.get("allowed", True))
+    intervention_required = confidence_risk or model_block or budget_block
+    blockers: list[str] = []
+    warnings: list[str] = []
+    if confidence_risk:
+        warnings.append("宪章 AI 置信度低于 0.80，建议补充关键事实后再推进。")
+    if model_block:
+        blockers.append("模型匹配分低于 70，禁止自动推进发布验证。")
+    elif not model_info["bound"]:
+        warnings.append("当前宪章未绑定模型版本，建议先完成模型能力矩阵校验。")
+    if budget_block:
+        blockers.append(f"预算管控阻断：{budget_check.get('reason') or 'budget_blocked'}")
+    elif str(budget_check.get("reason") or "") == "within_budget":
+        usage = budget_check.get("usage_percent")
+        if isinstance(usage, (int, float)) and float(usage) >= 80:
+            warnings.append("预算已接近阈值，建议执行成本优化方案。")
+
+    return {
+        "project_id": project_id,
+        "charter_id": charter_id_value,
+        "intervention_required": intervention_required,
+        "blockers": blockers,
+        "warnings": warnings,
+        "confidence": {
+            "score": confidence_score,
+            "risk": confidence_risk,
+        },
+        "model_binding": model_info,
+        "budget": budget_check,
+    }
+
+
+async def _assert_release_transition_allowed(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    project_id: int,
+    release_node: TruthNode,
+    target_status: str,
+) -> None:
+    current = str(release_node.status or "pending").lower()
+    target = str(target_status or "").lower()
+    if target == "verified":
+        raise HTTPException(status_code=409, detail="请通过验证完成接口标记为已验证")
+    if target not in RELEASE_STATUS_TRANSITIONS.get(current, set()):
+        raise HTTPException(
+            status_code=409,
+            detail=f"非法状态流转：{current} -> {target}",
+        )
+
+    content = _as_dict(release_node.content)
+    if target in {"releasing", "online"}:
+        task_ids = [int(tid) for tid in (content.get("task_ids") or []) if str(tid).isdigit()]
+        if not task_ids:
+            raise HTTPException(status_code=409, detail="发布计划未绑定任务")
+        task_result = await db.execute(
+            select(Task.id, Task.status).where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.project_id == project_id,
+                    Task.id.in_(task_ids),
+                )
+            )
+        )
+        task_rows = list(task_result.all())
+        if len(task_rows) != len(task_ids):
+            raise HTTPException(status_code=409, detail="发布计划存在无效任务绑定")
+        not_done = [row for row in task_rows if str(row[1] or "").lower() not in DONE_STATUSES]
+        if not_done:
+            raise HTTPException(status_code=409, detail="存在未完成任务，禁止进入发布流程")
+
+        governance = await _build_governance_snapshot(
+            db,
+            tenant_id=tenant_id,
+            project_id=project_id,
+            charter_id=int(content.get("charter_id") or 0) or None,
+        )
+        if governance.get("blockers"):
+            raise HTTPException(
+                status_code=409,
+                detail="；".join([str(item) for item in governance.get("blockers", [])]),
+            )
+
+        verification_items = content.get("verification_items")
+        if not isinstance(verification_items, list) or len(verification_items) == 0:
+            raise HTTPException(status_code=409, detail="请先生成发布验证清单")
+
+def _assert_charter_aligned(charter: TruthNode) -> None:
+    if (charter.status or "").lower() not in ALIGNED_STATUSES:
+        raise HTTPException(
+            status_code=409,
+            detail="宪章未对齐，任务环节未解锁",
+        )
+
+
+def _is_manager(user: User) -> bool:
+    return has_any_role(str(user.role or ""), ["BOSS", "ADMIN"])
+
+
+def _require_manager(user: User) -> None:
+    if not _is_manager(user):
+        raise HTTPException(status_code=403, detail="仅项目负责人或管理员可执行该操作")
+
+
+@router.get("/projects/{project_id}/sstc/overview")
+async def get_board_overview(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    overview = await svc.build_overview(pid)
+    stages = await svc.build_stages(pid)
+    risks = await svc.build_risks(pid)
+    return success_response(
+        {
+            "overview": overview,
+            "stages": stages,
+            "risks": risks,
+        }
+    )
+
+
+@router.get("/projects/{project_id}/sstc/stages")
+async def get_stages(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    return success_response(await svc.build_stages(pid))
+
+
+@router.get("/projects/{project_id}/sstc/risks")
+async def get_risks(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    return success_response(await svc.build_risks(pid))
+
+
+@router.get("/projects/{project_id}/sstc/governance")
+async def get_governance_snapshot(
+    project_id: str,
+    charter_id: Optional[int] = Query(default=None),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+    data = await _build_governance_snapshot(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        charter_id=charter_id,
+    )
+    return success_response(data)
+
+
+@router.get("/projects/{project_id}/sstc-chain")
+async def get_sstc_chain(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    return success_response(await svc.build_chain(pid))
+
+
+@router.get("/projects/{project_id}/signals")
+async def list_signals(
+    project_id: str,
+    status: Optional[str] = Query(default=None, description="pending/linked/closed/archived"),
+    keyword: Optional[str] = Query(default=None, description="按信号标题/内容/ID/提报人/关联宪章搜索"),
+    source: Optional[str] = Query(default=None, description="来源过滤"),
+    priority: Optional[str] = Query(default=None, description="优先级过滤"),
+    created_from: Optional[str] = Query(default=None, description="创建时间起始（北京时间 YYYY-MM-DD）"),
+    created_to: Optional[str] = Query(default=None, description="创建时间结束（北京时间 YYYY-MM-DD）"),
+    limit: int = Query(default=10, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    project = await _ensure_project(svc, pid)
+    project_creator_name = "-"
+    if project and getattr(project, "owner_id", None):
+        owner = (
+            await db.execute(
+                select(User.name).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id == int(project.owner_id),
+                    )
+                ).limit(1)
+            )
+        ).scalars().first()
+        if owner:
+            project_creator_name = str(owner)
+    signals = await svc.get_signals(pid)
+    await _backfill_feishu_sender_names(
+        db,
+        tenant_id=tenant_id,
+        signals=signals,
+    )
+    charters = await svc.get_charters(pid)
+    linked_signal_states: dict[int, str] = {}
+    linked_signal_charter_name: dict[int, str] = {}
+    for c in charters:
+        signal_ids = _extract_charter_related_signal_ids(c)
+        charter_status = (c.status or "").lower()
+        state = (
+            _SIGNAL_STATUS_CODE_CLOSED
+            if charter_status in {"aligned", "active", "online", "archived", "已对齐", "已上线", "已归档"}
+            else _SIGNAL_STATUS_CODE_LINKED
+        )
+        for sid in signal_ids:
+            if linked_signal_states.get(sid) == _SIGNAL_STATUS_CODE_CLOSED:
+                continue
+            linked_signal_states[sid] = state
+            linked_signal_charter_name[sid] = c.title
+    normalized_status_filter = _normalize_signal_status_filter(status)
+    created_from_utc, created_to_utc = _build_signal_created_at_range(
+        created_from=created_from,
+        created_to=created_to,
+    )
+    rows = []
+    for s in signals:
+        ai_analysis = _signal_ai_analysis(s)
+        source_context = _signal_source_context(s)
+        project_scope = _signal_project_scope(s)
+        proposer_name = _signal_proposer_name(s)
+        title_text = _signal_title_text(s)
+        normalized_status = str(s.status or "").strip()
+        sid = int(s.id)
+        status_code = _normalize_signal_status_code(
+            linked_state=linked_signal_states.get(sid),
+            signal_status=normalized_status,
+            processed=bool(s.processed),
+        )
+        row_status = _signal_status_label_from_code(status_code, normalized_status)
+        locked_by_charter = sid in linked_signal_states or ("关联" in normalized_status)
+        can_create_charter = not locked_by_charter
+        charter_lock_reason = None if can_create_charter else _SIGNAL_LOCK_REASON
+        row = {
+            "id": str(s.id),
+            "title": title_text[:80] or "未命名信号",
+            "content": s.content,
+            "priority": s.priority,
+            "signal_type": _signal_display_type(s),
+            "source": s.source,
+            "project_creator_name": project_creator_name,
+            "proposer_name": proposer_name,
+            "proposer_open_id": source_context.get("sender_open_id"),
+            "project_scope": project_scope,
+            "scope_reason": ai_analysis.get("reason"),
+            "target_project_hints": ai_analysis.get("target_project_hints", []),
+            "status": row_status,
+            "status_code": status_code,
+            "can_create_charter": can_create_charter,
+            "charter_lock_reason": charter_lock_reason,
+            "charter_name": linked_signal_charter_name.get(sid),
+            "created_at": s.created_at.isoformat() if s.created_at else None,
+            "attachments": _extract_signal_attachments(s),
+        }
+        created_at_dt = s.created_at
+        if created_at_dt is not None and created_at_dt.tzinfo is not None:
+            created_at_dt = created_at_dt.astimezone(timezone.utc).replace(tzinfo=None)
+        if not _match_signal_filters(
+            row=row,
+            status_filter=normalized_status_filter,
+            keyword=keyword,
+            source=source,
+            priority=priority,
+            created_from_utc=created_from_utc,
+            created_to_utc=created_to_utc,
+            created_at_dt=created_at_dt,
+        ):
+            continue
+        rows.append(row)
+    total = len(rows)
+    paged_rows = rows[offset : offset + limit]
+    return success_response(
+        {
+            "items": paged_rows,
+            "total": total,
+            "limit": limit,
+            "offset": offset,
+            "has_more": (offset + len(paged_rows)) < total,
+        }
+    )
+
+
+@router.post("/projects/{project_id}/signals/{signal_id}/attachments")
+async def upload_signal_attachment(
+    project_id: str,
+    signal_id: str,
+    file: UploadFile | None = File(default=None),
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    try:
+        pid = _parse_project_id(project_id)
+    except HTTPException:
+        return _error_json(status_code=400, message="project_id 不合法")
+    try:
+        sid = _parse_node_id(signal_id, "signal_id")
+    except HTTPException:
+        return _error_json(status_code=400, message="signal_id 不合法")
+    if file is None:
+        return _error_json(status_code=400, message="请选择要上传的附件")
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+
+    project = await svc.get_project(pid)
+    if project is None:
+        return _error_json(status_code=404, message="项目不存在")
+
+    signals = await svc.get_signals(pid)
+    signal = next((item for item in signals if int(item.id) == sid), None)
+    if signal is None:
+        return _error_json(status_code=404, message="信号不存在或不属于当前项目")
+
+    original_file_name = Path(str(file.filename or "")).name.strip()
+    if not original_file_name:
+        return _error_json(status_code=400, message="文件名不能为空")
+
+    extension = Path(original_file_name).suffix.lower()
+    attachment_type = _SIGNAL_ATTACHMENT_ALLOWED_EXTENSIONS.get(extension)
+    if attachment_type is None:
+        return _error_json(status_code=400, message="仅支持 jpg/png/gif/pdf/doc/docx/zip 文件")
+
+    mime_type = str(file.content_type or "").strip().lower() or None
+    if mime_type and mime_type not in _SIGNAL_ATTACHMENT_ALLOWED_MIME_TYPES:
+        return _error_json(status_code=400, message="文件类型不支持，请上传 jpg/png/gif/pdf/doc/docx/zip")
+
+    try:
+        content = await file.read()
+    except Exception:
+        return _error_json(status_code=500, message="读取上传文件失败")
+
+    if not content:
+        return _error_json(status_code=400, message="上传文件为空")
+    if len(content) > _SIGNAL_ATTACHMENT_MAX_BYTES:
+        return _error_json(status_code=400, message="单文件大小不能超过 20MB")
+
+    attachment_id = uuid4().hex
+    relative_dir = Path(f"tenant-{tenant_id}") / f"project-{pid}" / f"signal-{sid}"
+    storage_root = _signal_attachment_storage_root()
+    save_dir = storage_root / relative_dir
+    save_dir.mkdir(parents=True, exist_ok=True)
+    saved_file_name = f"{attachment_id}{extension}"
+    saved_path = save_dir / saved_file_name
+    try:
+        saved_path.write_bytes(content)
+    except Exception:
+        return _error_json(status_code=500, message="保存附件失败，请稍后重试")
+
+    file_key = f"local://{(relative_dir / saved_file_name).as_posix()}"
+    attachment_payload = {
+        "id": attachment_id,
+        "attachment_id": attachment_id,
+        "attachment_type": attachment_type,
+        "file_name": original_file_name,
+        "mime_type": mime_type or "application/octet-stream",
+        "file_size": len(content),
+        "file_key": file_key,
+        "image_key": None,
+        "media_id": None,
+        "source": "manual_upload",
+        "occurred_at": datetime.now(timezone.utc).isoformat(),
+        "download_url": _build_signal_attachment_download_url(pid, sid, attachment_id),
+    }
+
+    raw_payload = _as_dict(signal.strategic_analysis)
+    source_context = raw_payload.get("source_context")
+    if not isinstance(source_context, dict):
+        source_context = {}
+    attachments = source_context.get("attachments")
+    if not isinstance(attachments, list):
+        attachments = []
+    attachments.append(attachment_payload)
+    source_context["attachments"] = attachments
+    raw_payload["source_context"] = source_context
+    signal.strategic_analysis = json.dumps(raw_payload, ensure_ascii=False)
+
+    try:
+        await db.commit()
+        await db.refresh(signal)
+    except Exception:
+        await db.rollback()
+        try:
+            saved_path.unlink()
+        except Exception:
+            pass
+        return _error_json(status_code=500, message="上传附件失败，请稍后重试")
+
+    return success_response(attachment_payload)
+
+
+@router.get("/projects/{project_id}/signals/{signal_id}/attachments/{attachment_id}/download")
+async def download_signal_attachment(
+    project_id: str,
+    signal_id: str,
+    attachment_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    sid = _parse_node_id(signal_id, "signal_id")
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+    signals = await svc.get_signals(pid)
+    signal = next((item for item in signals if int(item.id) == sid), None)
+    if signal is None:
+        raise HTTPException(status_code=404, detail="Signal not found")
+    attachments = _extract_signal_attachments(signal)
+    target = next(
+        (
+            item
+            for item in attachments
+            if str(item.get("attachment_id") or "").strip() == attachment_id
+            or str(item.get("id") or "").strip() == attachment_id
+        ),
+        None,
+    )
+    if target is None:
+        raise HTTPException(status_code=404, detail="Attachment not found")
+
+    local_path = _resolve_local_attachment_path(str(target.get("file_key") or ""))
+    if local_path is not None:
+        if not local_path.exists() or not local_path.is_file():
+            raise HTTPException(status_code=404, detail="附件文件不存在")
+        try:
+            content = local_path.read_bytes()
+        except Exception as exc:
+            raise HTTPException(status_code=500, detail=f"附件读取失败: {str(exc)}") from exc
+        if not content:
+            raise HTTPException(status_code=500, detail="附件内容为空")
+        file_name = str(target.get("file_name") or f"attachment-{attachment_id}").strip().replace('"', "'")
+        media_type = str(target.get("mime_type") or "application/octet-stream")
+        return StreamingResponse(
+            BytesIO(content),
+            media_type=media_type,
+            headers={"Content-Disposition": f'attachment; filename="{file_name}"'},
+        )
+
+    feishu_service = get_feishu_service(db)
+    await feishu_service.load_config(tenant_id)
+    if not feishu_service.is_configured():
+        raise HTTPException(status_code=503, detail="飞书集成未配置，暂无法下载附件")
+
+    try:
+        payload = await feishu_service.download_attachment_bytes(target)
+    except ValueError as exc:
+        raise HTTPException(status_code=502, detail=str(exc)) from exc
+    except Exception as exc:
+        raise HTTPException(status_code=500, detail=f"附件下载失败: {str(exc)}") from exc
+
+    file_name = str(payload.get("file_name") or target.get("file_name") or f"attachment-{attachment_id}").strip()
+    file_name = file_name.replace('"', "'")
+    media_type = str(payload.get("content_type") or target.get("mime_type") or "application/octet-stream")
+    content = payload.get("content")
+    if not isinstance(content, (bytes, bytearray)):
+        raise HTTPException(status_code=500, detail="附件内容为空")
+    return StreamingResponse(
+        BytesIO(bytes(content)),
+        media_type=media_type,
+        headers={"Content-Disposition": f'attachment; filename="{file_name}"'},
+    )
+
+
+@router.post("/projects/{project_id}/signals/{signal_id}/archive")
+async def archive_signal(
+    project_id: str,
+    signal_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    sid = _parse_node_id(signal_id, "signal_id")
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    result = await db.execute(
+        select(Signal).where(
+            and_(Signal.id == sid, Signal.tenant_id == int(current_user.tenant_id))
+        )
+    )
+    signal = result.scalar_one_or_none()
+    if not signal:
+        raise HTTPException(status_code=404, detail="Signal not found")
+    before = {"processed": signal.processed}
+    signal.processed = True
+    signal.status = "已关闭"
+    await log_operation(
+        db,
+        tenant_id=int(current_user.tenant_id),
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="signal",
+        entity_id=str(sid),
+        action="archive",
+        before_data=before,
+        after_data={"processed": True},
+    )
+    await db.commit()
+    return success_response({"id": str(signal.id), "status": "已关闭"})
+
+
+class CharterCreateRequest(BaseModel):
+    title: str
+    content: dict[str, Any] | None = None
+    signal_ids: list[int]
+    skip_ai_prefill: bool = False
+
+
+class CharterAIPrefillRequest(BaseModel):
+    force: bool = False
+    raw_text: Optional[str] = None
+
+
+@router.get("/projects/{project_id}/charters")
+async def list_charters(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    charters = await svc.get_charters(pid)
+    payload = [
+        {
+            "id": str(c.id),
+            "title": c.title,
+            "content": _as_dict(c.content),
+            "signal_ids": _as_dict(c.content).get("signal_ids", []),
+            "status": c.status,
+            "version": c.version,
+            "parent_id": str(c.parent_id) if c.parent_id else None,
+            "created_at": c.created_at.isoformat() if c.created_at else None,
+        }
+        for c in charters
+    ]
+    return success_response(payload)
+
+
+@router.post("/projects/{project_id}/charters")
+async def create_charter(
+    project_id: str,
+    body: CharterCreateRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    if not body.signal_ids:
+        raise HTTPException(status_code=400, detail="宪章必须关联至少一个信号")
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+
+    link_result = await db.execute(
+        select(ProjectSignalLink.signal_id).where(
+            and_(
+                ProjectSignalLink.project_id == pid,
+                ProjectSignalLink.signal_id.in_(body.signal_ids),
+                ProjectSignalLink.tenant_id == tenant_id,
+            )
+        )
+    )
+    valid_signal_ids = {int(v) for v in link_result.scalars().all()}
+    if set(body.signal_ids) - valid_signal_ids:
+        raise HTTPException(status_code=400, detail="部分信号未关联到当前项目")
+
+    signal_ids = sorted(set(body.signal_ids))
+    existing_charters = await svc.get_charters(pid)
+    conflict_signal_ids, conflict_charters = _collect_locked_signal_conflicts(
+        charters=existing_charters,
+        candidate_signal_ids=signal_ids,
+    )
+    if conflict_signal_ids:
+        raise HTTPException(
+            status_code=409,
+            detail={
+                "message": _SIGNAL_LOCK_REASON,
+                "conflict_signal_ids": conflict_signal_ids,
+                "conflict_charters": conflict_charters,
+            },
+        )
+    signal_rows = (
+        await db.execute(
+            select(Signal).where(
+                and_(
+                    Signal.tenant_id == tenant_id,
+                    Signal.id.in_(signal_ids),
+                )
+            )
+        )
+    ).scalars().all()
+
+    content = dict(body.content or {})
+    raw_text = str(content.get("raw_source_text") or "").strip()
+    if not body.skip_ai_prefill and not _has_non_empty_module_data(content):
+        ai_prefill = await _build_ai_first_charter_content(
+            db=db,
+            tenant_id=tenant_id,
+            project_id=pid,
+            signal_rows=list(signal_rows),
+            signal_ids=signal_ids,
+            raw_text=raw_text,
+        )
+        content.update(ai_prefill)
+    content["signal_ids"] = signal_ids
+    primary_signal_id = content["signal_ids"][0] if content["signal_ids"] else None
+    charter = TruthNode(
+        node_type="charter",
+        parent_id=None,
+        project_id=pid,
+        signal_id=primary_signal_id,
+        created_by=int(current_user.id),
+        title=compact_title(body.title, max_len=24),
+        content=content,
+        version=1,
+        status="draft",
+        tenant_id=tenant_id,
+    )
+    db.add(charter)
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="charter",
+        entity_id="new",
+        action="create",
+        after_data={"title": charter.title, "signal_ids": content["signal_ids"]},
+    )
+
+    for signal in signal_rows:
+        signal.status = "已关联宪章"
+        signal.processed = False
+
+    await db.commit()
+    await db.refresh(charter)
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="charter",
+        record_id=int(charter.id),
+        project_id=pid,
+        title=charter.title,
+        content=_as_dict(charter.content),
+    )
+    return success_response(
+        {
+            "id": str(charter.id),
+            "title": charter.title,
+            "content": _as_dict(charter.content),
+            "signal_ids": _as_dict(charter.content).get("signal_ids", []),
+            "status": charter.status,
+            "version": charter.version,
+        }
+    )
+
+
+@router.get("/projects/{project_id}/charters/{charter_id}")
+async def get_charter_detail(
+    project_id: str,
+    charter_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+
+    content = _as_dict(charter.content)
+    signal_ids = [sid for sid in content.get("signal_ids", []) if isinstance(sid, int)]
+    signal_rows = []
+    if signal_ids:
+        signal_result = await db.execute(
+            select(Signal).where(
+                and_(Signal.id.in_(signal_ids), Signal.tenant_id == tenant_id)
+            )
+        )
+        signal_rows = list(signal_result.scalars().all())
+        await _backfill_feishu_sender_names(
+            db,
+            tenant_id=tenant_id,
+            signals=signal_rows,
+        )
+
+    task_result = await db.execute(
+        select(Task).where(
+            and_(
+                Task.truth_node_id == cid,
+                Task.project_id == pid,
+                Task.tenant_id == tenant_id,
+            )
+        )
+    )
+    tasks = list(task_result.scalars().all())
+
+    align_result = await db.execute(
+        select(AlignmentRecord).where(
+            and_(
+                AlignmentRecord.truth_node_id == cid,
+                AlignmentRecord.tenant_id == tenant_id,
+            )
+        )
+    )
+    alignments = list(align_result.scalars().all())
+    source_signal = None
+    if charter.signal_id:
+        for signal in signal_rows:
+            if int(signal.id) == int(charter.signal_id):
+                source_signal = signal
+                break
+    if source_signal is None and signal_rows:
+        source_signal = signal_rows[0]
+
+    return success_response(
+        {
+            "id": str(charter.id),
+            "title": charter.title,
+            "content": content,
+            "signal_ids": signal_ids,
+            "signals": [
+                {
+                    "id": str(s.id),
+                    "source": s.source,
+                    "signal_type": _signal_display_type(s),
+                    "project_scope": _signal_project_scope(s),
+                    "proposer_name": _signal_proposer_name(s),
+                    "title": _signal_title_text(s)[:80] or "未命名信号",
+                    "content": s.content,
+                    "priority": s.priority,
+                    "impact_score": s.impact_score,
+                    "attachments": _extract_signal_attachments(s),
+                }
+                for s in signal_rows
+            ],
+            "source_signal": (
+                {
+                    "id": str(source_signal.id),
+                    "title": _signal_title_text(source_signal)[:80] or "未命名信号",
+                    "signal_type": _signal_display_type(source_signal),
+                    "priority": source_signal.priority,
+                    "content": source_signal.content,
+                    "attachments": _extract_signal_attachments(source_signal),
+                }
+                if source_signal
+                else None
+            ),
+            "tasks": [
+                {
+                    "id": str(t.id),
+                    "title": t.title,
+                    "description": t.description,
+                    "status": t.status,
+                    "priority": t.priority,
+                    "estimated_hours": float(t.estimated_hours) if t.estimated_hours else None,
+                    "assignee_id": str(t.assignee_id) if t.assignee_id else None,
+                    "due_date": t.due_date.isoformat() if t.due_date else None,
+                }
+                for t in tasks
+            ],
+            "alignment_records": [
+                {
+                    "user_id": str(r.user_id),
+                    "score": r.alignment_score,
+                    "confirmed": bool(r.is_confirmed),
+                    "comments": r.comments,
+                }
+                for r in alignments
+            ],
+            "status": charter.status,
+            "version": charter.version,
+        }
+    )
+
+
+@router.patch("/projects/{project_id}/charters/{charter_id}")
+async def update_charter(
+    project_id: str,
+    charter_id: str,
+    body: CharterUpdateRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+
+    before = {
+        "title": charter.title,
+        "content": _as_dict(charter.content),
+        "version": charter.version,
+        "status": charter.status,
+    }
+    aligned = (charter.status or "").lower() in ALIGNED_STATUSES
+    if aligned and body.force_new_version:
+        new_node = TruthNode(
+            node_type="charter",
+            parent_id=int(charter.id),
+            project_id=pid,
+            created_by=int(current_user.id),
+            title=compact_title((body.title or charter.title), max_len=24),
+            content=body.content if body.content is not None else _as_dict(charter.content),
+            version=int(charter.version or 1) + 1,
+            status="draft",
+            tenant_id=tenant_id,
+        )
+        db.add(new_node)
+        await log_operation(
+            db,
+            tenant_id=tenant_id,
+            user_id=int(current_user.id),
+            project_id=pid,
+            entity_type="charter",
+            entity_id=str(charter.id),
+            action="new_version",
+            before_data=before,
+            after_data={"new_charter_id": "pending"},
+        )
+        await db.commit()
+        await db.refresh(new_node)
+        _sync_pmflow_rag_async(
+            tenant_id=tenant_id,
+            record_type="charter",
+            record_id=int(new_node.id),
+            project_id=pid,
+            title=new_node.title,
+            content=_as_dict(new_node.content),
+        )
+        return success_response(
+            {
+                "id": str(new_node.id),
+                "version": new_node.version,
+                "status": new_node.status,
+                "title": new_node.title,
+                "content": _as_dict(new_node.content),
+                "new_version_created": True,
+            }
+        )
+
+    if body.title is not None:
+        charter.title = compact_title(body.title, max_len=24)
+    if body.content is not None:
+        charter.content = body.content
+    charter.status = "draft"
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="charter",
+        entity_id=str(charter.id),
+        action="update",
+        before_data=before,
+        after_data={
+            "title": charter.title,
+            "content": _as_dict(charter.content),
+            "version": charter.version,
+            "status": charter.status,
+        },
+    )
+    await db.commit()
+    await db.refresh(charter)
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="charter",
+        record_id=int(charter.id),
+        project_id=pid,
+        title=charter.title,
+        content=_as_dict(charter.content),
+    )
+    return success_response(
+        {
+            "id": str(charter.id),
+            "title": charter.title,
+            "content": _as_dict(charter.content),
+            "status": charter.status,
+            "version": charter.version,
+            "new_version_created": False,
+        }
+    )
+
+
+@router.post("/projects/{project_id}/charters/{charter_id}/ai-prefill")
+async def ai_prefill_charter(
+    project_id: str,
+    charter_id: str,
+    body: CharterAIPrefillRequest = Body(default_factory=CharterAIPrefillRequest),
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+    before_content = _as_dict(charter.content)
+    before_status = str(charter.status or "")
+
+    if not body.force and _has_complete_module_data(before_content):
+        return success_response(
+            {
+                "id": str(charter.id),
+                "title": charter.title,
+                "content": before_content,
+                "status": charter.status,
+                "version": charter.version,
+                "prefilled": False,
+                "skipped": True,
+            }
+        )
+
+    signal_ids = _extract_charter_signal_ids(before_content, charter.signal_id)
+    signal_rows: list[Signal] = []
+    if signal_ids:
+        signal_rows = (
+            await db.execute(
+                select(Signal).where(
+                    and_(
+                        Signal.tenant_id == tenant_id,
+                        Signal.id.in_(signal_ids),
+                    )
+                )
+            )
+        ).scalars().all()
+
+    raw_text = str(body.raw_text or before_content.get("raw_source_text") or "").strip()
+    if not signal_rows and not raw_text:
+        raise HTTPException(status_code=400, detail="未找到可用于预填的信号源或原始文本")
+
+    generated = await _build_ai_first_charter_content(
+        db=db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        signal_rows=signal_rows,
+        signal_ids=signal_ids,
+        raw_text=raw_text,
+    )
+
+    merged = _merge_prefill_content(
+        before_content, generated, overwrite_existing=bool(body.force)
+    )
+    merged["signal_ids"] = signal_ids
+    if raw_text:
+        merged["raw_source_text"] = raw_text
+    charter.content = merged
+    charter.status = "draft"
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="charter",
+        entity_id=str(cid),
+        action="ai_prefill",
+        before_data={"content": before_content, "status": before_status},
+        after_data={"content": merged, "status": "draft"},
+    )
+
+    await db.commit()
+    await db.refresh(charter)
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="charter",
+        record_id=int(charter.id),
+        project_id=pid,
+        title=charter.title,
+        content=_as_dict(charter.content),
+    )
+    return success_response(
+        {
+            "id": str(charter.id),
+            "title": charter.title,
+            "content": _as_dict(charter.content),
+            "status": charter.status,
+            "version": charter.version,
+            "prefilled": True,
+            "skipped": False,
+        }
+    )
+
+
+@router.post("/projects/{project_id}/charters/{charter_id}/alignment/start")
+async def start_alignment(
+    project_id: str,
+    charter_id: str,
+    body: AlignmentStartRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+    before = {"status": charter.status}
+    charter.status = "aligning"
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="charter",
+        entity_id=str(cid),
+        action="start_alignment",
+        before_data=before,
+        after_data={"status": charter.status, "note": body.note},
+    )
+    await db.commit()
+    return success_response({"id": str(cid), "status": charter.status})
+
+
+@router.post("/projects/{project_id}/charters/{charter_id}/alignment/confirm")
+async def confirm_alignment(
+    project_id: str,
+    charter_id: str,
+    body: AlignmentConfirmRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+
+    row_result = await db.execute(
+        select(AlignmentRecord).where(
+            and_(
+                AlignmentRecord.truth_node_id == cid,
+                AlignmentRecord.user_id == int(current_user.id),
+                AlignmentRecord.tenant_id == tenant_id,
+            )
+        )
+    )
+    row = row_result.scalar_one_or_none()
+    if row is None:
+        row = AlignmentRecord(
+            truth_node_id=cid,
+            user_id=int(current_user.id),
+            alignment_score=body.score,
+            is_confirmed=body.confirmed,
+            comments=body.comments,
+            tenant_id=tenant_id,
+        )
+        db.add(row)
+    else:
+        row.alignment_score = body.score
+        row.is_confirmed = body.confirmed
+        row.comments = body.comments
+
+    member_ids = await svc.get_project_member_ids(pid)
+    confirmed_map: set[int] = set()
+    rows_result = await db.execute(
+        select(AlignmentRecord).where(
+            and_(
+                AlignmentRecord.truth_node_id == cid,
+                AlignmentRecord.tenant_id == tenant_id,
+                AlignmentRecord.is_confirmed.is_(True),
+            )
+        )
+    )
+    for r in rows_result.scalars().all():
+        confirmed_map.add(int(r.user_id))
+    all_confirmed = set(member_ids).issubset(confirmed_map) if member_ids else False
+    if all_confirmed:
+        charter.status = "aligned"
+    elif (charter.status or "").lower() not in {"aligned"}:
+        charter.status = "aligning"
+
+    if all_confirmed:
+        content = _as_dict(charter.content)
+        signal_ids = [sid for sid in content.get("signal_ids", []) if isinstance(sid, int)]
+        if charter.signal_id and int(charter.signal_id) not in signal_ids:
+            signal_ids.append(int(charter.signal_id))
+        if signal_ids:
+            signal_rows = (
+                await db.execute(
+                    select(Signal).where(
+                        and_(
+                            Signal.tenant_id == tenant_id,
+                            Signal.id.in_(signal_ids),
+                        )
+                    )
+                )
+            ).scalars().all()
+            for signal in signal_rows:
+                signal.status = "已关闭"
+                signal.processed = True
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="alignment",
+        entity_id=f"{cid}:{current_user.id}",
+        action="confirm",
+        after_data={
+            "score": body.score,
+            "confirmed": body.confirmed,
+            "all_confirmed": all_confirmed,
+            "charter_status": charter.status,
+        },
+    )
+    await db.commit()
+    return success_response(
+        {
+            "charter_id": str(cid),
+            "charter_status": charter.status,
+            "all_confirmed": all_confirmed,
+            "confirmed_count": len(confirmed_map),
+            "member_count": len(member_ids),
+        }
+    )
+
+
+class TaskCreateRequest(BaseModel):
+    title: str
+    description: Optional[str] = None
+    required_skill_tag_ids: Optional[list[int]] = None
+    priority: str = "medium"
+    estimated_hours: Optional[float] = None
+    due_date: Optional[str] = None
+
+
+class TaskUpdateRequest(BaseModel):
+    title: Optional[str] = None
+    description: Optional[str] = None
+    assignee_id: Optional[int] = None
+    required_skill_tag_ids: Optional[list[int]] = None
+    priority: Optional[str] = None
+    estimated_hours: Optional[float] = None
+    due_date: Optional[str] = None
+
+
+@router.post("/projects/{project_id}/charters/{charter_id}/tasks")
+async def create_task(
+    project_id: str,
+    charter_id: str,
+    body: TaskCreateRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+    _assert_charter_aligned(charter)
+
+    due_date = _parse_due_date(body.due_date)
+
+    task = await persist_task(
+        db,
+        tenant_id=tenant_id,
+        truth_node_id=cid,
+        project_id=pid,
+        draft=build_task_create_draft(
+            title=body.title,
+            description=body.description,
+            required_skill_tag_ids=body.required_skill_tag_ids or [],
+            priority=body.priority,
+            status="todo",
+            estimated_hours=body.estimated_hours,
+            due_date=due_date,
+        ),
+        actor_user_id=int(current_user.id),
+        action="task.create",
+        remark="sstc.create_task",
+        commit=False,
+        refresh_workload=False,
+    )
+    await db.commit()
+    await db.refresh(task)
+    if task.assignee_id is not None:
+        await workload_service.refresh_user_assigned_hours_from_tasks(
+            db,
+            tenant_id=tenant_id,
+            user_id=int(task.assignee_id),
+        )
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="task",
+        record_id=int(task.id),
+        project_id=pid,
+        title=task.title,
+        content=task.description or "",
+    )
+    return success_response(
+        {
+            "id": str(task.id),
+            "title": task.title,
+            "status": task.status,
+            "priority": task.priority,
+            "charter_id": str(cid),
+        }
+    )
+
+
+@router.post("/projects/{project_id}/charters/{charter_id}/tasks/ai-breakdown")
+async def ai_breakdown_tasks(
+    project_id: str,
+    charter_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    cid = _parse_node_id(charter_id, "charter_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    charter = await _ensure_charter(db, tenant_id=tenant_id, project_id=pid, charter_id=cid)
+    _assert_charter_aligned(charter)
+
+    prompt = f"""请基于以下需求宪章拆解 5-8 条任务，返回 JSON 数组：
+[
+  {{"task_name":"", "type":"backend/frontend/ai/devops", "estimated_hours": 4.0, "priority":"P0/P1/P2"}}
+]
+
+宪章标题：{charter.title}
+宪章内容：{_as_dict(charter.content)}
+"""
+    try:
+        rows = await llm_service.generate_task_breakdown_async(
+            prompt,
+            db=db,
+            tenant_id=tenant_id,
+            project_id=pid,
+        )
+    except RuntimeError as exc:
+        raise HTTPException(status_code=403, detail=str(exc)) from exc
+    except Exception as exc:
+        raise HTTPException(status_code=500, detail=f"AI 拆解失败: {exc}") from exc
+
+    created = []
+    prompt_tokens = max(len(prompt) // 4, 1)
+    completion_tokens = 0
+    for item in rows or []:
+        title = compact_title((item.get("task_name") or item.get("title") or ""), max_len=32)
+        if not title:
+            continue
+        priority = str(item.get("priority") or "P1").lower()
+        mapped_priority = "high" if priority == "p0" else "medium" if priority == "p1" else "low"
+        task = await persist_task(
+            db,
+            tenant_id=tenant_id,
+            truth_node_id=cid,
+            project_id=pid,
+            draft=build_task_create_draft(
+                title=title,
+                description=item.get("description"),
+                required_skill_tag_ids=[],
+                priority=mapped_priority,
+                status="todo",
+                estimated_hours=float(item.get("estimated_hours") or 0) or None,
+            ),
+            actor_user_id=int(current_user.id),
+            action="task.create",
+            remark="sstc.ai_breakdown",
+            commit=False,
+            refresh_workload=False,
+        )
+        created.append(task)
+        completion_tokens += max(len(title) // 2, 1)
+
+    db.add(
+        AICallLog(
+            tenant_id=tenant_id,
+            user_id=int(current_user.id),
+            function_name="sstc_ai_breakdown",
+            prompt=prompt,
+            response=f"created_count={len(created)}",
+            tokens_used=prompt_tokens + completion_tokens,
+            cost=((prompt_tokens * 0.00015 + completion_tokens * 0.0006) / 1000),
+            status="success",
+        )
+    )
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="task",
+        entity_id=str(cid),
+        action="ai_breakdown",
+        after_data={"created_count": len(created)},
+    )
+    await db.commit()
+    for task in created:
+        _sync_pmflow_rag_async(
+            tenant_id=tenant_id,
+            record_type="task",
+            record_id=int(task.id),
+            project_id=pid,
+            title=task.title,
+            content=task.description or "",
+        )
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    return success_response(
+        {
+            "created_count": len(created),
+            "items": [
+                {"id": str(t.id), "title": t.title, "priority": t.priority, "status": t.status}
+                for t in created
+            ],
+        }
+    )
+
+
+@router.get("/projects/{project_id}/tasks/board")
+async def get_task_board(
+    project_id: str,
+    assignee_id: Optional[int] = Query(default=None),
+    priority: Optional[str] = Query(default=None),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    result = await db.execute(
+        select(Task).where(
+            and_(
+                Task.project_id == pid,
+                Task.tenant_id == tenant_id,
+            )
+        )
+    )
+    tasks = list(result.scalars().all())
+    if assignee_id is not None:
+        tasks = [t for t in tasks if t.assignee_id == assignee_id]
+    if priority:
+        tasks = [t for t in tasks if (t.priority or "").lower() == priority.lower()]
+
+    rows = []
+    for t in tasks:
+        risk_count = 0
+        if (t.status or "").lower() == "blocked":
+            risk_count += 1
+        if t.due_date and t.due_date < datetime.now(timezone.utc).replace(tzinfo=None) and (t.status or "").lower() not in DONE_STATUSES:
+            risk_count += 1
+        rows.append(
+            {
+                "id": str(t.id),
+                "title": t.title,
+                "description": t.description,
+                "status": t.status,
+                "priority": t.priority,
+                "assignee_id": str(t.assignee_id) if t.assignee_id else None,
+                "due_date": t.due_date.isoformat() if t.due_date else None,
+                "estimated_hours": float(t.estimated_hours) if t.estimated_hours else None,
+                "risk_count": risk_count,
+                "truth_node_id": str(t.truth_node_id),
+            }
+        )
+    return success_response(rows)
+
+
+@router.patch("/projects/{project_id}/tasks/{task_id}/status")
+async def patch_task_status(
+    project_id: str,
+    task_id: str,
+    body: TaskStatusPatchRequest,
+    background_tasks: BackgroundTasks,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    tid = _parse_node_id(task_id, "task_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    task = await _ensure_task(db, tenant_id=tenant_id, project_id=pid, task_id=tid)
+    if not _is_manager(current_user) and int(task.assignee_id or 0) != int(current_user.id):
+        raise HTTPException(status_code=403, detail="仅负责人或任务执行人可更新状态")
+
+    before_status = (task.status or "").lower()
+    before = {"status": task.status}
+    task.status = body.status
+    if body.status in DONE_STATUSES:
+        task.completed_at = datetime.now(timezone.utc).replace(tzinfo=None)
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="task",
+        entity_id=str(tid),
+        action="status_change",
+        before_data=before,
+        after_data={"status": body.status},
+    )
+    if task.assignee_id:
+        await workload_service.refresh_user_assigned_hours_from_tasks(
+            db,
+            tenant_id=tenant_id,
+            user_id=int(task.assignee_id),
+        )
+    else:
+        await db.commit()
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    if task.assignee_id:
+        await context_cache_service.invalidate_user_context(
+            tenant_id=tenant_id,
+            user_id=int(task.assignee_id),
+        )
+    became_done = (body.status or "").lower() in DONE_STATUSES and before_status not in DONE_STATUSES
+    if became_done and task.assignee_id:
+        background_tasks.add_task(
+            skill_score_service.process_task_completion,
+            task_id=int(task.id),
+            tenant_id=tenant_id,
+            operator_user_id=int(current_user.id),
+        )
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="task",
+        record_id=int(task.id),
+        project_id=pid,
+        title=task.title,
+        content=(task.description or "") + f"\n状态: {task.status}",
+    )
+
+    return success_response({"id": str(task.id), "status": task.status})
+
+
+@router.patch("/projects/{project_id}/tasks/{task_id}")
+async def patch_task(
+    project_id: str,
+    task_id: str,
+    body: TaskUpdateRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    tid = _parse_node_id(task_id, "task_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    task = await _ensure_task(db, tenant_id=tenant_id, project_id=pid, task_id=tid)
+
+    before = {
+        "title": task.title,
+        "description": task.description,
+        "assignee_id": task.assignee_id,
+        "priority": task.priority,
+        "estimated_hours": float(task.estimated_hours) if task.estimated_hours is not None else None,
+        "due_date": task.due_date.isoformat() if task.due_date else None,
+    }
+
+    old_assignee_id = int(task.assignee_id) if task.assignee_id else None
+    if body.title is not None:
+        task.title = compact_title(body.title, max_len=32)
+    if body.description is not None:
+        task.description = body.description
+    if body.assignee_id is not None:
+        task.assignee_id = body.assignee_id
+    if body.required_skill_tag_ids is not None:
+        task.required_skill_tag_ids = body.required_skill_tag_ids
+    if body.priority is not None:
+        task.priority = body.priority
+    if body.estimated_hours is not None:
+        task.estimated_hours = body.estimated_hours
+    if body.due_date is not None:
+        task.due_date = _parse_due_date(body.due_date)
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="task",
+        entity_id=str(tid),
+        action="update",
+        before_data=before,
+        after_data={
+            "title": task.title,
+            "description": task.description,
+            "assignee_id": task.assignee_id,
+            "priority": task.priority,
+            "estimated_hours": float(task.estimated_hours) if task.estimated_hours is not None else None,
+            "due_date": task.due_date.isoformat() if task.due_date else None,
+        },
+    )
+
+    touched_user_ids: set[int] = set()
+    if old_assignee_id:
+        touched_user_ids.add(old_assignee_id)
+    if task.assignee_id:
+        touched_user_ids.add(int(task.assignee_id))
+
+    if touched_user_ids:
+        for uid in touched_user_ids:
+            await workload_service.refresh_user_assigned_hours_from_tasks(
+                db,
+                tenant_id=tenant_id,
+                user_id=uid,
+            )
+    else:
+        await db.commit()
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    for uid in touched_user_ids:
+        await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=uid)
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="task",
+        record_id=int(task.id),
+        project_id=pid,
+        title=task.title,
+        content=task.description or "",
+    )
+
+    return success_response({"id": str(task.id), "status": task.status})
+
+
+@router.delete("/projects/{project_id}/tasks/{task_id}")
+async def delete_task(
+    project_id: str,
+    task_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    tid = _parse_node_id(task_id, "task_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    task = await _ensure_task(db, tenant_id=tenant_id, project_id=pid, task_id=tid)
+    assignee_id = int(task.assignee_id) if task.assignee_id else None
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="task",
+        entity_id=str(tid),
+        action="delete",
+        before_data={
+            "assignee_id": assignee_id,
+            "estimated_hours": float(task.estimated_hours) if task.estimated_hours else None,
+            "status": task.status,
+        },
+    )
+    await db.execute(
+        delete(Task).where(
+            and_(
+                Task.id == tid,
+                Task.project_id == pid,
+                Task.tenant_id == tenant_id,
+            )
+        )
+    )
+
+    if assignee_id:
+        await workload_service.refresh_user_assigned_hours_from_tasks(
+            db,
+            tenant_id=tenant_id,
+            user_id=assignee_id,
+        )
+    else:
+        await db.commit()
+    await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+    if assignee_id:
+        await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=assignee_id)
+
+    return success_response({"id": str(tid), "deleted": True})
+
+
+@router.post("/projects/{project_id}/tasks/{task_id}/recommend-assignee")
+async def recommend_assignee_for_task(
+    project_id: str,
+    task_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    pid = _parse_project_id(project_id)
+    tid = _parse_node_id(task_id, "task_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    task = await _ensure_task(db, tenant_id=tenant_id, project_id=pid, task_id=tid)
+
+    required_skills = [str(x) for x in (task.required_skill_tag_ids or [])]
+    recommendations = await recommend_personnel(
+        db=db,
+        tenant_id=tenant_id,
+        required_skills=required_skills,
+        task_description=task.description or task.title,
+        task_id=str(task.id),
+        limit=5,
+    )
+
+    project_user_ids_result = await db.execute(
+        select(UserProject.user_id).where(
+            and_(
+                UserProject.tenant_id == tenant_id,
+                UserProject.project_id == pid,
+            )
+        )
+    )
+    project_user_ids = {int(uid) for uid in project_user_ids_result.scalars().all()}
+
+    active_users_result = await db.execute(
+        select(User.id).where(
+            and_(
+                User.tenant_id == tenant_id,
+                User.status == "active",
+                User.is_active.is_(True),
+            )
+        )
+    )
+    active_user_ids = {int(uid) for uid in active_users_result.scalars().all()}
+
+    filtered = []
+    for item in recommendations.get("recommendations", []):
+        try:
+            uid = int(item.get("user_id"))
+        except (TypeError, ValueError):
+            continue
+        if uid in project_user_ids and uid in active_user_ids:
+            filtered.append(item)
+
+    recommendations["recommendations"] = filtered
+    recommendations["total_count"] = len(filtered)
+    return success_response(recommendations)
+
+
+@router.post("/projects/{project_id}/tasks/optimize-workload")
+async def optimize_task_workload(
+    project_id: str,
+    body: WorkloadOptimizeRequest = Body(default_factory=WorkloadOptimizeRequest),
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    proposals = await workload_service.build_ai_optimize_preview(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        max_moves=body.max_moves,
+        unassigned_only=body.mode == "unassigned",
+    )
+    if not proposals:
+        return success_response(
+            {
+                "updated": 0,
+                "message": "no tasks to optimize",
+                "mode": body.mode,
+            }
+        )
+
+    optimize_result = await workload_service.apply_ai_optimize_plan(
+        db,
+        tenant_id=tenant_id,
+        proposals=proposals,
+    )
+    updated = int(optimize_result.get("applied_count") or 0)
+    affected_users = int(optimize_result.get("affected_users") or 0)
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="task",
+        entity_id=str(pid),
+        action="optimize_workload",
+        after_data={
+            "updated": updated,
+            "affected_users": affected_users,
+            "proposals": len(proposals),
+            "mode": body.mode,
+        },
+    )
+    await db.commit()
+    return success_response(
+        {
+            "updated": updated,
+            "affected_users": affected_users,
+            "proposals": len(proposals),
+            "mode": body.mode,
+        }
+    )
+
+
+@router.get("/projects/{project_id}/releases")
+async def list_releases(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    releases = await svc.get_releases(pid)
+    return success_response(
+        [
+            {
+                "id": str(r.id),
+                "title": r.title,
+                "status": r.status,
+                "content": _as_dict(r.content),
+                "version": r.version,
+                "created_at": r.created_at.isoformat() if r.created_at else None,
+            }
+            for r in releases
+        ]
+    )
+
+
+@router.post("/projects/{project_id}/releases")
+async def create_release_plan(
+    project_id: str,
+    body: ReleaseCreateRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    svc = SSTCService(db, tenant_id=tenant_id)
+    await _ensure_project(svc, pid)
+
+    charter = await _ensure_charter(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        charter_id=int(body.charter_id),
+    )
+
+    if not body.task_ids:
+        raise HTTPException(status_code=400, detail="发布计划必须绑定至少一个任务")
+
+    task_result = await db.execute(
+        select(Task).where(
+            and_(
+                Task.id.in_(body.task_ids),
+                Task.project_id == pid,
+                Task.tenant_id == tenant_id,
+            )
+        )
+    )
+    tasks = list(task_result.scalars().all())
+    if len(tasks) != len(body.task_ids):
+        raise HTTPException(status_code=400, detail="部分任务不存在")
+    not_done = [t for t in tasks if (t.status or "").lower() not in DONE_STATUSES]
+    if not_done:
+        raise HTTPException(status_code=409, detail="发布计划仅允许绑定已完成任务")
+
+    governance = await _build_governance_snapshot(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        charter_id=int(charter.id),
+    )
+    content = {
+        "charter_id": int(charter.id),
+        "task_ids": [int(t.id) for t in tasks],
+        "planned_at": body.planned_at,
+        "verification_items": [],
+        "verification_done": False,
+        "verification_progress": {"done": 0, "total": 0},
+        "governance_snapshot": governance,
+    }
+    release_node = TruthNode(
+        tenant_id=tenant_id,
+        node_type="release",
+        parent_id=int(charter.id),
+        project_id=pid,
+        created_by=int(current_user.id),
+        title=compact_title(body.title, max_len=24),
+        content=content,
+        version=1,
+        status="pending",
+    )
+    db.add(release_node)
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="release",
+        entity_id="new",
+        action="create",
+        after_data={"title": release_node.title, "task_ids": body.task_ids},
+    )
+    await db.commit()
+    await db.refresh(release_node)
+    return success_response(
+        {
+            "id": str(release_node.id),
+            "title": release_node.title,
+            "status": release_node.status,
+            "content": _as_dict(release_node.content),
+        }
+    )
+
+
+@router.patch("/projects/{project_id}/releases/{release_node_id}/status")
+async def patch_release_status(
+    project_id: str,
+    release_node_id: str,
+    body: ReleaseStatusPatchRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(release_node_id, "release_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    release_node = await _ensure_release(
+        db, tenant_id=tenant_id, project_id=pid, release_node_id=rid
+    )
+    await _assert_release_transition_allowed(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        release_node=release_node,
+        target_status=body.status,
+    )
+    before = {"status": release_node.status}
+    release_node.status = body.status
+    content = _as_dict(release_node.content)
+    content["status_note"] = body.note
+    content["last_status_changed_at"] = datetime.now(timezone.utc).isoformat()
+    release_node.content = content
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="release",
+        entity_id=str(rid),
+        action="status_change",
+        before_data=before,
+        after_data={"status": body.status, "note": body.note},
+    )
+    await db.commit()
+    return success_response({"id": str(release_node.id), "status": release_node.status})
+
+
+@router.post("/projects/{project_id}/releases/{release_node_id}/verification/generate")
+async def generate_release_verification(
+    project_id: str,
+    release_node_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(release_node_id, "release_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    release_node = await _ensure_release(
+        db, tenant_id=tenant_id, project_id=pid, release_node_id=rid
+    )
+
+    content = _as_dict(release_node.content)
+    charter_id = int(content.get("charter_id") or 0)
+    charter = await _ensure_charter(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        charter_id=charter_id,
+    )
+    charter_content = _as_dict(charter.content)
+    success_metrics = charter_content.get("success_metrics") or charter_content.get("core_metrics") or []
+    if not success_metrics:
+        success_metrics = ["功能验收通过", "核心流程无阻塞", "关键性能指标达标"]
+
+    items = [
+        {
+            "id": f"check-{i+1}",
+            "title": str(metric),
+            "expected": "达到预期",
+            "status": "pending",
+            "required": True,
+            "source": "charter_success_metric",
+            "verified_by": None,
+            "verified_at": None,
+            "note": None,
+        }
+        for i, metric in enumerate(success_metrics)
+    ]
+    content["verification_items"] = items
+    content["verification_progress"] = {"done": 0, "total": len(items)}
+    content["verification_done"] = False
+    release_node.content = content
+    await db.commit()
+    return success_response({"release_id": str(release_node.id), "verification_items": items})
+
+
+@router.patch("/projects/{project_id}/releases/{release_node_id}/verification/items/{item_id}")
+async def patch_release_verification_item(
+    project_id: str,
+    release_node_id: str,
+    item_id: str,
+    body: ReleaseVerificationItemPatchRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(release_node_id, "release_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    release_node = await _ensure_release(
+        db, tenant_id=tenant_id, project_id=pid, release_node_id=rid
+    )
+    content = _as_dict(release_node.content)
+    items = content.get("verification_items")
+    if not isinstance(items, list) or len(items) == 0:
+        raise HTTPException(status_code=404, detail="验证清单不存在，请先生成")
+
+    target = None
+    for item in items:
+        if str(item.get("id")) == item_id:
+            target = item
+            break
+    if target is None:
+        raise HTTPException(status_code=404, detail="验证项不存在")
+
+    target["status"] = body.status
+    if body.status == "done":
+        target["verified_by"] = str(current_user.id)
+        target["verified_at"] = datetime.now(timezone.utc).isoformat()
+    else:
+        target["verified_by"] = None
+        target["verified_at"] = None
+    target["note"] = body.note
+
+    done_count = len([item for item in items if str(item.get("status") or "").lower() == "done"])
+    content["verification_items"] = items
+    content["verification_progress"] = {"done": done_count, "total": len(items)}
+    content["verification_done"] = done_count == len(items) and len(items) > 0
+    release_node.content = content
+    await db.commit()
+    return success_response(
+        {
+            "release_id": str(release_node.id),
+            "item_id": item_id,
+            "verification_progress": content["verification_progress"],
+            "verification_done": content["verification_done"],
+        }
+    )
+
+
+@router.post("/projects/{project_id}/releases/{release_node_id}/verification/complete")
+async def complete_release_verification(
+    project_id: str,
+    release_node_id: str,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(release_node_id, "release_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    release_node = await _ensure_release(
+        db, tenant_id=tenant_id, project_id=pid, release_node_id=rid
+    )
+    status_raw = str(release_node.status or "").lower()
+    if status_raw not in {"online", "releasing"}:
+        raise HTTPException(status_code=409, detail="仅上线中的发布计划可完成验证")
+    content = _as_dict(release_node.content)
+    items = content.get("verification_items", [])
+    if not isinstance(items, list) or len(items) == 0:
+        raise HTTPException(status_code=409, detail="验证清单为空，请先生成并完成验证项")
+    incomplete = [item for item in items if str(item.get("status") or "").lower() != "done"]
+    if incomplete:
+        raise HTTPException(status_code=409, detail="仍有未完成验证项，请先逐项完成")
+    now_str = datetime.now(timezone.utc).isoformat()
+    content["verification_items"] = items
+    content["verification_done"] = True
+    content["verification_progress"] = {"done": len(items), "total": len(items)}
+    content["verified_at"] = now_str
+    release_node.content = content
+    release_node.status = "verified"
+    await db.commit()
+    return success_response({"id": str(release_node.id), "status": release_node.status})
+
+
+@router.get("/projects/{project_id}/reviews")
+async def list_reviews(
+    project_id: str,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    pid = _parse_project_id(project_id)
+    svc = SSTCService(db, tenant_id=int(current_user.tenant_id))
+    await _ensure_project(svc, pid)
+    reviews = await svc.get_reviews(pid)
+    return success_response(
+        [
+            {
+                "id": str(r.id),
+                "title": r.title,
+                "status": r.status,
+                "content": _as_dict(r.content),
+                "version": r.version,
+                "created_at": r.created_at.isoformat() if r.created_at else None,
+            }
+            for r in reviews
+        ]
+    )
+
+
+@router.post("/projects/{project_id}/reviews")
+async def create_review(
+    project_id: str,
+    body: ReviewCreateRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    release_node = await _ensure_release(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        release_node_id=int(body.release_node_id),
+    )
+    if (release_node.status or "").lower() not in RELEASE_VERIFIED_STATUSES:
+        raise HTTPException(status_code=409, detail="发布未验证，复盘环节未解锁")
+
+    normalized_content = _normalize_review_content(body.content)
+    if not normalized_content.get("summary"):
+        normalized_content["summary"] = "AI 自动生成：请补充目标达成、核心问题与改进项。"
+    normalized_content["release_node_id"] = int(release_node.id)
+    normalized_content["charter_id"] = int(release_node.parent_id) if release_node.parent_id else None
+
+    review_node = TruthNode(
+        tenant_id=tenant_id,
+        node_type="review",
+        parent_id=int(release_node.id),
+        project_id=pid,
+        created_by=int(current_user.id),
+        title=compact_title(body.title, max_len=24),
+        content=normalized_content,
+        version=1,
+        status="draft",
+    )
+    db.add(review_node)
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="review",
+        entity_id="new",
+        action="create",
+        after_data={"title": review_node.title, "release_node_id": body.release_node_id},
+    )
+    await db.commit()
+    await db.refresh(review_node)
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="review",
+        record_id=int(review_node.id),
+        project_id=pid,
+        title=review_node.title,
+        content=_as_dict(review_node.content),
+    )
+    return success_response(
+        {
+            "id": str(review_node.id),
+            "title": review_node.title,
+            "status": review_node.status,
+            "content": _as_dict(review_node.content),
+        }
+    )
+
+
+@router.patch("/projects/{project_id}/reviews/{review_node_id}/content")
+async def patch_review_content(
+    project_id: str,
+    review_node_id: str,
+    body: ReviewContentPatchRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(review_node_id, "review_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    review_node = await _ensure_review(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        review_node_id=rid,
+    )
+
+    existing = _as_dict(review_node.content)
+    merged = dict(existing)
+    merged.update(body.content or {})
+    normalized = _normalize_review_content(merged)
+    normalized["release_node_id"] = existing.get("release_node_id")
+    normalized["charter_id"] = existing.get("charter_id")
+    review_node.content = normalized
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="review",
+        entity_id=str(rid),
+        action="update_content",
+        after_data={
+            "improvement_count": len(normalized.get("improvements") or []),
+            "summary_len": len(str(normalized.get("summary") or "")),
+        },
+    )
+    await db.commit()
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="review",
+        record_id=int(review_node.id),
+        project_id=pid,
+        title=review_node.title,
+        content=_as_dict(review_node.content),
+    )
+    return success_response(
+        {
+            "id": str(review_node.id),
+            "status": review_node.status,
+            "content": _as_dict(review_node.content),
+        }
+    )
+
+
+@router.post("/projects/{project_id}/reviews/{review_node_id}/improvements/convert-task")
+async def convert_review_improvement_to_task(
+    project_id: str,
+    review_node_id: str,
+    body: ReviewImprovementToTaskRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(review_node_id, "review_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    review_node = await _ensure_review(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        review_node_id=rid,
+    )
+    review_content = _normalize_review_content(_as_dict(review_node.content))
+    improvements = review_content.get("improvements") or []
+    if body.improvement_index >= len(improvements):
+        raise HTTPException(status_code=400, detail="改进项索引超出范围")
+    item = improvements[body.improvement_index]
+    if not isinstance(item, dict):
+        raise HTTPException(status_code=400, detail="改进项数据无效")
+
+    release_node_id = int(review_node.parent_id or review_content.get("release_node_id") or 0)
+    if release_node_id <= 0:
+        raise HTTPException(status_code=409, detail="复盘未关联发布计划，无法生成任务")
+    release_node = await _ensure_release(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        release_node_id=release_node_id,
+    )
+
+    release_content = _as_dict(release_node.content)
+    charter_id = (
+        int(body.charter_id)
+        if body.charter_id
+        else int(release_content.get("charter_id") or release_node.parent_id or 0)
+    )
+    if charter_id <= 0:
+        raise HTTPException(status_code=409, detail="无法定位宪章，改进项无法转任务")
+    charter = await _ensure_charter(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        charter_id=charter_id,
+    )
+    _assert_charter_aligned(charter)
+
+    title = compact_title(
+        body.title or str(item.get("title") or "").strip() or "复盘改进项任务",
+        max_len=32,
+    )
+    if not title:
+        raise HTTPException(status_code=400, detail="改进项标题为空")
+    desc_parts = [
+        str(body.description or "").strip(),
+        str(item.get("description") or "").strip(),
+        f"来源复盘: {review_node.title}",
+    ]
+    description = "\n".join([part for part in desc_parts if part])
+    due_date = _parse_due_date(body.due_date)
+
+    task = await persist_task(
+        db,
+        tenant_id=tenant_id,
+        truth_node_id=charter_id,
+        project_id=pid,
+        draft=build_task_create_draft(
+            title=title,
+            description=description or None,
+            assignee_id=item.get("owner_id"),
+            required_skill_tag_ids=[],
+            status="todo",
+            priority=body.priority,
+            estimated_hours=body.estimated_hours,
+            due_date=due_date,
+        ),
+        actor_user_id=int(current_user.id),
+        action="create_from_review_improvement",
+        remark="sstc.review_improvement_to_task",
+        commit=False,
+        refresh_workload=False,
+    )
+
+    item["task_id"] = int(task.id)
+    item["status"] = "converted"
+    review_content["improvements"] = improvements
+    review_content["release_node_id"] = release_node_id
+    review_content["charter_id"] = charter_id
+    review_node.content = review_content
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="task",
+        entity_id=str(task.id),
+        action="create_from_review_improvement",
+        after_data={
+            "review_id": rid,
+            "improvement_index": body.improvement_index,
+            "charter_id": charter_id,
+        },
+    )
+    await db.commit()
+    await db.refresh(task)
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="task",
+        record_id=int(task.id),
+        project_id=pid,
+        title=task.title,
+        content=task.description or "",
+    )
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="review",
+        record_id=int(review_node.id),
+        project_id=pid,
+        title=review_node.title,
+        content=_as_dict(review_node.content),
+    )
+    return success_response(
+        {
+            "task": {
+                "id": str(task.id),
+                "title": task.title,
+                "status": task.status,
+                "priority": task.priority,
+                "charter_id": str(charter_id),
+            },
+            "review_id": str(review_node.id),
+            "improvement_index": body.improvement_index,
+        }
+    )
+
+
+@router.post("/projects/{project_id}/reviews/{review_node_id}/archive")
+async def archive_review(
+    project_id: str,
+    review_node_id: str,
+    body: ReviewArchiveRequest,
+    permission_user: User = Depends(require_project_member_or_manager),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    _ = permission_user
+    _require_manager(current_user)
+    pid = _parse_project_id(project_id)
+    rid = _parse_node_id(review_node_id, "review_node_id")
+    tenant_id = int(current_user.tenant_id)
+    await _ensure_project(SSTCService(db, tenant_id=tenant_id), pid)
+    review_node = await _ensure_review(
+        db,
+        tenant_id=tenant_id,
+        project_id=pid,
+        review_node_id=rid,
+    )
+
+    before = {"status": review_node.status}
+    review_node.status = "archived" if body.archive else "draft"
+    content = _as_dict(review_node.content)
+    content["archive_note"] = body.note
+    content["archived_at"] = datetime.now(timezone.utc).isoformat() if body.archive else None
+    review_node.content = content
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=pid,
+        entity_type="review",
+        entity_id=str(rid),
+        action="archive" if body.archive else "unarchive",
+        before_data=before,
+        after_data={"status": review_node.status},
+    )
+    await db.commit()
+    _sync_pmflow_rag_async(
+        tenant_id=tenant_id,
+        record_type="review",
+        record_id=int(review_node.id),
+        project_id=pid,
+        title=review_node.title,
+        content=_as_dict(review_node.content),
+    )
+    return success_response({"id": str(review_node.id), "status": review_node.status})
diff --git a/backend/app/api/users.py b/backend/app/api/users.py
new file mode 100644
index 00000000..2626e886
--- /dev/null
+++ b/backend/app/api/users.py
@@ -0,0 +1,1085 @@
+"""用户管理 API。"""
+
+from __future__ import annotations
+
+import csv
+from datetime import datetime
+from io import BytesIO, StringIO
+from typing import Any, Optional
+
+from fastapi import APIRouter, Depends, File, HTTPException, Query, UploadFile, status
+from pydantic import BaseModel
+from sqlalchemy import and_, delete, func, or_, select, update
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..database import get_db
+from ..dependencies import (
+    get_current_active_user,
+    get_current_user,
+    get_tenant_id,
+    require_roles,
+    require_self_or_manager,
+)
+from ..models import (
+    SkillTag,
+    Task,
+    User,
+    UserProject,
+    UserRole,
+    UserSkillBinding,
+    UserStatus,
+    UserWorkload,
+)
+from ..schemas import (
+    UserCreate,
+    UserResetPasswordRequest,
+    UserResponse,
+    UserStatusPatchRequest,
+    UserUpdate,
+)
+from ..services.auth import AuthService
+from ..services.operation_log_service import log_operation
+from ..services import skill_service
+
+router = APIRouter(prefix="/users", tags=["users"])
+
+COMPLETED_TASK_STATUSES = {"done", "completed", "closed", "resolved", "archived"}
+MANAGER_ROLE_SET = {"BOSS", "ADMIN"}
+VALID_ROLE_VALUES = {r.value for r in UserRole}
+
+
+class BulkDeleteUsersRequest(BaseModel):
+    user_ids: list[int]
+
+
+class BulkAssignUsersRequest(BaseModel):
+    user_ids: list[int]
+    role: Optional[str] = None
+    project_ids: Optional[list[int]] = None
+
+
+def _enum_value(value: Any) -> Any:
+    return value.value if hasattr(value, "value") else value
+
+
+def _parse_role_values(raw: Any) -> list[str]:
+    if raw is None:
+        return []
+    if isinstance(raw, (list, tuple, set)):
+        values = [str(_enum_value(item)).strip().upper() for item in raw]
+    else:
+        values = [token.strip().upper() for token in str(_enum_value(raw)).split(",")]
+    deduped: list[str] = []
+    for value in values:
+        if value and value not in deduped:
+            deduped.append(value)
+    return deduped
+
+
+def _normalize_role_input(raw: Any, *, allow_empty: bool = False) -> list[str]:
+    values = _parse_role_values(raw)
+    if not values and not allow_empty:
+        raise ValueError("role required")
+    invalid = [value for value in values if value not in VALID_ROLE_VALUES]
+    if invalid:
+        raise ValueError(f"Invalid role: {','.join(invalid)}")
+    return values
+
+
+def _pack_roles(raw: Any) -> str:
+    return ",".join(_normalize_role_input(raw))
+
+
+def _has_any_role(raw: Any, *roles: str) -> bool:
+    values = set(_parse_role_values(raw))
+    expected = {str(role).upper() for role in roles}
+    return bool(values.intersection(expected))
+
+
+def _build_role_filter(column: Any, roles: list[str]) -> Any:
+    filters = []
+    for role in roles:
+        filters.extend(
+            [
+                column == role,
+                column.like(f"{role},%"),
+                column.like(f"%,{role},%"),
+                column.like(f"%,{role}"),
+            ]
+        )
+    return or_(*filters)
+
+
+def _parse_int_list(raw: Optional[str]) -> list[int]:
+    if not raw:
+        return []
+    ids: list[int] = []
+    for part in raw.split(","):
+        token = part.strip()
+        if not token:
+            continue
+        try:
+            ids.append(int(token))
+        except ValueError:
+            continue
+    return ids
+
+
+def _translate_user_integrity_error(exc: IntegrityError) -> str:
+    message = str(getattr(exc, "orig", exc)).lower()
+    if "users.username" in message or "username" in message:
+        return "用户名已存在"
+    if "users.email" in message or "email" in message:
+        return "邮箱已存在"
+    return "用户数据冲突，请检查用户名或邮箱是否重复"
+
+
+async def _collect_user_enrichment(
+    db: AsyncSession,
+    tenant_id: int,
+    users: list[User],
+) -> tuple[dict[int, float], dict[int, list[str]], dict[int, list[int]]]:
+    if not users:
+        return {}, {}, {}
+
+    user_ids = [int(u.id) for u in users]
+
+    workloads_result = await db.execute(
+        select(UserWorkload).where(
+            and_(
+                UserWorkload.tenant_id == tenant_id,
+                UserWorkload.user_id.in_(user_ids),
+            )
+        )
+    )
+    workload_map: dict[int, float] = {
+        int(w.user_id): float(w.workload_ratio or 0.0)
+        for w in workloads_result.scalars().all()
+    }
+
+    skill_rows = await db.execute(
+        select(UserSkillBinding.user_id, SkillTag.name, UserSkillBinding.proficiency)
+        .join(SkillTag, SkillTag.id == UserSkillBinding.skill_tag_id)
+        .where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.user_id.in_(user_ids),
+            )
+        )
+        .order_by(UserSkillBinding.user_id, UserSkillBinding.proficiency.desc())
+    )
+    top_skills: dict[int, list[str]] = {}
+    for uid, skill_name, _prof in skill_rows.all():
+        key = int(uid)
+        top_skills.setdefault(key, [])
+        if len(top_skills[key]) < 3:
+            top_skills[key].append(skill_name)
+
+    project_rows = await db.execute(
+        select(UserProject.user_id, UserProject.project_id).where(
+            and_(
+                UserProject.tenant_id == tenant_id,
+                UserProject.user_id.in_(user_ids),
+            )
+        )
+    )
+    project_map: dict[int, list[int]] = {}
+    for uid, pid in project_rows.all():
+        project_map.setdefault(int(uid), []).append(int(pid))
+
+    return workload_map, top_skills, project_map
+
+
+def _serialize_user(
+    user: User,
+    workload_map: dict[int, float] | None = None,
+    top_skills: dict[int, list[str]] | None = None,
+    project_map: dict[int, list[int]] | None = None,
+) -> dict:
+    payload = user.to_dict()
+    uid = int(user.id)
+    payload["workload_ratio"] = (workload_map or {}).get(uid, float(user.workload or 0))
+    payload["top_skills"] = (top_skills or {}).get(uid, [])
+    payload["project_ids"] = (project_map or {}).get(uid, [])
+    return payload
+
+
+async def _replace_user_projects(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    project_ids: list[int],
+) -> None:
+    await db.execute(
+        delete(UserProject).where(
+            and_(
+                UserProject.tenant_id == tenant_id,
+                UserProject.user_id == user_id,
+            )
+        )
+    )
+    unique_project_ids = sorted(set(project_ids))
+    for pid in unique_project_ids:
+        db.add(UserProject(tenant_id=tenant_id, user_id=user_id, project_id=pid))
+
+    # bulk delete(UserProject) 不会触发 ORM 对象事件，这里手动注册缓存失效意图。
+    cache_payload = db.sync_session.info.get("_pmflow_cache_invalidation")
+    if not isinstance(cache_payload, dict):
+        cache_payload = {
+            "tenant_ids": set(),
+            "invalidate_system": False,
+            "invalidate_project_list": False,
+            "invalidate_signal_pending": False,
+            "user_ids": set(),
+        }
+        db.sync_session.info["_pmflow_cache_invalidation"] = cache_payload
+    cache_payload["tenant_ids"].add(int(tenant_id))
+    cache_payload["invalidate_system"] = True
+    cache_payload["invalidate_project_list"] = True
+    cache_payload["user_ids"].add(int(user_id))
+
+
+async def _create_user_with_defaults(
+    db: AsyncSession,
+    tenant_id: int,
+    user_data: UserCreate,
+    actor_id: int,
+    *,
+    commit: bool,
+) -> User:
+    username = user_data.username.strip()
+    if not username:
+        raise ValueError("用户名不能为空")
+
+    existing_username = await db.execute(
+        select(User).where(
+            and_(
+                User.tenant_id == tenant_id,
+                User.username == username,
+            )
+        )
+    )
+    if existing_username.scalar_one_or_none():
+        raise ValueError("用户名已存在")
+
+    if user_data.email:
+        existing_email = await db.execute(
+            select(User).where(
+                and_(
+                    User.tenant_id == tenant_id,
+                    User.email == user_data.email.strip(),
+                )
+            )
+        )
+        if existing_email.scalar_one_or_none():
+            raise ValueError("邮箱已存在")
+
+    user = User(
+        tenant_id=tenant_id,
+        username=username,
+        password_hash=AuthService.hash_password(user_data.password),
+        name=(user_data.name or username).strip(),
+        email=(user_data.email or f"{username}@placeholder.local").strip(),
+        role=_pack_roles(user_data.role),
+        assigned_domains=user_data.assigned_domains or [],
+        workload=0,
+        status=UserStatus.ACTIVE.value,
+        is_active=True,
+    )
+    db.add(user)
+    await db.flush()
+
+    db.add(
+        UserWorkload(
+            tenant_id=tenant_id,
+            user_id=int(user.id),
+            current_assigned_hours=0.0,
+            max_available_hours_per_week=40.0,
+            workload_ratio=0.0,
+        )
+    )
+    await skill_service.init_user_skill_bindings(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(user.id),
+        include_system_only=True,
+        commit=False,
+    )
+
+    if user_data.project_ids:
+        await _replace_user_projects(
+            db,
+            tenant_id=tenant_id,
+            user_id=int(user.id),
+            project_ids=user_data.project_ids,
+        )
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=actor_id,
+        project_id=None,
+        entity_type="user",
+        entity_id=str(user.id),
+        action="create",
+        before_data=None,
+        after_data=user.to_dict(),
+    )
+
+    if commit:
+        await db.commit()
+        await db.refresh(user)
+    return user
+
+
+@router.get("/me", response_model=UserResponse)
+async def get_current_user_info(current_user: User = Depends(get_current_active_user)):
+    """当前用户信息。"""
+    return UserResponse(**_serialize_user(current_user))
+
+
+@router.get("/")
+async def list_users(
+    page: int = Query(1, ge=1),
+    page_size: int = Query(20, ge=1, le=100),
+    search: Optional[str] = Query(None, description="用户名/姓名/邮箱关键字"),
+    role: Optional[str] = Query(None),
+    status_value: Optional[str] = Query(None, alias="status"),
+    is_active: Optional[bool] = Query(None),
+    project_id: Optional[int] = Query(None),
+    workload_state: Optional[str] = Query(None, description="idle|normal|overloaded"),
+    skill_ids: Optional[str] = Query(None, description="逗号分隔技能ID"),
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """用户分页列表（支持多维过滤）。"""
+    _ = current_user
+    skill_id_list = _parse_int_list(skill_ids)
+
+    query = select(User).where(User.tenant_id == tenant_id)
+    if role:
+        try:
+            role_values = _normalize_role_input(role, allow_empty=True)
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc))
+        if role_values:
+            query = query.where(_build_role_filter(User.role, role_values))
+    if status_value:
+        query = query.where(User.status == status_value)
+    if is_active is not None:
+        query = query.where(User.is_active == is_active)
+    if search:
+        q = f"%{search.strip()}%"
+        query = query.where(
+            or_(
+                User.username.ilike(q),
+                User.name.ilike(q),
+                User.email.ilike(q),
+            )
+        )
+
+    if project_id is not None:
+        query = query.join(
+            UserProject,
+            and_(
+                UserProject.user_id == User.id,
+                UserProject.tenant_id == tenant_id,
+            ),
+        ).where(UserProject.project_id == project_id)
+
+    if skill_id_list:
+        query = query.join(
+            UserSkillBinding,
+            and_(
+                UserSkillBinding.user_id == User.id,
+                UserSkillBinding.tenant_id == tenant_id,
+            ),
+        ).where(UserSkillBinding.skill_tag_id.in_(skill_id_list))
+
+    if workload_state:
+        query = query.outerjoin(
+            UserWorkload,
+            and_(
+                UserWorkload.user_id == User.id,
+                UserWorkload.tenant_id == tenant_id,
+            ),
+        )
+        if workload_state == "idle":
+            query = query.where(func.coalesce(UserWorkload.workload_ratio, 0.0) < 70.0)
+        elif workload_state == "normal":
+            query = query.where(
+                and_(
+                    func.coalesce(UserWorkload.workload_ratio, 0.0) >= 70.0,
+                    func.coalesce(UserWorkload.workload_ratio, 0.0) <= 100.0,
+                )
+            )
+        elif workload_state == "overloaded":
+            query = query.where(func.coalesce(UserWorkload.workload_ratio, 0.0) > 100.0)
+
+    total_result = await db.execute(
+        select(func.count()).select_from(query.distinct().subquery())
+    )
+    total = int(total_result.scalar() or 0)
+
+    result = await db.execute(
+        query.distinct()
+        .order_by(User.updated_at.desc(), User.id.desc())
+        .offset((page - 1) * page_size)
+        .limit(page_size)
+    )
+    users = list(result.scalars().all())
+    workload_map, top_skills, project_map = await _collect_user_enrichment(db, tenant_id, users)
+
+    return {
+        "items": [
+            _serialize_user(
+                u,
+                workload_map=workload_map,
+                top_skills=top_skills,
+                project_map=project_map,
+            )
+            for u in users
+        ],
+        "page": page,
+        "page_size": page_size,
+        "total": total,
+    }
+
+
+@router.get("/available")
+async def get_available_staff(
+    role: Optional[str] = Query(None, description="按角色过滤"),
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """获取可分配任务成员（启用 + active）。"""
+    _ = current_user
+    query = select(User).where(
+        and_(
+            User.tenant_id == tenant_id,
+            User.status == UserStatus.ACTIVE.value,
+            User.is_active == True,
+        )
+    )
+    if role:
+        try:
+            role_values = _normalize_role_input(role, allow_empty=True)
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc))
+        if role_values:
+            query = query.where(_build_role_filter(User.role, role_values))
+
+    result = await db.execute(query.order_by(User.workload.asc(), User.id.asc()))
+    users = result.scalars().all()
+    return [_serialize_user(u) for u in users]
+
+
+@router.get("/{user_id}")
+async def get_user(
+    user_id: int,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """获取单用户详情。"""
+    _ = current_user
+    result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(status_code=404, detail="User not found")
+
+    workloads, skills, projects = await _collect_user_enrichment(db, tenant_id, [user])
+    return _serialize_user(user, workloads, skills, projects)
+
+
+@router.post("/")
+async def create_user(
+    user_data: UserCreate,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    """创建用户（自动初始化 user_workloads）。"""
+    try:
+        user = await _create_user_with_defaults(
+            db,
+            tenant_id=tenant_id,
+            user_data=user_data,
+            actor_id=int(current_user.id),
+            commit=True,
+        )
+    except ValueError as exc:
+        await db.rollback()
+        raise HTTPException(status_code=400, detail=str(exc))
+    except IntegrityError as exc:
+        await db.rollback()
+        raise HTTPException(status_code=400, detail=_translate_user_integrity_error(exc))
+    except Exception:
+        await db.rollback()
+        raise
+
+    workloads, skills, projects = await _collect_user_enrichment(db, tenant_id, [user])
+    return _serialize_user(user, workloads, skills, projects)
+
+
+@router.put("/{user_id}")
+@router.patch("/{user_id}")
+async def update_user(
+    user_id: int,
+    user_data: UserUpdate,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """更新用户信息。"""
+    result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(status_code=404, detail="User not found")
+
+    current_role = _enum_value(current_user.role)
+    is_self = int(current_user.id) == user_id
+    is_manager = _has_any_role(current_role, *MANAGER_ROLE_SET)
+    if not is_self and not is_manager:
+        raise HTTPException(status_code=403, detail="No permission")
+
+    if is_self and not is_manager:
+        # 普通成员仅可修改自身基础资料。
+        if user_data.role is not None or user_data.status is not None or user_data.is_active is not None:
+            raise HTTPException(status_code=403, detail="Members can only edit own profile fields")
+        if user_data.project_ids is not None:
+            raise HTTPException(status_code=403, detail="Members cannot change project assignments")
+
+    before_data = user.to_dict()
+
+    if user_data.name is not None:
+        user.name = user_data.name.strip()
+    if user_data.email is not None:
+        user.email = user_data.email.strip()
+    if user_data.assigned_domains is not None:
+        user.assigned_domains = user_data.assigned_domains
+    if user_data.workload is not None:
+        user.workload = max(0, min(200, int(user_data.workload)))
+    if user_data.status is not None:
+        user.status = _enum_value(user_data.status)
+        user.is_active = user.status == UserStatus.ACTIVE.value
+    if user_data.is_active is not None:
+        user.is_active = bool(user_data.is_active)
+        if not user.is_active:
+            user.status = UserStatus.OFFLINE.value
+        elif user.status == UserStatus.OFFLINE.value:
+            user.status = UserStatus.ACTIVE.value
+    if user_data.role is not None and is_manager:
+        try:
+            user.role = _pack_roles(user_data.role)
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc))
+    if user_data.project_ids is not None:
+        await _replace_user_projects(
+            db,
+            tenant_id=tenant_id,
+            user_id=user_id,
+            project_ids=user_data.project_ids,
+        )
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=None,
+        entity_type="user",
+        entity_id=str(user.id),
+        action="update",
+        before_data=before_data,
+        after_data=user.to_dict(),
+    )
+    try:
+        await db.commit()
+        await db.refresh(user)
+    except IntegrityError as exc:
+        await db.rollback()
+        raise HTTPException(status_code=400, detail=_translate_user_integrity_error(exc))
+
+    workloads, skills, projects = await _collect_user_enrichment(db, tenant_id, [user])
+    return _serialize_user(user, workloads, skills, projects)
+
+
+@router.post("/{user_id}/reset-password")
+async def reset_password(
+    user_id: int,
+    body: UserResetPasswordRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """重置用户密码。"""
+    result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(status_code=404, detail="User not found")
+
+    current_role = _enum_value(current_user.role)
+    is_manager = _has_any_role(current_role, *MANAGER_ROLE_SET)
+    if not is_manager and int(current_user.id) != user_id:
+        raise HTTPException(status_code=403, detail="No permission")
+
+    try:
+        user.password_hash = AuthService.hash_password(body.new_password)
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+    user.updated_at = datetime.utcnow()
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=None,
+        entity_type="user",
+        entity_id=str(user.id),
+        action="reset_password",
+        before_data=None,
+        after_data={"updated_at": user.updated_at.isoformat()},
+    )
+    await db.commit()
+    return {"message": "Password reset successfully"}
+
+
+@router.patch("/{user_id}/status")
+async def patch_user_status(
+    user_id: int,
+    body: UserStatusPatchRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    """禁用/启用用户，禁用时清空任务指派并归零负载。"""
+    result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(status_code=404, detail="User not found")
+
+    before_data = user.to_dict()
+    user.is_active = body.is_active
+    user.status = UserStatus.ACTIVE.value if body.is_active else UserStatus.OFFLINE.value
+
+    if not body.is_active:
+        user.workload = 0
+        await db.execute(
+            update(Task)
+            .where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.assignee_id == user_id,
+                )
+            )
+            .values(assignee_id=None, updated_at=datetime.utcnow())
+        )
+        workload_result = await db.execute(
+            select(UserWorkload).where(
+                and_(
+                    UserWorkload.tenant_id == tenant_id,
+                    UserWorkload.user_id == user_id,
+                )
+            )
+        )
+        workload = workload_result.scalar_one_or_none()
+        if not workload:
+            workload = UserWorkload(tenant_id=tenant_id, user_id=user_id)
+            db.add(workload)
+        workload.current_assigned_hours = 0.0
+        workload.workload_ratio = 0.0
+
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=None,
+        entity_type="user",
+        entity_id=str(user.id),
+        action="status_patch",
+        before_data=before_data,
+        after_data=user.to_dict(),
+    )
+    await db.commit()
+    await db.refresh(user)
+    return {"message": "Status updated", "user": _serialize_user(user)}
+
+
+@router.delete("/{user_id}")
+async def delete_user(
+    user_id: int,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    """删除用户（有未完成任务时禁止）。"""
+    if int(current_user.id) == user_id:
+        raise HTTPException(status_code=400, detail="Cannot delete yourself")
+
+    result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(status_code=404, detail="User not found")
+
+    incomplete_count_result = await db.execute(
+        select(func.count(Task.id)).where(
+            and_(
+                Task.tenant_id == tenant_id,
+                Task.assignee_id == user_id,
+                func.lower(Task.status).notin_(COMPLETED_TASK_STATUSES),
+            )
+        )
+    )
+    incomplete_count = int(incomplete_count_result.scalar() or 0)
+    if incomplete_count > 0:
+        raise HTTPException(
+            status_code=400,
+            detail=f"User has {incomplete_count} incomplete tasks",
+        )
+
+    before_data = user.to_dict()
+    await db.execute(
+        delete(UserSkillBinding).where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.user_id == user_id,
+            )
+        )
+    )
+    await db.execute(
+        delete(UserWorkload).where(
+            and_(
+                UserWorkload.tenant_id == tenant_id,
+                UserWorkload.user_id == user_id,
+            )
+        )
+    )
+    await db.execute(
+        delete(UserProject).where(
+            and_(
+                UserProject.tenant_id == tenant_id,
+                UserProject.user_id == user_id,
+            )
+        )
+    )
+    await db.delete(user)
+    await log_operation(
+        db,
+        tenant_id=tenant_id,
+        user_id=int(current_user.id),
+        project_id=None,
+        entity_type="user",
+        entity_id=str(user_id),
+        action="delete",
+        before_data=before_data,
+        after_data=None,
+    )
+    await db.commit()
+    return {"message": "User deleted successfully"}
+
+
+@router.post("/bulk-delete")
+async def bulk_delete_users(
+    body: BulkDeleteUsersRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    """批量删除用户（带未完成任务保护）。"""
+    if not body.user_ids:
+        raise HTTPException(status_code=400, detail="user_ids required")
+
+    deleted: list[int] = []
+    skipped: list[dict[str, Any]] = []
+
+    for user_id in sorted(set(body.user_ids)):
+        if int(current_user.id) == user_id:
+            skipped.append({"user_id": user_id, "reason": "Cannot delete yourself"})
+            continue
+
+        user = (
+            await db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id == user_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not user:
+            skipped.append({"user_id": user_id, "reason": "User not found"})
+            continue
+
+        incomplete_count_result = await db.execute(
+            select(func.count(Task.id)).where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.assignee_id == user_id,
+                    func.lower(Task.status).notin_(COMPLETED_TASK_STATUSES),
+                )
+            )
+        )
+        incomplete_count = int(incomplete_count_result.scalar() or 0)
+        if incomplete_count > 0:
+            skipped.append(
+                {
+                    "user_id": user_id,
+                    "reason": f"User has {incomplete_count} incomplete tasks",
+                }
+            )
+            continue
+
+        await db.execute(
+            delete(UserSkillBinding).where(
+                and_(
+                    UserSkillBinding.tenant_id == tenant_id,
+                    UserSkillBinding.user_id == user_id,
+                )
+            )
+        )
+        await db.execute(
+            delete(UserWorkload).where(
+                and_(
+                    UserWorkload.tenant_id == tenant_id,
+                    UserWorkload.user_id == user_id,
+                )
+            )
+        )
+        await db.execute(
+            delete(UserProject).where(
+                and_(
+                    UserProject.tenant_id == tenant_id,
+                    UserProject.user_id == user_id,
+                )
+            )
+        )
+        await db.delete(user)
+        deleted.append(user_id)
+
+    await db.commit()
+    return {"deleted_count": len(deleted), "deleted_user_ids": deleted, "skipped": skipped}
+
+
+@router.post("/bulk-assign")
+async def bulk_assign_users(
+    body: BulkAssignUsersRequest,
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    """批量分配项目/角色。"""
+    user_ids = sorted(set(body.user_ids or []))
+    if not user_ids:
+        raise HTTPException(status_code=400, detail="user_ids required")
+    if body.role is None and body.project_ids is None:
+        raise HTTPException(status_code=400, detail="role or project_ids required")
+
+    updated = 0
+    skipped: list[dict[str, Any]] = []
+
+    for user_id in user_ids:
+        user = (
+            await db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id == user_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not user:
+            skipped.append({"user_id": user_id, "reason": "User not found"})
+            continue
+
+        before_data = user.to_dict()
+
+        if body.role is not None:
+            if not _has_any_role(_enum_value(current_user.role), "BOSS"):
+                skipped.append({"user_id": user_id, "reason": "Only BOSS can bulk assign roles"})
+                continue
+            try:
+                user.role = _pack_roles(body.role)
+            except ValueError as exc:
+                skipped.append({"user_id": user_id, "reason": str(exc)})
+                continue
+
+        if body.project_ids is not None:
+            await _replace_user_projects(
+                db,
+                tenant_id=tenant_id,
+                user_id=user_id,
+                project_ids=body.project_ids,
+            )
+
+        await log_operation(
+            db,
+            tenant_id=tenant_id,
+            user_id=int(current_user.id),
+            project_id=None,
+            entity_type="user",
+            entity_id=str(user_id),
+            action="bulk_assign",
+            before_data=before_data,
+            after_data=user.to_dict(),
+            remark=f"role={body.role}, projects={body.project_ids}",
+        )
+        updated += 1
+
+    await db.commit()
+    return {"updated_count": updated, "skipped": skipped}
+
+
+@router.put("/{user_id}/workload")
+async def update_user_workload(
+    user_id: int,
+    workload: int = Query(..., ge=0, le=200, description="Workload value (0-200)"),
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_self_or_manager),
+    db: AsyncSession = Depends(get_db),
+):
+    """手动更新用户 workload 百分比。"""
+    _ = current_user
+
+    result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(status_code=404, detail="User not found")
+
+    user.workload = max(0, min(200, int(workload)))
+    user.updated_at = datetime.utcnow()
+    await db.commit()
+    await db.refresh(user)
+    return _serialize_user(user)
+
+
+def _normalize_import_row(raw: dict[str, Any]) -> dict[str, Any]:
+    lower_map = {str(k).strip().lower(): v for k, v in raw.items()}
+    return {
+        "username": str(lower_map.get("username", "")).strip(),
+        "password": str(lower_map.get("password", "")).strip(),
+        "name": str(lower_map.get("name", "")).strip(),
+        "email": str(lower_map.get("email", "")).strip() or None,
+        "role": str(lower_map.get("role", "FE")).strip().upper(),
+        "project_ids": _parse_int_list(str(lower_map.get("project_ids", ""))),
+    }
+
+
+async def _iter_import_rows(file: UploadFile) -> list[dict[str, Any]]:
+    filename = (file.filename or "").lower()
+    content = await file.read()
+
+    if filename.endswith(".csv"):
+        text = content.decode("utf-8-sig")
+        reader = csv.DictReader(StringIO(text))
+        return [dict(row) for row in reader]
+
+    if filename.endswith(".xlsx") or filename.endswith(".xls"):
+        try:
+            from openpyxl import load_workbook  # type: ignore
+        except Exception as exc:
+            raise HTTPException(
+                status_code=400,
+                detail=f"Excel import requires openpyxl: {exc}",
+            )
+
+        wb = load_workbook(filename=BytesIO(content), read_only=True)
+        sheet = wb.active
+        rows = list(sheet.iter_rows(values_only=True))
+        if not rows:
+            return []
+        headers = [str(v).strip() if v is not None else "" for v in rows[0]]
+        data_rows: list[dict[str, Any]] = []
+        for row in rows[1:]:
+            item = {
+                headers[idx]: row[idx]
+                for idx in range(min(len(headers), len(row)))
+                if headers[idx]
+            }
+            data_rows.append(item)
+        return data_rows
+
+    raise HTTPException(status_code=400, detail="Only .csv/.xlsx/.xls are supported")
+
+
+@router.post("/import")
+async def import_users(
+    file: UploadFile = File(...),
+    tenant_id: int = Depends(get_tenant_id),
+    current_user: User = Depends(require_roles("BOSS", "ADMIN")),
+    db: AsyncSession = Depends(get_db),
+):
+    """批量导入用户（最多 50 行）。"""
+    rows = await _iter_import_rows(file)
+    if not rows:
+        raise HTTPException(status_code=400, detail="Import file is empty")
+    if len(rows) > 50:
+        raise HTTPException(status_code=400, detail="Import rows exceed max limit: 50")
+
+    results: list[dict[str, Any]] = []
+    success_count = 0
+
+    for idx, raw in enumerate(rows, start=1):
+        row = _normalize_import_row(raw)
+        try:
+            if row["role"] not in {r.value for r in UserRole}:
+                raise ValueError(f"Invalid role: {row['role']}")
+            if not row["username"] or not row["password"]:
+                raise ValueError("username/password required")
+
+            data = UserCreate(
+                username=row["username"],
+                password=row["password"],
+                name=row["name"] or row["username"],
+                email=row["email"],
+                role=UserRole(row["role"]),
+                assigned_domains=[],
+                project_ids=row["project_ids"],
+            )
+            user = await _create_user_with_defaults(
+                db,
+                tenant_id=tenant_id,
+                user_data=data,
+                actor_id=int(current_user.id),
+                commit=True,
+            )
+            success_count += 1
+            results.append(
+                {
+                    "row": idx,
+                    "status": "success",
+                    "user_id": str(user.id),
+                    "message": "created",
+                }
+            )
+        except Exception as exc:
+            await db.rollback()
+            results.append(
+                {
+                    "row": idx,
+                    "status": "failed",
+                    "message": str(exc),
+                }
+            )
+
+    return {
+        "total": len(rows),
+        "success": success_count,
+        "failed": len(rows) - success_count,
+        "results": results,
+    }
diff --git a/backend/app/api/wecom.py b/backend/app/api/wecom.py
new file mode 100644
index 00000000..b6cc25a2
--- /dev/null
+++ b/backend/app/api/wecom.py
@@ -0,0 +1,327 @@
+"""
+WeChat Work (企业微信) Integration API - 企业微信集成接口
+
+功能：
+1. 配置企业微信应用凭证
+2. 验证 URL 和处理事件
+3. 发送消息到企业微信群
+4. 查询集成状态
+"""
+from fastapi import APIRouter, Depends, HTTPException, Request, BackgroundTasks
+from sqlalchemy.ext.asyncio import AsyncSession
+from pydantic import BaseModel, Field
+from typing import List, Optional, Dict, Any
+
+from ..database import get_db
+from ..dependencies.auth import get_current_user
+from ..dependencies.tenant import get_tenant_id
+from ..models import User
+from ..services.auth import has_any_role
+from ..services.wecom_integration import get_wecom_service
+
+
+router = APIRouter(prefix="/wecom", tags=["WeChat Work Integration"])
+
+
+# ==================== Request/Response Models ====================
+
+class WeComConfigRequest(BaseModel):
+    """企业微信配置请求"""
+    corp_id: str = Field(..., description="企业 ID")
+    agent_id: str = Field(..., description="应用 ID")
+    app_secret: str = Field(..., description="应用 Secret")
+    token: str = Field(..., description="回调验证 Token")
+    encoding_aes_key: str = Field(..., description="消息加密密钥")
+
+
+class WeComSendMessageRequest(BaseModel):
+    """发送消息请求"""
+    chat_id: str = Field(..., description="群聊 ID")
+    message: str = Field(..., description="消息内容")
+
+
+# ==================== Configuration Endpoints ====================
+
+@router.post("/config")
+async def save_wecom_config(
+    request: WeComConfigRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    保存企业微信配置
+
+    配置信息说明：
+    - corp_id: 企业 ID，从企业微信管理后台获取
+    - agent_id: 应用 ID
+    - app_secret: 应用 Secret
+    - token: 事件订阅验证 Token
+    - encoding_aes_key: 消息加密密钥
+
+    获取步骤：
+    1. 访问 https://work.weixin.qq.com/
+    2. 进入应用管理 -> 自建应用
+    3. 在「开发配置」获取 Agent ID 和 Secret
+    4. 在「接收消息」配置 Token 和 EncodingAESKey
+    """
+    if not has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"]):
+        raise HTTPException(status_code=403, detail="只有管理员可以配置企业微信集成")
+
+    try:
+        service = get_wecom_service(db)
+        await service.save_config(
+            corp_id=request.corp_id,
+            agent_id=request.agent_id,
+            app_secret=request.app_secret,
+            token=request.token,
+            encoding_aes_key=request.encoding_aes_key,
+            tenant_id=tenant_id
+        )
+
+        return {
+            'status': 'success',
+            'message': '企业微信配置已保存',
+            'next_steps': [
+                '1. 在企业微信管理平台配置接收消息 URL',
+                f'2. URL: https://your-domain.com/api/wecom/webhook',
+                '3. 点击验证按钮完成 URL 验证'
+            ]
+        }
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"保存配置失败: {str(e)}"
+        )
+
+
+@router.get("/config")
+async def get_wecom_config(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """获取企业微信配置状态。管理员可获取完整配置用于回填修改表单。"""
+    try:
+        service = get_wecom_service(db)
+        await service.load_config(tenant_id)
+        can_view_secrets = has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"])
+
+        return {
+            'is_configured': service.is_configured(),
+            'corp_id': service.corp_id,
+            'agent_id': service.agent_id,
+            'app_secret': service.app_secret if can_view_secrets else None,
+            'has_token': bool(service.token),
+            'has_encoding_key': bool(service.encoding_aes_key),
+            'token_preview': f"{service.token[:4]}...{service.token[-4:]}" if service.token and len(service.token) > 8 else (service.token or None),
+            'encoding_aes_key_preview': f"{service.encoding_aes_key[:4]}...{service.encoding_aes_key[-4:]}" if service.encoding_aes_key and len(service.encoding_aes_key) > 8 else (service.encoding_aes_key or None),
+            'token': service.token if can_view_secrets else None,
+            'encoding_aes_key': service.encoding_aes_key if can_view_secrets else None,
+        }
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"获取配置失败: {str(e)}"
+        )
+
+
+@router.delete("/config")
+async def delete_wecom_config(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """删除企业微信配置"""
+    from ..models import Integration
+    from sqlalchemy.future import select
+
+    if not has_any_role(str(current_user.role or ""), ["ADMIN", "BOSS"]):
+        raise HTTPException(status_code=403, detail="只有管理员可以删除配置")
+
+    try:
+        result = await db.execute(
+            select(Integration).where(
+                Integration.tenant_id == tenant_id,
+                Integration.type == 'wecom'
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration:
+            integration.is_active = False
+            await db.commit()
+
+        return {'status': 'success', 'message': '企业微信配置已删除'}
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"删除配置失败: {str(e)}"
+        )
+
+
+# ==================== Webhook Endpoints ====================
+
+@router.get("/webhook")
+@router.get("/webhook/", include_in_schema=False)
+async def verify_webhook_url(
+    msg_signature: str,
+    timestamp: str,
+    nonce: str,
+    echostr: str,
+    db: AsyncSession = Depends(get_db)
+):
+    """
+    验证企业微信事件订阅 URL
+
+    企业微信会发送 GET 请求来验证 URL 有效性
+    """
+    try:
+        service = get_wecom_service(db)
+        await service.load_config(tenant_id=1)
+
+        if not service.is_configured():
+            raise HTTPException(
+                status_code=400,
+                detail="请先配置企业微信"
+            )
+
+        decrypted_echostr = service.verify_url(msg_signature, timestamp, nonce, echostr)
+
+        return {
+            "errcode": 0,
+            "errmsg": "ok",
+            "echostr": decrypted_echostr
+        }
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@router.post("/webhook")
+@router.post("/webhook/", include_in_schema=False)
+async def handle_wecom_event(
+    request: Request,
+    db: AsyncSession = Depends(get_db)
+):
+    """
+    处理企业微信事件推送
+
+    企业微信会发送 POST 请求推送事件
+    需要验证签名确保请求来自企业微信
+    """
+    try:
+        # 获取请求头
+        msg_signature = request.headers.get('msg_signature', '')
+        timestamp = request.headers.get('timestamp', '')
+        nonce = request.headers.get('nonce', '')
+
+        # 读取请求体
+        body_bytes = await request.body()
+        body = body_bytes.decode('utf-8')
+
+        # 加载企业微信配置
+        service = get_wecom_service(db)
+        await service.load_config(tenant_id=1)
+
+        # 验证签名
+        if not service.verify_callback(msg_signature, timestamp, nonce, body):
+            raise HTTPException(status_code=401, detail="签名验证失败")
+
+        # 解析事件
+        from ..services.wecom_integration import WeChatWorkIntegrationService
+        decrypted = service._decrypt(body)
+        event_data = json.loads(decrypted) if decrypted else {}
+
+        # 处理事件
+        result = await service.handle_message_event(event_data)
+
+        return {
+            "errcode": 0,
+            "errmsg": "ok"
+        }
+
+    except HTTPException:
+        raise
+    except Exception as e:
+        print(f"[WeCom] Webhook error: {e}")
+        # 企业微信要求返回成功状态，避免重复推送
+        return {'errcode': 0, 'errmsg': 'ok'}
+
+
+# ==================== Message Endpoints ====================
+
+@router.post("/send-message")
+async def send_wecom_message(
+    request: WeComSendMessageRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """
+    发送文本消息到企业微信群
+
+    用于发送简报、通知等
+    """
+    try:
+        service = get_wecom_service(db)
+        await service.load_config(tenant_id)
+
+        result = await service.send_message_to_group(
+            chat_id=request.chat_id,
+            message=request.message
+        )
+
+        return {
+            'status': 'success',
+            'result': result
+        }
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"发送消息失败: {str(e)}"
+        )
+
+
+# ==================== Test Endpoints ====================
+
+@router.post("/test-connection")
+async def test_wecom_connection(
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+    tenant_id: int = Depends(get_tenant_id)
+):
+    """测试企业微信连接是否正常"""
+    try:
+        service = get_wecom_service(db)
+        await service.load_config(tenant_id)
+
+        if not service.is_configured():
+            return {
+                'status': 'error',
+                'message': '企业微信未配置'
+            }
+
+        # 尝试获取 access_token 测试连接
+        try:
+            access_token = await service._get_access_token()
+            return {
+                'status': 'success',
+                'message': '企业微信连接正常',
+                'has_token': bool(access_token)
+            }
+        except Exception as token_error:
+            return {
+                'status': 'error',
+                'message': f'获取 Token 失败: {str(token_error)}',
+                'hint': '请检查 Corp ID、Agent ID 和 App Secret 是否正确'
+            }
+
+    except Exception as e:
+        raise HTTPException(
+            status_code=500,
+            detail=f"测试连接失败: {str(e)}"
+        )
diff --git a/backend/app/database.py b/backend/app/database.py
new file mode 100644
index 00000000..3d60e8b9
--- /dev/null
+++ b/backend/app/database.py
@@ -0,0 +1,73 @@
+"""数据库会话与引擎管理。"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncGenerator
+
+from sqlalchemy.ext.asyncio import (
+    AsyncEngine,
+    AsyncSession,
+    async_sessionmaker,
+    create_async_engine,
+)
+from sqlalchemy import event
+from sqlalchemy.pool import NullPool
+
+from .core.config import settings
+
+DATABASE_URL = settings.database_url
+
+
+def _is_sqlite() -> bool:
+    return DATABASE_URL.lower().startswith("sqlite")
+
+
+engine_kwargs = {
+    "echo": settings.debug,
+    "future": True,
+    "pool_pre_ping": not _is_sqlite(),
+}
+if _is_sqlite():
+    engine_kwargs["connect_args"] = {
+        "timeout": 30,
+        "check_same_thread": False,
+    }
+    engine_kwargs["poolclass"] = NullPool
+if not _is_sqlite():
+    engine_kwargs.update(
+        {
+            "pool_size": max(5, int(settings.db_pool_size)),
+            "max_overflow": max(0, int(settings.db_max_overflow)),
+            "pool_timeout": max(5, int(settings.db_pool_timeout)),
+            "pool_recycle": max(60, int(settings.db_pool_recycle)),
+        }
+    )
+
+engine: AsyncEngine = create_async_engine(
+    DATABASE_URL,
+    **engine_kwargs,
+)
+
+if _is_sqlite():
+    @event.listens_for(engine.sync_engine, "connect")
+    def _configure_sqlite_pragmas(dbapi_connection, _connection_record) -> None:  # pragma: no cover
+        cursor = dbapi_connection.cursor()
+        try:
+            cursor.execute("PRAGMA journal_mode=WAL")
+            cursor.execute("PRAGMA synchronous=NORMAL")
+            cursor.execute("PRAGMA busy_timeout=30000")
+        finally:
+            cursor.close()
+
+AsyncSessionLocal = async_sessionmaker(
+    bind=engine,
+    class_=AsyncSession,
+    expire_on_commit=False,
+    autoflush=False,
+)
+
+
+async def get_db() -> AsyncGenerator[AsyncSession, None]:
+    """FastAPI 依赖：提供数据库会话。"""
+    async with AsyncSessionLocal() as session:
+        yield session
diff --git a/backend/app/dispatcher.py b/backend/app/dispatcher.py
new file mode 100644
index 00000000..9515db88
--- /dev/null
+++ b/backend/app/dispatcher.py
@@ -0,0 +1,399 @@
+"""
+Neural Sync - Role Dispatcher
+
+根据任务的技术指纹自动路由到对应角色的工作台。
+"""
+
+from enum import Enum
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Any, Tuple
+from datetime import datetime
+import re
+import logging
+from pydantic import BaseModel, Field
+
+logger = logging.getLogger(__name__)
+
+
+class Role(str, Enum):
+    """系统角色定义"""
+
+    BOSS = "boss"  # 老板 - 战略决策
+    AI_ENGINEER = "ai_engineer"  # AI 工程师 - Prompt/模型/RAG
+    BACKEND = "backend"  # 后端开发 - API/数据库/爬虫
+    FRONTEND = "frontend"  # 前端开发 - UI/组件/交互
+
+
+@dataclass
+class RoutingRule:
+    """路由规则定义"""
+
+    role: Role
+    name: str
+    description: str
+    triggers: List[str] = field(default_factory=list)  # 触发关键词
+    keywords: List[str] = field(default_factory=list)  # 技术关键词
+    patterns: List[str] = field(default_factory=list)  # 正则模式
+    threshold: Optional[Dict[str, Any]] = None  # 阈值条件
+    priority: int = 0  # 优先级，数值越大越优先
+
+
+class RoleRouter:
+    """
+    角色路由引擎
+
+    根据信号的技术指纹自动路由到对应角色：
+    - Boss: 高成本、高风险、架构变更决策
+    - AI Engineer: LLM、Prompt、RAG、数据处理
+    - Backend: API、数据库、爬虫、性能优化
+    - Frontend: UI、可视化、交互、响应式
+    """
+
+    # 角色路由规则定义
+    ROLE_RULES: List[RoutingRule] = [
+        RoutingRule(
+            role=Role.BOSS,
+            name="战略决策",
+            description="需要老板决策的高影响任务",
+            triggers=["cost_estimate", "architecture_change", "high_risk", "security"],
+            keywords=["预算", "架构", "安全", "风险", "审批", "决策"],
+            threshold={"dev_days": 5, "risk_level": ["CRITICAL", "HIGH"]},
+            priority=100,
+        ),
+        RoutingRule(
+            role=Role.AI_ENGINEER,
+            name="AI 工程",
+            description="AI 模型和数据处理相关任务",
+            triggers=["prompt", "model", "embedding", "rag", "llm"],
+            keywords=[
+                "ollama",
+                "openai",
+                "claude",
+                "prompt",
+                "embedding",
+                "rag",
+                "llm",
+                "向量",
+                "模型",
+                "训练",
+                "数据清洗",
+                "token",
+                "context",
+            ],
+            patterns=[r"(prompt|embedding|rag|llm|vector|model)", r"(ai|machine.?learning|ml)"],
+            priority=50,
+        ),
+        RoutingRule(
+            role=Role.BACKEND,
+            name="后端开发",
+            description="API、数据库、爬虫等后端任务",
+            triggers=["api", "database", "crawler", "performance", "docker"],
+            keywords=[
+                "fastapi",
+                "sqlalchemy",
+                "postgres",
+                "sql",
+                "async",
+                "docker",
+                "kubernetes",
+                "redis",
+                "celery",
+                "爬虫",
+                "api",
+                "接口",
+                "数据库",
+                "后端",
+                "性能",
+                "缓存",
+            ],
+            patterns=[
+                r"(api|endpoint|route)",
+                r"(sql|database|db)",
+                r"(docker|k8s|kubernetes)",
+                r"(async|await|celery)",
+            ],
+            priority=30,
+        ),
+        RoutingRule(
+            role=Role.FRONTEND,
+            name="前端开发",
+            description="UI、可视化、交互等前端任务",
+            triggers=["ui", "visualization", "interaction", "responsive", "chart"],
+            keywords=[
+                "vue",
+                "react",
+                "typescript",
+                "component",
+                "chart",
+                "echarts",
+                "d3",
+                "可视化",
+                "ui",
+                "ux",
+                "样式",
+                "动画",
+                "响应式",
+                "组件",
+                "交互",
+                "页面",
+                "布局",
+            ],
+            patterns=[
+                r"(vue|react|tsx?|jsx?)",
+                r"(css|scss|tailwind)",
+                r"(component|props|emit)",
+                r"(chart|graph|visualization)",
+            ],
+            priority=30,
+        ),
+    ]
+
+    # 技术栈关键词映射
+    TECH_KEYWORDS = {
+        "frontend": [
+            "vue",
+            "react",
+            "typescript",
+            "javascript",
+            "css",
+            "scss",
+            "tailwind",
+            "component",
+            "props",
+            "emit",
+            "template",
+            "style",
+        ],
+        "backend": [
+            "fastapi",
+            "flask",
+            "django",
+            "sqlalchemy",
+            "postgres",
+            "mysql",
+            "redis",
+            "celery",
+            "async",
+            "await",
+            "docker",
+            "kubernetes",
+            "k8s",
+            "api",
+            "endpoint",
+            "route",
+        ],
+        "ai": [
+            "llm",
+            "ollama",
+            "openai",
+            "claude",
+            "prompt",
+            "embedding",
+            "rag",
+            "vector",
+            "model",
+            "training",
+            "inference",
+            "token",
+            "context",
+        ],
+    }
+
+    def __init__(self):
+        # 编译正则模式
+        self._compiled_patterns: Dict[Role, List[re.Pattern]] = {}
+        for rule in self.ROLE_RULES:
+            patterns = [re.compile(p, re.IGNORECASE) for p in rule.patterns]
+            self._compiled_patterns[rule.role] = patterns
+
+    def route(
+        self,
+        signal_content: str,
+        analysis: Optional[Dict[str, Any]] = None,
+        source: Optional[str] = None,
+    ) -> Tuple[Role, float, Dict[str, Any]]:
+        """
+        根据信号内容和分析结果决定路由角色
+
+        Args:
+            signal_content: 信号内容
+            analysis: AI 分析结果（可选）
+            source: 信号来源（可选）
+
+        Returns:
+            (角色, 置信度, 路由详情)
+        """
+        analysis = analysis or {}
+        details = {
+            "matched_rules": [],
+            "scores": {},
+            "reasons": [],
+        }
+
+        # 1. 首先检查是否需要升级到老板决策
+        if self._should_escalate_to_boss(analysis):
+            details["reasons"].append("触发老板决策阈值")
+            details["matched_rules"].append("threshold_check")
+            return Role.BOSS, 1.0, details
+
+        # 2. 计算每个角色的匹配分数
+        scores: Dict[Role, float] = {rule.role: 0.0 for rule in self.ROLE_RULES}
+
+        # 基于关键词匹配
+        content_lower = signal_content.lower()
+        for rule in self.ROLE_RULES:
+            keyword_score = sum(
+                1 for kw in rule.keywords if kw.lower() in content_lower
+            )
+            if keyword_score > 0:
+                scores[rule.role] += keyword_score * 0.3
+                details["scores"][f"{rule.role}_keywords"] = keyword_score
+
+        # 基于正则模式匹配
+        for rule in self.ROLE_RULES:
+            if rule.role in self._compiled_patterns:
+                pattern_score = sum(
+                    1 for pattern in self._compiled_patterns[rule.role]
+                    if pattern.search(signal_content)
+                )
+                if pattern_score > 0:
+                    scores[rule.role] += pattern_score * 0.5
+                    details["scores"][f"{rule.role}_patterns"] = pattern_score
+
+        # 基于触发词匹配
+        for rule in self.ROLE_RULES:
+            trigger_score = sum(
+                1 for trigger in rule.triggers if trigger.lower() in content_lower
+            )
+            if trigger_score > 0:
+                scores[rule.role] += trigger_score * 0.7
+                details["scores"][f"{rule.role}_triggers"] = trigger_score
+
+        # 基于优先级加权
+        for rule in self.ROLE_RULES:
+            if scores[rule.role] > 0:
+                scores[rule.role] *= (1 + rule.priority * 0.01)
+
+        # 3. 找出最高分的角色
+        max_role = Role.AI_ENGINEER  # 默认 AI 工程师
+        max_score = 0.0
+
+        for role, score in scores.items():
+            if score > max_score:
+                max_score = score
+                max_role = role
+            elif score == max_score and score > 0:
+                # 分数相同时，优先级高的胜出
+                for rule in self.ROLE_RULES:
+                    if rule.role == role:
+                        role_priority = rule.priority
+                    if rule.role == max_role:
+                        max_role_priority = rule.priority
+                if role_priority > max_role_priority:
+                    max_role = role
+
+        # 归一化置信度
+        confidence = min(max_score / 3.0, 1.0) if max_score > 0 else 0.5
+
+        # 记录详情
+        details["matched_rules"] = [
+            rule.name for rule in self.ROLE_RULES if scores[rule.role] > 0
+        ]
+        details["final_scores"] = {role.value: score for role, score in scores.items()}
+        details["reasons"].append(f"基于关键词和模式匹配")
+
+        return max_role, confidence, details
+
+    def _should_escalate_to_boss(self, analysis: Dict[str, Any]) -> bool:
+        """检查是否需要升级到老板决策"""
+        # 检查开发成本
+        dev_days = analysis.get("resource_cost", {}).get("dev_days", 0)
+        if dev_days >= 5:
+            return True
+
+        # 检查风险等级
+        risk_level = analysis.get("risk_level", "")
+        if risk_level in ["CRITICAL", "HIGH"]:
+            return True
+
+        # 检查是否涉及架构变更
+        is_architecture = analysis.get("is_architecture_change", False)
+        if is_architecture:
+            return True
+
+        return False
+
+    def get_role_info(self, role: Role) -> Dict[str, Any]:
+        """获取角色信息"""
+        for rule in self.ROLE_RULES:
+            if rule.role == role:
+                return {
+                    "role": role.value,
+                    "name": rule.name,
+                    "description": rule.description,
+                    "keywords": rule.keywords,
+                    "priority": rule.priority,
+                }
+        return {}
+
+    def get_all_roles(self) -> List[Dict[str, Any]]:
+        """获取所有角色信息"""
+        return [
+            {
+                "role": rule.role.value,
+                "name": rule.name,
+                "description": rule.description,
+                "priority": rule.priority,
+            }
+            for rule in sorted(self.ROLE_RULES, key=lambda r: -r.priority)
+        ]
+
+
+class RoutingResult(BaseModel):
+    """路由结果模型"""
+
+    role: str
+    confidence: float
+    details: Dict[str, Any]
+    signal_id: Optional[str] = None
+    timestamp: float = Field(default_factory=lambda: datetime.utcnow().timestamp())
+
+
+# 全局路由器实例
+_global_router: Optional[RoleRouter] = None
+
+
+def get_router() -> RoleRouter:
+    """获取全局路由器实例"""
+    global _global_router
+    if _global_router is None:
+        _global_router = RoleRouter()
+    return _global_router
+
+
+# 便捷函数
+def route_signal(
+    signal_content: str,
+    analysis: Optional[Dict[str, Any]] = None,
+    signal_id: Optional[str] = None,
+) -> RoutingResult:
+    """
+    路由信号到对应角色
+
+    Args:
+        signal_content: 信号内容
+        analysis: AI 分析结果
+        signal_id: 信号 ID
+
+    Returns:
+        路由结果
+    """
+    router = get_router()
+    role, confidence, details = router.route(signal_content, analysis)
+
+    return RoutingResult(
+        role=role.value,
+        confidence=confidence,
+        details=details,
+        signal_id=signal_id,
+    )
diff --git a/backend/app/events.py b/backend/app/events.py
new file mode 100644
index 00000000..fdf55e86
--- /dev/null
+++ b/backend/app/events.py
@@ -0,0 +1,372 @@
+"""
+Neural Sync - Event Bus System
+
+实现事件驱动架构的核心组件，支持异步事件发布/订阅模式。
+"""
+
+from enum import Enum
+from dataclasses import dataclass, field
+from datetime import datetime
+from typing import Callable, Dict, List, Any, Optional
+import asyncio
+import logging
+from contextlib import asynccontextmanager
+
+logger = logging.getLogger(__name__)
+
+
+class EventType(Enum):
+    """系统事件类型定义"""
+
+    # 信号事件
+    SIGNAL_INGESTED = "signal.ingested"
+    SIGNAL_ANALYZED = "signal.analyzed"
+    SIGNAL_FEEDBACK = "signal.feedback"
+    SIGNAL_PRIORITY_CHANGED = "signal.priority_changed"
+
+    # 任务事件
+    TASK_ASSIGNED = "task.assigned"
+    TASK_STARTED = "task.started"
+    TASK_COMPLETED = "task.completed"
+    TASK_FAILED = "task.failed"
+
+    # 代码事件
+    CODE_COMMIT = "code.commit"
+    CODE_REVIEW_REQUESTED = "code.review_requested"
+    CODE_DRAFT_GENERATED = "code.draft_generated"
+
+    # 审计事件
+    AUDIT_STARTED = "audit.started"
+    AUDIT_COMPLETED = "audit.completed"
+    AUDIT_FAILED = "audit.failed"
+
+    # 合并事件
+    AUTO_MERGE_REQUESTED = "auto_merge.requested"
+    AUTO_MERGE_SUCCESS = "auto_merge.success"
+    AUTO_MERGE_FAILED = "auto_merge.failed"
+    AUTO_MERGE_COOLDOWN = "auto_merge.cooldown"
+
+    # 集成事件
+    PLANE_SYNC_SUCCESS = "integration.plane.success"
+    PLANE_SYNC_FAILED = "integration.plane.failed"
+    LARK_NOTIFY_SUCCESS = "integration.lark.success"
+    LARK_NOTIFY_FAILED = "integration.lark.failed"
+
+    # 项目域事件
+    DOMAIN_CREATED = "domain.created"
+    DOMAIN_UPDATED = "domain.updated"
+    DOMAIN_DELETED = "domain.deleted"
+    DOMAIN_SWITCHED = "domain.switched"
+    DOMAIN_SYNCING = "domain.syncing"
+    DOMAIN_SYNCED = "domain.synced"
+
+    # 项目事件
+    PROJECT_CREATED = "project.created"
+    PROJECT_UPDATED = "project.updated"
+    PROJECT_DELETED = "project.deleted"
+    SIGNAL_LINKED = "signal.linked"
+    SIGNAL_UNLINKED = "signal.unlinked"
+
+
+@dataclass
+class Event:
+    """事件数据结构"""
+
+    type: EventType
+    data: Dict[str, Any]
+    timestamp: float = field(default_factory=lambda: datetime.utcnow().timestamp())
+    source: Optional[str] = None
+    correlation_id: Optional[str] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+    def to_dict(self) -> Dict[str, Any]:
+        """转换为字典格式"""
+        return {
+            "type": self.type.value,
+            "data": self.data,
+            "timestamp": self.timestamp,
+            "source": self.source,
+            "correlation_id": self.correlation_id,
+            "metadata": self.metadata,
+        }
+
+
+class EventBus:
+    """
+    事件总线 - 核心事件分发系统
+
+    特性:
+    - 异步事件发布
+    - 多订阅者支持
+    - 事件过滤
+    - 错误处理
+    - 性能监控
+    """
+
+    def __init__(self):
+        self._subscribers: Dict[EventType, List[Callable]] = {}
+        self._wildcard_subscribers: List[Callable] = []
+        self._event_history: List[Event] = []
+        self._max_history = 1000
+        self._metrics = {
+            "published": 0,
+            "delivered": 0,
+            "failed": 0,
+        }
+        self._lock = asyncio.Lock()
+        self._running = False
+
+    @property
+    def metrics(self) -> Dict[str, int]:
+        """获取事件指标"""
+        return self._metrics.copy()
+
+    @property
+    def history(self) -> List[Dict[str, Any]]:
+        """获取事件历史"""
+        return [e.to_dict() for e in self._event_history]
+
+    async def start(self):
+        """启动事件总线"""
+        self._running = True
+        logger.info("EventBus started")
+
+    async def stop(self):
+        """停止事件总线"""
+        self._running = False
+        logger.info("EventBus stopped")
+
+    def subscribe(
+        self,
+        event_type: EventType,
+        handler: Callable,
+        filter_func: Optional[Callable[[Event], bool]] = None,
+    ) -> Callable:
+        """
+        订阅事件
+
+        Args:
+            event_type: 事件类型
+            handler: 事件处理函数 (async def handler(event: Event) -> None)
+            filter_func: 事件过滤函数 (def filter(event: Event) -> bool)
+
+        Returns:
+            取消订阅的函数
+        """
+
+        async def wrapped_handler(event: Event):
+            if filter_func and not filter_func(event):
+                return
+            try:
+                await handler(event)
+                self._metrics["delivered"] += 1
+            except Exception as e:
+                logger.error(
+                    f"Error in event handler for {event_type.value}: {e}",
+                    exc_info=True,
+                )
+                self._metrics["failed"] += 1
+
+        # 包装为原始 handler 用于取消订阅
+        wrapped_handler.original = handler  # type: ignore
+
+        async def _subscribe():
+            async with self._lock:
+                if event_type not in self._subscribers:
+                    self._subscribers[event_type] = []
+                self._subscribers[event_type].append(wrapped_handler)
+
+        asyncio.create_task(_subscribe())
+
+        def unsubscribe():
+            asyncio.create_task(self._unsubscribe(event_type, wrapped_handler))
+
+        return unsubscribe
+
+    def subscribe_wildcard(
+        self,
+        handler: Callable,
+        filter_func: Optional[Callable[[Event], bool]] = None,
+    ) -> Callable:
+        """订阅所有事件（通配符订阅）"""
+
+        async def wrapped_handler(event: Event):
+            if filter_func and not filter_func(event):
+                return
+            try:
+                await handler(event)
+                self._metrics["delivered"] += 1
+            except Exception as e:
+                logger.error(
+                    f"Error in wildcard event handler: {e}",
+                    exc_info=True,
+                )
+                self._metrics["failed"] += 1
+
+        wrapped_handler.original = handler  # type: ignore
+
+        async def _subscribe():
+            async with self._lock:
+                self._wildcard_subscribers.append(wrapped_handler)
+
+        asyncio.create_task(_subscribe())
+
+        def unsubscribe():
+            asyncio.create_task(self._unsubscribe_wildcard(wrapped_handler))
+
+        return unsubscribe
+
+    async def _unsubscribe(self, event_type: EventType, handler: Callable):
+        """取消订阅"""
+        async with self._lock:
+            if event_type in self._subscribers:
+                self._subscribers[event_type] = [
+                    h for h in self._subscribers[event_type] if h != handler
+                ]
+
+    async def _unsubscribe_wildcard(self, handler: Callable):
+        """取消通配符订阅"""
+        async with self._lock:
+            self._wildcard_subscribers = [
+                h for h in self._wildcard_subscribers if h != handler
+            ]
+
+    async def publish(self, event: Event) -> bool:
+        """
+        发布事件
+
+        Args:
+            event: 事件对象
+
+        Returns:
+            是否成功发布
+        """
+        if not self._running:
+            logger.warning(f"EventBus not running, discarding event: {event.type.value}")
+            return False
+
+        self._metrics["published"] += 1
+
+        # 记录历史
+        self._event_history.append(event)
+        if len(self._event_history) > self._max_history:
+            self._event_history.pop(0)
+
+        # 通知订阅者
+        tasks = []
+
+        # 特定事件订阅者
+        if event.type in self._subscribers:
+            for handler in self._subscribers[event.type]:
+                tasks.append(handler(event))
+
+        # 通配符订阅者
+        for handler in self._wildcard_subscribers:
+            tasks.append(handler(event))
+
+        if tasks:
+            results = await asyncio.gather(*tasks, return_exceptions=True)
+            failed = sum(1 for r in results if isinstance(r, Exception))
+            if failed > 0:
+                self._metrics["failed"] += failed
+
+        logger.debug(
+            f"Published event {event.type.value} to {len(tasks)} subscribers"
+        )
+        return True
+
+    async def publish_and_wait(
+        self, event: Event, timeout: float = 5.0
+    ) -> List[Any]:
+        """
+        发布事件并等待所有处理完成
+
+        Args:
+            event: 事件对象
+            timeout: 超时时间（秒）
+
+        Returns:
+            所有处理器的返回值列表
+        """
+        if not self._running:
+            raise RuntimeError("EventBus not running")
+
+        self._metrics["published"] += 1
+        self._event_history.append(event)
+
+        tasks = []
+
+        if event.type in self._subscribers:
+            tasks.extend(self._subscribers[event.type])
+
+        for handler in self._wildcard_subscribers:
+            tasks.append(handler(event))
+
+        if not tasks:
+            return []
+
+        try:
+            results = await asyncio.wait_for(
+                asyncio.gather(*tasks, return_exceptions=True),
+                timeout=timeout,
+            )
+            return results
+        except asyncio.TimeoutError:
+            logger.warning(
+                f"Event {event.type.value} processing timed out after {timeout}s"
+            )
+            raise
+
+    def clear_history(self):
+        """清空事件历史"""
+        self._event_history.clear()
+
+    def get_subscriber_count(self, event_type: Optional[EventType] = None) -> int:
+        """获取订阅者数量"""
+        if event_type:
+            return len(self._subscribers.get(event_type, []))
+        return sum(len(subs) for subs in self._subscribers.values()) + len(
+            self._wildcard_subscribers
+        )
+
+
+# 全局事件总线实例
+_global_event_bus: Optional[EventBus] = None
+
+
+def get_event_bus() -> EventBus:
+    """获取全局事件总线实例"""
+    global _global_event_bus
+    if _global_event_bus is None:
+        _global_event_bus = EventBus()
+    return _global_event_bus
+
+
+@asynccontextmanager
+async def event_bus_context():
+    """事件总线上下文管理器"""
+    bus = get_event_bus()
+    await bus.start()
+    try:
+        yield bus
+    finally:
+        await bus.stop()
+
+
+# 便捷函数
+async def publish_event(event_type: EventType, data: Dict[str, Any], **kwargs) -> bool:
+    """发布事件的便捷函数"""
+    event = Event(type=event_type, data=data, **kwargs)
+    return await get_event_bus().publish(event)
+
+
+def subscribe_event(
+    event_type: EventType, handler: Callable
+) -> Callable:
+    """订阅事件的便捷函数"""
+    return get_event_bus().subscribe(event_type, handler)
+
+
+def subscribe_all_events(handler: Callable) -> Callable:
+    """订阅所有事件的便捷函数"""
+    return get_event_bus().subscribe_wildcard(handler)
diff --git a/backend/app/seed_projects.py b/backend/app/seed_projects.py
new file mode 100644
index 00000000..9be6f4a0
--- /dev/null
+++ b/backend/app/seed_projects.py
@@ -0,0 +1,401 @@
+"""
+Seed Projects and Signals
+
+初始化示例项目和信号数据（新模型：BigInteger id）
+含 SSTC 宪章：Pixelle-Video 项目关联宪章与任务
+"""
+
+import asyncio
+from datetime import datetime
+from typing import Optional
+from sqlalchemy import func, select, or_, text
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import sessionmaker
+
+from app.database import _is_sqlite
+from app.models import Signal, Project, ProjectSignalLink, TruthNode, Task, User
+from app.services.auth import AuthService
+
+PIXELLE_NAME = "Pixelle-Video 集成"
+
+
+async def seed_pixelle_project(db: Optional[AsyncSession] = None):
+    """
+    创建或恢复 Pixelle-Video 示例项目：无项目则全量创建；有项目但无宪章/无信号链接则补全 SSTC 数据。
+    """
+    from app.database import engine
+
+    if db is None:
+        async_session_maker = sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
+        async with async_session_maker() as session:
+            await _seed_projects(session)
+    else:
+        await _seed_projects(db)
+
+
+async def _ensure_pixelle_sstc_data(session: AsyncSession, pixelle_project: Project) -> bool:
+    """
+    若「Pixelle-Video 集成」已存在但无宪章或无项目-信号链接，则补全信号、链接、宪章与任务。返回是否做了变更。
+    """
+    import json
+
+    pid = pixelle_project.id
+    # 是否已有项目-信号链接
+    link_count = await session.execute(
+        select(func.count()).select_from(ProjectSignalLink).where(ProjectSignalLink.project_id == pid)
+    )
+    has_links = (link_count.scalar() or 0) > 0
+    # 是否已有宪章
+    charter_count = await session.execute(
+        select(func.count()).select_from(TruthNode).where(
+            TruthNode.project_id == pid, TruthNode.node_type == "charter"
+        )
+    )
+    has_charters = (charter_count.scalar() or 0) > 0
+    if has_links and has_charters:
+        return False
+
+    changed = False
+    pixelle_signals = []
+    if not has_links:
+        pixelle_signals = [
+            Signal(
+                source="HackerNews",
+                content="Pixelle-Video: An AI-powered automated short video generation platform with script-to-video automation",
+                impact_score=9.2,
+                priority="critical",
+                processed=True,
+            ),
+            Signal(
+                source="ProductHunt",
+                content="Pixelle-Video launches API for programmatic video content generation",
+                impact_score=8.5,
+                priority="high",
+                processed=True,
+            ),
+            Signal(
+                source="TechCrunch",
+                content="Pixelle-Video raises Series A funding to scale AI video infrastructure",
+                impact_score=7.8,
+                priority="high",
+                processed=True,
+            ),
+        ]
+        for signal in pixelle_signals:
+            signal.strategic_analysis = json.dumps({
+                "decision_proposal": {
+                    "is_strategic": True,
+                    "threat_level": "HIGH",
+                    "lead_time_days": 14,
+                    "impact_score": getattr(signal, "impact_score", 0),
+                    "recommendation": "立即集成 Pixelle-Video API 到主平台"
+                },
+                "competitiveness_radar": {
+                    "innovation": 85,
+                    "market_potential": 90,
+                    "technical_complexity": 75,
+                    "resource_intensity": 70
+                },
+                "detected_tech_stacks": ["AI/ML", "Video Processing", "REST API"],
+                "related_knowledge_nodes": [],
+                "rag_context_used": 0
+            })
+            signal.tasks_generated = json.dumps([
+                {"task_name": "Integrate Pixelle-Video API", "type": "backend", "estimated_hours": 8.0},
+                {"task_name": "Design Video Workflow UI", "type": "frontend", "estimated_hours": 12.0},
+                {"task_name": "Optimize Video Pipeline", "type": "ai_engineer", "estimated_hours": 16.0},
+            ])
+        if _is_sqlite():
+            next_id = (await session.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM signals"))).scalar() or 1
+            for i, s in enumerate(pixelle_signals):
+                s.id = next_id + i
+        for signal in pixelle_signals:
+            session.add(signal)
+        await session.flush()
+        next_link_id = None
+        if _is_sqlite():
+            next_link_id = (await session.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM project_signal_links"))).scalar() or 1
+        for i, signal in enumerate(pixelle_signals):
+            link = ProjectSignalLink(
+                project_id=pid,
+                signal_id=signal.id,
+                link_type="related",
+                relevance_score=1.0,
+            )
+            if _is_sqlite() and next_link_id is not None:
+                link.id = next_link_id + i
+            session.add(link)
+        pixelle_project.signal_count = len(pixelle_signals)
+        changed = True
+        await session.flush()
+
+    if not has_charters:
+        user_result = await session.execute(select(User).limit(1))
+        seed_user = user_result.scalar_one_or_none()
+        if not seed_user:
+            seed_user = User(
+                username="boss",
+                email="boss@pmflow.local",
+                password_hash=AuthService.hash_password("boss123"),
+                name="负责人",
+                role="BOSS",
+                status="active",
+            )
+            session.add(seed_user)
+            await session.flush()
+        if pixelle_signals:
+            signal_ids = [s.id for s in pixelle_signals]
+        else:
+            link_rows = await session.execute(
+                select(ProjectSignalLink.signal_id).where(ProjectSignalLink.project_id == pid)
+            )
+            signal_ids = [r[0] for r in link_rows.fetchall()]
+        charter = TruthNode(
+            node_type="charter",
+            parent_id=None,
+            project_id=pid,
+            created_by=seed_user.id,
+            title="Pixelle-Video 集成需求宪章",
+            content={
+                "rationale": "基于 HackerNews/ProductHunt/TechCrunch 信号，决定集成 Pixelle-Video API 到主平台",
+                "signal_ids": signal_ids,
+                "decision": "立即集成 Pixelle-Video API，支持脚本到视频的自动化生成",
+            },
+            version=1,
+            status="aligned",
+        )
+        if _is_sqlite():
+            charter.id = (await session.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM truth_nodes"))).scalar() or 1
+        session.add(charter)
+        await session.flush()
+        next_task_id = None
+        if _is_sqlite():
+            next_task_id = (await session.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM tasks"))).scalar() or 1
+        for idx, task_item in enumerate([
+            {"title": "Integrate Pixelle-Video API", "priority": "high", "hours": 8.0},
+            {"title": "Design Video Workflow UI", "priority": "medium", "hours": 12.0},
+            {"title": "Optimize Video Pipeline", "priority": "medium", "hours": 16.0},
+        ]):
+            task = Task(
+                truth_node_id=charter.id,
+                project_id=pid,
+                title=task_item["title"],
+                required_skill_tag_ids=[],
+                status="todo",
+                priority=task_item["priority"],
+                estimated_hours=task_item["hours"],
+            )
+            if _is_sqlite() and next_task_id is not None:
+                task.id = next_task_id + idx
+            session.add(task)
+        changed = True
+        print(f"[SSTC] 已为「{PIXELLE_NAME}」恢复宪章 1 个、任务 3 个")
+    return changed
+
+
+async def _seed_projects(session: AsyncSession):
+    import json
+
+    result = await session.execute(
+        select(Project).where(
+            or_(Project.title == PIXELLE_NAME, Project.name == PIXELLE_NAME)
+        )
+    )
+    existing = result.scalar_one_or_none()
+    if existing:
+        restored = await _ensure_pixelle_sstc_data(session, existing)
+        if restored:
+            await session.commit()
+        else:
+            print("示例项目已存在且 SSTC 数据完整，跳过")
+        return
+
+    pixelle_signals = [
+        Signal(
+            source="HackerNews",
+            content="Pixelle-Video: An AI-powered automated short video generation platform with script-to-video automation",
+            impact_score=9.2,
+            priority="critical",
+            processed=True,
+        ),
+        Signal(
+            source="ProductHunt",
+            content="Pixelle-Video launches API for programmatic video content generation",
+            impact_score=8.5,
+            priority="high",
+            processed=True,
+        ),
+        Signal(
+            source="TechCrunch",
+            content="Pixelle-Video raises Series A funding to scale AI video infrastructure",
+            impact_score=7.8,
+            priority="high",
+            processed=True,
+        ),
+    ]
+
+    for signal in pixelle_signals:
+        signal.strategic_analysis = json.dumps({
+            "decision_proposal": {
+                "is_strategic": True,
+                "threat_level": "HIGH",
+                "lead_time_days": 14,
+                "impact_score": signal.impact_score,
+                "recommendation": "立即集成 Pixelle-Video API 到主平台"
+            },
+            "competitiveness_radar": {
+                "innovation": 85,
+                "market_potential": 90,
+                "technical_complexity": 75,
+                "resource_intensity": 70
+            },
+            "detected_tech_stacks": ["AI/ML", "Video Processing", "REST API"],
+            "related_knowledge_nodes": [],
+            "rag_context_used": 0
+        })
+        signal.tasks_generated = json.dumps([
+            {"task_name": "Integrate Pixelle-Video API", "type": "backend", "estimated_hours": 8.0},
+            {"task_name": "Design Video Workflow UI", "type": "frontend", "estimated_hours": 12.0},
+            {"task_name": "Optimize Video Pipeline", "type": "ai_engineer", "estimated_hours": 16.0},
+        ])
+        session.add(signal)
+
+    await session.flush()
+
+    pixelle_project = Project(
+        name="Pixelle-Video 集成",
+        title="Pixelle-Video 集成",
+        description="集成 Pixelle-Video AI 视频生成平台到主系统",
+        status="active",
+        assigned_role="backend",
+        signal_count=len(pixelle_signals),
+    )
+    session.add(pixelle_project)
+    await session.flush()
+
+    for signal in pixelle_signals:
+        link = ProjectSignalLink(
+            project_id=pixelle_project.id,
+            signal_id=signal.id,
+            link_type="related",
+            relevance_score=1.0,
+        )
+        session.add(link)
+
+    ai_analyst_project = Project(
+        name="AI 分析引擎优化",
+        title="AI 分析引擎优化",
+        description="优化 LLM 服务性能，实现多模型投票机制和 RAG 检索增强",
+        status="active",
+        assigned_role="ai_engineer",
+        signal_count=1,
+    )
+    session.add(ai_analyst_project)
+    await session.flush()
+
+    ai_signal = Signal(
+        source="ArXiv",
+        content="New paper: Multi-Agent LLM Voting System improves accuracy by 23%",
+        impact_score=7.5,
+        priority="medium",
+        processed=True,
+    )
+    ai_signal.strategic_analysis = json.dumps({
+        "decision_proposal": {
+            "is_strategic": True,
+            "threat_level": "MEDIUM",
+            "lead_time_days": 30,
+            "impact_score": 7.5,
+            "recommendation": "实现多模型投票机制提升 AI 分析准确性"
+        },
+        "detected_tech_stacks": ["LLM", "Voting", "Multi-Agent"]
+    })
+    session.add(ai_signal)
+    await session.flush()
+
+    ai_link = ProjectSignalLink(
+        project_id=ai_analyst_project.id,
+        signal_id=ai_signal.id,
+        link_type="related",
+        relevance_score=0.9,
+    )
+    session.add(ai_link)
+
+    await session.flush()
+
+    # --- SSTC: 获取或创建默认用户，用于宪章 created_by ---
+    user_result = await session.execute(select(User).limit(1))
+    seed_user = user_result.scalar_one_or_none()
+    if not seed_user:
+        seed_user = User(
+            username="boss",
+            email="boss@pmflow.local",
+            password_hash=AuthService.hash_password("boss123"),
+            name="负责人",
+            role="BOSS",
+            status="active",
+        )
+        session.add(seed_user)
+        await session.flush()
+
+    # --- SSTC: 为 Pixelle-Video 创建宪章（关联信号）---
+    pixelle_signal_ids = [s.id for s in pixelle_signals]
+    charter = TruthNode(
+        node_type="charter",
+        parent_id=None,
+        project_id=pixelle_project.id,
+        created_by=seed_user.id,
+        title="Pixelle-Video 集成需求宪章",
+        content={
+            "rationale": "基于 HackerNews/ProductHunt/TechCrunch 信号，决定集成 Pixelle-Video API 到主平台",
+            "signal_ids": pixelle_signal_ids,
+            "decision": "立即集成 Pixelle-Video API，支持脚本到视频的自动化生成",
+        },
+        version=1,
+        status="aligned",
+    )
+    session.add(charter)
+    await session.flush()
+
+    # --- SSTC: 创建任务（绑定到宪章）---
+    for task_item in [
+        {"title": "Integrate Pixelle-Video API", "priority": "high", "hours": 8.0},
+        {"title": "Design Video Workflow UI", "priority": "medium", "hours": 12.0},
+        {"title": "Optimize Video Pipeline", "priority": "medium", "hours": 16.0},
+    ]:
+        task = Task(
+            truth_node_id=charter.id,
+            project_id=pixelle_project.id,
+            title=task_item["title"],
+            required_skill_tag_ids=[],
+            status="todo",
+            priority=task_item["priority"],
+            estimated_hours=task_item["hours"],
+        )
+        session.add(task)
+
+    await session.commit()
+
+    print(f"创建示例项目:")
+    print(f"  - {pixelle_project.title} ({pixelle_project.assigned_role})")
+    print(f"  - {ai_analyst_project.title} ({ai_analyst_project.assigned_role})")
+    print(f"\n关联信号: {len(pixelle_signals) + 1} 条")
+    print(f"[SSTC] Pixelle-Video 宪章 1 个，任务 3 个")
+
+
+async def main():
+    print("开始初始化示例数据...")
+
+    from app.database import engine
+    from app.models import Base
+
+    async with engine.begin() as conn:
+        await conn.run_sync(Base.metadata.create_all)
+
+    await seed_pixelle_project()
+
+    print("\n初始化完成!")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/backend/app/services/adaptive_rag_service.py b/backend/app/services/adaptive_rag_service.py
new file mode 100644
index 00000000..e1fb526d
--- /dev/null
+++ b/backend/app/services/adaptive_rag_service.py
@@ -0,0 +1,6598 @@
+"""全自适应 RAG 服务。
+
+能力：
+1. 意图识别（question / statement / mixed）
+2. 多域路由（结构化 DB + 向量检索）
+3. 回答生成与来源引用
+4. QA 记录、反馈、JSONL 导出
+5. 知识库文档/URL/飞书消息入库向量化
+"""
+
+from __future__ import annotations
+
+import ast
+import asyncio
+import hashlib
+import json
+import math
+import os
+from pathlib import Path
+import re
+import time
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+import httpx
+from sqlalchemy import and_, delete, func, or_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import PROMPT_RAG_ANSWER, PROMPT_RAG_DB_PROJECT_QA, PROMPT_RAG_INTENT_CLASSIFICATION
+from ..database import AsyncSessionLocal
+from ..models import (
+    KnowledgeBaseChunk,
+    KnowledgeBaseDoc,
+    KGEvidence,
+    KGEntity,
+    KGRelation,
+    Project,
+    ProjectSignalLink,
+    QAFeedback,
+    QARecord,
+    QASession,
+    QASessionState,
+    Signal,
+    Task,
+    TruthNode,
+    User,
+    UserProject,
+    UserWorkload,
+)
+from .context_cache import context_cache_service
+from .llm import LLMService
+from .signal_source_orchestration_service import SignalSourceOrchestrationService
+from .symbol_index_service import SymbolIndexService
+
+
+def _env_int(name: str, default: int, *, minimum: int, maximum: int) -> int:
+    raw = str(os.getenv(name, "") or "").strip()
+    if not raw:
+        value = int(default)
+    else:
+        try:
+            value = int(raw)
+        except Exception:
+            value = int(default)
+    return max(minimum, min(maximum, value))
+
+QUESTION_HINTS = (
+    "吗",
+    "么",
+    "呢",
+    "呀",
+    "啊",
+    "？",
+    "?",
+    "如何",
+    "怎么",
+    "为什么",
+    "为何",
+    "谁",
+    "哪里",
+    "有没有",
+    "有无",
+    "有吗",
+    "是否",
+    "是否有",
+    "是否存在",
+    "哪些",
+    "哪个",
+    "什么",
+    "是什么",
+    "多少",
+    "几",
+)
+PROJECT_EXISTENCE_RE = re.compile(r"(有没有|是否有|有).*?(项目|系统)|项目.*?(有没有|是否有|有).*(吗|\?)")
+PROJECT_PROGRESS_RE = re.compile(r"(进度|阶段|完成度|完成率|推进|里程碑|状态)")
+QUESTION_NOISE_RE = re.compile(
+    r"(目前|当前|现在|请问|一下|系统中|项目中|系统内|项目内|是否|有没有|有无|有|吗|呢|呀|啊|什么|哪个|多少|查询|看看)"
+)
+STATEMENT_HINTS = (
+    "需要",
+    "请",
+    "新增",
+    "优化",
+    "修复",
+    "安排",
+    "立项",
+    "开发",
+    "上线",
+)
+PROJECT_KEYWORDS = (
+    "项目",
+    "信号",
+    "任务",
+    "宪章",
+    "进度",
+    "状态",
+    "工作量",
+    "负责人",
+)
+PEOPLE_KEYWORDS = (
+    "成员",
+    "人员",
+    "同事",
+    "负责人",
+    "owner",
+    "assignee",
+    "谁负责",
+    "谁在做",
+    "名单",
+)
+DOC_KEYWORDS = (
+    "文档",
+    "规范",
+    "spec",
+    "说明",
+    "流程",
+    "模板",
+    "手册",
+    "how to",
+    "怎么做",
+)
+FEATURE_QUERY_KEYWORDS = (
+    "功能",
+    "用途",
+    "是干嘛",
+    "做什么",
+    "有什么用",
+    "页面",
+    "菜单",
+    "入口",
+    "在哪",
+    "ai配置",
+    "智能配置",
+    "配置中心",
+    "能力管理",
+    "用户管理",
+    "项目列表",
+    "团队负载",
+    "ai管家",
+)
+FEATURE_PURPOSE_HINTS = (
+    "用途",
+    "作用",
+    "干嘛",
+    "做什么",
+    "是什么",
+    "入口",
+    "在哪",
+    "页面",
+)
+CORRECTION_CUE_KEYWORDS = (
+    "我说的是",
+    "我说的不是",
+    "不是这个",
+    "不是系统架构",
+    "不是啊",
+    "不对",
+    "错了",
+    "答错",
+    "跑偏",
+    "偏了",
+    "不是这个意思",
+    "你理解错",
+    "你答偏了",
+    "答非所问",
+    "请纠正",
+    "纠正一下",
+    "请更正",
+    "上一条",
+    "上一个回答",
+    "不要模板",
+    "模板化",
+    "我要的是",
+    "我问的是",
+    "不是在问",
+)
+ENGINE_PROCESS_KEYWORDS = (
+    "ai引擎",
+    "引擎驱动",
+    "驱动方式",
+    "调用链",
+    "调用流程",
+    "调用路径",
+    "模型路由",
+    "模型调用",
+    "路由策略",
+    "openclaw",
+    "网关",
+    "ai架构",
+    "ai 管家",
+    "ai管家",
+    "第一层",
+    "第二层",
+    "第三层",
+    "fallback",
+    "兜底",
+)
+DEEP_EXPLICIT_KEYWORDS = (
+    "深入",
+    "深度",
+    "详细分析",
+    "证据链",
+    "调用链",
+    "逐层",
+    "逐步",
+    "详细展开",
+    "deep",
+)
+BRIEF_NEGATIVE_FEEDBACK_RE = re.compile(
+    r"^(不是啊?|不对|错了|答错了?|跑偏了?|偏了|不是这个(意思)?|不(是|对)啊)$"
+)
+HISTORY_KEYWORDS = ("之前", "上次", "历史", "曾经", "之前说", "last time")
+PMFLOW_KEYWORDS = ("复盘", "charter", "宪章", "task", "任务描述")
+FOLLOWUP_REFERENTIAL_KEYWORDS = (
+    "这些",
+    "它们",
+    "这个",
+    "那个",
+    "上面",
+    "前面",
+    "刚才",
+    "分别",
+    "具体",
+    "都有哪些",
+    "分别是什么",
+)
+FOLLOWUP_ENUMERATION_KEYWORDS = (
+    "分别是什么",
+    "分别是啥",
+    "都有哪些",
+    "具体有哪些",
+    "列出来",
+    "清单",
+    "名单",
+)
+IDENTITY_KEYWORDS = (
+    "你是谁",
+    "你是",
+    "自我介绍",
+    "介绍你自己",
+    "你能做什么",
+    "你会什么",
+    "你负责什么",
+    "你叫什么",
+)
+ARCHITECTURE_KEYWORDS = (
+    "系统架构",
+    "架构",
+    "运行机制",
+    "技术栈",
+    "模块",
+    "服务组成",
+    "docker",
+    "数据库",
+    "redis",
+    "后端",
+    "前端",
+)
+CODE_KEYWORDS = (
+    "代码",
+    "函数",
+    "方法",
+    "类",
+    "模块",
+    "文件",
+    "api",
+    "接口实现",
+    "service",
+    "model",
+    "schema",
+    "where defined",
+)
+URL_RE = re.compile(r"https?://\S+", flags=re.IGNORECASE)
+SENTENCE_SPLIT_RE = re.compile(r"(?<=[。！？!?\n])")
+PAGE_MARKER_RE = re.compile(r"\[\[PAGE:(\d{1,6})\]\]")
+
+MAX_CONTEXT_TOKENS = 8000
+DEFAULT_TOP_K = 5
+DB_QUERY_TIMEOUT_SECONDS = 5.0
+VECTOR_SEARCH_TIMEOUT_SECONDS = 10.0
+GRAPH_QUERY_TIMEOUT_SECONDS = 4.0
+AGENT_MAX_RETRIEVAL_STEPS = 3
+SESSION_GAP_MINUTES = 30
+SESSION_DEFAULT_TITLE = "新会话"
+SESSION_TITLE_MIN_LEN = 5
+
+THINKING_MODE_NORMAL = "normal"
+THINKING_MODE_DEEP = "deep"
+DEEP_RETRIEVAL_STEPS = _env_int("RAG_DEEP_RETRIEVAL_STEPS", 5, minimum=2, maximum=8)
+DEEP_VECTOR_TOP_K = _env_int("RAG_DEEP_VECTOR_TOP_K", 12, minimum=6, maximum=40)
+DEEP_CONTEXT_BLOCKS = _env_int("RAG_DEEP_CONTEXT_BLOCKS", 16, minimum=8, maximum=32)
+DEEP_ANSWER_MAX_TOKENS = _env_int("RAG_DEEP_ANSWER_MAX_TOKENS", 1400, minimum=900, maximum=3000)
+
+QUERY_TYPE_COUNT = "count"
+QUERY_TYPE_LIST = "list"
+QUERY_TYPE_STATUS = "status"
+QUERY_TYPE_EXPLAIN = "explain"
+QUERY_TYPE_UNKNOWN = "unknown"
+
+ACTION_MODE_AUTO = "auto"
+ACTION_MODE_READ = "read"
+ACTION_MODE_WRITE = "write"
+ACTION_MODE_READ_WRITE = "read_write"
+
+REQUIREMENT_WRITE_VERB_KEYWORDS = (
+    "新增",
+    "增加",
+    "添加",
+    "接入",
+    "支持",
+    "实现",
+    "创建",
+    "新建",
+    "记录",
+    "入库",
+    "同步",
+    "补充",
+)
+REQUIREMENT_WRITE_OBJECT_KEYWORDS = (
+    "信号",
+    "需求",
+    "任务",
+    "项目",
+    "功能",
+    "接口",
+    "页面",
+    "按钮",
+    "规则",
+    "模块",
+    "卡路里",
+    "食谱",
+)
+WRITE_EXPRESSIONS = (
+    "给",
+    "帮我",
+    "请帮",
+    "请给",
+    "麻烦",
+)
+
+COUNT_HINTS = ("多少", "几个", "数量", "总数", "count")
+LIST_HINTS = ("哪些", "列表", "清单", "列出", "分别", "名单", "都有哪些")
+STATUS_HINTS = ("进度", "状态", "阶段", "完成率", "完成度", "情况")
+
+ENTITY_ALIAS_MAP: dict[str, str] = {
+    "能力": "skill",
+    "技能": "skill",
+    "ability": "skill",
+    "abilities": "skill",
+    "skill": "skill",
+    "skills": "skill",
+    "负责人": "owner",
+    "owner": "owner",
+    "assignee": "owner",
+    "成员": "member",
+    "人员": "member",
+    "member": "member",
+    "任务": "task",
+    "task": "task",
+    "tasks": "task",
+    "项目": "project",
+    "project": "project",
+    "projects": "project",
+    "信号": "signal",
+    "signal": "signal",
+    "signals": "signal",
+    "函数": "function",
+    "function": "function",
+    "method": "function",
+    "类": "class",
+    "class": "class",
+    "模块": "module",
+    "module": "module",
+    "智能配置": "ai_config",
+    "ai配置": "ai_config",
+    "配置中心": "ai_config",
+    "能力管理": "skills",
+    "技能管理": "skills",
+    "用户管理": "users",
+    "项目列表": "projects",
+    "团队负载": "workloads",
+    "ai管家": "ai_steward",
+}
+TS_SYMBOL_RE = re.compile(
+    r"(?:export\s+)?(?:async\s+)?(?:function|class|interface|type|const)\s+([A-Za-z_][A-Za-z0-9_]*)"
+)
+
+
+@dataclass(slots=True)
+class RetrievedItem:
+    source: str
+    content: str
+    score: float
+    source_label: str
+    metadata: dict[str, Any]
+
+
+class AdaptiveRAGService:
+    """项目内统一 RAG 服务。"""
+
+    _symbol_snapshot_cache: dict[str, tuple[float, dict[str, Any]]] = {}
+
+    def __init__(self, db: AsyncSession, *, tenant_id: int) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+        self.llm = LLMService()
+        self.context_cache = context_cache_service
+
+    @staticmethod
+    def _normalize_thinking_mode(value: Optional[str]) -> str:
+        text = str(value or "").strip().lower()
+        if text == THINKING_MODE_DEEP:
+            return THINKING_MODE_DEEP
+        return THINKING_MODE_NORMAL
+
+    @staticmethod
+    def _normalize_conversation_key(value: Optional[str]) -> str:
+        raw = str(value or "").strip()
+        if not raw:
+            return ""
+        cleaned = re.sub(r"\s+", "", raw.lower())
+        cleaned = re.sub(r"[^a-z0-9:_\-\.]", "", cleaned)
+        return cleaned[:160]
+
+    def _build_external_conversation_key(
+        self,
+        *,
+        asked_via: str,
+        source_chat_id: Optional[str],
+        conversation_key: Optional[str],
+    ) -> str:
+        explicit = self._normalize_conversation_key(conversation_key)
+        if explicit:
+            return explicit
+        channel = str(asked_via or "").strip().lower()
+        if channel == "feishu":
+            chat = str(source_chat_id or "").strip()
+            if chat:
+                return self._normalize_conversation_key(f"feishu:{chat}")
+        return ""
+
+    @staticmethod
+    def _external_session_id_from_key(conversation_key: str) -> Optional[int]:
+        key = str(conversation_key or "").strip()
+        if not key:
+            return None
+        digest = hashlib.sha1(key.encode("utf-8")).hexdigest()
+        bucket = int(digest[:10], 16) % 500_000_000
+        return 1_500_000_000 + bucket
+
+    # ------------------------------------------------------------------
+    # Intent
+    # ------------------------------------------------------------------
+    async def classify_intent(self, message: str) -> dict[str, Any]:
+        text = str(message or "").strip()
+        if not text:
+            return {
+                "intent": "statement",
+                "confidence": 1.0,
+                "question_part": "",
+                "statement_part": "",
+            }
+
+        heuristic = self._classify_intent_rule(text)
+        if heuristic["confidence"] >= 0.86:
+            return heuristic
+
+        prompt = PROMPT_RAG_INTENT_CLASSIFICATION.format(message=text[:1200]).strip()
+        try:
+            raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=self.tenant_id,
+                function_name="rag_intent_classification",
+                temperature=0.0,
+                max_tokens=220,
+                use_prompt_profile=False,
+            )
+            parsed = self._parse_json_object(raw)
+            intent = str(parsed.get("intent") or heuristic["intent"]).strip().lower()
+            if intent not in {"question", "statement", "mixed"}:
+                intent = heuristic["intent"]
+            confidence = self._to_float(parsed.get("confidence"), heuristic["confidence"])
+            question_part = str(parsed.get("question_part") or "").strip()
+            statement_part = str(parsed.get("statement_part") or "").strip()
+            if not question_part and intent in {"question", "mixed"}:
+                question_part = heuristic["question_part"]
+            if not statement_part and intent in {"statement", "mixed"}:
+                statement_part = heuristic["statement_part"]
+            return {
+                "intent": intent,
+                "confidence": max(0.0, min(1.0, confidence)),
+                "question_part": question_part,
+                "statement_part": statement_part,
+            }
+        except Exception:
+            return heuristic
+
+    def _classify_intent_rule(self, text: str) -> dict[str, Any]:
+        question_score = 0.0
+        statement_score = 0.0
+        lowered = text.lower()
+
+        if "?" in text or "？" in text:
+            question_score += 0.35
+        if lowered.endswith(("吗", "么", "呢", "呀", "啊")):
+            question_score += 0.24
+        for hint in QUESTION_HINTS:
+            if hint in lowered:
+                question_score += 0.08
+        if PROJECT_EXISTENCE_RE.search(text):
+            question_score += 0.26
+        if re.search(r"(现在|当前|目前|请问).*(项目|任务|负责人|状态|进度|列表|清单)", text):
+            question_score += 0.18
+        for hint in STATEMENT_HINTS:
+            if hint in lowered:
+                statement_score += 0.10
+
+        sentences = [s.strip() for s in SENTENCE_SPLIT_RE.split(text) if s.strip()]
+        question_sentences = [s for s in sentences if "?" in s or "？" in s or any(k in s.lower() for k in QUESTION_HINTS)]
+        statement_sentences = [s for s in sentences if s not in question_sentences]
+
+        if question_sentences and statement_sentences:
+            return {
+                "intent": "mixed",
+                "confidence": 0.88,
+                "question_part": " ".join(question_sentences)[:1000],
+                "statement_part": " ".join(statement_sentences)[:1000],
+            }
+
+        if question_score >= 0.28 and question_score >= statement_score:
+            return {
+                "intent": "question",
+                "confidence": min(0.9, 0.62 + question_score),
+                "question_part": text[:1000],
+                "statement_part": "",
+            }
+
+        return {
+            "intent": "statement",
+            "confidence": min(0.9, 0.66 + statement_score),
+            "question_part": "",
+            "statement_part": text[:1000],
+        }
+
+    # ------------------------------------------------------------------
+    # Public ask flow
+    # ------------------------------------------------------------------
+    async def answer_question(
+        self,
+        *,
+        question: str,
+        asked_by: Optional[int],
+        asked_via: str,
+        project_id: Optional[int] = None,
+        source_chat_id: Optional[str] = None,
+        session_id: Optional[int] = None,
+        conversation_key: Optional[str] = None,
+        force_new_session: bool = False,
+        thinking_mode: str = THINKING_MODE_NORMAL,
+        action_mode: str = ACTION_MODE_AUTO,
+        sync_requirement: bool = True,
+    ) -> dict[str, Any]:
+        started = time.perf_counter()
+        clean_q = str(question or "").strip()
+        external_conversation_key = self._build_external_conversation_key(
+            asked_via=asked_via,
+            source_chat_id=source_chat_id,
+            conversation_key=conversation_key,
+        )
+        active_thinking_mode = self._normalize_thinking_mode(thinking_mode)
+        deep_thinking = active_thinking_mode == THINKING_MODE_DEEP
+        asked_via_lower = str(asked_via or "").strip().lower()
+        deep_retrieval_steps = DEEP_RETRIEVAL_STEPS
+        deep_vector_top_k = DEEP_VECTOR_TOP_K
+        if asked_via_lower == "feishu":
+            # 飞书链路优先低时延，保留 deep 能力但收紧预算。
+            deep_retrieval_steps = min(DEEP_RETRIEVAL_STEPS, 3)
+            deep_vector_top_k = min(DEEP_VECTOR_TOP_K, 8)
+        if not clean_q:
+            return self._build_response_payload(
+                answer="问题为空，无法检索。",
+                sources=[],
+                domains_used=[],
+                response_time_ms=0,
+                db_query_ms=0,
+                vector_search_ms=0,
+                db_timeout=False,
+                vector_timeout=False,
+                qa_id=None,
+                session_id=None,
+                turn_index=1,
+                relevance_max=0.0,
+                query_type=QUERY_TYPE_UNKNOWN,
+                confidence=0.0,
+                missing_fields=["question"],
+                thinking_mode=active_thinking_mode,
+            )
+
+        session_id, turn_index = await self._touch_session_for_new_record(
+            asked_by=asked_by,
+            asked_via=asked_via,
+            project_id=project_id,
+            question=clean_q,
+            preferred_session_id=session_id,
+            force_new_session=bool(force_new_session),
+        )
+        context_session_id: Optional[int]
+        if session_id is not None:
+            context_session_id = int(session_id)
+        elif asked_by is None:
+            context_session_id = self._external_session_id_from_key(external_conversation_key)
+        else:
+            context_session_id = None
+        conversation = await self._load_recent_conversation(
+            asked_by=asked_by,
+            asked_via=asked_via,
+            project_id=project_id,
+            session_id=session_id,
+            external_session_id=context_session_id,
+            limit=10,
+        )
+        state_row = await self._ensure_session_state(session_id=session_id)
+        effective_q, rewrite_meta = self._rewrite_question_with_state(
+            question=clean_q,
+            conversation=conversation,
+            session_state=state_row,
+        )
+        query_type = self._classify_query_type(effective_q)
+        requested_action_mode = self._normalize_action_mode(action_mode)
+        effective_action_mode = self._resolve_action_mode(
+            requested_mode=requested_action_mode,
+            question=clean_q,
+            effective_question=effective_q,
+        )
+        if (
+            bool(sync_requirement)
+            and asked_via_lower != "feishu"
+            and effective_action_mode in {ACTION_MODE_WRITE, ACTION_MODE_READ_WRITE}
+        ):
+            ack_result = await self._handle_ack_first_requirement(
+                question=clean_q,
+                effective_question=effective_q,
+                asked_by=asked_by,
+                asked_via=asked_via,
+                project_id=project_id,
+                session_id=session_id,
+                turn_index=turn_index,
+                state_row=state_row,
+                rewrite_meta=rewrite_meta,
+                context_session_id=context_session_id,
+                query_type=query_type,
+                action_mode=effective_action_mode,
+                started_at=started,
+            )
+            if ack_result is not None:
+                return ack_result
+        forced_deep_reason = ""
+        if not deep_thinking:
+            forced_deep_reason = self._forced_deep_reason(
+                question=clean_q,
+                effective_question=effective_q,
+                conversation=conversation,
+            )
+            if forced_deep_reason and self._should_force_deep_for_channel(
+                asked_via=asked_via,
+                question=clean_q,
+                effective_question=effective_q,
+                reason=forced_deep_reason,
+            ):
+                deep_thinking = True
+                active_thinking_mode = THINKING_MODE_DEEP
+            else:
+                forced_deep_reason = ""
+
+        clarification_answer = self._build_clarification_question(
+            original_question=clean_q,
+            effective_question=effective_q,
+            session_state=state_row,
+        )
+        if clarification_answer:
+            elapsed = int((time.perf_counter() - started) * 1000)
+            clarify_sources = [
+                {
+                    "source": "system",
+                    "label": "Session State",
+                    "score": 0.95,
+                    "metadata": {"mode": "clarification", "reason": "missing_subject"},
+                }
+            ]
+            qa = QARecord(
+                tenant_id=self.tenant_id,
+                question=clean_q,
+                answer=clarification_answer,
+                sources=clarify_sources,
+                intent="question",
+                domains_used=["system_context"],
+                response_time_ms=elapsed,
+                asked_by=asked_by,
+                asked_via=(asked_via or "pmflow")[:20],
+                project_id=project_id,
+                session_id=session_id,
+                turn_index=turn_index,
+            )
+            self.db.add(qa)
+            await self._update_session_state_after_turn(
+                session_state=state_row,
+                question=clean_q,
+                effective_question=effective_q,
+                answer=clarification_answer,
+                query_type=query_type,
+                ambiguity="needs_clarification",
+                rewrite_meta=rewrite_meta,
+                sources=clarify_sources,
+            )
+            await self.db.commit()
+            await self.db.refresh(qa)
+            await self.context_cache.append_session_messages(
+                tenant_id=self.tenant_id,
+                session_id=context_session_id,
+                question=clean_q,
+                answer=clarification_answer,
+            )
+            return self._build_response_payload(
+                answer=clarification_answer,
+                sources=clarify_sources,
+                domains_used=["system_context"],
+                response_time_ms=elapsed,
+                db_query_ms=0,
+                vector_search_ms=0,
+                db_timeout=False,
+                vector_timeout=False,
+                qa_id=int(qa.id),
+                session_id=int(session_id) if session_id is not None else None,
+                turn_index=int(turn_index),
+                relevance_max=0.95,
+                query_type=query_type,
+                confidence=0.92,
+                missing_fields=["subject"],
+                thinking_mode=active_thinking_mode,
+            )
+
+        system_snapshot = await self.context_cache.get_system_snapshot(
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        user_snapshot = await self.context_cache.get_user_snapshot(
+            db=self.db,
+            tenant_id=self.tenant_id,
+            user_id=asked_by,
+        )
+
+        if not deep_thinking:
+            deterministic_result = await self._run_deterministic_query(
+                question=effective_q,
+                project_id=project_id,
+                query_type=query_type,
+            )
+            if deterministic_result is not None:
+                return await self._persist_fast_answer(
+                    question=clean_q,
+                    asked_by=asked_by,
+                    asked_via=asked_via,
+                    project_id=project_id,
+                    fast_result=deterministic_result,
+                    session_id=session_id,
+                    external_session_id=context_session_id,
+                    turn_index=turn_index,
+                    query_type=str(deterministic_result.get("query_type") or query_type),
+                    missing_fields=list(deterministic_result.get("missing_fields") or []),
+                    state_row=state_row,
+                    effective_question=effective_q,
+                    rewrite_meta=rewrite_meta,
+                    thinking_mode=active_thinking_mode,
+                )
+
+        direct_answer = None
+        if not deep_thinking:
+            direct_answer = self._build_snapshot_direct_answer(
+                effective_q,
+                system_snapshot=system_snapshot,
+                user_snapshot=user_snapshot,
+            )
+        if direct_answer:
+            elapsed = int((time.perf_counter() - started) * 1000)
+            direct_sources = [
+                {
+                    "source": "system",
+                    "label": "System Snapshot",
+                    "score": 0.99,
+                    "metadata": {"mode": "snapshot_direct", "topic": "identity_or_architecture"},
+                }
+            ]
+            if user_snapshot:
+                direct_sources.append(
+                    {
+                        "source": "system",
+                        "label": "User Snapshot",
+                        "score": 0.98,
+                        "metadata": {"mode": "snapshot_direct", "topic": "current_user"},
+                    }
+                )
+
+            qa = QARecord(
+                tenant_id=self.tenant_id,
+                question=clean_q,
+                answer=direct_answer,
+                sources=direct_sources,
+                intent="question",
+                domains_used=["system_context"],
+                response_time_ms=elapsed,
+                asked_by=asked_by,
+                asked_via=(asked_via or "pmflow")[:20],
+                project_id=project_id,
+                session_id=session_id,
+                turn_index=turn_index,
+            )
+            self.db.add(qa)
+            await self._update_session_state_after_turn(
+                session_state=state_row,
+                question=clean_q,
+                effective_question=effective_q,
+                answer=direct_answer,
+                query_type=QUERY_TYPE_EXPLAIN,
+                ambiguity="clear",
+                rewrite_meta=rewrite_meta,
+                sources=direct_sources,
+            )
+            await self.db.commit()
+            await self.db.refresh(qa)
+            await self.context_cache.append_session_messages(
+                tenant_id=self.tenant_id,
+                session_id=context_session_id,
+                question=clean_q,
+                answer=direct_answer,
+            )
+            return self._build_response_payload(
+                answer=direct_answer,
+                sources=direct_sources,
+                domains_used=["system_context"],
+                response_time_ms=elapsed,
+                db_query_ms=0,
+                vector_search_ms=0,
+                db_timeout=False,
+                vector_timeout=False,
+                qa_id=int(qa.id),
+                session_id=int(session_id) if session_id is not None else None,
+                turn_index=int(turn_index),
+                relevance_max=1.0,
+                query_type=QUERY_TYPE_EXPLAIN,
+                confidence=0.99,
+                missing_fields=[],
+                thinking_mode=active_thinking_mode,
+            )
+
+        is_snapshot_direct_question = self._is_snapshot_direct_question(effective_q)
+        if not deep_thinking and not is_snapshot_direct_question:
+            glm_semantic_result = await self._try_glm_project_semantic_answer(clean_q=effective_q, project_id=project_id)
+            if glm_semantic_result is not None:
+                return await self._persist_fast_answer(
+                    question=clean_q,
+                    asked_by=asked_by,
+                    asked_via=asked_via,
+                    project_id=project_id,
+                    fast_result=glm_semantic_result,
+                    session_id=session_id,
+                    external_session_id=context_session_id,
+                    turn_index=turn_index,
+                    query_type=query_type if query_type != QUERY_TYPE_UNKNOWN else QUERY_TYPE_EXPLAIN,
+                    missing_fields=[],
+                    state_row=state_row,
+                    effective_question=effective_q,
+                    rewrite_meta=rewrite_meta,
+                    thinking_mode=active_thinking_mode,
+                )
+
+        domains = self._route_domains(effective_q)
+        if str(asked_via or "").strip().lower() == "feishu" and source_chat_id and "feishu" not in domains:
+            domains.append("feishu")
+        retrieved, search_meta = await self._multi_step_search(
+            question=effective_q,
+            domains=domains,
+            project_id=project_id,
+            source_chat_id=source_chat_id,
+            session_state=state_row,
+            max_steps_override=deep_retrieval_steps if deep_thinking else None,
+            vector_top_k=deep_vector_top_k if deep_thinking else None,
+        )
+        db_query_ms = int(search_meta.get("db_query_ms") or 0)
+        vector_search_ms = int(search_meta.get("vector_search_ms") or 0)
+        graph_query_ms = int(search_meta.get("graph_query_ms") or 0)
+        db_timeout = bool(search_meta.get("db_timeout"))
+        vector_timeout = bool(search_meta.get("vector_timeout"))
+        graph_timeout = bool(search_meta.get("graph_timeout"))
+        tool_trace = list(search_meta.get("tool_trace") or [])
+        tool_trace.insert(
+            0,
+            {
+                "step": 0,
+                "tool": "thinking_mode",
+                "mode": active_thinking_mode,
+                "fast_path_bypassed": bool(deep_thinking),
+                "reason": forced_deep_reason or ("manual_deep" if self._normalize_thinking_mode(thinking_mode) == THINKING_MODE_DEEP else "normal"),
+            },
+        )
+        final_domains = [str(x) for x in (search_meta.get("domains_used") or domains)]
+
+        max_score = retrieved[0].score if retrieved else 0.0
+        answer_text = await self._generate_answer(
+            effective_q,
+            retrieved,
+            conversation,
+            project_id,
+            system_snapshot=system_snapshot,
+            user_snapshot=user_snapshot,
+            relevance_max=max_score,
+            thinking_mode=active_thinking_mode,
+            context_block_limit=DEEP_CONTEXT_BLOCKS if deep_thinking else None,
+            answer_max_tokens=DEEP_ANSWER_MAX_TOKENS if deep_thinking else None,
+        )
+        used_sources = self._dedupe_sources(retrieved)
+        if not used_sources:
+            used_sources = [
+                {
+                    "source": "system",
+                    "label": "System Snapshot",
+                    "score": round(max(max_score, 0.6), 4),
+                    "metadata": {"mode": "snapshot_only", "relevance_max": round(max_score, 4)},
+                }
+            ]
+            if user_snapshot:
+                used_sources.append(
+                    {
+                        "source": "system",
+                        "label": "User Snapshot",
+                        "score": round(max(max_score, 0.58), 4),
+                        "metadata": {"mode": "snapshot_only", "relevance_max": round(max_score, 4)},
+                    }
+                )
+
+        auto_retry_reason = ""
+        if not deep_thinking:
+            auto_retry_reason = self._should_auto_deep_retry(
+                question=clean_q,
+                effective_question=effective_q,
+                sources=used_sources,
+                relevance_max=max_score,
+            )
+            if auto_retry_reason and not self._allow_auto_deep_retry_for_channel(
+                asked_via=asked_via,
+                question=clean_q,
+                effective_question=effective_q,
+                reason=auto_retry_reason,
+            ):
+                auto_retry_reason = ""
+        if auto_retry_reason:
+            deep_retrieved, deep_meta = await self._multi_step_search(
+                question=effective_q,
+                domains=domains,
+                project_id=project_id,
+                source_chat_id=source_chat_id,
+                session_state=state_row,
+                max_steps_override=deep_retrieval_steps,
+                vector_top_k=deep_vector_top_k,
+            )
+            deep_max_score = deep_retrieved[0].score if deep_retrieved else 0.0
+            deep_answer_text = await self._generate_answer(
+                effective_q,
+                deep_retrieved,
+                conversation,
+                project_id,
+                system_snapshot=system_snapshot,
+                user_snapshot=user_snapshot,
+                relevance_max=deep_max_score,
+                thinking_mode=THINKING_MODE_DEEP,
+                context_block_limit=DEEP_CONTEXT_BLOCKS,
+                answer_max_tokens=DEEP_ANSWER_MAX_TOKENS,
+            )
+            deep_sources = self._dedupe_sources(deep_retrieved)
+            if not deep_sources:
+                deep_sources = [
+                    {
+                        "source": "system",
+                        "label": "System Snapshot",
+                        "score": round(max(deep_max_score, 0.6), 4),
+                        "metadata": {"mode": "snapshot_only", "relevance_max": round(deep_max_score, 4)},
+                    }
+                ]
+                if user_snapshot:
+                    deep_sources.append(
+                        {
+                            "source": "system",
+                            "label": "User Snapshot",
+                            "score": round(max(deep_max_score, 0.58), 4),
+                            "metadata": {"mode": "snapshot_only", "relevance_max": round(deep_max_score, 4)},
+                        }
+                    )
+
+            merged_metrics = self._merge_search_metrics(
+                {
+                    "db_query_ms": db_query_ms,
+                    "vector_search_ms": vector_search_ms,
+                    "graph_query_ms": graph_query_ms,
+                    "db_timeout": db_timeout,
+                    "vector_timeout": vector_timeout,
+                    "graph_timeout": graph_timeout,
+                },
+                deep_meta,
+            )
+            db_query_ms = int(merged_metrics.get("db_query_ms") or 0)
+            vector_search_ms = int(merged_metrics.get("vector_search_ms") or 0)
+            graph_query_ms = int(merged_metrics.get("graph_query_ms") or 0)
+            db_timeout = bool(merged_metrics.get("db_timeout"))
+            vector_timeout = bool(merged_metrics.get("vector_timeout"))
+            graph_timeout = bool(merged_metrics.get("graph_timeout"))
+
+            deep_trace = list(deep_meta.get("tool_trace") or [])
+            tool_trace.append(
+                {
+                    "step": 98,
+                    "tool": "auto_deep_retry",
+                    "reason": auto_retry_reason,
+                    "prev_relevance_max": round(float(max_score or 0.0), 4),
+                    "prev_snapshot_only": self._is_snapshot_only_sources(used_sources),
+                }
+            )
+            for row in deep_trace:
+                tool_trace.append(row)
+
+            final_domains = [str(x) for x in (deep_meta.get("domains_used") or final_domains or domains)]
+            retrieved = deep_retrieved
+            max_score = deep_max_score
+            answer_text = deep_answer_text
+            used_sources = deep_sources
+            deep_thinking = True
+            active_thinking_mode = THINKING_MODE_DEEP
+
+        evidence_guard_triggered = self._should_block_answer_with_guard(
+            question=clean_q,
+            effective_question=effective_q,
+            sources=used_sources,
+            relevance_max=max_score,
+        )
+        if evidence_guard_triggered:
+            answer_text = self._build_evidence_guard_answer(clean_q)
+            used_sources = [
+                {
+                    "source": "system",
+                    "label": "Evidence Guard",
+                    "score": 0.99,
+                    "metadata": {"mode": "guard", "reason": "snapshot_only_process_query"},
+                }
+            ]
+            max_score = min(float(max_score or 0.0), 0.35)
+            if "system_context" not in final_domains:
+                final_domains.append("system_context")
+            tool_trace.append(
+                {
+                    "step": 99,
+                    "tool": "answer_guard",
+                    "reason": "insufficient_call_chain_evidence",
+                }
+            )
+
+        elapsed = int((time.perf_counter() - started) * 1000)
+        qa = QARecord(
+            tenant_id=self.tenant_id,
+            question=clean_q,
+            answer=answer_text,
+            sources=used_sources,
+            intent="question",
+            domains_used=final_domains,
+            response_time_ms=elapsed,
+            asked_by=asked_by,
+            asked_via=(asked_via or "pmflow")[:20],
+            project_id=project_id,
+            session_id=session_id,
+            turn_index=turn_index,
+        )
+        self.db.add(qa)
+        confidence = round(max(0.2, min(1.0, max_score if max_score > 0 else 0.55)), 4)
+        if evidence_guard_triggered:
+            confidence = min(confidence, 0.35)
+        missing_fields = []
+        if max_score < 0.45:
+            missing_fields = ["high_relevance_evidence"]
+        if evidence_guard_triggered:
+            missing_fields = sorted(set([*missing_fields, "call_chain_evidence", "target_scope"]))
+        await self._update_session_state_after_turn(
+            session_state=state_row,
+            question=clean_q,
+            effective_question=effective_q,
+            answer=answer_text,
+            query_type=query_type if query_type != QUERY_TYPE_UNKNOWN else QUERY_TYPE_EXPLAIN,
+            ambiguity="clear" if not missing_fields else "low_evidence",
+            rewrite_meta=rewrite_meta,
+            sources=used_sources,
+        )
+        await self.db.commit()
+        await self.db.refresh(qa)
+        await self.context_cache.append_session_messages(
+            tenant_id=self.tenant_id,
+            session_id=context_session_id,
+            question=clean_q,
+            answer=answer_text,
+        )
+
+        return self._build_response_payload(
+            answer=answer_text,
+            sources=used_sources,
+            domains_used=final_domains,
+            response_time_ms=elapsed,
+            db_query_ms=db_query_ms,
+            vector_search_ms=vector_search_ms,
+            graph_query_ms=graph_query_ms,
+            db_timeout=db_timeout,
+            vector_timeout=vector_timeout,
+            graph_timeout=graph_timeout,
+            tool_trace=tool_trace,
+            qa_id=int(qa.id),
+            session_id=int(session_id) if session_id is not None else None,
+            turn_index=int(turn_index),
+            relevance_max=round(max_score, 4),
+            query_type=query_type if query_type != QUERY_TYPE_UNKNOWN else QUERY_TYPE_EXPLAIN,
+            confidence=confidence,
+            missing_fields=missing_fields,
+            thinking_mode=active_thinking_mode,
+        )
+
+    def _build_response_payload(
+        self,
+        *,
+        answer: str,
+        sources: list[dict[str, Any]],
+        domains_used: list[str],
+        response_time_ms: int,
+        db_query_ms: int,
+        vector_search_ms: int,
+        graph_query_ms: int = 0,
+        db_timeout: bool = False,
+        vector_timeout: bool = False,
+        graph_timeout: bool = False,
+        qa_id: Optional[int] = None,
+        session_id: Optional[int] = None,
+        turn_index: int = 1,
+        relevance_max: float = 0.0,
+        query_type: str = QUERY_TYPE_UNKNOWN,
+        confidence: float = 0.0,
+        missing_fields: Optional[list[str]] = None,
+        tool_trace: Optional[list[dict[str, Any]]] = None,
+        thinking_mode: str = THINKING_MODE_NORMAL,
+        action_taken: Optional[str] = None,
+        signal_id: Optional[int] = None,
+        reply_required: bool = False,
+        reply_text: Optional[str] = None,
+        ack_latency_ms: Optional[int] = None,
+        trace_id: Optional[str] = None,
+        processing_stage: Optional[str] = None,
+    ) -> dict[str, Any]:
+        evidence: list[dict[str, Any]] = []
+        for item in sources[:8]:
+            if not isinstance(item, dict):
+                continue
+            evidence.append(
+                {
+                    "source": str(item.get("source") or "unknown"),
+                    "label": str(item.get("label") or "unknown"),
+                    "score": round(self._to_float(item.get("score"), 0.0), 4),
+                    "metadata": item.get("metadata") if isinstance(item.get("metadata"), dict) else {},
+                }
+            )
+        return {
+            "answer": str(answer or "").strip(),
+            "evidence": evidence,
+            "confidence": round(max(0.0, min(1.0, self._to_float(confidence, 0.0))), 4),
+            "thinking_mode": self._normalize_thinking_mode(thinking_mode),
+            "missing_fields": [str(x) for x in (missing_fields or []) if str(x).strip()],
+            "query_type": str(query_type or QUERY_TYPE_UNKNOWN),
+            "sources": sources,
+            "domains_used": domains_used,
+            "response_time_ms": int(response_time_ms or 0),
+            "retrieval_metrics": {
+                "db_query_ms": int(db_query_ms or 0),
+                "vector_search_ms": int(vector_search_ms or 0),
+                "graph_query_ms": int(graph_query_ms or 0),
+                "db_timeout": bool(db_timeout),
+                "vector_timeout": bool(vector_timeout),
+                "graph_timeout": bool(graph_timeout),
+            },
+            "tool_trace": tool_trace if isinstance(tool_trace, list) else [],
+            "qa_id": int(qa_id) if qa_id is not None else None,
+            "session_id": int(session_id) if session_id is not None else None,
+            "turn_index": int(turn_index or 1),
+            "relevance_max": round(max(0.0, self._to_float(relevance_max, 0.0)), 4),
+            "action_taken": str(action_taken).strip() if action_taken else None,
+            "signal_id": int(signal_id) if signal_id is not None else None,
+            "reply_required": bool(reply_required),
+            "reply_text": str(reply_text or "").strip() if reply_text else None,
+            "ack_latency_ms": int(ack_latency_ms) if ack_latency_ms is not None else None,
+            "trace_id": str(trace_id or "").strip() or None,
+            "processing_stage": str(processing_stage or "").strip() or None,
+        }
+
+    @staticmethod
+    def _extract_ack_metadata_from_sources(
+        sources: Any,
+        *,
+        answer: str = "",
+    ) -> dict[str, Any]:
+        rows = sources if isinstance(sources, list) else []
+        action_taken = ""
+        signal_id: Optional[int] = None
+        reply_required = False
+        ack_latency_ms: Optional[int] = None
+        trace_id = ""
+        processing_stage = ""
+        for source in rows:
+            if not isinstance(source, dict):
+                continue
+            meta = source.get("metadata") if isinstance(source.get("metadata"), dict) else {}
+            candidate_action = str(meta.get("action_taken") or "").strip()
+            if not candidate_action:
+                continue
+            action_taken = candidate_action
+            try:
+                signal_id = int(meta.get("signal_id")) if meta.get("signal_id") is not None else None
+            except Exception:
+                signal_id = None
+            trace_id = str(meta.get("trace_id") or "").strip()
+            processing_stage = str(meta.get("processing_stage") or meta.get("action_mode") or "").strip()
+            try:
+                ack_latency_ms = (
+                    int(meta.get("ack_latency_ms")) if meta.get("ack_latency_ms") is not None else None
+                )
+            except Exception:
+                ack_latency_ms = None
+            reply_required = True
+            break
+        reply_text = str(answer or "").strip() if reply_required else ""
+        normalized_stage = processing_stage
+        if normalized_stage == "write":
+            normalized_stage = "ack_only"
+        return {
+            "action_taken": action_taken or None,
+            "signal_id": signal_id,
+            "reply_required": bool(reply_required),
+            "reply_text": reply_text or None,
+            "ack_latency_ms": ack_latency_ms,
+            "trace_id": trace_id or None,
+            "processing_stage": normalized_stage or None,
+        }
+
+    @staticmethod
+    def _normalize_action_mode(value: Optional[str]) -> str:
+        text = str(value or "").strip().lower()
+        if text in {ACTION_MODE_READ, ACTION_MODE_WRITE, ACTION_MODE_READ_WRITE}:
+            return text
+        return ACTION_MODE_AUTO
+
+    @classmethod
+    def _looks_like_write_intent(cls, question: str) -> bool:
+        text = str(question or "").strip().lower()
+        if not text:
+            return False
+        has_verb = any(k in text for k in REQUIREMENT_WRITE_VERB_KEYWORDS)
+        has_object = any(k in text for k in REQUIREMENT_WRITE_OBJECT_KEYWORDS)
+        if not has_verb or not has_object:
+            return False
+        if any(prefix in text for prefix in WRITE_EXPRESSIONS):
+            return True
+        if "新增" in text or "增加" in text or "记录" in text or "入库" in text:
+            return True
+        return False
+
+    @classmethod
+    def _resolve_action_mode(
+        cls,
+        *,
+        requested_mode: str,
+        question: str,
+        effective_question: str,
+    ) -> str:
+        mode = cls._normalize_action_mode(requested_mode)
+        if mode != ACTION_MODE_AUTO:
+            return mode
+
+        merged = f"{question} {effective_question}".strip()
+        if not cls._looks_like_write_intent(merged):
+            return ACTION_MODE_READ
+        has_question_form = ("?" in merged) or ("？" in merged) or cls._contains_any(merged, QUESTION_HINTS)
+        if has_question_form:
+            return ACTION_MODE_READ_WRITE
+        return ACTION_MODE_WRITE
+
+    @staticmethod
+    def _normalize_requirement_summary(text: str) -> str:
+        normalized = re.sub(r"\s+", " ", str(text or "").strip())
+        normalized = normalized.strip("，,。.!！？?：:;；")
+        return normalized[:300]
+
+    @staticmethod
+    def _build_ack_trace_id(question: str) -> str:
+        base = str(question or "").strip()
+        digest = hashlib.sha1(base.encode("utf-8")).hexdigest()[:10] if base else "empty"
+        return f"rag-ack-{int(time.time() * 1000)}-{digest}"
+
+    async def _ensure_project_signal_link(
+        self,
+        *,
+        signal_id: int,
+        project_id: int,
+        link_type: str = "related",
+    ) -> None:
+        existing = (
+            await self.db.execute(
+                select(ProjectSignalLink.id).where(
+                    and_(
+                        ProjectSignalLink.tenant_id == self.tenant_id,
+                        ProjectSignalLink.project_id == int(project_id),
+                        ProjectSignalLink.signal_id == int(signal_id),
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+        if existing is not None:
+            return
+        self.db.add(
+            ProjectSignalLink(
+                tenant_id=self.tenant_id,
+                project_id=int(project_id),
+                signal_id=int(signal_id),
+                link_type=link_type,
+                relevance_score=1.0,
+            )
+        )
+
+    async def _create_or_reuse_requirement_signal(
+        self,
+        *,
+        question: str,
+        effective_question: str,
+        asked_by: Optional[int],
+        project_id: Optional[int],
+    ) -> tuple[str, Optional[int], Optional[int], str]:
+        summary = self._normalize_requirement_summary(effective_question or question)
+        resolved_project_id: Optional[int] = project_id
+        if resolved_project_id is None:
+            scoped_id, _, _ = await self._resolve_project_scope_for_question(
+                question=effective_question or question,
+                project_id=project_id,
+            )
+            resolved_project_id = scoped_id
+
+        existing_stmt = select(Signal).where(
+            and_(
+                Signal.tenant_id == self.tenant_id,
+                Signal.source == "PMFlow",
+                Signal.summary == summary,
+            )
+        )
+        if resolved_project_id is None:
+            existing_stmt = existing_stmt.where(Signal.matched_project_id.is_(None))
+        else:
+            existing_stmt = existing_stmt.where(Signal.matched_project_id == int(resolved_project_id))
+        existing = (
+            await self.db.execute(
+                existing_stmt.order_by(Signal.created_at.desc(), Signal.id.desc()).limit(1)
+            )
+        ).scalar_one_or_none()
+        if existing is not None:
+            if resolved_project_id is not None:
+                try:
+                    await self._ensure_project_signal_link(
+                        signal_id=int(existing.id),
+                        project_id=int(resolved_project_id),
+                    )
+                    await self.db.commit()
+                except Exception as exc:
+                    await self.db.rollback()
+                    return ("signal_failed", None, resolved_project_id, str(exc)[:240])
+            return ("signal_deduped", int(existing.id), resolved_project_id, "")
+
+        actor = f"PMFlow用户#{int(asked_by)}" if asked_by is not None else "PMFlow用户"
+        content = f"[{actor}]: {summary}"
+        strategic_payload = {
+            "source": "pmflow_qa",
+            "original_question": str(question or "")[:4000],
+            "normalized_question": summary,
+            "write_mode": "ack_first",
+            "created_at": datetime.now(timezone.utc).isoformat(),
+        }
+        signal = Signal(
+            source="PMFlow",
+            content=content[:2000],
+            impact_score=0.65,
+            priority="medium",
+            processed=False,
+            strategic_analysis=json.dumps(strategic_payload, ensure_ascii=False),
+            tasks_generated=json.dumps([], ensure_ascii=False),
+            matched_project_id=int(resolved_project_id) if resolved_project_id is not None else None,
+            title=summary[:120] or "PMFlow 新需求信号",
+            signal_type="新需求",
+            summary=summary,
+            reasoning=json.dumps(["来自 PMFlow 问答写入意图，按 ACK-first 策略快速入库。"], ensure_ascii=False),
+            confidence=0.74,
+            recommended_action="enter_only",
+            status="待处理",
+            tenant_id=self.tenant_id,
+        )
+        try:
+            self.db.add(signal)
+            await self.db.flush()
+            if resolved_project_id is not None:
+                await self._ensure_project_signal_link(
+                    signal_id=int(signal.id),
+                    project_id=int(resolved_project_id),
+                )
+            await self.db.commit()
+            await self.db.refresh(signal)
+            return ("signal_created", int(signal.id), resolved_project_id, "")
+        except Exception as exc:
+            await self.db.rollback()
+            return ("signal_failed", None, resolved_project_id, str(exc)[:240])
+
+    @staticmethod
+    def _build_requirement_ack_text(
+        *,
+        action_taken: str,
+        signal_id: Optional[int],
+        project_id: Optional[int],
+        error_reason: str,
+    ) -> str:
+        project_text = f"项目#{int(project_id)}" if project_id is not None else "未指定项目"
+        if action_taken == "signal_created" and signal_id is not None:
+            return f"已入库：已为{project_text}创建信号 #{int(signal_id)}，并进入待处理队列。"
+        if action_taken == "signal_deduped" and signal_id is not None:
+            return f"已入库：检测到重复需求，已复用{project_text}的信号 #{int(signal_id)}。"
+        if action_taken == "signal_failed":
+            reason = error_reason or "写入异常"
+            return f"未完成入库：{reason}。我已记录本次失败，请稍后重试。"
+        return "已收到需求，正在处理中。"
+
+    async def _handle_ack_first_requirement(
+        self,
+        *,
+        question: str,
+        effective_question: str,
+        asked_by: Optional[int],
+        asked_via: str,
+        project_id: Optional[int],
+        session_id: Optional[int],
+        turn_index: int,
+        state_row: Optional[QASessionState],
+        rewrite_meta: dict[str, Any],
+        context_session_id: Optional[int],
+        query_type: str,
+        action_mode: str,
+        started_at: float,
+    ) -> Optional[dict[str, Any]]:
+        action_taken, linked_signal_id, resolved_project_id, error_reason = await self._create_or_reuse_requirement_signal(
+            question=question,
+            effective_question=effective_question,
+            asked_by=asked_by,
+            project_id=project_id,
+        )
+        if not action_taken:
+            return None
+
+        ack_latency_ms = int((time.perf_counter() - started_at) * 1000)
+        trace_id = self._build_ack_trace_id(question)
+        reply_text = self._build_requirement_ack_text(
+            action_taken=action_taken,
+            signal_id=linked_signal_id,
+            project_id=resolved_project_id,
+            error_reason=error_reason,
+        )
+        source_orchestration_result: Optional[dict[str, Any]] = None
+        if (
+            linked_signal_id is not None
+            and resolved_project_id is not None
+            and action_taken in {"signal_created", "signal_deduped"}
+            and SignalSourceOrchestrationService.looks_like_signal_source_request(effective_question or question)
+        ):
+            try:
+                source_orchestration_result = await SignalSourceOrchestrationService(self.db).orchestrate_from_signal(
+                    tenant_id=self.tenant_id,
+                    project_id=int(resolved_project_id),
+                    signal_id=int(linked_signal_id),
+                    initiating_text=effective_question or question,
+                    requested_via="pmflow",
+                    operator_user_id=asked_by,
+                    dry_run=False,
+                )
+            except Exception as exc:
+                source_orchestration_result = {
+                    "status": "failed",
+                    "reason": str(exc),
+                }
+            orchestration_status = str(source_orchestration_result.get("status") or "").lower()
+            if orchestration_status in {"completed", "partial", "planned", "pending_manual"}:
+                source_payload = source_orchestration_result.get("signal_source") if isinstance(source_orchestration_result.get("signal_source"), dict) else {}
+                worker_executions = (
+                    source_orchestration_result.get("worker_executions")
+                    if isinstance(source_orchestration_result.get("worker_executions"), list)
+                    else []
+                )
+                if orchestration_status == "pending_manual":
+                    reply_text = (
+                        f"{reply_text}"
+                        f" 已记录信号源“{str(source_payload.get('source_name') or '未命名信号源')}”的接入请求，"
+                        "但当前仍缺少可直接执行的接入配置，已转待确认。"
+                    )
+                else:
+                    reply_text = (
+                        f"{reply_text}"
+                        f" 已登记信号源“{str(source_payload.get('source_name') or '未命名信号源')}”，"
+                        f"并已下发 {len(worker_executions)} 个 AI worker 执行。"
+                    )
+        answer_text = reply_text
+        sources = [
+            {
+                "source": "pmflow",
+                "label": "Signal Queue",
+                "score": 0.98 if action_taken != "signal_failed" else 0.42,
+                "metadata": {
+                    "action_taken": action_taken,
+                    "signal_id": linked_signal_id,
+                    "project_id": resolved_project_id,
+                    "action_mode": action_mode,
+                    "processing_stage": "ack_only",
+                    "ack_latency_ms": ack_latency_ms,
+                    "trace_id": trace_id,
+                    "source_orchestration": source_orchestration_result,
+                },
+            }
+        ]
+
+        qa_id: Optional[int] = None
+        try:
+            qa = QARecord(
+                tenant_id=self.tenant_id,
+                question=question,
+                answer=answer_text,
+                sources=sources,
+                intent="statement",
+                domains_used=["signal_write"],
+                response_time_ms=ack_latency_ms,
+                asked_by=asked_by,
+                asked_via=(asked_via or "pmflow")[:20],
+                project_id=resolved_project_id if resolved_project_id is not None else project_id,
+                session_id=session_id,
+                turn_index=int(turn_index or 1),
+            )
+            self.db.add(qa)
+            await self._update_session_state_after_turn(
+                session_state=state_row,
+                question=question,
+                effective_question=effective_question,
+                answer=answer_text,
+                query_type=query_type if query_type != QUERY_TYPE_UNKNOWN else QUERY_TYPE_STATUS,
+                ambiguity="clear" if action_taken != "signal_failed" else "low_evidence",
+                rewrite_meta=rewrite_meta,
+                sources=sources,
+            )
+            await self.db.commit()
+            await self.db.refresh(qa)
+            qa_id = int(qa.id)
+        except Exception:
+            await self.db.rollback()
+
+        try:
+            await self.context_cache.append_session_messages(
+                tenant_id=self.tenant_id,
+                session_id=context_session_id,
+                question=question,
+                answer=answer_text,
+            )
+        except Exception:
+            pass
+
+        confidence = 0.94 if action_taken in {"signal_created", "signal_deduped"} else 0.35
+        missing_fields = [] if resolved_project_id is not None else ["project_id"]
+        return self._build_response_payload(
+            answer=answer_text,
+            sources=sources,
+            domains_used=["signal_write"],
+            response_time_ms=ack_latency_ms,
+            db_query_ms=0,
+            vector_search_ms=0,
+            db_timeout=False,
+            vector_timeout=False,
+            qa_id=qa_id,
+            session_id=int(session_id) if session_id is not None else None,
+            turn_index=int(turn_index or 1),
+            relevance_max=0.99 if action_taken != "signal_failed" else 0.30,
+            query_type=query_type if query_type != QUERY_TYPE_UNKNOWN else QUERY_TYPE_STATUS,
+            confidence=confidence,
+            missing_fields=missing_fields,
+            thinking_mode=THINKING_MODE_NORMAL,
+            action_taken=action_taken,
+            signal_id=linked_signal_id,
+            reply_required=True,
+            reply_text=reply_text,
+            ack_latency_ms=ack_latency_ms,
+            trace_id=trace_id,
+            processing_stage="ack_only",
+        )
+
+    def _classify_query_type(self, question: str) -> str:
+        text = str(question or "").strip().lower()
+        if not text:
+            return QUERY_TYPE_UNKNOWN
+        # 若同一句同时出现“多少/几个”和“列出/分别”，优先按 list 处理，避免追问被误判为 count。
+        if any(k in text for k in LIST_HINTS):
+            return QUERY_TYPE_LIST
+        if any(k in text for k in COUNT_HINTS):
+            return QUERY_TYPE_COUNT
+        if any(k in text for k in STATUS_HINTS):
+            return QUERY_TYPE_STATUS
+        if "为什么" in text or "如何" in text or "怎么" in text:
+            return QUERY_TYPE_EXPLAIN
+        return QUERY_TYPE_UNKNOWN
+
+    async def _run_deterministic_query(
+        self,
+        *,
+        question: str,
+        project_id: Optional[int],
+        query_type: str,
+    ) -> Optional[dict[str, Any]]:
+        q_type = str(query_type or QUERY_TYPE_UNKNOWN)
+        if q_type == QUERY_TYPE_COUNT:
+            return await self._run_count_query(question=question, project_id=project_id)
+        if q_type == QUERY_TYPE_LIST:
+            return await self._run_list_query(question=question, project_id=project_id)
+        if q_type == QUERY_TYPE_STATUS:
+            progress = await self._try_fast_project_progress_answer(clean_q=question, project_id=project_id)
+            if progress:
+                progress["query_type"] = QUERY_TYPE_STATUS
+                progress.setdefault("confidence", 0.98)
+                progress.setdefault("missing_fields", [])
+                return progress
+            existence = await self._try_fast_project_existence_answer(clean_q=question, project_id=project_id)
+            if existence:
+                existence["query_type"] = QUERY_TYPE_STATUS
+                existence.setdefault("confidence", 0.96)
+                existence.setdefault("missing_fields", [])
+                return existence
+        return None
+
+    async def _resolve_project_scope_for_question(
+        self,
+        *,
+        question: str,
+        project_id: Optional[int],
+    ) -> tuple[Optional[int], Optional[str], str]:
+        if project_id is not None:
+            row = (
+                await self.db.execute(
+                    select(Project).where(
+                        and_(
+                            Project.tenant_id == self.tenant_id,
+                            Project.id == int(project_id),
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if row is None:
+                return (None, None, "")
+            return (int(row.id), str(row.title or row.name or f"项目#{int(row.id)}"), "")
+
+        direct_project_id, direct_project_name = await self._match_project_name_in_question(question=question)
+        if direct_project_id is not None and direct_project_name:
+            return (int(direct_project_id), str(direct_project_name), str(direct_project_name)[:24])
+
+        keyword = self._extract_project_lookup_keyword(question)
+        terms = self._extract_query_terms(question, preferred=keyword)
+        for term in terms:
+            matched = await self._search_projects_by_keyword(keyword=term, project_id=None, limit=5)
+            if matched:
+                target = matched[0]
+                return (int(target.id), str(target.title or target.name or f"项目#{int(target.id)}"), term)
+        return (None, None, keyword[:24] if keyword else "")
+
+    @staticmethod
+    def _normalize_project_subject_text(text: str) -> str:
+        clean = str(text or "").strip().lower()
+        if not clean:
+            return ""
+        clean = clean.replace("—", "-").replace("_", " ")
+        clean = re.sub(r"[^\w\u4e00-\u9fff\- ]+", " ", clean)
+        clean = re.sub(
+            r"(我说的是|我说的|请问|请|帮我|一下|这个里面|这里面|里面|这个|那个|一下子|目前|当前|现在|有多少个|多少个|信号源|信号|项目|系统)",
+            " ",
+            clean,
+        )
+        clean = re.sub(r"[-]+", " ", clean)
+        clean = re.sub(r"\s+", " ", clean).strip()
+        return clean
+
+    async def _match_project_name_in_question(self, *, question: str) -> tuple[Optional[int], Optional[str]]:
+        normalized_question = self._normalize_project_subject_text(question)
+        if not normalized_question:
+            return (None, None)
+
+        projects = (
+            await self.db.execute(
+                select(Project)
+                .where(Project.tenant_id == self.tenant_id)
+                .order_by(Project.updated_at.desc(), Project.id.desc())
+                .limit(200)
+            )
+        ).scalars().all()
+
+        best_id: Optional[int] = None
+        best_name: Optional[str] = None
+        best_score = 0
+        for row in projects:
+            candidates = [
+                str(row.title or "").strip(),
+                str(row.name or "").strip(),
+            ]
+            for name in candidates:
+                if len(name) < 2:
+                    continue
+                normalized_name = self._normalize_project_subject_text(name)
+                if len(normalized_name) < 2:
+                    continue
+
+                score = 0
+                if normalized_name in normalized_question:
+                    score = len(normalized_name)
+                else:
+                    tokens = [t for t in normalized_name.split(" ") if len(t) >= 2]
+                    if tokens and all(t in normalized_question for t in tokens):
+                        score = sum(len(t) for t in tokens)
+
+                if score > best_score:
+                    best_score = score
+                    best_id = int(row.id)
+                    best_name = str(row.title or row.name or f"项目#{int(row.id)}")
+
+        return (best_id, best_name)
+
+    async def _run_count_query(self, *, question: str, project_id: Optional[int]) -> Optional[dict[str, Any]]:
+        text = str(question or "").lower()
+        started = time.perf_counter()
+        has_project = "项目" in text
+        has_task = "任务" in text
+        has_signal = "信号" in text or "信号源" in text
+        has_people = self._is_people_query(text)
+
+        # 先处理更具体的对象（信号/任务），避免“xxx项目里有多少信号”误落到项目总数分支。
+        if has_signal:
+            scoped_project_id, scoped_project_name, scoped_keyword = await self._resolve_project_scope_for_question(
+                question=text,
+                project_id=project_id,
+            )
+            if has_project and scoped_project_id is None and scoped_keyword:
+                elapsed = int((time.perf_counter() - started) * 1000)
+                return {
+                    "answer": f"未检索到名称或描述包含“{scoped_keyword}”的项目，无法统计信号数。",
+                    "sources": [
+                        {
+                            "source": "db",
+                            "label": "Project database",
+                            "score": 0.94,
+                            "metadata": {"table": "projects", "query_keyword": scoped_keyword, "match_count": 0},
+                        }
+                    ],
+                    "db_query_ms": elapsed,
+                    "response_time_ms": elapsed,
+                    "query_type": QUERY_TYPE_COUNT,
+                    "confidence": 0.94,
+                    "missing_fields": ["project_id"],
+                }
+
+            stmt = select(func.count(Signal.id)).where(Signal.tenant_id == self.tenant_id)
+            if scoped_project_id is not None:
+                stmt = stmt.where(Signal.matched_project_id == int(scoped_project_id))
+            total = int((await self.db.execute(stmt)).scalar() or 0)
+            elapsed = int((time.perf_counter() - started) * 1000)
+            if scoped_project_id is not None and scoped_project_name:
+                answer = f"项目「{scoped_project_name}」当前关联信号总数为 {total} 条。"
+            else:
+                answer = f"当前可见信号总数为 {total} 条。"
+            return {
+                "answer": answer,
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {
+                            "table": "signals",
+                            "count": total,
+                            "project_id": scoped_project_id,
+                            "project_name": scoped_project_name,
+                        },
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_COUNT,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+
+        if has_task:
+            scoped_project_id, scoped_project_name, scoped_keyword = await self._resolve_project_scope_for_question(
+                question=text,
+                project_id=project_id,
+            )
+            if has_project and scoped_project_id is None and scoped_keyword:
+                elapsed = int((time.perf_counter() - started) * 1000)
+                return {
+                    "answer": f"未检索到名称或描述包含“{scoped_keyword}”的项目，无法统计任务数。",
+                    "sources": [
+                        {
+                            "source": "db",
+                            "label": "Project database",
+                            "score": 0.94,
+                            "metadata": {"table": "projects", "query_keyword": scoped_keyword, "match_count": 0},
+                        }
+                    ],
+                    "db_query_ms": elapsed,
+                    "response_time_ms": elapsed,
+                    "query_type": QUERY_TYPE_COUNT,
+                    "confidence": 0.94,
+                    "missing_fields": ["project_id"],
+                }
+
+            stmt = select(func.count(Task.id)).where(Task.tenant_id == self.tenant_id)
+            if scoped_project_id is not None:
+                stmt = stmt.where(Task.project_id == int(scoped_project_id))
+            total = int((await self.db.execute(stmt)).scalar() or 0)
+            elapsed = int((time.perf_counter() - started) * 1000)
+            if scoped_project_id is not None and scoped_project_name:
+                answer = f"项目「{scoped_project_name}」当前任务总数为 {total} 条。"
+            else:
+                answer = f"当前可见任务总数为 {total} 条。"
+            return {
+                "answer": answer,
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {
+                            "table": "tasks",
+                            "count": total,
+                            "project_id": scoped_project_id,
+                            "project_name": scoped_project_name,
+                        },
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_COUNT,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+
+        if has_people:
+            scoped_project_id: Optional[int] = None
+            scoped_project_name: Optional[str] = None
+            scoped_keyword = ""
+            if has_project or project_id is not None:
+                scoped_project_id, scoped_project_name, scoped_keyword = await self._resolve_project_scope_for_question(
+                    question=text,
+                    project_id=project_id,
+                )
+                if has_project and scoped_project_id is None and scoped_keyword and not self._is_people_query(scoped_keyword):
+                    elapsed = int((time.perf_counter() - started) * 1000)
+                    return {
+                        "answer": f"未检索到名称或描述包含“{scoped_keyword}”的项目，无法统计成员人数。",
+                        "sources": [
+                            {
+                                "source": "db",
+                                "label": "Project database",
+                                "score": 0.94,
+                                "metadata": {"table": "projects", "query_keyword": scoped_keyword, "match_count": 0},
+                            }
+                        ],
+                        "db_query_ms": elapsed,
+                        "response_time_ms": elapsed,
+                        "query_type": QUERY_TYPE_COUNT,
+                        "confidence": 0.94,
+                        "missing_fields": ["project_id"],
+                    }
+
+            user_stmt = select(func.count(User.id)).where(
+                and_(
+                    User.tenant_id == self.tenant_id,
+                    User.is_active.is_(True),
+                )
+            )
+            if scoped_project_id is not None:
+                member_subq = select(UserProject.user_id).where(
+                    and_(
+                        UserProject.tenant_id == self.tenant_id,
+                        UserProject.project_id == int(scoped_project_id),
+                    )
+                )
+                user_stmt = user_stmt.where(User.id.in_(member_subq))
+            total = int((await self.db.execute(user_stmt)).scalar() or 0)
+            elapsed = int((time.perf_counter() - started) * 1000)
+            if scoped_project_id is not None and scoped_project_name:
+                answer = f"项目「{scoped_project_name}」当前活跃成员总数为 {total} 人。"
+            else:
+                answer = f"当前系统里活跃成员总数为 {total} 人。"
+            return {
+                "answer": answer,
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "User database",
+                        "score": 0.99,
+                        "metadata": {
+                            "table": "users",
+                            "count": total,
+                            "project_id": scoped_project_id,
+                            "project_name": scoped_project_name,
+                        },
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_COUNT,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+
+        if has_project:
+            stmt = select(func.count(Project.id)).where(Project.tenant_id == self.tenant_id)
+            if project_id is not None:
+                stmt = stmt.where(Project.id == int(project_id))
+            total = int((await self.db.execute(stmt)).scalar() or 0)
+            elapsed = int((time.perf_counter() - started) * 1000)
+            return {
+                "answer": f"当前系统里项目总数为 {total} 个。",
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {"table": "projects", "count": total},
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_COUNT,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+        return None
+
+    async def _run_list_query(self, *, question: str, project_id: Optional[int]) -> Optional[dict[str, Any]]:
+        text = str(question or "").lower()
+        started = time.perf_counter()
+        if self._is_people_query(text):
+            scoped_project_id: Optional[int] = None
+            scoped_project_name: Optional[str] = None
+            scoped_keyword = ""
+            if "项目" in text or project_id is not None:
+                scoped_project_id, scoped_project_name, scoped_keyword = await self._resolve_project_scope_for_question(
+                    question=text,
+                    project_id=project_id,
+                )
+                if "项目" in text and scoped_project_id is None and scoped_keyword and not self._is_people_query(scoped_keyword):
+                    elapsed = int((time.perf_counter() - started) * 1000)
+                    return {
+                        "answer": f"未检索到名称或描述包含“{scoped_keyword}”的项目，无法列出成员清单。",
+                        "sources": [
+                            {
+                                "source": "db",
+                                "label": "Project database",
+                                "score": 0.94,
+                                "metadata": {"table": "projects", "query_keyword": scoped_keyword, "match_count": 0},
+                            }
+                        ],
+                        "db_query_ms": elapsed,
+                        "response_time_ms": elapsed,
+                        "query_type": QUERY_TYPE_LIST,
+                        "confidence": 0.94,
+                        "missing_fields": ["project_id"],
+                    }
+
+            user_stmt = select(User).where(
+                and_(
+                    User.tenant_id == self.tenant_id,
+                    User.is_active.is_(True),
+                )
+            )
+            if scoped_project_id is not None:
+                member_subq = select(UserProject.user_id).where(
+                    and_(
+                        UserProject.tenant_id == self.tenant_id,
+                        UserProject.project_id == int(scoped_project_id),
+                    )
+                )
+                user_stmt = user_stmt.where(User.id.in_(member_subq))
+            members = (
+                await self.db.execute(
+                    user_stmt.order_by(User.updated_at.desc(), User.id.desc()).limit(30)
+                )
+            ).scalars().all()
+            elapsed = int((time.perf_counter() - started) * 1000)
+            if not members:
+                answer = (
+                    f"项目「{scoped_project_name}」当前没有可列举的成员。"
+                    if scoped_project_name
+                    else "当前没有可列举的成员。"
+                )
+                return {
+                    "answer": answer,
+                    "sources": [
+                        {
+                            "source": "db",
+                            "label": "User database",
+                            "score": 0.95,
+                            "metadata": {
+                                "table": "users",
+                                "count": 0,
+                                "project_id": scoped_project_id,
+                                "project_name": scoped_project_name,
+                            },
+                        }
+                    ],
+                    "db_query_ms": elapsed,
+                    "response_time_ms": elapsed,
+                    "query_type": QUERY_TYPE_LIST,
+                    "confidence": 0.95,
+                    "missing_fields": [],
+                }
+
+            member_ids = [int(row.id) for row in members]
+            workload_rows = (
+                await self.db.execute(
+                    select(UserWorkload.user_id, UserWorkload.workload_ratio).where(
+                        and_(
+                            UserWorkload.tenant_id == self.tenant_id,
+                            UserWorkload.user_id.in_(member_ids),
+                        )
+                    )
+                )
+            ).all()
+            workload_map = {int(uid): float(ratio or 0.0) for uid, ratio in workload_rows}
+
+            project_count_rows = (
+                await self.db.execute(
+                    select(UserProject.user_id, func.count(UserProject.project_id))
+                    .where(
+                        and_(
+                            UserProject.tenant_id == self.tenant_id,
+                            UserProject.user_id.in_(member_ids),
+                        )
+                    )
+                    .group_by(UserProject.user_id)
+                )
+            ).all()
+            project_count_map = {int(uid): int(cnt or 0) for uid, cnt in project_count_rows}
+
+            lines = []
+            for row in members:
+                uid = int(row.id)
+                ratio = workload_map.get(uid)
+                if ratio is None:
+                    ratio = round(float(row.workload or 0) / 100.0, 2)
+                role = str(row.role or "unknown")
+                project_count = project_count_map.get(uid, 0)
+                lines.append(
+                    f"#{uid} {str(row.name or f'User#{uid}')}（角色:{role}，负载:{float(ratio):.2f}，关联项目:{project_count}）"
+                )
+            elapsed = int((time.perf_counter() - started) * 1000)
+            header = f"项目「{scoped_project_name}」成员清单如下：" if scoped_project_name else "当前团队成员清单如下："
+            return {
+                "answer": header + "\n" + "\n".join(lines),
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "User database",
+                        "score": 0.99,
+                        "metadata": {
+                            "table": "users",
+                            "count": len(members),
+                            "project_id": scoped_project_id,
+                            "project_name": scoped_project_name,
+                        },
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_LIST,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+
+        if "信号" in text or "信号源" in text:
+            scoped_project_id, scoped_project_name, scoped_keyword = await self._resolve_project_scope_for_question(
+                question=text,
+                project_id=project_id,
+            )
+            if "项目" in text and scoped_project_id is None and scoped_keyword:
+                elapsed = int((time.perf_counter() - started) * 1000)
+                return {
+                    "answer": f"未检索到名称或描述包含“{scoped_keyword}”的项目，无法列出信号清单。",
+                    "sources": [
+                        {
+                            "source": "db",
+                            "label": "Project database",
+                            "score": 0.94,
+                            "metadata": {"table": "projects", "query_keyword": scoped_keyword, "match_count": 0},
+                        }
+                    ],
+                    "db_query_ms": elapsed,
+                    "response_time_ms": elapsed,
+                    "query_type": QUERY_TYPE_LIST,
+                    "confidence": 0.94,
+                    "missing_fields": ["project_id"],
+                }
+
+            stmt = select(Signal).where(Signal.tenant_id == self.tenant_id)
+            if scoped_project_id is not None:
+                stmt = stmt.where(Signal.matched_project_id == int(scoped_project_id))
+            signals = (
+                await self.db.execute(stmt.order_by(Signal.updated_at.desc(), Signal.id.desc()).limit(30))
+            ).scalars().all()
+            elapsed = int((time.perf_counter() - started) * 1000)
+            if not signals:
+                answer = (
+                    f"项目「{scoped_project_name}」当前没有可列举的信号。"
+                    if scoped_project_name
+                    else "当前没有可列举的信号。"
+                )
+                return {
+                    "answer": answer,
+                    "sources": [
+                        {
+                            "source": "db",
+                            "label": "Project database",
+                            "score": 0.95,
+                            "metadata": {
+                                "table": "signals",
+                                "count": 0,
+                                "project_id": scoped_project_id,
+                                "project_name": scoped_project_name,
+                            },
+                        }
+                    ],
+                    "db_query_ms": elapsed,
+                    "response_time_ms": elapsed,
+                    "query_type": QUERY_TYPE_LIST,
+                    "confidence": 0.95,
+                    "missing_fields": [],
+                }
+
+            lines = []
+            for row in signals:
+                title = str(row.title or row.summary or row.content or f"Signal#{int(row.id)}").replace("\n", " ").strip()
+                status = str(row.status or "unknown")
+                lines.append(f"#{int(row.id)} {title[:90]}（{status}）")
+            header = f"项目「{scoped_project_name}」信号清单如下：" if scoped_project_name else "当前信号清单如下："
+            return {
+                "answer": header + "\n" + "\n".join(lines),
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {
+                            "table": "signals",
+                            "count": len(signals),
+                            "project_id": scoped_project_id,
+                            "project_name": scoped_project_name,
+                        },
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_LIST,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+
+        if "项目" in text:
+            stmt = select(Project).where(Project.tenant_id == self.tenant_id)
+            if project_id is not None:
+                stmt = stmt.where(Project.id == int(project_id))
+            projects = (
+                await self.db.execute(stmt.order_by(Project.updated_at.desc(), Project.id.desc()).limit(20))
+            ).scalars().all()
+            if not projects:
+                elapsed = int((time.perf_counter() - started) * 1000)
+                return {
+                    "answer": "当前没有可列举的项目。",
+                    "sources": [
+                        {
+                            "source": "db",
+                            "label": "Project database",
+                            "score": 0.95,
+                            "metadata": {"table": "projects", "count": 0},
+                        }
+                    ],
+                    "db_query_ms": elapsed,
+                    "response_time_ms": elapsed,
+                    "query_type": QUERY_TYPE_LIST,
+                    "confidence": 0.95,
+                    "missing_fields": [],
+                }
+
+            lines = [f"#{int(p.id)} {str(p.title or p.name or '未命名项目')}（{str(p.status or 'unknown')}）" for p in projects]
+            elapsed = int((time.perf_counter() - started) * 1000)
+            return {
+                "answer": "当前项目列表如下：\n" + "\n".join(lines),
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {"table": "projects", "count": len(projects)},
+                    }
+                ],
+                "db_query_ms": elapsed,
+                "response_time_ms": elapsed,
+                "query_type": QUERY_TYPE_LIST,
+                "confidence": 0.99,
+                "missing_fields": [],
+            }
+        return None
+
+    async def _ensure_session_state(self, *, session_id: Optional[int]) -> Optional[QASessionState]:
+        if session_id is None:
+            return None
+        row = (
+            await self.db.execute(
+                select(QASessionState).where(
+                    and_(
+                        QASessionState.tenant_id == self.tenant_id,
+                        QASessionState.session_id == int(session_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            row = QASessionState(
+                tenant_id=self.tenant_id,
+                session_id=int(session_id),
+                last_subject=None,
+                last_entity_ids=[],
+                last_intent=None,
+                ambiguity="unknown",
+                rewrite_history=[],
+            )
+            self.db.add(row)
+            await self.db.flush()
+        return row
+
+    def _rewrite_question_with_state(
+        self,
+        *,
+        question: str,
+        conversation: list[dict[str, str]],
+        session_state: Optional[QASessionState],
+    ) -> tuple[str, dict[str, Any]]:
+        expanded = self._expand_followup_question(question, conversation)
+        current = str(expanded or question or "").strip()
+        if not current:
+            return current, {"mode": "empty"}
+
+        lower = current.lower()
+        has_referential = any(k in lower for k in FOLLOWUP_REFERENTIAL_KEYWORDS)
+        if not has_referential:
+            return current, {"mode": "direct"}
+
+        state_subject = ""
+        state_entity_ids: list[int] = []
+        if session_state is not None:
+            state_subject = str(session_state.last_subject or "").strip()
+            if isinstance(session_state.last_entity_ids, list):
+                for value in session_state.last_entity_ids:
+                    try:
+                        state_entity_ids.append(int(value))
+                    except Exception:
+                        continue
+        if state_subject:
+            rewritten = f"围绕“{state_subject}”回答：{current}"
+            return rewritten[:1200], {"mode": "state_subject", "subject": state_subject}
+        if state_entity_ids:
+            anchor = ",".join(str(x) for x in state_entity_ids[:6])
+            rewritten = f"围绕会话实体ID[{anchor}]回答：{current}"
+            return rewritten[:1200], {"mode": "state_entities", "entity_ids": state_entity_ids[:6]}
+        return current, {"mode": "referential_without_subject"}
+
+    def _build_clarification_question(
+        self,
+        *,
+        original_question: str,
+        effective_question: str,
+        session_state: Optional[QASessionState],
+    ) -> Optional[str]:
+        text = str(original_question or "").strip().lower()
+        if not text:
+            return None
+        if self._is_correction_query(text):
+            if self._has_explicit_target_scope(text):
+                return None
+            return (
+                "收到，你在纠正上一条回答。请直接说你要的目标范围，例如："
+                "`我问的是 AI 引擎驱动方式，第一层是否调用 openclaw？`"
+            )
+        has_referential = any(k in text for k in FOLLOWUP_REFERENTIAL_KEYWORDS)
+        if not has_referential:
+            return None
+        state_subject = str(getattr(session_state, "last_subject", "") or "").strip()
+        state_entity_ids = getattr(session_state, "last_entity_ids", None)
+        has_state_entities = isinstance(state_entity_ids, list) and any(str(x).strip() for x in state_entity_ids)
+        if state_subject or has_state_entities:
+            return None
+        if str(effective_question or "").strip() != text and "补充问题" in str(effective_question):
+            return None
+        return "你这句缺少明确主语。请先指明对象，例如：`分别是哪些项目？` 或 `请列出这些项目的名称和ID`。"
+
+    async def _update_session_state_after_turn(
+        self,
+        *,
+        session_state: Optional[QASessionState],
+        question: str,
+        effective_question: str,
+        answer: str,
+        query_type: str,
+        ambiguity: str,
+        rewrite_meta: dict[str, Any],
+        sources: list[dict[str, Any]],
+    ) -> None:
+        if session_state is None:
+            return
+        subject = self._extract_state_subject(question=question, answer=answer, sources=sources)
+        if subject:
+            session_state.last_subject = subject[:255]
+        session_state.last_intent = "question"
+        session_state.ambiguity = str(ambiguity or "clear")[:20]
+        session_state.last_entity_ids = self._collect_entity_ids_from_sources(sources)
+
+        history = session_state.rewrite_history if isinstance(session_state.rewrite_history, list) else []
+        history.append(
+            {
+                "at": datetime.now(timezone.utc).isoformat(),
+                "question": question[:500],
+                "effective_question": effective_question[:500],
+                "query_type": str(query_type or QUERY_TYPE_UNKNOWN),
+                "ambiguity": str(ambiguity or "clear"),
+                "rewrite_meta": rewrite_meta if isinstance(rewrite_meta, dict) else {},
+            }
+        )
+        session_state.rewrite_history = history[-20:]
+        await self.db.flush()
+
+    def _collect_entity_ids_from_sources(self, sources: list[dict[str, Any]]) -> list[int]:
+        ids: list[int] = []
+        for source in sources:
+            if not isinstance(source, dict):
+                continue
+            meta = source.get("metadata") if isinstance(source.get("metadata"), dict) else {}
+            for key in (
+                "project_id",
+                "task_id",
+                "signal_id",
+                "user_id",
+                "entity_id",
+                "subject_entity_id",
+                "object_entity_id",
+            ):
+                value = meta.get(key)
+                try:
+                    if value is not None:
+                        ids.append(int(value))
+                except Exception:
+                    continue
+            matched = meta.get("matched_project_ids")
+            if isinstance(matched, list):
+                for value in matched:
+                    try:
+                        ids.append(int(value))
+                    except Exception:
+                        continue
+        uniq: list[int] = []
+        for value in ids:
+            if value not in uniq:
+                uniq.append(value)
+        return uniq[:20]
+
+    def _extract_state_subject(
+        self,
+        *,
+        question: str,
+        answer: str,
+        sources: list[dict[str, Any]],
+    ) -> str:
+        for source in sources:
+            if not isinstance(source, dict):
+                continue
+            meta = source.get("metadata") if isinstance(source.get("metadata"), dict) else {}
+            project_keyword = str(meta.get("query_keyword") or "").strip()
+            if project_keyword:
+                return project_keyword[:255]
+            subject_name = str(meta.get("subject_name") or "").strip()
+            object_name = str(meta.get("object_name") or "").strip()
+            if subject_name and object_name:
+                return f"{subject_name}->{object_name}"[:255]
+
+        text = str(question or "").strip()
+        keyword = self._extract_project_lookup_keyword(text)
+        if len(keyword) >= 2:
+            return keyword[:255]
+
+        answer_text = str(answer or "").strip()
+        m = re.search(r"项目[「\"]?([^」\"\n]{2,40})", answer_text)
+        if m:
+            return str(m.group(1)).strip()[:255]
+        return ""
+
+    async def get_session_state_replay(self, *, session_id: int, user_id: int) -> dict[str, Any]:
+        session = (
+            await self.db.execute(
+                select(QASession).where(
+                    and_(
+                        QASession.tenant_id == self.tenant_id,
+                        QASession.id == int(session_id),
+                        QASession.user_id == int(user_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if session is None:
+            raise ValueError("qa session 不存在")
+
+        state = await self._ensure_session_state(session_id=int(session_id))
+        history = state.rewrite_history if isinstance(state.rewrite_history, list) else []
+        return {
+            "session_id": int(session_id),
+            "last_subject": str(state.last_subject or "") if state else "",
+            "last_entity_ids": state.last_entity_ids if state else [],
+            "last_intent": str(state.last_intent or "") if state else "",
+            "ambiguity": str(state.ambiguity or "") if state else "",
+            "rewrite_history": history[-20:],
+            "updated_at": state.updated_at.isoformat() if state and state.updated_at else None,
+        }
+
+    async def _try_glm_project_semantic_answer(
+        self,
+        *,
+        clean_q: str,
+        project_id: Optional[int],
+    ) -> Optional[dict[str, Any]]:
+        question = str(clean_q or "").strip()
+        if not question:
+            return None
+        if self._is_people_query(question):
+            return None
+        if not any(k in question for k in ("项目", "系统", "进度", "状态", "有没有", "是否", "存在", "立项", "任务")):
+            return None
+
+        db_started = time.perf_counter()
+        stmt = select(Project).where(Project.tenant_id == self.tenant_id)
+        if project_id is not None:
+            stmt = stmt.where(Project.id == int(project_id))
+        projects = (await self.db.execute(stmt.order_by(Project.updated_at.desc()).limit(120))).scalars().all()
+        if not projects:
+            db_query_ms = int((time.perf_counter() - db_started) * 1000)
+            return {
+                "answer": "当前项目库为空，无法检索到匹配项目。",
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {"table": "projects", "match_count": 0, "retrieval_mode": "glm_semantic"},
+                    }
+                ],
+                "db_query_ms": db_query_ms,
+                "response_time_ms": db_query_ms,
+            }
+
+        project_ids = [int(p.id) for p in projects]
+        task_rows = []
+        if project_ids:
+            task_rows = (
+                await self.db.execute(
+                    select(Task).where(
+                        and_(
+                            Task.tenant_id == self.tenant_id,
+                            Task.project_id.in_(project_ids),
+                        )
+                    )
+                )
+            ).scalars().all()
+        db_query_ms = int((time.perf_counter() - db_started) * 1000)
+
+        done_statuses = {"done", "completed", "已完成", "完成", "closed", "关闭"}
+        doing_statuses = {"in_progress", "doing", "进行中", "开发中", "处理中", "processing"}
+        pending_statuses = {"todo", "pending", "待处理", "未开始", "open"}
+
+        task_stats: dict[int, dict[str, Any]] = {}
+        for pid in project_ids:
+            task_stats[pid] = {
+                "task_total": 0,
+                "done_count": 0,
+                "doing_count": 0,
+                "pending_count": 0,
+                "latest_task_updated_at": None,
+            }
+
+        for row in task_rows:
+            pid = int(row.project_id)
+            if pid not in task_stats:
+                continue
+            stat = task_stats[pid]
+            stat["task_total"] += 1
+            status = str(row.status or "").strip().lower()
+            if status in done_statuses:
+                stat["done_count"] += 1
+            elif status in doing_statuses:
+                stat["doing_count"] += 1
+            elif status in pending_statuses:
+                stat["pending_count"] += 1
+            else:
+                stat["pending_count"] += 1
+            row_time = row.updated_at or row.created_at
+            latest = stat.get("latest_task_updated_at")
+            if row_time and (latest is None or row_time > latest):
+                stat["latest_task_updated_at"] = row_time
+
+        project_index: list[dict[str, Any]] = []
+        for p in projects:
+            pid = int(p.id)
+            stats = task_stats.get(pid) or {}
+            project_index.append(
+                {
+                    "id": pid,
+                    "title": str(p.title or p.name or "")[:120],
+                    "name": str(p.name or "")[:120],
+                    "status": str(p.status or "unknown"),
+                    "description": str(p.description or "")[:220],
+                    "task_total": int(stats.get("task_total") or 0),
+                    "done_count": int(stats.get("done_count") or 0),
+                    "doing_count": int(stats.get("doing_count") or 0),
+                    "pending_count": int(stats.get("pending_count") or 0),
+                    "latest_task_updated_at": (
+                        stats.get("latest_task_updated_at").isoformat()
+                        if stats.get("latest_task_updated_at") is not None
+                        else None
+                    ),
+                }
+            )
+
+        prompt = PROMPT_RAG_DB_PROJECT_QA.format(
+            question=question[:600],
+            project_index=json.dumps(project_index, ensure_ascii=False),
+        ).strip()
+        llm_started = time.perf_counter()
+        try:
+            raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=self.tenant_id,
+                function_name="rag_db_project_semantic_retrieval",
+                temperature=0.0,
+                max_tokens=520,
+                use_prompt_profile=False,
+            )
+            parsed = self._parse_json_object(raw)
+            if not parsed:
+                return None
+            handled = bool(parsed.get("handled"))
+            answer = str(parsed.get("answer") or "").strip()
+            confidence = self._to_float(parsed.get("confidence"), 0.0)
+            matched_ids_raw = parsed.get("matched_project_ids")
+            matched_ids: list[int] = []
+            if isinstance(matched_ids_raw, list):
+                for x in matched_ids_raw:
+                    try:
+                        matched_ids.append(int(x))
+                    except Exception:
+                        continue
+            matched_ids = matched_ids[:8]
+            if not handled or not answer:
+                return None
+            response_ms = int((time.perf_counter() - llm_started) * 1000) + db_query_ms
+            return {
+                "answer": answer,
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": round(max(0.0, min(1.0, confidence or 0.9)), 4),
+                        "metadata": {
+                            "table": "projects",
+                            "matched_project_ids": matched_ids,
+                            "project_index_count": len(project_index),
+                            "retrieval_mode": "glm_semantic",
+                        },
+                    }
+                ],
+                "db_query_ms": db_query_ms,
+                "response_time_ms": response_ms,
+            }
+        except Exception:
+            return None
+
+    async def _persist_fast_answer(
+        self,
+        *,
+        question: str,
+        asked_by: Optional[int],
+        asked_via: str,
+        project_id: Optional[int],
+        fast_result: dict[str, Any],
+        session_id: Optional[int] = None,
+        external_session_id: Optional[int] = None,
+        turn_index: Optional[int] = None,
+        query_type: str = QUERY_TYPE_EXPLAIN,
+        missing_fields: Optional[list[str]] = None,
+        state_row: Optional[QASessionState] = None,
+        effective_question: Optional[str] = None,
+        rewrite_meta: Optional[dict[str, Any]] = None,
+        thinking_mode: str = THINKING_MODE_NORMAL,
+    ) -> dict[str, Any]:
+        started = time.perf_counter()
+        answer_text = str(fast_result.get("answer") or "未命中结果")
+        sources = list(fast_result.get("sources") or [])
+        active_session_id = session_id
+        active_turn_index = turn_index
+        if active_session_id is None or active_turn_index is None:
+            active_session_id, active_turn_index = await self._touch_session_for_new_record(
+                asked_by=asked_by,
+                asked_via=asked_via,
+                project_id=project_id,
+                question=question,
+            )
+        active_turn_index = int(active_turn_index or 1)
+        active_session_id_int = int(active_session_id) if active_session_id is not None else None
+
+        confidence = self._to_float(fast_result.get("confidence"), 0.96)
+        if missing_fields is None:
+            missing_fields = list(fast_result.get("missing_fields") or [])
+        final_query_type = str(fast_result.get("query_type") or query_type or QUERY_TYPE_EXPLAIN)
+
+        qa = QARecord(
+            tenant_id=self.tenant_id,
+            question=question,
+            answer=answer_text,
+            sources=sources,
+            intent="question",
+            domains_used=["db"],
+            response_time_ms=int(fast_result.get("response_time_ms") or 0),
+            asked_by=asked_by,
+            asked_via=(asked_via or "pmflow")[:20],
+            project_id=project_id,
+            session_id=active_session_id_int,
+            turn_index=active_turn_index,
+        )
+        self.db.add(qa)
+        if state_row is None and active_session_id_int is not None:
+            state_row = await self._ensure_session_state(session_id=active_session_id_int)
+        await self._update_session_state_after_turn(
+            session_state=state_row,
+            question=question,
+            effective_question=str(effective_question or question),
+            answer=answer_text,
+            query_type=final_query_type,
+            ambiguity="clear",
+            rewrite_meta=rewrite_meta or {},
+            sources=sources,
+        )
+        await self.db.commit()
+        await self.db.refresh(qa)
+        cache_session_id = active_session_id_int if active_session_id_int is not None else (
+            int(external_session_id) if external_session_id is not None else None
+        )
+        await self.context_cache.append_session_messages(
+            tenant_id=self.tenant_id,
+            session_id=cache_session_id,
+            question=question,
+            answer=answer_text,
+        )
+        elapsed = int((time.perf_counter() - started) * 1000)
+        return self._build_response_payload(
+            answer=answer_text,
+            sources=sources,
+            domains_used=["db"],
+            response_time_ms=int(fast_result.get("response_time_ms") or elapsed),
+            db_query_ms=int(fast_result.get("db_query_ms") or 0),
+            vector_search_ms=0,
+            db_timeout=False,
+            vector_timeout=False,
+            qa_id=int(qa.id),
+            session_id=active_session_id_int,
+            turn_index=active_turn_index,
+            relevance_max=0.99,
+            query_type=final_query_type,
+            confidence=confidence,
+            missing_fields=list(missing_fields or []),
+            thinking_mode=thinking_mode,
+        )
+
+    async def _try_fast_project_progress_answer(
+        self,
+        *,
+        clean_q: str,
+        project_id: Optional[int],
+    ) -> Optional[dict[str, Any]]:
+        question = str(clean_q or "").strip()
+        if not question:
+            return None
+        if not PROJECT_PROGRESS_RE.search(question):
+            return None
+        if "项目" not in question and "系统" not in question:
+            return None
+
+        primary_keyword = self._extract_project_lookup_keyword(question)
+        query_terms = self._extract_query_terms(question, preferred=primary_keyword)
+        if not query_terms:
+            return None
+
+        started = time.perf_counter()
+        matched: list[Project] = []
+        used_keyword = query_terms[0]
+        for term in query_terms:
+            rows = await self._search_projects_by_keyword(keyword=term, project_id=project_id, limit=3)
+            if rows:
+                matched = rows
+                used_keyword = term
+                break
+        if not matched:
+            db_query_ms = int((time.perf_counter() - started) * 1000)
+            return {
+                "answer": f"未检索到名称或描述包含“{used_keyword}”的项目，无法判断进度。",
+                "sources": [
+                    {
+                        "source": "db",
+                        "label": "Project database",
+                        "score": 0.99,
+                        "metadata": {"table": "projects", "query_keyword": used_keyword, "match_count": 0},
+                    }
+                ],
+                "db_query_ms": db_query_ms,
+                "response_time_ms": db_query_ms,
+            }
+
+        target = matched[0]
+        task_rows = (
+            await self.db.execute(
+                select(Task).where(
+                    and_(
+                        Task.tenant_id == self.tenant_id,
+                        Task.project_id == int(target.id),
+                    )
+                )
+            )
+        ).scalars().all()
+        db_query_ms = int((time.perf_counter() - started) * 1000)
+
+        total = len(task_rows)
+        done_statuses = {"done", "completed", "已完成", "完成", "closed", "关闭"}
+        doing_statuses = {"in_progress", "doing", "进行中", "开发中", "处理中", "processing"}
+        pending_statuses = {"todo", "pending", "待处理", "未开始", "open"}
+
+        done = 0
+        doing = 0
+        pending = 0
+        latest_update: Optional[datetime] = None
+        for row in task_rows:
+            status = str(row.status or "").strip().lower()
+            if status in done_statuses:
+                done += 1
+            elif status in doing_statuses:
+                doing += 1
+            elif status in pending_statuses:
+                pending += 1
+            else:
+                pending += 1
+
+            row_time = row.updated_at or row.created_at
+            if row_time and (latest_update is None or row_time > latest_update):
+                latest_update = row_time
+
+        project_name = str(target.title or target.name or f"项目#{int(target.id)}")
+        project_status = str(target.status or "unknown")
+        if total == 0:
+            answer = (
+                f"项目「{project_name}」当前状态为 {project_status}。"
+                "目前未检索到该项目的任务拆解记录，因此只能确认项目状态，无法量化阶段进度。"
+            )
+        else:
+            completion = round((done / total) * 100, 1)
+            latest_text = latest_update.strftime("%Y-%m-%d %H:%M") if latest_update else "未知"
+            answer = (
+                f"项目「{project_name}」当前状态为 {project_status}。"
+                f"任务总数 {total}，已完成 {done}，进行中 {doing}，待处理 {pending}，完成率 {completion}%。"
+                f"最近任务更新时间：{latest_text}。"
+            )
+
+        if len(matched) > 1:
+            others = "、".join([str(x.title or x.name or f'#{x.id}') for x in matched[1:]])
+            answer += f" 另有 {len(matched) - 1} 个同名候选项目：{others}。"
+
+        return {
+            "answer": answer,
+            "sources": [
+                {
+                    "source": "db",
+                    "label": "Project database",
+                    "score": 0.99,
+                    "metadata": {
+                        "table": "projects",
+                        "project_id": int(target.id),
+                        "query_keyword": used_keyword,
+                        "match_count": len(matched),
+                    },
+                },
+                {
+                    "source": "db",
+                    "label": "Project database",
+                    "score": 0.98,
+                    "metadata": {
+                        "table": "tasks",
+                        "project_id": int(target.id),
+                        "task_total": total,
+                        "done_count": done,
+                        "doing_count": doing,
+                        "pending_count": pending,
+                    },
+                },
+            ],
+            "db_query_ms": db_query_ms,
+            "response_time_ms": db_query_ms,
+        }
+
+    async def _try_fast_project_existence_answer(
+        self,
+        *,
+        clean_q: str,
+        project_id: Optional[int],
+    ) -> Optional[dict[str, Any]]:
+        question = str(clean_q or "").strip()
+        if not question:
+            return None
+        if "项目" not in question and "系统" not in question:
+            return None
+        if (
+            "吗" not in question
+            and "?" not in question
+            and "？" not in question
+            and "是否有" not in question
+            and "有没有" not in question
+            and "有无" not in question
+            and not PROJECT_EXISTENCE_RE.search(question)
+        ):
+            return None
+
+        primary_keyword = self._extract_project_lookup_keyword(question)
+        query_terms = self._extract_query_terms(question, preferred=primary_keyword)
+        if not query_terms:
+            return None
+
+        started = time.perf_counter()
+        matched: list[Project] = []
+        used_keyword = query_terms[0]
+        for term in query_terms:
+            rows = await self._search_projects_by_keyword(keyword=term, project_id=project_id, limit=5)
+            if rows:
+                matched = rows
+                used_keyword = term
+                break
+        db_query_ms = int((time.perf_counter() - started) * 1000)
+
+        if matched:
+            lines = [
+                f"#{int(p.id)} {str(p.title or p.name or '未命名项目')}（状态:{str(p.status or 'unknown')}）"
+                for p in matched
+            ]
+            answer = f"有，当前匹配到 {len(matched)} 个相关项目：\n" + "\n".join(lines)
+        else:
+            answer = f"暂未检索到名称或描述包含“{used_keyword}”的项目。"
+
+        return {
+            "answer": answer,
+            "sources": [
+                {
+                    "source": "db",
+                    "label": "Project database",
+                    "score": 0.99,
+                    "metadata": {
+                        "table": "projects",
+                        "query_keyword": used_keyword,
+                        "match_count": len(matched),
+                    },
+                }
+            ],
+            "db_query_ms": db_query_ms,
+        }
+
+    async def _search_projects_by_keyword(
+        self,
+        *,
+        keyword: str,
+        project_id: Optional[int],
+        limit: int,
+    ) -> list[Project]:
+        term = str(keyword or "").strip()
+        if not term:
+            return []
+        stmt = select(Project).where(Project.tenant_id == self.tenant_id)
+        if project_id is not None:
+            stmt = stmt.where(Project.id == int(project_id))
+        like = f"%{term}%"
+        stmt = stmt.where(
+            or_(
+                Project.name.ilike(like),
+                Project.title.ilike(like),
+                Project.description.ilike(like),
+            )
+        )
+        return (await self.db.execute(stmt.limit(max(1, int(limit))))).scalars().all()
+
+    @staticmethod
+    def _extract_project_lookup_keyword(question: str) -> str:
+        text = str(question or "").strip()
+        if not text:
+            return ""
+        text = text.replace("？", "?")
+
+        direct_patterns = [
+            r"(?:是否有|有没有|有无)([\u4e00-\u9fffA-Za-z0-9_-]{1,24})(?:项目|系统)?",
+            r"(?:是否存在|存在)([\u4e00-\u9fffA-Za-z0-9_-]{1,24})(?:项目|系统)?",
+        ]
+        for pattern in direct_patterns:
+            m = re.search(pattern, text)
+            if m:
+                value = str(m.group(1) or "").strip(" ，。,.!?？：:;；“”\"'()（）")
+                value = QUESTION_NOISE_RE.sub("", value).strip()
+                if len(value) >= 2:
+                    return value[:24]
+
+        anchored_matches = list(re.finditer(r"([\u4e00-\u9fffA-Za-z0-9_-]{1,24})(?:项目|系统)", text))
+        for m in reversed(anchored_matches):
+            value = str(m.group(1) or "").strip(" ，。,.!?？：:;；“”\"'()（）")
+            value = QUESTION_NOISE_RE.sub("", value).strip()
+            if len(value) >= 2:
+                return value[:24]
+
+        patterns = [
+            r"有(?:没有)?(.{1,24}?)(?:项目|系统).*\?",
+            r"项目.*?有(?:没有)?(.{1,24}?)(?:吗|\?)",
+            r"是否有(.{1,24}?)(?:项目|系统)",
+            r"(.{1,24}?)(?:项目|系统)(?:有吗|有没有|是否有)",
+        ]
+        for pattern in patterns:
+            m = re.search(pattern, text)
+            if not m:
+                continue
+            kw = str(m.group(1) or "").strip(" ，。,.!?？：:;；“”\"'()（）")
+            kw = QUESTION_NOISE_RE.sub("", kw).strip()
+            if kw:
+                return kw[:24]
+
+        fallback = re.sub(r"[?？]", " ", text)
+        fallback = re.sub(
+            r"(项目|系统|目前|当前|现在|我们|有没有|是否有|有吗|有|吗|什么|进度|层面|阶段|是在|在|是|请问|一下|多少)",
+            " ",
+            fallback,
+        )
+        fallback = re.sub(r"\s+", " ", fallback).strip(" ，。,.!?？：:;；")
+        return fallback[:24]
+
+    @staticmethod
+    def _extract_query_terms(question: str, *, preferred: str = "") -> list[str]:
+        base = str(question or "").strip()
+        terms: list[str] = []
+        pref = str(preferred or "").strip()
+        if len(pref) >= 2:
+            terms.append(pref[:24])
+
+        clean = re.sub(r"[?？,，。!！:：;；()（）\"“”']", " ", base)
+        clean = QUESTION_NOISE_RE.sub(" ", clean)
+        clean = re.sub(r"(项目|系统|进度|状态|阶段|完成率|层面)", " ", clean)
+        clean = re.sub(r"\s+", " ", clean).strip()
+
+        token_candidates: list[str] = []
+        for token in clean.split(" "):
+            t = token.strip()
+            if len(t) >= 2:
+                token_candidates.append(t[:24])
+
+        for m in re.finditer(r"[\u4e00-\u9fff]{2,10}", clean):
+            token_candidates.append(m.group(0)[:24])
+
+        for t in sorted(token_candidates, key=len, reverse=True):
+            if t and t not in terms:
+                terms.append(t)
+        return terms[:6]
+
+    @staticmethod
+    def _is_people_query(question: str) -> bool:
+        text = str(question or "").strip().lower()
+        if not text:
+            return False
+        return any(k in text for k in PEOPLE_KEYWORDS)
+
+    # ------------------------------------------------------------------
+    # Manifest / Entity resolution / Agentic retrieval
+    # ------------------------------------------------------------------
+    async def get_project_manifest(self, *, project_id: Optional[int] = None) -> dict[str, Any]:
+        system_snapshot = await self.context_cache.get_system_snapshot(
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        symbol_snapshot = await self._get_symbol_snapshot_cached()
+
+        project_info: dict[str, Any] = {}
+        if project_id is not None:
+            row = (
+                await self.db.execute(
+                    select(Project).where(
+                        and_(
+                            Project.tenant_id == self.tenant_id,
+                            Project.id == int(project_id),
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if row is not None:
+                project_info = {
+                    "id": int(row.id),
+                    "title": str(row.title or row.name or f"项目#{int(row.id)}"),
+                    "status": str(row.status or "unknown"),
+                }
+
+        kg_overview = {"entity_total": 0, "relation_total": 0, "evidence_total": 0}
+        try:
+            kg_overview = {
+                "entity_total": int(
+                    (await self.db.execute(select(func.count(KGEntity.id)).where(KGEntity.tenant_id == self.tenant_id)))
+                    .scalar_one()
+                    or 0
+                ),
+                "relation_total": int(
+                    (
+                        await self.db.execute(select(func.count(KGRelation.id)).where(KGRelation.tenant_id == self.tenant_id))
+                    ).scalar_one()
+                    or 0
+                ),
+                "evidence_total": int(
+                    (
+                        await self.db.execute(select(func.count(KGEvidence.id)).where(KGEvidence.tenant_id == self.tenant_id))
+                    ).scalar_one()
+                    or 0
+                ),
+            }
+        except Exception:
+            kg_overview = {"entity_total": 0, "relation_total": 0, "evidence_total": 0}
+
+        platform = (
+            system_snapshot.get("platform_profile")
+            if isinstance(system_snapshot.get("platform_profile"), dict)
+            else {}
+        )
+        backend = platform.get("backend") if isinstance(platform.get("backend"), dict) else {}
+        module_counts = backend.get("module_counts") if isinstance(backend.get("module_counts"), dict) else {}
+        feature_catalog = platform.get("feature_catalog") if isinstance(platform.get("feature_catalog"), list) else []
+        feature_summary = {
+            "feature_total": len(feature_catalog),
+            "top_features": [
+                {
+                    "name": str(item.get("name") or ""),
+                    "route": str(item.get("route") or ""),
+                }
+                for item in feature_catalog[:10]
+                if isinstance(item, dict)
+            ],
+        }
+
+        return {
+            "tenant_id": self.tenant_id,
+            "project": project_info,
+            "module_counts": module_counts,
+            "feature_summary": feature_summary,
+            "symbol_summary": symbol_snapshot,
+            "kg_overview": kg_overview,
+            "manifest_text": self._compose_manifest_text(
+                project_info=project_info,
+                module_counts=module_counts,
+                feature_summary=feature_summary,
+                symbol_snapshot=symbol_snapshot,
+                kg_overview=kg_overview,
+            ),
+            "generated_at": datetime.now(timezone.utc).isoformat(),
+        }
+
+    async def entity_resolution(
+        self,
+        *,
+        question: str,
+        project_id: Optional[int] = None,
+        limit: int = 10,
+    ) -> dict[str, Any]:
+        normalized = self._normalize_entity_text(question)
+        alias_terms = self._collect_alias_terms(normalized)
+        query_terms = self._extract_query_terms(question, preferred=self._extract_project_lookup_keyword(question))
+        for term in alias_terms:
+            if term not in query_terms:
+                query_terms.append(term)
+
+        candidates: list[dict[str, Any]] = []
+        if query_terms:
+            try:
+                max_terms = query_terms[:6]
+                conditions = []
+                for term in max_terms:
+                    like = f"%{term}%"
+                    conditions.append(KGEntity.display_name.ilike(like))
+                    conditions.append(KGEntity.entity_key.ilike(like))
+                rows = (
+                    await self.db.execute(
+                        select(KGEntity)
+                        .where(
+                            and_(
+                                KGEntity.tenant_id == self.tenant_id,
+                                or_(*conditions),
+                            )
+                        )
+                        .order_by(KGEntity.last_seen_at.desc(), KGEntity.id.desc())
+                        .limit(max(20, int(limit) * 4))
+                    )
+                ).scalars().all()
+                for row in rows:
+                    payload = self._safe_dict(row.payload)
+                    row_project_id = payload.get("project_id")
+                    if project_id is not None and row_project_id is not None and int(row_project_id) != int(project_id):
+                        continue
+                    score = 0.0
+                    lowered_name = str(row.display_name or "").lower()
+                    lowered_key = str(row.entity_key or "").lower()
+                    for term in max_terms:
+                        term_lower = term.lower()
+                        if term_lower in lowered_name:
+                            score += 0.34
+                        if term_lower in lowered_key:
+                            score += 0.22
+                    if score <= 0:
+                        continue
+                    candidates.append(
+                        {
+                            "entity_id": int(row.id),
+                            "entity_type": str(row.entity_type or ""),
+                            "entity_key": str(row.entity_key or ""),
+                            "display_name": str(row.display_name or ""),
+                            "score": round(min(score, 0.99), 4),
+                        }
+                    )
+            except Exception:
+                candidates = []
+        candidates.sort(key=lambda x: float(x.get("score") or 0.0), reverse=True)
+
+        return {
+            "normalized_question": normalized,
+            "canonical_terms": alias_terms,
+            "query_terms": query_terms[:8],
+            "matched_entities": candidates[: max(1, int(limit))],
+        }
+
+    async def get_feature_catalog(
+        self,
+        *,
+        query: Optional[str] = None,
+        limit: int = 20,
+    ) -> dict[str, Any]:
+        system_snapshot = await self.context_cache.get_system_snapshot(
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        catalog = self._feature_catalog_from_snapshot(system_snapshot)
+        if query and str(query).strip():
+            rows = self._match_feature_catalog_entries(
+                question=str(query),
+                system_snapshot=system_snapshot,
+                limit=max(1, min(int(limit), 100)),
+            )
+            return {
+                "query": str(query),
+                "count": len(rows),
+                "items": rows,
+            }
+        max_rows = max(1, min(int(limit), 200))
+        return {
+            "query": str(query or ""),
+            "count": min(len(catalog), max_rows),
+            "items": catalog[:max_rows],
+        }
+
+    def _compose_manifest_text(
+        self,
+        *,
+        project_info: dict[str, Any],
+        module_counts: dict[str, Any],
+        feature_summary: dict[str, Any],
+        symbol_snapshot: dict[str, Any],
+        kg_overview: dict[str, Any],
+    ) -> str:
+        lines: list[str] = []
+        if project_info:
+            lines.append(
+                f"当前项目：#{int(project_info.get('id') or 0)} {str(project_info.get('title') or 'unknown')} "
+                f"(status={str(project_info.get('status') or 'unknown')})"
+            )
+        lines.append(
+            "后端模块规模："
+            f"api={int(module_counts.get('api', 0))},"
+            f"services={int(module_counts.get('services', 0))},"
+                f"models={int(module_counts.get('models', 0))},"
+                f"schemas={int(module_counts.get('schemas', 0))}"
+        )
+        lines.append(f"产品功能：total={int(feature_summary.get('feature_total', 0))}")
+        top_features = feature_summary.get("top_features")
+        if isinstance(top_features, list) and top_features:
+            feature_brief = "；".join(
+                f"{str(item.get('name') or '')}({str(item.get('route') or '')})"
+                for item in top_features[:6]
+                if isinstance(item, dict) and str(item.get("name") or "").strip()
+            )
+            if feature_brief:
+                lines.append(f"核心入口：{feature_brief}")
+        lines.append(
+            "符号索引："
+            f"files={int(symbol_snapshot.get('indexed_files', 0))},"
+            f"symbols={int(symbol_snapshot.get('symbol_total', 0))},"
+            f"python={int(symbol_snapshot.get('python_symbols', 0))},"
+            f"typescript={int(symbol_snapshot.get('typescript_symbols', 0))}"
+        )
+        lines.append(
+            "知识图谱："
+            f"entities={int(kg_overview.get('entity_total', 0))},"
+            f"relations={int(kg_overview.get('relation_total', 0))},"
+            f"evidences={int(kg_overview.get('evidence_total', 0))}"
+        )
+        top_modules = symbol_snapshot.get("top_modules")
+        if isinstance(top_modules, list) and top_modules:
+            brief = "；".join(
+                f"{str(item.get('module') or '')}:{int(item.get('symbols') or 0)}"
+                for item in top_modules[:5]
+                if isinstance(item, dict)
+            )
+            if brief:
+                lines.append(f"热点模块：{brief}")
+        return "\n".join(lines)[:2200]
+
+    @staticmethod
+    def _normalize_entity_text(text: str) -> str:
+        normalized = str(text or "").strip().lower()
+        normalized = normalized.replace("—", "-").replace("_", " ")
+        normalized = re.sub(r"[^\w\u4e00-\u9fff\- ]+", " ", normalized)
+        normalized = re.sub(r"\s+", " ", normalized).strip()
+        return normalized
+
+    def _collect_alias_terms(self, normalized_question: str) -> list[str]:
+        aliases: list[str] = []
+        for token, canonical in ENTITY_ALIAS_MAP.items():
+            if token in normalized_question and canonical not in aliases:
+                aliases.append(canonical)
+        return aliases[:8]
+
+    @staticmethod
+    def _feature_catalog_from_snapshot(system_snapshot: Optional[dict[str, Any]]) -> list[dict[str, Any]]:
+        if not isinstance(system_snapshot, dict):
+            return []
+        platform = system_snapshot.get("platform_profile")
+        if not isinstance(platform, dict):
+            return []
+        catalog = platform.get("feature_catalog")
+        if not isinstance(catalog, list):
+            return []
+        return [item for item in catalog if isinstance(item, dict)]
+
+    def _match_feature_catalog_entries(
+        self,
+        *,
+        question: str,
+        system_snapshot: Optional[dict[str, Any]],
+        limit: int = 4,
+    ) -> list[dict[str, Any]]:
+        text = str(question or "").strip().lower()
+        if not text:
+            return []
+        catalog = self._feature_catalog_from_snapshot(system_snapshot)
+        if not catalog:
+            return []
+        query_terms = [x.lower() for x in self._extract_query_terms(question)[:8] if str(x).strip()]
+        matched: list[dict[str, Any]] = []
+        for item in catalog:
+            name = str(item.get("name") or "").strip()
+            route = str(item.get("route") or "").strip()
+            component = str(item.get("page_component") or "").strip()
+            aliases_raw = item.get("aliases")
+            alias_values = aliases_raw if isinstance(aliases_raw, list) else []
+            alias_pool = [name, route, component, *[str(x) for x in alias_values]]
+            alias_pool = [a.strip().lower() for a in alias_pool if str(a).strip()]
+            if not alias_pool:
+                continue
+
+            score = 0.0
+            for alias in alias_pool:
+                if alias and alias in text:
+                    score += 0.45
+            for term in query_terms:
+                if term and any(term in alias or alias in term for alias in alias_pool):
+                    score += 0.18
+            if score <= 0:
+                continue
+            matched.append(
+                {
+                    **item,
+                    "match_score": round(min(score, 0.99), 4),
+                }
+            )
+
+        matched.sort(key=lambda x: float(x.get("match_score") or 0.0), reverse=True)
+        return matched[: max(1, int(limit))]
+
+    def _retrieve_feature_catalog_context(
+        self,
+        *,
+        question: str,
+        system_snapshot: Optional[dict[str, Any]],
+        max_items: int = 4,
+    ) -> list[RetrievedItem]:
+        rows = self._match_feature_catalog_entries(
+            question=question,
+            system_snapshot=system_snapshot,
+            limit=max_items,
+        )
+        out: list[RetrievedItem] = []
+        for row in rows:
+            name = str(row.get("name") or "unknown_feature")
+            route = str(row.get("route") or "")
+            summary = str(row.get("summary") or "")
+            page_file = str(row.get("page_file") or "")
+            page_component = str(row.get("page_component") or "")
+            apis = row.get("apis") if isinstance(row.get("apis"), list) else []
+            api_text = "、".join(str(x) for x in apis[:4] if str(x).strip())
+            content = f"功能: {name}\n用途: {summary}\n入口: {route}\n页面: {page_file}"
+            if api_text:
+                content += f"\n相关接口: {api_text}"
+            score = min(0.99, max(0.62, float(row.get("match_score") or 0.62)))
+            out.append(
+                RetrievedItem(
+                    source="feature_catalog",
+                    content=content[:1200],
+                    score=score,
+                    source_label="Feature Catalog",
+                    metadata={
+                        "feature_name": name,
+                        "route": route,
+                        "page_file": page_file,
+                        "page_component": page_component,
+                        "apis": [str(x) for x in apis[:8] if str(x).strip()],
+                        "match_score": round(score, 4),
+                    },
+                )
+            )
+        return out
+
+    def _collect_feature_evidence(
+        self,
+        *,
+        question: str,
+        retrieved: list[RetrievedItem],
+        system_snapshot: Optional[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        rows: list[dict[str, Any]] = []
+        for item in retrieved:
+            if str(item.source or "").strip().lower() != "feature_catalog":
+                continue
+            meta = item.metadata if isinstance(item.metadata, dict) else {}
+            route = str(meta.get("route") or "").strip()
+            page_file = str(meta.get("page_file") or "").strip()
+            feature_name = str(meta.get("feature_name") or "").strip()
+            apis = meta.get("apis") if isinstance(meta.get("apis"), list) else []
+            clean_apis = [str(x).strip() for x in apis if str(x).strip()]
+            if not feature_name and not route and not page_file:
+                continue
+            rows.append(
+                {
+                    "feature_name": feature_name,
+                    "route": route,
+                    "page_file": page_file,
+                    "apis": clean_apis[:6],
+                }
+            )
+        if rows:
+            return rows[:4]
+
+        # 若检索结果里没有 feature_catalog，回退到 snapshot 匹配结果
+        fallback = self._match_feature_catalog_entries(
+            question=question,
+            system_snapshot=system_snapshot,
+            limit=3,
+        )
+        for row in fallback:
+            apis = row.get("apis") if isinstance(row.get("apis"), list) else []
+            rows.append(
+                {
+                    "feature_name": str(row.get("name") or ""),
+                    "route": str(row.get("route") or ""),
+                    "page_file": str(row.get("page_file") or ""),
+                    "apis": [str(x).strip() for x in apis[:6] if str(x).strip()],
+                }
+            )
+        return rows[:3]
+
+    def _enforce_feature_evidence_in_answer(
+        self,
+        *,
+        question: str,
+        answer_text: str,
+        retrieved: list[RetrievedItem],
+        system_snapshot: Optional[dict[str, Any]],
+    ) -> str:
+        text = str(question or "").strip().lower()
+        has_feature_token = any(k in text for k in FEATURE_QUERY_KEYWORDS)
+        ask_feature = has_feature_token and any(k in text for k in FEATURE_PURPOSE_HINTS)
+        if not ask_feature:
+            return answer_text
+
+        evidence_rows = self._collect_feature_evidence(
+            question=question,
+            retrieved=retrieved,
+            system_snapshot=system_snapshot,
+        )
+        if not evidence_rows:
+            return answer_text
+
+        # 已包含明确入口与接口信息时不重复追加
+        lowered_answer = str(answer_text or "").lower()
+        if ("/" in lowered_answer and "接口" in lowered_answer) or "证据补充" in lowered_answer:
+            return answer_text
+
+        lines: list[str] = [str(answer_text or "").strip(), "", "证据补充（Feature Catalog）："]
+        for idx, row in enumerate(evidence_rows, start=1):
+            feature_name = str(row.get("feature_name") or "未知功能")
+            route = str(row.get("route") or "")
+            page_file = str(row.get("page_file") or "")
+            apis = row.get("apis") if isinstance(row.get("apis"), list) else []
+            api_text = "、".join(str(x) for x in apis[:3] if str(x).strip())
+            line = f"{idx}. {feature_name}"
+            if route:
+                line += f" | 路由: {route}"
+            if page_file:
+                line += f" | 页面: {page_file}"
+            if api_text:
+                line += f" | API: {api_text}"
+            lines.append(line)
+        return "\n".join(lines).strip()[:3200]
+
+    def _symbol_cache_key(self) -> str:
+        return f"{self.tenant_id}:backend_symbol_snapshot"
+
+    async def _get_symbol_snapshot_cached(self, *, ttl_seconds: int = 300) -> dict[str, Any]:
+        now = time.time()
+        key = self._symbol_cache_key()
+        cached = self._symbol_snapshot_cache.get(key)
+        if cached and cached[0] > now:
+            return dict(cached[1])
+        try:
+            service = SymbolIndexService(self.db, tenant_id=self.tenant_id)
+            built = await service.get_overview(top_n_files=10, auto_sync_if_empty=True)
+        except Exception:
+            built = self._build_symbol_index_snapshot()
+        self._symbol_snapshot_cache[key] = (now + max(60, int(ttl_seconds)), dict(built))
+        return built
+
+    def _build_symbol_index_snapshot(
+        self,
+        *,
+        max_files: int = 240,
+        max_symbols: int = 4000,
+    ) -> dict[str, Any]:
+        backend_root = Path(__file__).resolve().parents[2]
+        app_root = backend_root / "app"
+        if not app_root.exists():
+            return {
+                "root": str(app_root),
+                "indexed_files": 0,
+                "symbol_total": 0,
+                "python_symbols": 0,
+                "typescript_symbols": 0,
+                "top_modules": [],
+                "generated_at": datetime.now(timezone.utc).isoformat(),
+            }
+
+        candidates = [
+            p
+            for p in app_root.rglob("*")
+            if p.is_file() and p.suffix.lower() in {".py", ".ts", ".tsx", ".js", ".jsx"}
+        ]
+        candidates.sort(key=lambda p: str(p))
+        file_paths = candidates[: max(1, int(max_files))]
+
+        python_symbols = 0
+        ts_symbols = 0
+        total_symbols = 0
+        module_counter: dict[str, int] = {}
+
+        for path in file_paths:
+            rel = str(path.relative_to(backend_root))
+            module = rel.split("/", 2)[:2]
+            module_name = "/".join(module) if module else rel
+            try:
+                text = path.read_text(encoding="utf-8", errors="ignore")
+            except Exception:
+                continue
+
+            symbols_in_file = 0
+            if path.suffix.lower() == ".py":
+                extracted = self._extract_python_symbols(text)
+                symbols_in_file = len(extracted)
+                python_symbols += symbols_in_file
+            else:
+                extracted = self._extract_text_symbols(text)
+                symbols_in_file = len(extracted)
+                ts_symbols += symbols_in_file
+            if symbols_in_file <= 0:
+                continue
+            module_counter[module_name] = int(module_counter.get(module_name, 0)) + symbols_in_file
+            total_symbols += symbols_in_file
+            if total_symbols >= max_symbols:
+                break
+
+        top_modules = [
+            {"module": name, "symbols": int(count)}
+            for name, count in sorted(module_counter.items(), key=lambda item: item[1], reverse=True)[:10]
+        ]
+        return {
+            "root": str(app_root),
+            "indexed_files": len(file_paths),
+            "symbol_total": int(total_symbols),
+            "python_symbols": int(python_symbols),
+            "typescript_symbols": int(ts_symbols),
+            "top_modules": top_modules,
+            "generated_at": datetime.now(timezone.utc).isoformat(),
+        }
+
+    @staticmethod
+    def _extract_python_symbols(text: str) -> list[str]:
+        raw = str(text or "")
+        if not raw.strip():
+            return []
+        try:
+            tree = ast.parse(raw)
+        except Exception:
+            return []
+        out: list[str] = []
+        for node in tree.body:
+            if isinstance(node, ast.ClassDef):
+                out.append(f"class:{node.name}")
+            elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                out.append(f"function:{node.name}")
+        return out
+
+    @staticmethod
+    def _extract_text_symbols(text: str) -> list[str]:
+        raw = str(text or "")
+        if not raw.strip():
+            return []
+        out: list[str] = []
+        for match in TS_SYMBOL_RE.finditer(raw):
+            name = str(match.group(1) or "").strip()
+            if name:
+                out.append(name)
+        return out
+
+    async def _run_single_retrieval_pass(
+        self,
+        *,
+        question: str,
+        domains: list[str],
+        project_id: Optional[int],
+        source_chat_id: Optional[str],
+        resolved_entities: Optional[dict[str, Any]] = None,
+        vector_top_k: Optional[int] = None,
+    ) -> tuple[list[RetrievedItem], dict[str, Any]]:
+        async def _run_db_retrieval() -> tuple[list[RetrievedItem], int, bool]:
+            if "db" not in domains:
+                return ([], 0, False)
+            started_at = time.perf_counter()
+            try:
+                rows = await asyncio.wait_for(
+                    self._retrieve_structured_db(question, project_id=project_id),
+                    timeout=DB_QUERY_TIMEOUT_SECONDS,
+                )
+                return (rows, int((time.perf_counter() - started_at) * 1000), False)
+            except asyncio.TimeoutError:
+                return ([], int((time.perf_counter() - started_at) * 1000), True)
+
+        async def _run_vector_retrieval() -> tuple[list[RetrievedItem], int, bool]:
+            if not any(x in domains for x in {"doc", "feishu", "pmflow", "url"}):
+                return ([], 0, False)
+            started_at = time.perf_counter()
+            try:
+                rows = await asyncio.wait_for(
+                    self._retrieve_vector_context(
+                        question,
+                        domains=domains,
+                        project_id=project_id,
+                        source_chat_id=source_chat_id,
+                        top_k=vector_top_k,
+                    ),
+                    timeout=VECTOR_SEARCH_TIMEOUT_SECONDS,
+                )
+                return (rows, int((time.perf_counter() - started_at) * 1000), False)
+            except asyncio.TimeoutError:
+                return ([], int((time.perf_counter() - started_at) * 1000), True)
+
+        async def _run_graph_retrieval() -> tuple[list[RetrievedItem], int, bool]:
+            if "graph" not in domains:
+                return ([], 0, False)
+            started_at = time.perf_counter()
+            try:
+                rows = await asyncio.wait_for(
+                    self._retrieve_graph_context(
+                        question,
+                        project_id=project_id,
+                        resolved_entities=resolved_entities,
+                    ),
+                    timeout=GRAPH_QUERY_TIMEOUT_SECONDS,
+                )
+                return (rows, int((time.perf_counter() - started_at) * 1000), False)
+            except asyncio.TimeoutError:
+                return ([], int((time.perf_counter() - started_at) * 1000), True)
+
+        (db_rows, db_ms, db_timeout), (vec_rows, vec_ms, vec_timeout), (graph_rows, graph_ms, graph_timeout) = (
+            await asyncio.gather(
+                _run_db_retrieval(),
+                _run_vector_retrieval(),
+                _run_graph_retrieval(),
+            )
+        )
+        merged = [*db_rows, *vec_rows, *graph_rows]
+        merged.sort(key=lambda x: x.score, reverse=True)
+        return (
+            merged,
+            {
+                "db_query_ms": int(db_ms),
+                "vector_search_ms": int(vec_ms),
+                "graph_query_ms": int(graph_ms),
+                "db_timeout": bool(db_timeout),
+                "vector_timeout": bool(vec_timeout),
+                "graph_timeout": bool(graph_timeout),
+            },
+        )
+
+    async def _build_second_pass_query(
+        self,
+        *,
+        question: str,
+        project_id: Optional[int],
+    ) -> tuple[str, dict[str, Any]]:
+        resolved = await self.entity_resolution(question=question, project_id=project_id, limit=5)
+        manifest = await self.get_project_manifest(project_id=project_id)
+        hints: list[str] = []
+        for term in resolved.get("canonical_terms") or []:
+            if str(term).strip():
+                hints.append(str(term).strip())
+        for row in resolved.get("matched_entities") or []:
+            if not isinstance(row, dict):
+                continue
+            name = str(row.get("display_name") or "").strip()
+            if name:
+                hints.append(name[:30])
+
+        symbol_names: list[str] = []
+        try:
+            symbol_service = SymbolIndexService(self.db, tenant_id=self.tenant_id)
+            symbol_hits = await symbol_service.search_symbols(query=question, limit=8)
+            for row in symbol_hits.get("items") or []:
+                if not isinstance(row, dict):
+                    continue
+                name = str(row.get("qualified_name") or row.get("symbol_name") or "").strip()
+                if name:
+                    symbol_names.append(name[:60])
+        except Exception:
+            symbol_names = []
+        for name in symbol_names[:6]:
+            hints.append(name)
+
+        feature_names: list[str] = []
+        feature_routes: list[str] = []
+        try:
+            system_snapshot = await self.context_cache.get_system_snapshot(
+                db=self.db,
+                tenant_id=self.tenant_id,
+            )
+            feature_hits = self._match_feature_catalog_entries(
+                question=question,
+                system_snapshot=system_snapshot,
+                limit=6,
+            )
+            for row in feature_hits:
+                name = str(row.get("name") or "").strip()
+                route = str(row.get("route") or "").strip()
+                if name:
+                    feature_names.append(name[:50])
+                if route:
+                    feature_routes.append(route[:80])
+        except Exception:
+            feature_names = []
+            feature_routes = []
+        for name in feature_names[:4]:
+            hints.append(name)
+        for route in feature_routes[:3]:
+            hints.append(route)
+
+        symbol_top = manifest.get("symbol_summary") if isinstance(manifest.get("symbol_summary"), dict) else {}
+        modules = symbol_top.get("top_modules") if isinstance(symbol_top.get("top_modules"), list) else []
+        module_hint = "；".join(
+            str(item.get("module") or "")
+            for item in modules[:3]
+            if isinstance(item, dict) and str(item.get("module") or "").strip()
+        )
+        clean_hints = []
+        for item in hints:
+            if item and item not in clean_hints:
+                clean_hints.append(item)
+        if not clean_hints and not module_hint:
+            return question[:1200], {"mode": "no_hint"}
+
+        revised = str(question or "").strip()
+        if clean_hints:
+            revised = f"{revised}。实体线索：{', '.join(clean_hints[:6])}"
+        if module_hint:
+            revised = f"{revised}。候选模块：{module_hint}"
+        if symbol_names:
+            revised = f"{revised}。候选符号：{', '.join(symbol_names[:4])}"
+        if feature_names:
+            revised = f"{revised}。候选功能：{', '.join(feature_names[:3])}"
+        if feature_routes:
+            revised = f"{revised}。候选入口：{', '.join(feature_routes[:2])}"
+        return revised[:1200], {
+            "mode": "second_pass_query",
+            "hint_count": len(clean_hints),
+            "module_hint": module_hint,
+            "symbol_hit_count": len(symbol_names),
+            "feature_hit_count": len(feature_names),
+        }
+
+    @staticmethod
+    def _merge_retrieved_items(
+        base: list[RetrievedItem],
+        extra: list[RetrievedItem],
+        *,
+        limit: Optional[int] = None,
+    ) -> list[RetrievedItem]:
+        seen: set[str] = set()
+        merged: list[RetrievedItem] = []
+        for item in [*base, *extra]:
+            key = f"{item.source}|{item.source_label}|{item.content[:120]}"
+            if key in seen:
+                continue
+            seen.add(key)
+            merged.append(item)
+        merged.sort(key=lambda x: x.score, reverse=True)
+        cap = max(DEFAULT_TOP_K * 2, 8)
+        if limit is not None:
+            cap = max(4, min(int(limit), 80))
+        return merged[:cap]
+
+    @staticmethod
+    def _should_second_pass(retrieved: list[RetrievedItem]) -> bool:
+        if not retrieved:
+            return True
+        top_score = float(retrieved[0].score or 0.0)
+        if top_score < 0.55:
+            return True
+        return len(retrieved) < 3
+
+    @staticmethod
+    def _merge_search_metrics(base: dict[str, Any], delta: dict[str, Any]) -> dict[str, Any]:
+        return {
+            "db_query_ms": int(base.get("db_query_ms", 0)) + int(delta.get("db_query_ms", 0)),
+            "vector_search_ms": int(base.get("vector_search_ms", 0)) + int(delta.get("vector_search_ms", 0)),
+            "graph_query_ms": int(base.get("graph_query_ms", 0)) + int(delta.get("graph_query_ms", 0)),
+            "db_timeout": bool(base.get("db_timeout")) or bool(delta.get("db_timeout")),
+            "vector_timeout": bool(base.get("vector_timeout")) or bool(delta.get("vector_timeout")),
+            "graph_timeout": bool(base.get("graph_timeout")) or bool(delta.get("graph_timeout")),
+        }
+
+    @staticmethod
+    def _summarize_retrieval_hints(rows: list[RetrievedItem], *, max_items: int = 3) -> str:
+        hints: list[str] = []
+        for row in rows[: max(1, int(max_items))]:
+            label = str(row.source_label or "").strip()
+            if label:
+                hints.append(label[:60])
+                continue
+            content = str(row.content or "").strip()
+            if content:
+                hints.append(content.replace("\n", " ")[:60])
+        uniq: list[str] = []
+        for item in hints:
+            if item and item not in uniq:
+                uniq.append(item)
+        return "；".join(uniq[:3])
+
+    async def _multi_step_search(
+        self,
+        *,
+        question: str,
+        domains: list[str],
+        project_id: Optional[int],
+        source_chat_id: Optional[str],
+        session_state: Optional[QASessionState],
+        max_steps_override: Optional[int] = None,
+        vector_top_k: Optional[int] = None,
+    ) -> tuple[list[RetrievedItem], dict[str, Any]]:
+        resolved = await self.entity_resolution(question=question, project_id=project_id, limit=6)
+        manifest = await self.get_project_manifest(project_id=project_id)
+        system_snapshot = await self.context_cache.get_system_snapshot(
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        symbol_hits: dict[str, Any] = {"count": 0, "items": []}
+        try:
+            symbol_service = SymbolIndexService(self.db, tenant_id=self.tenant_id)
+            symbol_hits = await symbol_service.search_symbols(query=question, limit=6)
+        except Exception:
+            symbol_hits = {"count": 0, "items": []}
+        feature_hits = self._match_feature_catalog_entries(
+            question=question,
+            system_snapshot=system_snapshot,
+            limit=6,
+        )
+        max_steps = max(1, min(int(max_steps_override or AGENT_MAX_RETRIEVAL_STEPS), 8))
+        tool_trace: list[dict[str, Any]] = [
+            {
+                "step": 0,
+                "tool": "entity_resolution",
+                "matched_entities": len(resolved.get("matched_entities") or []),
+                "canonical_terms": resolved.get("canonical_terms") if isinstance(resolved.get("canonical_terms"), list) else [],
+            },
+            {
+                "step": 0,
+                "tool": "get_project_manifest",
+                "symbol_total": int(
+                    (
+                        manifest.get("symbol_summary")
+                        if isinstance(manifest.get("symbol_summary"), dict)
+                        else {}
+                    ).get("symbol_total", 0)
+                ),
+                "kg_entity_total": int(
+                    (
+                        manifest.get("kg_overview")
+                        if isinstance(manifest.get("kg_overview"), dict)
+                        else {}
+                    ).get("entity_total", 0)
+                ),
+            },
+            {
+                "step": 0,
+                "tool": "search_symbols",
+                "hit_count": int(symbol_hits.get("count") or 0),
+            },
+            {
+                "step": 0,
+                "tool": "feature_catalog_lookup",
+                "hit_count": len(feature_hits),
+                "feature_names": [str(x.get("name") or "") for x in feature_hits[:4]],
+            },
+        ]
+        if session_state is not None and isinstance(session_state.last_entity_ids, list):
+            anchor_ids = [int(x) for x in session_state.last_entity_ids if str(x).isdigit()][:6]
+        else:
+            anchor_ids = []
+        current_question = str(question or "").strip()[:1200]
+        current_domains: list[str] = []
+        for item in domains:
+            if item not in current_domains:
+                current_domains.append(item)
+        merged_rows: list[RetrievedItem] = []
+        merged_metrics: dict[str, Any] = {
+            "db_query_ms": 0,
+            "vector_search_ms": 0,
+            "graph_query_ms": 0,
+            "db_timeout": False,
+            "vector_timeout": False,
+            "graph_timeout": False,
+        }
+        merge_limit = max(DEFAULT_TOP_K * 2, 8)
+        if vector_top_k is not None:
+            merge_limit = max(merge_limit, max(2, int(vector_top_k)) * 3)
+
+        for step in range(1, max_steps + 1):
+            resolved_payload = {
+                **resolved,
+                "anchor_entity_ids": anchor_ids,
+            }
+            step_rows, step_metrics = await self._run_single_retrieval_pass(
+                question=current_question,
+                domains=current_domains,
+                project_id=project_id,
+                source_chat_id=source_chat_id,
+                resolved_entities=resolved_payload,
+                vector_top_k=vector_top_k,
+            )
+            feature_rows = self._retrieve_feature_catalog_context(
+                question=current_question,
+                system_snapshot=system_snapshot,
+                max_items=4,
+            )
+            if feature_rows:
+                step_rows = self._merge_retrieved_items(feature_rows, step_rows, limit=merge_limit)
+            merged_rows = self._merge_retrieved_items(merged_rows, step_rows, limit=merge_limit)
+            merged_metrics = self._merge_search_metrics(merged_metrics, step_metrics)
+            top_score = round(float(step_rows[0].score), 4) if step_rows else 0.0
+            tool_trace.append(
+                {
+                    "step": step,
+                    "tool": "parallel_retrieval",
+                    "query": current_question[:240],
+                    "domains": current_domains,
+                    "items": len(step_rows),
+                    "feature_items": len(feature_rows),
+                    "top_score": top_score,
+                    "merged_items": len(merged_rows),
+                }
+            )
+
+            if step >= max_steps:
+                tool_trace.append(
+                    {
+                        "step": step,
+                        "tool": "planner_stop",
+                        "reason": "max_steps_reached",
+                        "max_steps": max_steps,
+                    }
+                )
+                break
+            if not self._should_second_pass(merged_rows):
+                tool_trace.append(
+                    {
+                        "step": step,
+                        "tool": "planner_stop",
+                        "reason": "enough_evidence",
+                        "top_score": round(float(merged_rows[0].score), 4) if merged_rows else 0.0,
+                    }
+                )
+                break
+
+            next_question, rewrite_meta = await self._build_second_pass_query(
+                question=current_question,
+                project_id=project_id,
+            )
+            row_hint = self._summarize_retrieval_hints(step_rows)
+            if row_hint and row_hint not in next_question:
+                next_question = f"{next_question}。已命中线索：{row_hint}"
+            next_question = next_question[:1200]
+            if next_question.strip() == current_question.strip():
+                tool_trace.append(
+                    {
+                        "step": step,
+                        "tool": "planner_stop",
+                        "reason": "query_not_changed",
+                        "rewrite_meta": rewrite_meta,
+                    }
+                )
+                break
+
+            next_domains: list[str] = []
+            for item in [*current_domains, "graph", "db", "doc", "pmflow", "feishu"]:
+                if item not in next_domains:
+                    next_domains.append(item)
+            tool_trace.append(
+                {
+                    "step": step + 1,
+                    "tool": "query_refine",
+                    "rewrite_meta": rewrite_meta,
+                    "next_query": next_question[:240],
+                    "next_domains": next_domains,
+                    "anchor_entity_ids": anchor_ids,
+                }
+            )
+            current_question = next_question
+            current_domains = next_domains
+
+        result_cap = max(DEFAULT_TOP_K * 2, 8)
+        if vector_top_k is not None:
+            result_cap = max(result_cap, max(2, int(vector_top_k)) * 2)
+
+        return merged_rows[:result_cap], {
+            **merged_metrics,
+            "domains_used": current_domains,
+            "tool_trace": tool_trace,
+        }
+
+    async def multi_step_search(
+        self,
+        *,
+        question: str,
+        project_id: Optional[int] = None,
+        source_chat_id: Optional[str] = None,
+        session_state: Optional[QASessionState] = None,
+    ) -> dict[str, Any]:
+        domains = self._route_domains(question)
+        rows, metrics = await self._multi_step_search(
+            question=question,
+            domains=domains,
+            project_id=project_id,
+            source_chat_id=source_chat_id,
+            session_state=session_state,
+        )
+        return {
+            "question": question,
+            "domains": metrics.get("domains_used") or domains,
+            "items": self._dedupe_sources(rows),
+            "retrieval_metrics": {
+                "db_query_ms": int(metrics.get("db_query_ms") or 0),
+                "vector_search_ms": int(metrics.get("vector_search_ms") or 0),
+                "graph_query_ms": int(metrics.get("graph_query_ms") or 0),
+                "db_timeout": bool(metrics.get("db_timeout")),
+                "vector_timeout": bool(metrics.get("vector_timeout")),
+                "graph_timeout": bool(metrics.get("graph_timeout")),
+            },
+            "tool_trace": metrics.get("tool_trace") if isinstance(metrics.get("tool_trace"), list) else [],
+        }
+
+    # ------------------------------------------------------------------
+    # Retrieval
+    # ------------------------------------------------------------------
+    def _route_domains(self, question: str) -> list[str]:
+        text = question.lower()
+        domains: list[str] = []
+        is_referential = any(k in text for k in FOLLOWUP_REFERENTIAL_KEYWORDS)
+
+        if self._is_people_query(text):
+            domains.append("db")
+        if any(k in text for k in PROJECT_KEYWORDS):
+            domains.append("db")
+            domains.append("graph")
+        if is_referential:
+            domains.append("graph")
+        if any(k in text for k in CODE_KEYWORDS):
+            domains.append("graph")
+            domains.append("doc")
+        if any(k in text for k in FEATURE_QUERY_KEYWORDS):
+            domains.append("doc")
+            domains.append("graph")
+            domains.append("pmflow")
+        if any(k in text for k in DOC_KEYWORDS):
+            domains.append("doc")
+        if any(k in text for k in HISTORY_KEYWORDS):
+            domains.append("feishu")
+        if any(k in text for k in PMFLOW_KEYWORDS):
+            domains.append("pmflow")
+        if self._is_engine_process_query(text):
+            domains.append("graph")
+            domains.append("doc")
+            domains.append("pmflow")
+            domains.append("db")
+        if URL_RE.search(text) or "根据" in text and "http" in text:
+            domains.append("url")
+
+        if not domains:
+            # 模糊问题：并行检索所有向量域
+            domains.extend(["graph", "doc", "feishu", "pmflow", "url"])
+
+        # 去重并保持顺序
+        uniq: list[str] = []
+        for item in domains:
+            if item not in uniq:
+                uniq.append(item)
+        return uniq
+
+    async def _retrieve_structured_db(self, question: str, *, project_id: Optional[int]) -> list[RetrievedItem]:
+        text = question.lower()
+        rows: list[RetrievedItem] = []
+        ask_people = self._is_people_query(text)
+
+        if ask_people:
+            scoped_project_id: Optional[int] = None
+            scoped_project_name: Optional[str] = None
+            if "项目" in text or project_id is not None:
+                scoped_project_id, scoped_project_name, _ = await self._resolve_project_scope_for_question(
+                    question=text,
+                    project_id=project_id,
+                )
+            user_stmt = select(User).where(
+                and_(
+                    User.tenant_id == self.tenant_id,
+                    User.is_active.is_(True),
+                )
+            )
+            if scoped_project_id is not None:
+                member_subq = select(UserProject.user_id).where(
+                    and_(
+                        UserProject.tenant_id == self.tenant_id,
+                        UserProject.project_id == int(scoped_project_id),
+                    )
+                )
+                user_stmt = user_stmt.where(User.id.in_(member_subq))
+            members = (
+                await self.db.execute(
+                    user_stmt.order_by(User.updated_at.desc(), User.id.desc()).limit(20)
+                )
+            ).scalars().all()
+            if members:
+                preview = "；".join(
+                    f"#{int(row.id)} {str(row.name or f'User#{int(row.id)}')}（角色:{str(row.role or 'unknown')}）"
+                    for row in members[:10]
+                )
+                rows.append(
+                    RetrievedItem(
+                        source="db",
+                        content=(
+                            f"成员清单（{len(members)}人）：{preview}"
+                            if scoped_project_name is None
+                            else f"项目「{scoped_project_name}」成员清单（{len(members)}人）：{preview}"
+                        ),
+                        score=0.95,
+                        source_label="User database",
+                        metadata={
+                            "table": "users",
+                            "count": len(members),
+                            "project_id": scoped_project_id,
+                            "project_name": scoped_project_name,
+                        },
+                    )
+                )
+
+        if "项目" in text and not ask_people:
+            stmt = select(Project).where(Project.tenant_id == self.tenant_id)
+            if project_id:
+                stmt = stmt.where(Project.id == project_id)
+            projects = (await self.db.execute(stmt.limit(8))).scalars().all()
+            if projects:
+                summary = "；".join(
+                    f"#{p.id} {p.title or p.name}（状态:{p.status}）" for p in projects
+                )
+                rows.append(
+                    RetrievedItem(
+                        source="db",
+                        content=f"项目列表：{summary}",
+                        score=0.93,
+                        source_label="Project database",
+                        metadata={"table": "projects", "count": len(projects)},
+                    )
+                )
+
+        if "信号" in text or "待处理" in text:
+            signal_stmt = select(Signal).where(Signal.tenant_id == self.tenant_id)
+            if project_id:
+                signal_stmt = signal_stmt.where(
+                    or_(
+                        Signal.matched_project_id == project_id,
+                        Signal.content.ilike(f"%#{project_id}%"),
+                    )
+                )
+            pending = (
+                await self.db.execute(signal_stmt.order_by(Signal.created_at.desc()).limit(10))
+            ).scalars().all()
+            if pending:
+                preview = "；".join((s.summary or s.content or "")[:80] for s in pending[:5])
+                rows.append(
+                    RetrievedItem(
+                        source="db",
+                        content=f"最近信号（{len(pending)}条）：{preview}",
+                        score=0.9,
+                        source_label="Project database",
+                        metadata={"table": "signals", "count": len(pending)},
+                    )
+                )
+
+        if "任务" in text or "进度" in text:
+            task_stmt = select(Task).where(Task.tenant_id == self.tenant_id)
+            if project_id:
+                task_stmt = task_stmt.where(Task.project_id == project_id)
+            tasks = (await self.db.execute(task_stmt.order_by(Task.updated_at.desc()).limit(12))).scalars().all()
+            if tasks:
+                open_tasks = [t for t in tasks if str(t.status or "").lower() not in {"done", "completed", "已完成"}]
+                owners = [str(t.assignee_id) for t in open_tasks if t.assignee_id is not None]
+                rows.append(
+                    RetrievedItem(
+                        source="db",
+                        content=(
+                            f"任务总数 {len(tasks)}，未完成 {len(open_tasks)}。"
+                            f"当前负责人ID（去重）: {', '.join(sorted(set(owners))[:8]) or '暂无'}"
+                        ),
+                        score=0.91,
+                        source_label="Project database",
+                        metadata={"table": "tasks", "open_count": len(open_tasks)},
+                    )
+                )
+
+        if "宪章" in text or "charter" in text:
+            charter_stmt = select(TruthNode).where(TruthNode.tenant_id == self.tenant_id)
+            if project_id:
+                charter_stmt = charter_stmt.where(TruthNode.project_id == project_id)
+            charters = (await self.db.execute(charter_stmt.order_by(TruthNode.updated_at.desc()).limit(6))).scalars().all()
+            if charters:
+                rows.append(
+                    RetrievedItem(
+                        source="db",
+                        content="\n".join(
+                            f"宪章#{c.id}: {(c.title or '无标题')} / {str(c.content or '')[:200]}" for c in charters
+                        ),
+                        score=0.92,
+                        source_label="Project database",
+                        metadata={"table": "truth_nodes", "count": len(charters)},
+                    )
+                )
+
+        if "工作量" in text or "负载" in text:
+            workload_stmt = select(UserWorkload).where(UserWorkload.tenant_id == self.tenant_id)
+            workloads = (await self.db.execute(workload_stmt.order_by(UserWorkload.updated_at.desc()).limit(10))).scalars().all()
+            if workloads:
+                content = "；".join(
+                    f"user#{w.user_id} 已分配 {float(w.current_assigned_hours or 0):.1f}h / "
+                    f"最大 {float(w.max_available_hours_per_week or 0):.1f}h，负载比 {float(w.workload_ratio or 0):.2f}"
+                    for w in workloads
+                )
+                rows.append(
+                    RetrievedItem(
+                        source="db",
+                        content=f"团队工作量：{content}",
+                        score=0.88,
+                        source_label="Project database",
+                        metadata={"table": "user_workloads", "count": len(workloads)},
+                    )
+                )
+
+        return rows[:DEFAULT_TOP_K]
+
+    async def _retrieve_graph_context(
+        self,
+        question: str,
+        *,
+        project_id: Optional[int],
+        resolved_entities: Optional[dict[str, Any]] = None,
+    ) -> list[RetrievedItem]:
+        try:
+            resolution = resolved_entities or {}
+            matched_entities = (
+                resolution.get("matched_entities") if isinstance(resolution.get("matched_entities"), list) else []
+            )
+            anchor_ids_raw = (
+                resolution.get("anchor_entity_ids") if isinstance(resolution.get("anchor_entity_ids"), list) else []
+            )
+            anchor_ids: list[int] = []
+            for value in anchor_ids_raw:
+                try:
+                    anchor_ids.append(int(value))
+                except Exception:
+                    continue
+
+            seed_entities: list[KGEntity] = []
+            if matched_entities:
+                ids: list[int] = []
+                for row in matched_entities:
+                    if not isinstance(row, dict):
+                        continue
+                    try:
+                        ids.append(int(row.get("entity_id")))
+                    except Exception:
+                        continue
+                if ids:
+                    seed_entities = (
+                        await self.db.execute(
+                            select(KGEntity).where(
+                                and_(
+                                    KGEntity.tenant_id == self.tenant_id,
+                                    KGEntity.id.in_(ids),
+                                )
+                            )
+                        )
+                    ).scalars().all()
+
+            if not seed_entities:
+                terms = self._extract_query_terms(question, preferred=self._extract_project_lookup_keyword(question))
+                if terms:
+                    conds = []
+                    for term in terms[:6]:
+                        like = f"%{term}%"
+                        conds.append(KGEntity.display_name.ilike(like))
+                        conds.append(KGEntity.entity_key.ilike(like))
+                    seed_entities = (
+                        await self.db.execute(
+                            select(KGEntity)
+                            .where(
+                                and_(
+                                    KGEntity.tenant_id == self.tenant_id,
+                                    or_(*conds),
+                                )
+                            )
+                            .order_by(KGEntity.last_seen_at.desc(), KGEntity.id.desc())
+                            .limit(15)
+                        )
+                    ).scalars().all()
+            if not seed_entities and anchor_ids:
+                seed_entities = (
+                    await self.db.execute(
+                        select(KGEntity).where(
+                            and_(
+                                KGEntity.tenant_id == self.tenant_id,
+                                KGEntity.id.in_(anchor_ids[:10]),
+                            )
+                        )
+                    )
+                ).scalars().all()
+            if not seed_entities:
+                return []
+
+            seed_ids = [int(x.id) for x in seed_entities]
+            relation_rows = (
+                await self.db.execute(
+                    select(KGRelation)
+                    .where(
+                        and_(
+                            KGRelation.tenant_id == self.tenant_id,
+                            or_(
+                                KGRelation.subject_entity_id.in_(seed_ids),
+                                KGRelation.object_entity_id.in_(seed_ids),
+                            ),
+                        )
+                    )
+                    .order_by(KGRelation.weight.desc(), KGRelation.id.desc())
+                    .limit(120)
+                )
+            ).scalars().all()
+
+            entity_ids: set[int] = set(seed_ids)
+            for rel in relation_rows:
+                entity_ids.add(int(rel.subject_entity_id))
+                entity_ids.add(int(rel.object_entity_id))
+            entities = (
+                await self.db.execute(
+                    select(KGEntity).where(
+                        and_(
+                            KGEntity.tenant_id == self.tenant_id,
+                            KGEntity.id.in_(list(entity_ids)),
+                        )
+                    )
+                )
+            ).scalars().all()
+            entity_map = {int(row.id): row for row in entities}
+
+            relation_ids = [int(row.id) for row in relation_rows]
+            evidence_rows: list[KGEvidence] = []
+            if relation_ids or entity_ids:
+                evidence_rows = (
+                    await self.db.execute(
+                        select(KGEvidence)
+                        .where(
+                            and_(
+                                KGEvidence.tenant_id == self.tenant_id,
+                                or_(
+                                    KGEvidence.relation_id.in_(relation_ids or [-1]),
+                                    KGEvidence.entity_id.in_(list(entity_ids) or [-1]),
+                                ),
+                            )
+                        )
+                        .order_by(KGEvidence.occurred_at.desc(), KGEvidence.id.desc())
+                        .limit(220)
+                    )
+                ).scalars().all()
+            evidence_by_relation: dict[int, str] = {}
+            evidence_by_entity: dict[int, str] = {}
+            for ev in evidence_rows:
+                excerpt = str(ev.content_excerpt or "").strip()
+                if not excerpt:
+                    continue
+                if ev.relation_id is not None and int(ev.relation_id) not in evidence_by_relation:
+                    evidence_by_relation[int(ev.relation_id)] = excerpt[:320]
+                if ev.entity_id is not None and int(ev.entity_id) not in evidence_by_entity:
+                    evidence_by_entity[int(ev.entity_id)] = excerpt[:320]
+
+            out: list[RetrievedItem] = []
+            for rel in relation_rows:
+                subject = entity_map.get(int(rel.subject_entity_id))
+                target = entity_map.get(int(rel.object_entity_id))
+                if subject is None or target is None:
+                    continue
+                subject_payload = self._safe_dict(subject.payload)
+                target_payload = self._safe_dict(target.payload)
+                if project_id is not None:
+                    subject_pid = subject_payload.get("project_id")
+                    target_pid = target_payload.get("project_id")
+                    if subject_pid is not None and int(subject_pid) != int(project_id):
+                        if target_pid is not None and int(target_pid) != int(project_id):
+                            continue
+                evidence_text = evidence_by_relation.get(int(rel.id))
+                if not evidence_text:
+                    evidence_text = evidence_by_entity.get(int(subject.id)) or evidence_by_entity.get(int(target.id)) or ""
+                content = (
+                    f"图路径: {str(subject.display_name or subject.entity_key)} --[{str(rel.relation_type)}]--> "
+                    f"{str(target.display_name or target.entity_key)}\n"
+                    f"证据: {evidence_text or '无显式证据摘录'}"
+                )
+                weight = self._to_float(rel.weight, 1.0)
+                score = min(0.99, max(0.45, 0.62 + min(weight, 1.5) * 0.18))
+                out.append(
+                    RetrievedItem(
+                        source="graph",
+                        content=content[:1200],
+                        score=score,
+                        source_label="Knowledge Graph",
+                        metadata={
+                            "relation_id": int(rel.id),
+                            "relation_type": str(rel.relation_type or ""),
+                            "subject_entity_id": int(subject.id),
+                            "object_entity_id": int(target.id),
+                            "subject_name": str(subject.display_name or subject.entity_key),
+                            "object_name": str(target.display_name or target.entity_key),
+                            "path": f"{subject.id}->{str(rel.relation_type)}->{target.id}",
+                        },
+                    )
+                )
+
+            if not out:
+                for seed in seed_entities[:DEFAULT_TOP_K]:
+                    payload = self._safe_dict(seed.payload)
+                    if project_id is not None and payload.get("project_id") is not None:
+                        if int(payload.get("project_id")) != int(project_id):
+                            continue
+                    excerpt = evidence_by_entity.get(int(seed.id)) or ""
+                    out.append(
+                        RetrievedItem(
+                            source="graph",
+                            content=f"实体: {str(seed.display_name or seed.entity_key)}\n证据: {excerpt or '无显式证据摘录'}",
+                            score=0.58,
+                            source_label="Knowledge Graph",
+                            metadata={
+                                "entity_id": int(seed.id),
+                                "entity_type": str(seed.entity_type or ""),
+                                "entity_key": str(seed.entity_key or ""),
+                            },
+                        )
+                    )
+            out.sort(key=lambda x: x.score, reverse=True)
+            return out[: max(DEFAULT_TOP_K, 6)]
+        except Exception:
+            return []
+
+    async def _retrieve_vector_context(
+        self,
+        question: str,
+        *,
+        domains: list[str],
+        project_id: Optional[int],
+        source_chat_id: Optional[str],
+        top_k: Optional[int] = None,
+    ) -> list[RetrievedItem]:
+        question_embedding = await self.llm.get_embedding_async(
+            question,
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        if not question_embedding:
+            return []
+
+        source_types = self._map_domains_to_source_types(domains)
+        if not source_types:
+            return []
+
+        stmt = select(KnowledgeBaseChunk).where(
+            and_(
+                KnowledgeBaseChunk.tenant_id == self.tenant_id,
+                KnowledgeBaseChunk.source_type.in_(source_types),
+            )
+        )
+
+        chunks = (await self.db.execute(stmt.limit(500))).scalars().all()
+        scored: list[RetrievedItem] = []
+        for chunk in chunks:
+            meta = self._safe_dict(chunk.chunk_metadata).copy()
+            chunk_project_id = int(chunk.project_id) if chunk.project_id is not None else None
+            if chunk_project_id is not None:
+                meta.setdefault("project_id", chunk_project_id)
+
+            if project_id is not None:
+                # 防止跨项目泄漏：Feishu/PMFlow 来源必须严格匹配项目组；
+                # 文档/URL 来源可读取全局（project_id 为空）和当前项目数据。
+                if chunk.source_type in {"feishu", "pmflow"}:
+                    if chunk_project_id != int(project_id):
+                        continue
+                elif chunk_project_id is not None and chunk_project_id != int(project_id):
+                    continue
+
+            if chunk.source_type == "feishu" and source_chat_id:
+                chat_id = str(meta.get("chat_id") or "").strip()
+                if chat_id and chat_id != str(source_chat_id):
+                    continue
+            if chunk.source_type == "pmflow":
+                target_url = str(meta.get("target_url") or "").strip()
+                if not target_url:
+                    target_url = self._build_pmflow_target_url(
+                        record_type=str(meta.get("record_type") or ""),
+                        record_id=meta.get("record_id"),
+                        project_id=meta.get("project_id"),
+                    )
+                    if target_url:
+                        meta["target_url"] = target_url
+            score = self._cosine_similarity(question_embedding, self._embedding_to_list(chunk.embedding))
+            if score <= 0:
+                continue
+            label = self._source_label(chunk.source_type, meta)
+            scored.append(
+                RetrievedItem(
+                    source=str(chunk.source_type or "doc"),
+                    content=str(chunk.content or "")[:1200],
+                    score=score,
+                    source_label=label,
+                    metadata=meta,
+                )
+            )
+        scored.sort(key=lambda x: x.score, reverse=True)
+        limit = max(1, min(int(top_k or DEFAULT_TOP_K), 40))
+        return scored[:limit]
+
+    # ------------------------------------------------------------------
+    # Knowledge indexing
+    # ------------------------------------------------------------------
+    async def create_knowledge_doc(
+        self,
+        *,
+        title: str,
+        file_type: str,
+        raw_content: str,
+        uploaded_by: Optional[int],
+        source_type: str = "file",
+        file_url: Optional[str] = None,
+        project_id: Optional[int] = None,
+    ) -> dict[str, Any]:
+        doc = await self.create_knowledge_doc_placeholder(
+            title=title,
+            file_type=file_type,
+            uploaded_by=uploaded_by,
+            source_type=source_type,
+            file_url=file_url,
+        )
+        await self._index_doc_content(
+            doc=doc,
+            raw_content=raw_content,
+            project_id=project_id,
+            source_type=source_type,
+        )
+        return self._serialize_doc(doc)
+
+    async def create_knowledge_doc_placeholder(
+        self,
+        *,
+        title: str,
+        file_type: str,
+        uploaded_by: Optional[int],
+        source_type: str = "file",
+        file_url: Optional[str] = None,
+    ) -> KnowledgeBaseDoc:
+        doc = KnowledgeBaseDoc(
+            tenant_id=self.tenant_id,
+            title=(title or "未命名文档")[:255],
+            file_type=(file_type or "txt")[:32],
+            source_type=(source_type or "file")[:32],
+            file_url=(file_url or "")[:1024] or None,
+            uploaded_by=uploaded_by,
+            status="processing",
+            last_fetched_at=datetime.now(timezone.utc) if source_type == "url" else None,
+        )
+        self.db.add(doc)
+        await self.db.commit()
+        await self.db.refresh(doc)
+        return doc
+
+    async def process_knowledge_doc_content(
+        self,
+        *,
+        doc_id: int,
+        raw_content: str,
+        project_id: Optional[int] = None,
+    ) -> dict[str, Any]:
+        doc = (
+            await self.db.execute(
+                select(KnowledgeBaseDoc).where(
+                    and_(
+                        KnowledgeBaseDoc.id == int(doc_id),
+                        KnowledgeBaseDoc.tenant_id == self.tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not doc:
+            raise ValueError("doc_not_found")
+        await self._index_doc_content(
+            doc=doc,
+            raw_content=raw_content,
+            project_id=project_id,
+            source_type=str(doc.source_type or "file"),
+        )
+        return self._serialize_doc(doc)
+
+    async def _index_doc_content(
+        self,
+        *,
+        doc: KnowledgeBaseDoc,
+        raw_content: str,
+        project_id: Optional[int],
+        source_type: str,
+    ) -> None:
+        clean_content = str(raw_content or "").strip()
+        if not clean_content:
+            doc.status = "failed"
+            doc.error_message = "文档内容为空，无法向量化"
+            await self.db.commit()
+            return
+
+        try:
+            await self.db.execute(
+                delete(KnowledgeBaseChunk).where(
+                    and_(
+                        KnowledgeBaseChunk.tenant_id == self.tenant_id,
+                        KnowledgeBaseChunk.doc_id == int(doc.id),
+                    )
+                )
+            )
+
+            chunk_items = self._chunk_text_with_metadata(clean_content, chunk_size=500, overlap=50)
+            for idx, (part, extra_meta) in enumerate(chunk_items):
+                emb = await self.llm.get_embedding_async(part, db=self.db, tenant_id=self.tenant_id)
+                chunk_metadata = {
+                    "title": doc.title,
+                    "file_type": doc.file_type,
+                    "source_type": source_type,
+                    "file_url": doc.file_url,
+                }
+                chunk_metadata.update(extra_meta)
+                self.db.add(
+                    KnowledgeBaseChunk(
+                        tenant_id=self.tenant_id,
+                        doc_id=int(doc.id),
+                        source_type="url" if source_type == "url" else "doc",
+                        project_id=project_id,
+                        content=part,
+                        embedding=emb if emb else None,
+                        chunk_index=idx,
+                        chunk_metadata=chunk_metadata,
+                    )
+                )
+            doc.status = "complete"
+            doc.error_message = None
+            if source_type == "url":
+                doc.last_fetched_at = datetime.now(timezone.utc)
+            await self.db.commit()
+        except Exception as exc:
+            doc.status = "failed"
+            doc.error_message = str(exc)[:500]
+            await self.db.commit()
+            raise
+
+    async def list_knowledge_docs(self) -> list[dict[str, Any]]:
+        docs = (
+            await self.db.execute(
+                select(KnowledgeBaseDoc)
+                .where(KnowledgeBaseDoc.tenant_id == self.tenant_id)
+                .order_by(KnowledgeBaseDoc.created_at.desc())
+                .limit(200)
+            )
+        ).scalars().all()
+        return [
+            self._serialize_doc(d)
+            for d in docs
+        ]
+
+    async def get_knowledge_doc(self, *, doc_id: int) -> dict[str, Any]:
+        doc = (
+            await self.db.execute(
+                select(KnowledgeBaseDoc).where(
+                    and_(
+                        KnowledgeBaseDoc.id == int(doc_id),
+                        KnowledgeBaseDoc.tenant_id == self.tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not doc:
+            raise ValueError("doc_not_found")
+        return self._serialize_doc(doc)
+
+    async def delete_knowledge_doc(self, *, doc_id: int) -> bool:
+        doc = (
+            await self.db.execute(
+                select(KnowledgeBaseDoc).where(
+                    and_(
+                        KnowledgeBaseDoc.id == int(doc_id),
+                        KnowledgeBaseDoc.tenant_id == self.tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not doc:
+            return False
+
+        await self.db.execute(
+            delete(KnowledgeBaseChunk).where(
+                and_(
+                    KnowledgeBaseChunk.doc_id == int(doc_id),
+                    KnowledgeBaseChunk.tenant_id == self.tenant_id,
+                )
+            )
+        )
+        await self.db.delete(doc)
+        await self.db.commit()
+        return True
+
+    async def ingest_feishu_message(
+        self,
+        *,
+        content: str,
+        sender: str,
+        chat_id: str,
+        message_id: Optional[str],
+        captured_at: Optional[datetime] = None,
+        project_id: Optional[int] = None,
+    ) -> None:
+        text = str(content or "").strip()
+        if not text:
+            return
+        chunks = self._chunk_text(text, chunk_size=320, overlap=40)
+        for idx, part in enumerate(chunks):
+            emb = await self.llm.get_embedding_async(part, db=self.db, tenant_id=self.tenant_id)
+            self.db.add(
+                KnowledgeBaseChunk(
+                    tenant_id=self.tenant_id,
+                    doc_id=None,
+                    source_type="feishu",
+                    project_id=project_id,
+                    content=part,
+                    embedding=emb if emb else None,
+                    chunk_index=idx,
+                    chunk_metadata={
+                        "sender": sender,
+                        "chat_id": chat_id,
+                        "message_id": message_id,
+                        "captured_at": (captured_at or datetime.now(timezone.utc)).isoformat(),
+                    },
+                )
+            )
+        await self.db.commit()
+
+    async def mark_doc_processing(self, *, doc_id: int) -> dict[str, Any]:
+        doc = (
+            await self.db.execute(
+                select(KnowledgeBaseDoc).where(
+                    and_(
+                        KnowledgeBaseDoc.id == int(doc_id),
+                        KnowledgeBaseDoc.tenant_id == self.tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not doc:
+            raise ValueError("doc_not_found")
+        doc.status = "processing"
+        doc.error_message = None
+        await self.db.commit()
+        await self.db.refresh(doc)
+        return self._serialize_doc(doc)
+
+    @staticmethod
+    async def ingest_feishu_message_async(
+        *,
+        tenant_id: int,
+        content: str,
+        sender: str,
+        chat_id: str,
+        message_id: Optional[str],
+        captured_at: Optional[datetime] = None,
+        project_id: Optional[int] = None,
+    ) -> None:
+        async with AsyncSessionLocal() as session:
+            service = AdaptiveRAGService(session, tenant_id=tenant_id)
+            try:
+                await service.ingest_feishu_message(
+                    content=content,
+                    sender=sender,
+                    chat_id=chat_id,
+                    message_id=message_id,
+                    captured_at=captured_at,
+                    project_id=project_id,
+                )
+            except Exception:
+                await session.rollback()
+
+    @staticmethod
+    async def process_knowledge_doc_content_async(
+        *,
+        tenant_id: int,
+        doc_id: int,
+        raw_content: str,
+        project_id: Optional[int] = None,
+    ) -> None:
+        async with AsyncSessionLocal() as session:
+            service = AdaptiveRAGService(session, tenant_id=tenant_id)
+            try:
+                await service.process_knowledge_doc_content(
+                    doc_id=doc_id,
+                    raw_content=raw_content,
+                    project_id=project_id,
+                )
+            except Exception:
+                await session.rollback()
+
+    async def refresh_url_doc(self, *, doc_id: int, project_id: Optional[int] = None) -> dict[str, Any]:
+        doc = (
+            await self.db.execute(
+                select(KnowledgeBaseDoc).where(
+                    and_(
+                        KnowledgeBaseDoc.id == int(doc_id),
+                        KnowledgeBaseDoc.tenant_id == self.tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not doc:
+            raise ValueError("doc_not_found")
+        if str(doc.source_type or "").lower() != "url":
+            raise ValueError("doc_is_not_url_source")
+        if not str(doc.file_url or "").strip():
+            raise ValueError("doc_url_missing")
+        text = await self._fetch_url_text(str(doc.file_url))
+        await self._index_doc_content(
+            doc=doc,
+            raw_content=text,
+            project_id=project_id,
+            source_type="url",
+        )
+        await self.db.refresh(doc)
+        return self._serialize_doc(doc)
+
+    @staticmethod
+    async def refresh_url_doc_async(
+        *,
+        tenant_id: int,
+        doc_id: int,
+        project_id: Optional[int] = None,
+    ) -> None:
+        async with AsyncSessionLocal() as session:
+            service = AdaptiveRAGService(session, tenant_id=tenant_id)
+            try:
+                await service.refresh_url_doc(doc_id=doc_id, project_id=project_id)
+            except Exception:
+                await session.rollback()
+
+    async def upsert_pmflow_record(
+        self,
+        *,
+        record_type: str,
+        record_id: int,
+        project_id: Optional[int],
+        title: str,
+        content: str,
+    ) -> None:
+        safe_type = str(record_type or "").strip().lower()[:32]
+        safe_id = int(record_id)
+        if not safe_type or safe_id <= 0:
+            return
+
+        stmt = select(KnowledgeBaseChunk).where(
+            and_(
+                KnowledgeBaseChunk.tenant_id == self.tenant_id,
+                KnowledgeBaseChunk.source_type == "pmflow",
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(KnowledgeBaseChunk.project_id == int(project_id))
+        existing = (await self.db.execute(stmt.limit(2000))).scalars().all()
+        for row in existing:
+            meta = self._safe_dict(row.chunk_metadata)
+            if str(meta.get("record_type") or "").strip().lower() != safe_type:
+                continue
+            if str(meta.get("record_id") or "").strip() != str(safe_id):
+                continue
+            await self.db.delete(row)
+
+        text = f"{title}\n\n{content}".strip()
+        chunks = self._chunk_text(text, chunk_size=500, overlap=50)
+        for idx, part in enumerate(chunks):
+            emb = await self.llm.get_embedding_async(part, db=self.db, tenant_id=self.tenant_id)
+            target_url = self._build_pmflow_target_url(
+                record_type=safe_type,
+                record_id=safe_id,
+                project_id=project_id,
+            )
+            self.db.add(
+                KnowledgeBaseChunk(
+                    tenant_id=self.tenant_id,
+                    doc_id=None,
+                    source_type="pmflow",
+                    project_id=project_id,
+                    content=part,
+                    embedding=emb if emb else None,
+                    chunk_index=idx,
+                    chunk_metadata={
+                        "record_type": safe_type,
+                        "record_id": safe_id,
+                        "project_id": int(project_id) if project_id is not None else None,
+                        "target_url": target_url,
+                        "title": title[:255],
+                        "updated_at": datetime.now(timezone.utc).isoformat(),
+                    },
+                )
+            )
+        await self.db.commit()
+
+    @staticmethod
+    async def upsert_pmflow_record_async(
+        *,
+        tenant_id: int,
+        record_type: str,
+        record_id: int,
+        project_id: Optional[int],
+        title: str,
+        content: str,
+    ) -> None:
+        async with AsyncSessionLocal() as session:
+            service = AdaptiveRAGService(session, tenant_id=tenant_id)
+            try:
+                await service.upsert_pmflow_record(
+                    record_type=record_type,
+                    record_id=record_id,
+                    project_id=project_id,
+                    title=title,
+                    content=content,
+                )
+            except Exception:
+                await session.rollback()
+
+    @staticmethod
+    def _build_intent_correction_record(
+        *,
+        user_utterance: str,
+        wrong_interpretation: str,
+        correct_intent: str,
+        canonical_action: str,
+        clarification_template: str,
+        risk_level: str,
+        project_id: Optional[int],
+    ) -> tuple[int, str, str, dict[str, Any]]:
+        utterance = " ".join(str(user_utterance or "").split()).strip()[:500]
+        wrong = " ".join(str(wrong_interpretation or "").split()).strip()[:300]
+        correct = " ".join(str(correct_intent or "").split()).strip().lower()[:60]
+        action = " ".join(str(canonical_action or "").split()).strip().lower()[:80]
+        clarification = " ".join(str(clarification_template or "").split()).strip()[:500]
+        risk = " ".join(str(risk_level or "medium").split()).strip().lower()[:20] or "medium"
+
+        fingerprint = hashlib.sha1(
+            "||".join([utterance, wrong, correct, action, clarification, risk]).encode("utf-8")
+        ).hexdigest()
+        record_id = int(fingerprint[:12], 16)
+        title = f"意图纠错/{correct or 'unknown'}"
+        content = (
+            f"用户原话: {utterance}\n"
+            f"错误理解: {wrong}\n"
+            f"正确意图: {correct}\n"
+            f"标准动作: {action}\n"
+            f"澄清模板: {clarification}\n"
+            f"风险等级: {risk}"
+        )[:4000]
+        metadata = {
+            "record_type": "intent_correction",
+            "record_id": str(record_id),
+            "project_id": int(project_id) if project_id is not None else None,
+            "user_utterance": utterance,
+            "wrong_interpretation": wrong,
+            "correct_intent": correct,
+            "canonical_action": action,
+            "clarification_template": clarification,
+            "risk_level": risk,
+            "target_url": None,
+            "title": title,
+            "updated_at": datetime.now(timezone.utc).isoformat(),
+        }
+        return record_id, title, content, metadata
+
+    async def upsert_intent_correction_sample(
+        self,
+        *,
+        user_utterance: str,
+        wrong_interpretation: str,
+        correct_intent: str,
+        canonical_action: str,
+        clarification_template: str,
+        risk_level: str,
+        project_id: Optional[int] = None,
+    ) -> dict[str, Any]:
+        record_id, title, content, metadata = self._build_intent_correction_record(
+            user_utterance=user_utterance,
+            wrong_interpretation=wrong_interpretation,
+            correct_intent=correct_intent,
+            canonical_action=canonical_action,
+            clarification_template=clarification_template,
+            risk_level=risk_level,
+            project_id=project_id,
+        )
+        if not metadata["user_utterance"] or not metadata["correct_intent"]:
+            raise ValueError("intent correction sample missing required fields")
+
+        stmt = select(KnowledgeBaseChunk).where(
+            and_(
+                KnowledgeBaseChunk.tenant_id == self.tenant_id,
+                KnowledgeBaseChunk.source_type == "pmflow",
+            )
+        )
+        existing = (await self.db.execute(stmt.limit(2000))).scalars().all()
+        for row in existing:
+            meta = self._safe_dict(row.chunk_metadata)
+            if str(meta.get("record_type") or "").strip().lower() != "intent_correction":
+                continue
+            if str(meta.get("record_id") or "").strip() != str(record_id):
+                continue
+            await self.db.delete(row)
+
+        embedding = await self.llm.get_embedding_async(content, db=self.db, tenant_id=self.tenant_id)
+        self.db.add(
+            KnowledgeBaseChunk(
+                tenant_id=self.tenant_id,
+                doc_id=None,
+                source_type="pmflow",
+                project_id=int(project_id) if project_id is not None else None,
+                content=content,
+                embedding=embedding if embedding else None,
+                chunk_index=0,
+                chunk_metadata=metadata,
+            )
+        )
+        await self.db.commit()
+        return metadata
+
+    async def retrieve_intent_correction_samples(
+        self,
+        *,
+        utterance: str,
+        top_k: int = 3,
+        project_id: Optional[int] = None,
+    ) -> list[dict[str, Any]]:
+        text = " ".join(str(utterance or "").split()).strip()
+        if len(text) < 2:
+            return []
+        question_embedding = await self.llm.get_embedding_async(
+            text,
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        if not question_embedding:
+            return []
+
+        rows = (
+            await self.db.execute(
+                select(KnowledgeBaseChunk).where(
+                    and_(
+                        KnowledgeBaseChunk.tenant_id == self.tenant_id,
+                        KnowledgeBaseChunk.source_type == "pmflow",
+                    )
+                ).limit(500)
+            )
+        ).scalars().all()
+        scored: list[dict[str, Any]] = []
+        for row in rows:
+            meta = self._safe_dict(row.chunk_metadata)
+            if str(meta.get("record_type") or "").strip().lower() != "intent_correction":
+                continue
+            try:
+                chunk_project_id = int(meta.get("project_id")) if meta.get("project_id") is not None else None
+            except Exception:
+                chunk_project_id = None
+            if project_id is not None and chunk_project_id is not None and chunk_project_id != int(project_id):
+                continue
+            score = self._cosine_similarity(question_embedding, self._embedding_to_list(row.embedding))
+            if score <= 0:
+                continue
+            scored.append(
+                {
+                    "score": round(float(score), 4),
+                    "record_id": str(meta.get("record_id") or ""),
+                    "user_utterance": str(meta.get("user_utterance") or ""),
+                    "wrong_interpretation": str(meta.get("wrong_interpretation") or ""),
+                    "correct_intent": str(meta.get("correct_intent") or ""),
+                    "canonical_action": str(meta.get("canonical_action") or ""),
+                    "clarification_template": str(meta.get("clarification_template") or ""),
+                    "risk_level": str(meta.get("risk_level") or "medium"),
+                    "project_id": chunk_project_id,
+                }
+            )
+        scored.sort(key=lambda x: float(x.get("score") or 0.0), reverse=True)
+        return scored[: max(1, min(int(top_k or 3), 10))]
+
+    # ------------------------------------------------------------------
+    # QA session helpers
+    # ------------------------------------------------------------------
+    @staticmethod
+    def _to_naive_utc(value: Optional[datetime]) -> Optional[datetime]:
+        if value is None:
+            return None
+        if value.tzinfo is None:
+            return value
+        return value.astimezone(timezone.utc).replace(tzinfo=None)
+
+    @staticmethod
+    def _normalize_session_title_raw(text: str, *, max_len: int = 30) -> str:
+        compact = " ".join(str(text or "").replace("\n", " ").split()).strip()
+        if not compact:
+            return ""
+        return compact[:max_len]
+
+    @classmethod
+    def _build_auto_session_title(cls, question: str) -> Optional[str]:
+        normalized = cls._normalize_session_title_raw(question, max_len=30)
+        if not normalized:
+            return None
+        compact = re.sub(r"\s+", "", normalized)
+        if len(compact) < SESSION_TITLE_MIN_LEN:
+            return None
+        return normalized
+
+    @staticmethod
+    def _is_default_session_title(title: Optional[str]) -> bool:
+        clean = str(title or "").strip()
+        return (not clean) or clean == SESSION_DEFAULT_TITLE
+
+    async def _touch_session_for_new_record(
+        self,
+        *,
+        asked_by: Optional[int],
+        asked_via: str,
+        project_id: Optional[int],
+        question: str,
+        preferred_session_id: Optional[int] = None,
+        force_new_session: bool = False,
+    ) -> tuple[Optional[int], int]:
+        if asked_by is None:
+            return (None, 1)
+
+        now_utc = datetime.now(timezone.utc).replace(tzinfo=None)
+        chosen: Optional[QASession] = None
+        preferred_id = int(preferred_session_id) if preferred_session_id else 0
+        if preferred_id > 0:
+            chosen = (
+                await self.db.execute(
+                    select(QASession).where(
+                        and_(
+                            QASession.tenant_id == self.tenant_id,
+                            QASession.user_id == int(asked_by),
+                            QASession.id == preferred_id,
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+
+        # 无显式 session_id 时，仅在非强制新会话模式下复用同渠道（asked_via）且时间窗口内的最近会话。
+        if chosen is None and not force_new_session:
+            latest_record = (
+                await self.db.execute(
+                    select(QARecord.session_id, QARecord.created_at)
+                    .where(
+                        and_(
+                            QARecord.tenant_id == self.tenant_id,
+                            QARecord.asked_by == int(asked_by),
+                            QARecord.asked_via == (asked_via or "pmflow")[:20],
+                            QARecord.session_id.is_not(None),
+                        )
+                    )
+                    .order_by(QARecord.created_at.desc(), QARecord.id.desc())
+                    .limit(1)
+                )
+            ).first()
+            if latest_record and latest_record.session_id is not None:
+                last_at = self._to_naive_utc(latest_record.created_at)
+                if last_at is not None:
+                    gap_seconds = abs((now_utc - last_at).total_seconds())
+                    is_same_day = now_utc.date() == last_at.date()
+                    if is_same_day and gap_seconds <= SESSION_GAP_MINUTES * 60:
+                        chosen = (
+                            await self.db.execute(
+                                select(QASession).where(
+                                    and_(
+                                        QASession.tenant_id == self.tenant_id,
+                                        QASession.user_id == int(asked_by),
+                                        QASession.id == int(latest_record.session_id),
+                                    )
+                                )
+                            )
+                        ).scalar_one_or_none()
+
+        if chosen is None:
+            title = self._build_auto_session_title(question) or SESSION_DEFAULT_TITLE
+            chosen = QASession(
+                tenant_id=self.tenant_id,
+                user_id=int(asked_by),
+                title=title[:120],
+                project_id=project_id,
+                started_at=now_utc,
+                last_message_at=now_utc,
+                message_count=0,
+            )
+            self.db.add(chosen)
+            await self.db.flush()
+        else:
+            if chosen.project_id is None and project_id is not None:
+                chosen.project_id = int(project_id)
+            chosen.last_message_at = now_utc
+
+        chosen.message_count = int(chosen.message_count or 0) + 1
+        if self._is_default_session_title(chosen.title):
+            generated = self._build_auto_session_title(question)
+            if generated:
+                chosen.title = generated[:120]
+        await self.db.flush()
+        return (int(chosen.id), int(chosen.message_count))
+
+    # ------------------------------------------------------------------
+    # QA feedback/list/export
+    # ------------------------------------------------------------------
+    async def _latest_feedback_map(self, record_ids: list[int]) -> dict[int, QAFeedback]:
+        ids = [int(x) for x in record_ids if x is not None]
+        if not ids:
+            return {}
+        feedback_rows = (
+            await self.db.execute(
+                select(QAFeedback).where(
+                    and_(
+                        QAFeedback.tenant_id == self.tenant_id,
+                        QAFeedback.qa_id.in_(ids),
+                    )
+                )
+            )
+        ).scalars().all()
+
+        latest_feedback: dict[int, QAFeedback] = {}
+        for fb in feedback_rows:
+            key = int(fb.qa_id)
+            curr = latest_feedback.get(key)
+            fb_at = self._to_naive_utc(fb.created_at) or datetime.min
+            curr_at = self._to_naive_utc(curr.created_at) if curr else None
+            if curr is None or curr_at is None or fb_at >= curr_at:
+                latest_feedback[key] = fb
+        return latest_feedback
+
+    async def list_qa_sessions(
+        self,
+        *,
+        user_id: int,
+        date_from: Optional[datetime],
+        date_to: Optional[datetime],
+        page: int,
+        page_size: int,
+    ) -> dict[str, Any]:
+        page = max(1, int(page))
+        page_size = max(1, min(100, int(page_size)))
+        date_from_naive = self._to_naive_utc(date_from)
+        date_to_naive = self._to_naive_utc(date_to)
+
+        base_stmt = select(QASession).where(
+            and_(
+                QASession.tenant_id == self.tenant_id,
+                QASession.user_id == int(user_id),
+            )
+        )
+        if date_from_naive is not None:
+            base_stmt = base_stmt.where(QASession.last_message_at >= date_from_naive)
+        if date_to_naive is not None:
+            base_stmt = base_stmt.where(QASession.last_message_at <= date_to_naive)
+
+        total = int((await self.db.execute(select(func.count()).select_from(base_stmt.subquery()))).scalar() or 0)
+        rows = (
+            await self.db.execute(
+                base_stmt.order_by(QASession.last_message_at.desc(), QASession.id.desc())
+                .offset((page - 1) * page_size)
+                .limit(page_size)
+            )
+        ).scalars().all()
+        if not rows:
+            return {
+                "items": [],
+                "page": page,
+                "page_size": page_size,
+                "total": total,
+                "total_pages": 0,
+                "count": 0,
+            }
+
+        session_ids = [int(x.id) for x in rows]
+        record_rows = (
+            await self.db.execute(
+                select(QARecord).where(
+                    and_(
+                        QARecord.tenant_id == self.tenant_id,
+                        QARecord.session_id.in_(session_ids),
+                    )
+                )
+            )
+        ).scalars().all()
+
+        records_by_session: dict[int, list[QARecord]] = {}
+        for row in record_rows:
+            sid = int(row.session_id or 0)
+            if sid <= 0:
+                continue
+            records_by_session.setdefault(sid, []).append(row)
+
+        feedback_map = await self._latest_feedback_map([int(x.id) for x in record_rows])
+        items: list[dict[str, Any]] = []
+        for row in rows:
+            sid = int(row.id)
+            session_records = sorted(
+                records_by_session.get(sid, []),
+                key=lambda r: (int(r.turn_index or 0), self._to_naive_utc(r.created_at) or datetime.min),
+            )
+            positive = 0
+            negative = 0
+            corrected = 0
+            last_asked_via = "pmflow"
+            for rec in session_records:
+                last_asked_via = str(rec.asked_via or last_asked_via)
+                fb = feedback_map.get(int(rec.id))
+                if not fb:
+                    continue
+                rating = str(fb.rating or "").strip().lower()
+                if rating == "positive":
+                    positive += 1
+                elif rating == "negative":
+                    negative += 1
+                if str(fb.correct_answer or "").strip():
+                    corrected += 1
+
+            message_count = int(row.message_count or len(session_records))
+            unrated = max(message_count - positive - negative, 0)
+            first_question = session_records[0].question if session_records else ""
+            items.append(
+                {
+                    "id": sid,
+                    "title": str(row.title or SESSION_DEFAULT_TITLE),
+                    "project_id": row.project_id,
+                    "started_at": row.started_at.isoformat() if row.started_at else None,
+                    "last_message_at": row.last_message_at.isoformat() if row.last_message_at else None,
+                    "message_count": message_count,
+                    "turn_count": message_count,
+                    "rating_summary": {
+                        "positive": positive,
+                        "negative": negative,
+                        "unrated": unrated,
+                        "corrected": corrected,
+                    },
+                    "last_asked_via": last_asked_via,
+                    "preview_question": str(first_question or "")[:120],
+                }
+            )
+
+        total_pages = math.ceil(total / page_size) if total > 0 else 0
+        return {
+            "items": items,
+            "page": page,
+            "page_size": page_size,
+            "total": total,
+            "total_pages": total_pages,
+            "count": len(items),
+        }
+
+    async def get_qa_session_detail(self, *, session_id: int, user_id: int) -> dict[str, Any]:
+        session_row = (
+            await self.db.execute(
+                select(QASession).where(
+                    and_(
+                        QASession.id == int(session_id),
+                        QASession.tenant_id == self.tenant_id,
+                        QASession.user_id == int(user_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not session_row:
+            raise ValueError("qa session 不存在")
+
+        records = (
+            await self.db.execute(
+                select(QARecord).where(
+                    and_(
+                        QARecord.tenant_id == self.tenant_id,
+                        QARecord.session_id == int(session_id),
+                    )
+                ).order_by(QARecord.turn_index.asc(), QARecord.created_at.asc())
+            )
+        ).scalars().all()
+        feedback_map = await self._latest_feedback_map([int(x.id) for x in records])
+
+        messages: list[dict[str, Any]] = []
+        positive = 0
+        negative = 0
+        corrected = 0
+        for row in records:
+            fb = feedback_map.get(int(row.id))
+            ack_meta = self._extract_ack_metadata_from_sources(row.sources, answer=row.answer)
+            if fb:
+                rating = str(fb.rating or "").strip().lower()
+                if rating == "positive":
+                    positive += 1
+                elif rating == "negative":
+                    negative += 1
+                if str(fb.correct_answer or "").strip():
+                    corrected += 1
+            messages.append(
+                {
+                    "id": int(row.id),
+                    "question": row.question,
+                    "answer": row.answer,
+                    "sources": row.sources or [],
+                    "intent": row.intent,
+                    "domains_used": row.domains_used or [],
+                    "response_time_ms": row.response_time_ms,
+                    "asked_by": row.asked_by,
+                    "asked_via": row.asked_via,
+                    "project_id": row.project_id,
+                    "session_id": row.session_id,
+                    "turn_index": int(row.turn_index or 1),
+                    "created_at": row.created_at.isoformat() if row.created_at else None,
+                    "feedback": {
+                        "id": int(fb.id),
+                        "rating": fb.rating,
+                        "helpfulness": fb.helpfulness,
+                        "error_types": fb.error_types or [],
+                        "correct_answer": fb.correct_answer,
+                    }
+                    if fb
+                    else None,
+                    **ack_meta,
+                }
+            )
+
+        message_count = int(session_row.message_count or len(records))
+        unrated = max(message_count - positive - negative, 0)
+        return {
+            "session": {
+                "id": int(session_row.id),
+                "title": str(session_row.title or SESSION_DEFAULT_TITLE),
+                "project_id": session_row.project_id,
+                "started_at": session_row.started_at.isoformat() if session_row.started_at else None,
+                "last_message_at": session_row.last_message_at.isoformat() if session_row.last_message_at else None,
+                "message_count": message_count,
+                "turn_count": message_count,
+                "rating_summary": {
+                    "positive": positive,
+                    "negative": negative,
+                    "unrated": unrated,
+                    "corrected": corrected,
+                },
+            },
+            "messages": messages,
+        }
+
+    async def rename_qa_session(self, *, session_id: int, user_id: int, title: str) -> dict[str, Any]:
+        session_row = (
+            await self.db.execute(
+                select(QASession).where(
+                    and_(
+                        QASession.id == int(session_id),
+                        QASession.tenant_id == self.tenant_id,
+                        QASession.user_id == int(user_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not session_row:
+            raise ValueError("qa session 不存在")
+
+        cleaned = self._normalize_session_title_raw(title, max_len=120)
+        if not cleaned:
+            raise ValueError("会话标题不能为空")
+        session_row.title = cleaned
+        await self.db.commit()
+        return {"id": int(session_row.id), "title": cleaned}
+
+    async def delete_qa_session(self, *, session_id: int, user_id: int) -> bool:
+        session_row = (
+            await self.db.execute(
+                select(QASession).where(
+                    and_(
+                        QASession.id == int(session_id),
+                        QASession.tenant_id == self.tenant_id,
+                        QASession.user_id == int(user_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not session_row:
+            return False
+
+        qa_ids = (
+            await self.db.execute(
+                select(QARecord.id).where(
+                    and_(
+                        QARecord.tenant_id == self.tenant_id,
+                        QARecord.session_id == int(session_id),
+                    )
+                )
+            )
+        ).scalars().all()
+        id_list = [int(x) for x in qa_ids]
+        if id_list:
+            await self.db.execute(
+                delete(QAFeedback).where(
+                    and_(
+                        QAFeedback.tenant_id == self.tenant_id,
+                        QAFeedback.qa_id.in_(id_list),
+                    )
+                )
+            )
+            await self.db.execute(
+                delete(QARecord).where(
+                    and_(
+                        QARecord.tenant_id == self.tenant_id,
+                        QARecord.session_id == int(session_id),
+                    )
+                )
+            )
+        await self.db.delete(session_row)
+        await self.db.commit()
+        return True
+
+    async def submit_feedback(
+        self,
+        *,
+        qa_id: int,
+        rating: str,
+        submitted_by: Optional[int],
+        helpfulness: Optional[str] = None,
+        error_types: Optional[list[str]] = None,
+        correct_answer: Optional[str] = None,
+    ) -> dict[str, Any]:
+        qa = (
+            await self.db.execute(
+                select(QARecord).where(
+                    and_(QARecord.id == int(qa_id), QARecord.tenant_id == self.tenant_id)
+                )
+            )
+        ).scalar_one_or_none()
+        if not qa:
+            raise ValueError("qa record 不存在")
+
+        feedback = QAFeedback(
+            tenant_id=self.tenant_id,
+            qa_id=int(qa_id),
+            rating=(rating or "").strip().lower()[:16],
+            helpfulness=(helpfulness or "").strip()[:32] or None,
+            error_types=[str(x).strip() for x in (error_types or []) if str(x).strip()] or None,
+            correct_answer=(correct_answer or "").strip()[:4000] or None,
+            submitted_by=submitted_by,
+        )
+        self.db.add(feedback)
+        await self.db.commit()
+        await self.db.refresh(feedback)
+        if str(feedback.rating or "").strip().lower() == "negative" and (
+            str(feedback.correct_answer or "").strip() or [str(x).strip() for x in (feedback.error_types or []) if str(x).strip()]
+        ):
+            try:
+                await self.upsert_intent_correction_sample(
+                    user_utterance=str(qa.question or ""),
+                    wrong_interpretation=str(qa.answer or ""),
+                    correct_intent="query",
+                    canonical_action="clarify_before_answer",
+                    clarification_template=str(feedback.correct_answer or "").strip() or "收到纠正后先澄清范围再回答。",
+                    risk_level="medium",
+                    project_id=int(qa.project_id) if qa.project_id is not None else None,
+                )
+            except Exception:
+                await self.db.rollback()
+        return {
+            "id": int(feedback.id),
+            "qa_id": int(feedback.qa_id),
+            "rating": feedback.rating,
+            "helpfulness": feedback.helpfulness,
+            "error_types": feedback.error_types or [],
+            "correct_answer": feedback.correct_answer,
+        }
+
+    async def list_qa_records(
+        self,
+        *,
+        project_id: Optional[int],
+        date_from: Optional[datetime],
+        date_to: Optional[datetime],
+        rating: Optional[str],
+        page: int,
+        page_size: int,
+    ) -> dict[str, Any]:
+        page = max(1, int(page))
+        page_size = max(1, min(100, int(page_size)))
+        date_from_naive = self._to_naive_utc(date_from)
+        date_to_naive = self._to_naive_utc(date_to)
+
+        stmt = select(QARecord).where(QARecord.tenant_id == self.tenant_id)
+        if project_id is not None:
+            stmt = stmt.where(QARecord.project_id == int(project_id))
+        if date_from_naive is not None:
+            stmt = stmt.where(QARecord.created_at >= date_from_naive)
+        if date_to_naive is not None:
+            stmt = stmt.where(QARecord.created_at <= date_to_naive)
+
+        records = (
+            await self.db.execute(
+                stmt.order_by(QARecord.created_at.desc())
+                .offset((page - 1) * page_size)
+                .limit(page_size)
+            )
+        ).scalars().all()
+
+        ids = [int(r.id) for r in records]
+        feedback_rows = []
+        if ids:
+            fb_stmt = select(QAFeedback).where(
+                and_(
+                    QAFeedback.tenant_id == self.tenant_id,
+                    QAFeedback.qa_id.in_(ids),
+                )
+            )
+            if rating:
+                fb_stmt = fb_stmt.where(QAFeedback.rating == rating)
+            feedback_rows = (await self.db.execute(fb_stmt)).scalars().all()
+
+        latest_feedback: dict[int, QAFeedback] = {}
+        for fb in feedback_rows:
+            key = int(fb.qa_id)
+            curr = latest_feedback.get(key)
+            if curr is None or (fb.created_at and curr.created_at and fb.created_at > curr.created_at):
+                latest_feedback[key] = fb
+
+        items: list[dict[str, Any]] = []
+        for row in records:
+            fb = latest_feedback.get(int(row.id))
+            if rating and not fb:
+                continue
+            ack_meta = self._extract_ack_metadata_from_sources(row.sources, answer=row.answer)
+            items.append(
+                {
+                    "id": int(row.id),
+                    "question": row.question,
+                    "answer": row.answer,
+                    "sources": row.sources or [],
+                    "intent": row.intent,
+                    "domains_used": row.domains_used or [],
+                    "response_time_ms": row.response_time_ms,
+                    "asked_by": row.asked_by,
+                    "asked_via": row.asked_via,
+                    "project_id": row.project_id,
+                    "session_id": row.session_id,
+                    "turn_index": int(row.turn_index or 1),
+                    "created_at": row.created_at.isoformat() if row.created_at else None,
+                    "feedback": {
+                        "id": int(fb.id),
+                        "rating": fb.rating,
+                        "helpfulness": fb.helpfulness,
+                        "error_types": fb.error_types or [],
+                        "correct_answer": fb.correct_answer,
+                    }
+                    if fb
+                    else None,
+                    **ack_meta,
+                }
+            )
+
+        return {
+            "items": items,
+            "page": page,
+            "page_size": page_size,
+            "count": len(items),
+        }
+
+    async def get_qa_stats(
+        self,
+        *,
+        project_id: Optional[int],
+        date_from: Optional[datetime],
+        date_to: Optional[datetime],
+    ) -> dict[str, Any]:
+        date_from_naive = self._to_naive_utc(date_from)
+        date_to_naive = self._to_naive_utc(date_to)
+        record_stmt = select(QARecord).where(QARecord.tenant_id == self.tenant_id)
+        if project_id is not None:
+            record_stmt = record_stmt.where(QARecord.project_id == int(project_id))
+        if date_from_naive is not None:
+            record_stmt = record_stmt.where(QARecord.created_at >= date_from_naive)
+        if date_to_naive is not None:
+            record_stmt = record_stmt.where(QARecord.created_at <= date_to_naive)
+
+        records = (await self.db.execute(record_stmt)).scalars().all()
+        if not records:
+            return {
+                "total_sessions": 0,
+                "total_qa": 0,
+                "rated_count": 0,
+                "rated_ratio": 0.0,
+                "average_helpfulness": 0.0,
+                "exportable_pairs": 0,
+            }
+
+        record_ids = [int(r.id) for r in records]
+        feedback_rows = (
+            await self.db.execute(
+                select(QAFeedback).where(
+                    and_(
+                        QAFeedback.tenant_id == self.tenant_id,
+                        QAFeedback.qa_id.in_(record_ids),
+                    )
+                )
+            )
+        ).scalars().all()
+
+        latest_feedback: dict[int, QAFeedback] = {}
+        for fb in feedback_rows:
+            key = int(fb.qa_id)
+            curr = latest_feedback.get(key)
+            fb_at = fb.created_at or datetime.min
+            curr_at = curr.created_at if curr and curr.created_at else datetime.min
+            if curr is None or fb_at >= curr_at:
+                latest_feedback[key] = fb
+
+        rated_count = len(latest_feedback)
+        rated_ratio = round(rated_count / max(len(records), 1), 4)
+
+        helpful_scores: list[float] = []
+        helpful_map = {
+            "fully": 5.0,
+            "could_be_better": 4.0,
+            "partially": 3.0,
+        }
+        for fb in latest_feedback.values():
+            rating = str(fb.rating or "").strip().lower()
+            if rating == "positive":
+                score = helpful_map.get(str(fb.helpfulness or "").strip().lower(), 4.0)
+            elif rating == "negative":
+                score = 1.0
+            else:
+                continue
+            helpful_scores.append(score)
+
+        avg_helpfulness = round(sum(helpful_scores) / len(helpful_scores), 2) if helpful_scores else 0.0
+
+        exportable_pairs = 0
+        for fb in latest_feedback.values():
+            rating = str(fb.rating or "").strip().lower()
+            if rating == "positive" and str(fb.helpfulness or "").strip():
+                exportable_pairs += 1
+                continue
+            if rating == "negative":
+                exportable_pairs += 1
+
+        return {
+            "total_sessions": len({int(r.session_id) for r in records if r.session_id is not None}),
+            "total_qa": len(records),
+            "rated_count": rated_count,
+            "rated_ratio": rated_ratio,
+            "average_helpfulness": avg_helpfulness,
+            "exportable_pairs": exportable_pairs,
+        }
+
+    async def export_qa_jsonl(
+        self,
+        *,
+        include_unrated: bool,
+        project_id: Optional[int],
+        rating: Optional[str],
+        date_from: Optional[datetime],
+        date_to: Optional[datetime],
+        domain: Optional[str],
+    ) -> str:
+        date_from_naive = self._to_naive_utc(date_from)
+        date_to_naive = self._to_naive_utc(date_to)
+        record_stmt = select(QARecord).where(QARecord.tenant_id == self.tenant_id)
+        if project_id is not None:
+            record_stmt = record_stmt.where(QARecord.project_id == int(project_id))
+        if date_from_naive is not None:
+            record_stmt = record_stmt.where(QARecord.created_at >= date_from_naive)
+        if date_to_naive is not None:
+            record_stmt = record_stmt.where(QARecord.created_at <= date_to_naive)
+        if domain:
+            record_stmt = record_stmt.where(QARecord.domains_used.is_not(None))
+
+        records = (await self.db.execute(record_stmt.order_by(QARecord.created_at.asc()))).scalars().all()
+        if not records:
+            return ""
+
+        feedback_map = await self._latest_feedback_map([int(r.id) for r in records])
+        session_ids = sorted({int(r.session_id) for r in records if r.session_id is not None})
+        session_map: dict[int, QASession] = {}
+        if session_ids:
+            rows = (
+                await self.db.execute(
+                    select(QASession).where(
+                        and_(
+                            QASession.tenant_id == self.tenant_id,
+                            QASession.id.in_(session_ids),
+                        )
+                    )
+                )
+            ).scalars().all()
+            session_map = {int(x.id): x for x in rows}
+
+        lines: list[str] = []
+        normalized_rating = str(rating or "").strip().lower() or None
+        for record in records:
+            feedback = feedback_map.get(int(record.id))
+            if normalized_rating == "unrated":
+                if feedback:
+                    continue
+            else:
+                if not feedback and not include_unrated:
+                    continue
+                if normalized_rating in {"positive", "negative"}:
+                    if not feedback or str(feedback.rating or "").strip().lower() != normalized_rating:
+                        continue
+            if domain:
+                domains = [str(x).lower() for x in (record.domains_used or [])]
+                if domain.lower() not in domains:
+                    continue
+
+            session = session_map.get(int(record.session_id)) if record.session_id is not None else None
+            row = {
+                "session_id": int(record.session_id) if record.session_id is not None else None,
+                "session_title": str(session.title or SESSION_DEFAULT_TITLE) if session else SESSION_DEFAULT_TITLE,
+                "turn_index": int(record.turn_index or 1),
+                "question": record.question,
+                "answer": record.answer,
+                "sources": record.sources or [],
+                "domains_used": record.domains_used or [],
+                "asked_via": record.asked_via,
+                "project_id": record.project_id,
+                "created_at": record.created_at.isoformat() if record.created_at else None,
+                # 兼容历史导出字段
+                "prompt": record.question,
+                "response": record.answer,
+                "rating": str(feedback.rating) if feedback else "unrated",
+            }
+            if feedback:
+                if feedback.helpfulness:
+                    row["helpfulness"] = feedback.helpfulness
+                errors = [str(x) for x in (feedback.error_types or []) if str(x).strip()]
+                if errors:
+                    row["error_type"] = errors[0]
+                    row["error_types"] = errors
+                if feedback.correct_answer:
+                    row["correct_answer"] = feedback.correct_answer
+            lines.append(json.dumps(row, ensure_ascii=False))
+
+        return "\n".join(lines)
+
+    # ------------------------------------------------------------------
+    # Internal helpers
+    # ------------------------------------------------------------------
+    @staticmethod
+    def _contains_any(text: str, keywords: tuple[str, ...]) -> bool:
+        lowered = str(text or "").strip().lower()
+        if not lowered:
+            return False
+        return any(keyword in lowered for keyword in keywords)
+
+    @classmethod
+    def _is_correction_query(cls, question: str) -> bool:
+        text = str(question or "").strip().lower()
+        if not text:
+            return False
+        if BRIEF_NEGATIVE_FEEDBACK_RE.match(text):
+            return True
+        if cls._contains_any(text, CORRECTION_CUE_KEYWORDS):
+            return True
+        if re.search(r"(不是|别|不要).{0,8}(系统架构|架构)", text):
+            return True
+        return False
+
+    @classmethod
+    def _has_explicit_target_scope(cls, question: str) -> bool:
+        text = str(question or "").strip().lower()
+        if not text:
+            return False
+        explicit_groups: tuple[tuple[str, ...], ...] = (
+            ENGINE_PROCESS_KEYWORDS,
+            FEATURE_QUERY_KEYWORDS,
+            ARCHITECTURE_KEYWORDS,
+            PROJECT_KEYWORDS,
+            PEOPLE_KEYWORDS,
+            DOC_KEYWORDS,
+        )
+        for group in explicit_groups:
+            if cls._contains_any(text, group):
+                return True
+        if any(token in text for token in ("openclaw", "llm", "model", "engine", "api")):
+            return True
+        return False
+
+    @classmethod
+    def _is_engine_process_query(cls, question: str) -> bool:
+        return cls._contains_any(question, ENGINE_PROCESS_KEYWORDS)
+
+    @classmethod
+    def _is_explicit_deep_request(cls, question: str) -> bool:
+        return cls._contains_any(question, DEEP_EXPLICIT_KEYWORDS)
+
+    @classmethod
+    def _forced_deep_reason(
+        cls,
+        *,
+        question: str,
+        effective_question: str,
+        conversation: list[dict[str, str]],
+    ) -> str:
+        if cls._is_correction_query(question) or cls._is_correction_query(effective_question):
+            if conversation:
+                return "correction_turn"
+        if cls._is_engine_process_query(question) or cls._is_engine_process_query(effective_question):
+            return "engine_process_query"
+        return ""
+
+    @classmethod
+    def _should_force_deep_for_channel(
+        cls,
+        *,
+        asked_via: str,
+        question: str,
+        effective_question: str,
+        reason: str,
+    ) -> bool:
+        if not str(reason or "").strip():
+            return False
+        channel = str(asked_via or "").strip().lower()
+        if channel != "feishu":
+            return True
+
+        if cls._is_explicit_deep_request(question) or cls._is_explicit_deep_request(effective_question):
+            return True
+        if str(reason or "").strip() == "engine_process_query":
+            combined = f"{question} {effective_question}".strip()
+            return cls._has_explicit_target_scope(combined)
+        return False
+
+    @classmethod
+    def _is_snapshot_only_sources(cls, sources: list[dict[str, Any]]) -> bool:
+        if not isinstance(sources, list) or not sources:
+            return True
+        allowed_labels = {"System Snapshot", "User Snapshot", "Session State"}
+        for row in sources:
+            if not isinstance(row, dict):
+                continue
+            source = str(row.get("source") or "").strip().lower()
+            label = str(row.get("label") or "").strip()
+            if source != "system":
+                return False
+            if label and label not in allowed_labels:
+                return False
+        return True
+
+    @classmethod
+    def _should_auto_deep_retry(
+        cls,
+        *,
+        question: str,
+        effective_question: str,
+        sources: list[dict[str, Any]],
+        relevance_max: float,
+    ) -> str:
+        process_sensitive = cls._is_engine_process_query(question) or cls._is_engine_process_query(effective_question)
+        correction_sensitive = cls._is_correction_query(question) or cls._is_correction_query(effective_question)
+        if not process_sensitive and not correction_sensitive:
+            return ""
+        snapshot_only = cls._is_snapshot_only_sources(sources)
+        if snapshot_only:
+            return "snapshot_only"
+        if float(relevance_max or 0.0) < 0.62:
+            return "low_relevance"
+        return ""
+
+    @classmethod
+    def _allow_auto_deep_retry_for_channel(
+        cls,
+        *,
+        asked_via: str,
+        question: str,
+        effective_question: str,
+        reason: str,
+    ) -> bool:
+        if not str(reason or "").strip():
+            return False
+        channel = str(asked_via or "").strip().lower()
+        if channel != "feishu":
+            return True
+        if cls._is_explicit_deep_request(question) or cls._is_explicit_deep_request(effective_question):
+            return True
+        return False
+
+    @classmethod
+    def _should_block_answer_with_guard(
+        cls,
+        *,
+        question: str,
+        effective_question: str,
+        sources: list[dict[str, Any]],
+        relevance_max: float,
+    ) -> bool:
+        if not (cls._is_engine_process_query(question) or cls._is_engine_process_query(effective_question)):
+            return False
+        if not cls._is_snapshot_only_sources(sources):
+            return False
+        return float(relevance_max or 0.0) < 0.68
+
+    @staticmethod
+    def _build_evidence_guard_answer(question: str) -> str:
+        raw = str(question or "").strip()
+        target = raw[:80] if raw else "该问题"
+        return (
+            f"我不能对“{target}”直接给结论：当前只命中系统快照，缺少调用链证据。"
+            "请补充目标对象（如 openclaw 路由规则、函数名、入口层级），我会按调用链返回可核验答案。"
+        )
+
+    @staticmethod
+    def _is_snapshot_direct_question(question: str) -> bool:
+        text = str(question or "").strip().lower()
+        if not text:
+            return False
+        if AdaptiveRAGService._is_correction_query(text):
+            return False
+        if AdaptiveRAGService._is_engine_process_query(text):
+            return False
+        has_feature_token = any(k in text for k in FEATURE_QUERY_KEYWORDS)
+        ask_feature_purpose = has_feature_token and any(k in text for k in FEATURE_PURPOSE_HINTS)
+        return (
+            any(k in text for k in IDENTITY_KEYWORDS)
+            or any(k in text for k in ARCHITECTURE_KEYWORDS)
+            or ask_feature_purpose
+        )
+
+    @staticmethod
+    def _expand_followup_question(question: str, conversation: list[dict[str, str]]) -> str:
+        current = str(question or "").strip()
+        if not current:
+            return current
+
+        lowered = current.lower()
+        # 短句不等于追问：若已包含明确业务对象，则按独立问题处理，避免串错上下文。
+        has_explicit_subject = any(
+            marker in lowered for marker in ("项目", "信号", "信号源", "任务", "系统", "架构", "数据库", "redis", "接口", "用户")
+        )
+        is_short_followup = len(current) <= 16 and not has_explicit_subject
+        has_referential = any(k in lowered for k in FOLLOWUP_REFERENTIAL_KEYWORDS)
+        if not (is_short_followup or has_referential):
+            return current
+
+        anchor_question = ""
+        for turn in reversed(conversation[-5:]):
+            candidate = str(turn.get("question") or "").strip()
+            if not candidate:
+                continue
+            if candidate == current:
+                continue
+            if any(
+                marker in candidate
+                for marker in ("项目", "系统", "任务", "信号", "接口", "数据库", "redis", "架构", "模块")
+            ):
+                anchor_question = candidate
+                break
+
+        if not anchor_question:
+            return current
+
+        is_enumeration = any(k in lowered for k in FOLLOWUP_ENUMERATION_KEYWORDS)
+        if is_enumeration and ("信号" in lowered or "信号" in anchor_question):
+            return f"{anchor_question}。请列出该项目关联的信号标题和ID。"
+        if is_enumeration and ("任务" in lowered or "任务" in anchor_question):
+            return f"{anchor_question}。请列出对应任务的标题和状态。"
+        if is_enumeration and "项目" in anchor_question:
+            return f"{anchor_question}。请列出对应项目的名称和ID。"
+        return f"{anchor_question}。补充问题：{current}"
+
+    def _build_snapshot_direct_answer(
+        self,
+        question: str,
+        *,
+        system_snapshot: Optional[dict[str, Any]],
+        user_snapshot: Optional[dict[str, Any]],
+    ) -> Optional[str]:
+        text = str(question or "").strip().lower()
+        if not text:
+            return None
+        if self._is_correction_query(text):
+            return None
+        if self._is_engine_process_query(text):
+            return None
+
+        ask_identity = any(k in text for k in IDENTITY_KEYWORDS)
+        ask_architecture = any(k in text for k in ARCHITECTURE_KEYWORDS)
+        has_feature_token = any(k in text for k in FEATURE_QUERY_KEYWORDS)
+        ask_feature = has_feature_token and any(k in text for k in FEATURE_PURPOSE_HINTS)
+        if not ask_identity and not ask_architecture and not ask_feature:
+            return None
+
+        system_data = system_snapshot if isinstance(system_snapshot, dict) else {}
+        assistant = system_data.get("assistant_profile") if isinstance(system_data.get("assistant_profile"), dict) else {}
+        platform = system_data.get("platform_profile") if isinstance(system_data.get("platform_profile"), dict) else {}
+        backend = platform.get("backend") if isinstance(platform.get("backend"), dict) else {}
+        infra = platform.get("infrastructure") if isinstance(platform.get("infrastructure"), dict) else {}
+        module_counts = backend.get("module_counts") if isinstance(backend.get("module_counts"), dict) else {}
+
+        lines: list[str] = []
+        if ask_identity:
+            name = str(assistant.get("name") or "PMFlow AI 管家")
+            identity = str(assistant.get("identity") or "PMFlow 内置 AI 助手")
+            lines.append(f"我是{name}，{identity}。")
+            responsibilities = assistant.get("responsibilities") if isinstance(assistant.get("responsibilities"), list) else []
+            if responsibilities:
+                joined = "；".join(str(x) for x in responsibilities[:5] if str(x).strip())
+                if joined:
+                    lines.append(f"职责范围：{joined}。")
+
+        if ask_architecture:
+            runtime = str(platform.get("runtime") or "unknown")
+            backend_framework = str(backend.get("framework") or "FastAPI")
+            orm_name = str(backend.get("orm") or "SQLAlchemy 2.0")
+            db_name = str(infra.get("database") or "PostgreSQL")
+            cache_name = str(infra.get("cache") or "Redis")
+            lines.append(
+                f"当前系统采用前后端分层架构：后端 {backend_framework} + {orm_name}，数据库 {db_name}，缓存 {cache_name}，运行环境 {runtime}。"
+            )
+            if module_counts:
+                lines.append(
+                    "后端模块规模："
+                    f"api {int(module_counts.get('api', 0))}、"
+                    f"services {int(module_counts.get('services', 0))}、"
+                    f"models {int(module_counts.get('models', 0))}、"
+                    f"schemas {int(module_counts.get('schemas', 0))}。"
+                )
+
+        if ask_feature:
+            feature_hits = self._match_feature_catalog_entries(
+                question=question,
+                system_snapshot=system_data,
+                limit=4,
+            )
+            if feature_hits:
+                lines.append("命中的产品功能：")
+                for idx, row in enumerate(feature_hits, start=1):
+                    name = str(row.get("name") or "未知功能")
+                    summary = str(row.get("summary") or "暂无说明")
+                    route = str(row.get("route") or "")
+                    page_file = str(row.get("page_file") or "")
+                    apis = row.get("apis") if isinstance(row.get("apis"), list) else []
+                    api_text = "、".join(str(x) for x in apis[:3] if str(x).strip())
+                    line = f"{idx}. {name}：{summary}"
+                    if route:
+                        line += f" 入口 {route}。"
+                    if page_file:
+                        line += f" 页面文件 {page_file}。"
+                    if api_text:
+                        line += f" 相关接口 {api_text}。"
+                    lines.append(line)
+            else:
+                catalog = self._feature_catalog_from_snapshot(system_data)
+                if catalog:
+                    brief = "、".join(
+                        str(row.get("name") or "")
+                        for row in catalog[:8]
+                        if isinstance(row, dict) and str(row.get("name") or "").strip()
+                    )
+                    if brief:
+                        lines.append(f"当前功能目录包括：{brief}。你可以继续指定其中一个功能名称，我会返回用途与入口。")
+                else:
+                    lines.append("当前 System Snapshot 未包含 feature_catalog，建议先刷新上下文缓存后再提问。")
+
+        if ask_identity and isinstance(user_snapshot, dict) and user_snapshot:
+            projects = user_snapshot.get("projects") if isinstance(user_snapshot.get("projects"), list) else []
+            tasks = user_snapshot.get("assigned_tasks") if isinstance(user_snapshot.get("assigned_tasks"), list) else []
+            lines.append(f"你的当前上下文：关联项目 {len(projects)} 个，未完成任务 {len(tasks)} 个。")
+
+        if not lines:
+            return None
+        lines.append("来源: System Snapshot / User Snapshot")
+        return "\n".join(lines)[:3200]
+
+    async def _generate_answer(
+        self,
+        question: str,
+        retrieved: list[RetrievedItem],
+        conversation: list[dict[str, str]],
+        project_id: Optional[int],
+        *,
+        system_snapshot: Optional[dict[str, Any]] = None,
+        user_snapshot: Optional[dict[str, Any]] = None,
+        relevance_max: float = 0.0,
+        thinking_mode: str = THINKING_MODE_NORMAL,
+        context_block_limit: Optional[int] = None,
+        answer_max_tokens: Optional[int] = None,
+    ) -> str:
+        active_thinking_mode = self._normalize_thinking_mode(thinking_mode)
+        deep_thinking = active_thinking_mode == THINKING_MODE_DEEP
+        context_blocks: list[str] = []
+        used_tokens = 0
+        max_blocks = max(4, min(int(context_block_limit or DEFAULT_TOP_K * 2), 32))
+        manifest = await self.get_project_manifest(project_id=project_id)
+        manifest_text = str(manifest.get("manifest_text") or "").strip()
+        if manifest_text:
+            manifest_block = f"[M] 项目画像\n{manifest_text}\n"
+            manifest_tokens = max(len(manifest_block) // 4, 1)
+            if used_tokens + manifest_tokens <= MAX_CONTEXT_TOKENS:
+                context_blocks.append(manifest_block)
+                used_tokens += manifest_tokens
+
+        for idx, item in enumerate(retrieved[:max_blocks]):
+            block = f"[{idx + 1}] 来源: {item.source_label}\n相关度: {item.score:.3f}\n内容: {item.content}\n"
+            block_tokens = max(len(block) // 4, 1)
+            if used_tokens + block_tokens > MAX_CONTEXT_TOKENS:
+                break
+            used_tokens += block_tokens
+            context_blocks.append(block)
+
+        history_lines = []
+        for row in conversation[-10:]:
+            history_lines.append(f"Q: {row.get('question', '')}\nA: {row.get('answer', '')}")
+
+        system_snapshot_text = json.dumps(system_snapshot or {}, ensure_ascii=False)[:12000]
+        user_snapshot_text = json.dumps(user_snapshot or {}, ensure_ascii=False)[:6000]
+
+        prompt = PROMPT_RAG_ANSWER.format(
+            project_id=project_id if project_id is not None else "unknown",
+            question=question,
+            conversation=chr(10).join(history_lines) if history_lines else "无",
+            context_blocks=chr(10).join(context_blocks) if context_blocks else "无",
+            system_snapshot=system_snapshot_text or "{}",
+            user_snapshot=user_snapshot_text or "{}",
+            relevance_max=f"{max(0.0, float(relevance_max or 0.0)):.3f}",
+        ).strip()
+        if deep_thinking:
+            prompt = (
+                f"{prompt}\n\n"
+                "补充要求（深入思考模式）：\n"
+                "1. 先给结论，再给证据链；\n"
+                "2. 明确区分已验证事实与推断；\n"
+                "3. 若存在不确定项，给出待确认清单。"
+            )
+
+        effective_max_tokens = int(
+            answer_max_tokens
+            if answer_max_tokens is not None
+            else (DEEP_ANSWER_MAX_TOKENS if deep_thinking else 900)
+        )
+        effective_max_tokens = max(256, min(effective_max_tokens, 3000))
+        effective_temperature = 0.15 if deep_thinking else 0.2
+        function_name = "adaptive_rag_answer_deep" if deep_thinking else "adaptive_rag_answer_normal"
+
+        answer = await self.llm.generate(
+            prompt,
+            db=self.db,
+            tenant_id=self.tenant_id,
+            function_name=function_name,
+            temperature=effective_temperature,
+            max_tokens=effective_max_tokens,
+            use_prompt_profile=False,
+        )
+        text = str(answer or "").strip()
+        if not text:
+            return "我未能从当前上下文生成有效答案，请补充问题细节后重试。"
+        enforced = self._enforce_feature_evidence_in_answer(
+            question=question,
+            answer_text=text,
+            retrieved=retrieved,
+            system_snapshot=system_snapshot,
+        )
+        return str(enforced or text)[:3200]
+
+    async def _load_recent_conversation(
+        self,
+        *,
+        asked_by: Optional[int],
+        asked_via: str,
+        project_id: Optional[int],
+        session_id: Optional[int],
+        external_session_id: Optional[int],
+        limit: int,
+    ) -> list[dict[str, str]]:
+        active_session_id = session_id if session_id is not None else external_session_id
+        if active_session_id is not None:
+            cached_messages = await self.context_cache.get_session_messages(
+                tenant_id=self.tenant_id,
+                session_id=active_session_id,
+            )
+            if cached_messages:
+                cached_turns = self.context_cache.messages_to_turns(cached_messages)
+                if cached_turns:
+                    return cached_turns[-limit:]
+
+        if asked_by is None:
+            return []
+
+        stmt = select(QARecord).where(
+            and_(
+                QARecord.tenant_id == self.tenant_id,
+                QARecord.asked_by == int(asked_by),
+                QARecord.asked_via == (asked_via or "pmflow")[:20],
+            )
+        )
+        if session_id is not None:
+            stmt = stmt.where(QARecord.session_id == int(session_id))
+        if project_id is not None:
+            stmt = stmt.where(QARecord.project_id == int(project_id))
+        rows = (await self.db.execute(stmt.order_by(QARecord.created_at.desc()).limit(limit))).scalars().all()
+        return [{"question": str(x.question or ""), "answer": str(x.answer or "")} for x in reversed(rows)]
+
+    def _map_domains_to_source_types(self, domains: list[str]) -> list[str]:
+        mapped: list[str] = []
+        if "doc" in domains:
+            mapped.append("doc")
+        if "feishu" in domains:
+            mapped.append("feishu")
+        if "pmflow" in domains:
+            mapped.append("pmflow")
+        if "url" in domains:
+            mapped.append("url")
+        return mapped
+
+    def _source_label(self, source_type: str, metadata: dict[str, Any]) -> str:
+        if source_type == "doc":
+            title = metadata.get("title") or "unknown"
+            page_number = metadata.get("page_number")
+            if page_number is not None:
+                return f"Document[{title}] p.{page_number}"
+            return f"Document[{title}]"
+        if source_type == "url":
+            return f"URL[{metadata.get('file_url') or metadata.get('url') or 'external'}]"
+        if source_type == "feishu":
+            return "Feishu history"
+        if source_type == "pmflow":
+            record_type = str(metadata.get("record_type") or "").strip()
+            record_id = metadata.get("record_id")
+            if record_type and record_id is not None:
+                return f"PMFlow[{record_type}#{record_id}]"
+            return "PMFlow records"
+        return "Knowledge base"
+
+    async def _noop_results(self) -> list[RetrievedItem]:
+        return []
+
+    async def _fetch_url_text(self, url: str) -> str:
+        target = str(url or "").strip()
+        if not target:
+            raise ValueError("url_empty")
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False, follow_redirects=True) as client:
+            resp = await client.get(target, headers={"User-Agent": "PMFlow-RAG/1.0"})
+            resp.raise_for_status()
+            content_type = str(resp.headers.get("content-type") or "").lower()
+            body = resp.text
+        if "html" in content_type:
+            return self._html_to_text(body)
+        return body[:300000]
+
+    @staticmethod
+    def _html_to_text(html: str) -> str:
+        text = str(html or "")
+        text = re.sub(r"<script.*?>.*?</script>", " ", text, flags=re.IGNORECASE | re.DOTALL)
+        text = re.sub(r"<style.*?>.*?</style>", " ", text, flags=re.IGNORECASE | re.DOTALL)
+        text = re.sub(r"<[^>]+>", " ", text)
+        text = re.sub(r"&nbsp;|&#160;", " ", text)
+        text = re.sub(r"&amp;", "&", text)
+        text = re.sub(r"&lt;", "<", text)
+        text = re.sub(r"&gt;", ">", text)
+        text = re.sub(r"\s+", " ", text).strip()
+        return text[:300000]
+
+    def _chunk_text(self, text: str, *, chunk_size: int = 500, overlap: int = 50) -> list[str]:
+        clean = str(text or "").strip()
+        if not clean:
+            return []
+        if len(clean) <= chunk_size:
+            return [clean]
+
+        chunks: list[str] = []
+        start = 0
+        step = max(1, chunk_size - overlap)
+        while start < len(clean):
+            end = min(len(clean), start + chunk_size)
+            part = clean[start:end].strip()
+            if part:
+                chunks.append(part)
+            if end >= len(clean):
+                break
+            start += step
+        return chunks
+
+    def _chunk_text_with_metadata(
+        self,
+        text: str,
+        *,
+        chunk_size: int = 500,
+        overlap: int = 50,
+    ) -> list[tuple[str, dict[str, Any]]]:
+        segments = self._split_text_by_page_markers(text)
+        if not segments:
+            return [(part, {}) for part in self._chunk_text(text, chunk_size=chunk_size, overlap=overlap)]
+
+        out: list[tuple[str, dict[str, Any]]] = []
+        for page_number, segment in segments:
+            for part in self._chunk_text(segment, chunk_size=chunk_size, overlap=overlap):
+                meta: dict[str, Any] = {}
+                if page_number is not None:
+                    meta["page_number"] = int(page_number)
+                out.append((part, meta))
+        return out
+
+    def _split_text_by_page_markers(self, text: str) -> list[tuple[Optional[int], str]]:
+        clean = str(text or "")
+        matches = list(PAGE_MARKER_RE.finditer(clean))
+        if not matches:
+            return []
+
+        segments: list[tuple[Optional[int], str]] = []
+        prefix = clean[: matches[0].start()].strip()
+        if prefix:
+            segments.append((None, prefix))
+
+        for idx, marker in enumerate(matches):
+            try:
+                page_number = int(marker.group(1))
+            except Exception:
+                page_number = None
+            start = marker.end()
+            end = matches[idx + 1].start() if idx + 1 < len(matches) else len(clean)
+            segment = clean[start:end].strip()
+            if segment:
+                segments.append((page_number, segment))
+        return segments
+
+    @staticmethod
+    def _to_float(value: Any, default: float) -> float:
+        try:
+            return float(value)
+        except Exception:
+            return float(default)
+
+    @staticmethod
+    def _safe_dict(value: Any) -> dict[str, Any]:
+        if isinstance(value, dict):
+            return value
+        if isinstance(value, str):
+            try:
+                parsed = json.loads(value)
+                return parsed if isinstance(parsed, dict) else {}
+            except Exception:
+                return {}
+        return {}
+
+    @staticmethod
+    def _parse_json_object(raw: str) -> dict[str, Any]:
+        text = str(raw or "").strip()
+        if text.startswith("```"):
+            parts = text.split("```")
+            if len(parts) >= 2:
+                text = parts[1].strip()
+            if text.startswith("json"):
+                text = text[4:].strip()
+        parsed: dict[str, Any] = {}
+        try:
+            value = json.loads(text)
+            if isinstance(value, dict):
+                parsed = value
+        except Exception:
+            pass
+        return parsed
+
+    @staticmethod
+    def _embedding_to_list(raw: Any) -> list[float]:
+        if raw is None:
+            return []
+        if isinstance(raw, list):
+            return [float(x) for x in raw]
+        if isinstance(raw, str):
+            try:
+                value = json.loads(raw)
+                if isinstance(value, list):
+                    return [float(x) for x in value]
+            except Exception:
+                return []
+        return []
+
+    @staticmethod
+    def _cosine_similarity(v1: list[float], v2: list[float]) -> float:
+        if not v1 or not v2:
+            return 0.0
+        size = min(len(v1), len(v2))
+        if size == 0:
+            return 0.0
+        a = v1[:size]
+        b = v2[:size]
+        dot = sum(x * y for x, y in zip(a, b))
+        norm_a = math.sqrt(sum(x * x for x in a))
+        norm_b = math.sqrt(sum(y * y for y in b))
+        if norm_a == 0 or norm_b == 0:
+            return 0.0
+        sim = dot / (norm_a * norm_b)
+        return (sim + 1) / 2
+
+    def _dedupe_sources(self, retrieved: list[RetrievedItem]) -> list[dict[str, Any]]:
+        seen: set[str] = set()
+        out: list[dict[str, Any]] = []
+        for item in retrieved:
+            meta_identity = (
+                item.metadata.get("message_id")
+                or item.metadata.get("file_url")
+                or item.metadata.get("target_url")
+                or item.metadata.get("path")
+                or item.metadata.get("relation_id")
+                or item.metadata.get("entity_id")
+                or item.metadata.get("subject_entity_id")
+                or item.metadata.get("object_entity_id")
+                or item.content[:80]
+            )
+            key = f"{item.source}|{item.source_label}|{meta_identity}"
+            if key in seen:
+                continue
+            seen.add(key)
+            out.append(
+                {
+                    "source": item.source,
+                    "label": item.source_label,
+                    "score": round(item.score, 4),
+                    "metadata": item.metadata,
+                }
+            )
+            if len(out) >= 8:
+                break
+        return out
+
+    @staticmethod
+    def _build_pmflow_target_url(*, record_type: str, record_id: Any, project_id: Any) -> str:
+        safe_type = str(record_type or "").strip().lower()
+        try:
+            pid = int(project_id) if project_id is not None else None
+        except Exception:
+            pid = None
+        try:
+            rid = int(record_id) if record_id is not None else None
+        except Exception:
+            rid = None
+        if pid is None:
+            return ""
+        if safe_type == "charter" and rid is not None:
+            return f"/projects/{pid}/charter/{rid}"
+        if safe_type == "task" and rid is not None:
+            return f"/projects/{pid}/board?focus=task&taskId={rid}"
+        if safe_type == "review" and rid is not None:
+            return f"/projects/{pid}/board?focus=review&reviewId={rid}"
+        return f"/projects/{pid}/board"
+
+    @staticmethod
+    def _serialize_doc(doc: KnowledgeBaseDoc) -> dict[str, Any]:
+        return {
+            "id": int(doc.id),
+            "title": doc.title,
+            "file_type": doc.file_type,
+            "source_type": doc.source_type,
+            "file_url": doc.file_url,
+            "status": doc.status,
+            "uploaded_by": doc.uploaded_by,
+            "created_at": doc.created_at.isoformat() if doc.created_at else None,
+            "last_fetched_at": doc.last_fetched_at.isoformat() if doc.last_fetched_at else None,
+            "error_message": doc.error_message,
+        }
+
+
+__all__ = ["AdaptiveRAGService"]
diff --git a/backend/app/services/ai_orchestration_run_service.py b/backend/app/services/ai_orchestration_run_service.py
new file mode 100644
index 00000000..3b9b44d1
--- /dev/null
+++ b/backend/app/services/ai_orchestration_run_service.py
@@ -0,0 +1,179 @@
+"""AI 编排运行账本服务。"""
+
+from __future__ import annotations
+
+from datetime import datetime, timedelta, timezone
+from typing import Any, Optional
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import AIOrchestrationRun
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+class AIOrchestrationRunService:
+    def __init__(self, db: AsyncSession) -> None:
+        self.db = db
+
+    async def create_run(
+        self,
+        *,
+        tenant_id: int,
+        trace_id: str,
+        run_type: str,
+        project_id: int,
+        signal_id: Optional[int],
+        response_mode: str,
+        status: str,
+        planner_status: Optional[str],
+        openclaw_status: Optional[str],
+        worker_total: int,
+        worker_completed: int,
+        worker_failed: int,
+        summary: str,
+        planner_diagnostics: Optional[dict[str, Any]],
+        result_payload: Optional[dict[str, Any]],
+        created_by_user_id: Optional[int],
+        started_at: Optional[datetime] = None,
+        completed_at: Optional[datetime] = None,
+    ) -> AIOrchestrationRun:
+        row = AIOrchestrationRun(
+            tenant_id=tenant_id,
+            trace_id=trace_id,
+            run_type=run_type,
+            project_id=project_id,
+            signal_id=signal_id,
+            response_mode=response_mode,
+            status=status,
+            planner_status=planner_status,
+            openclaw_status=openclaw_status,
+            worker_total=max(0, int(worker_total)),
+            worker_completed=max(0, int(worker_completed)),
+            worker_failed=max(0, int(worker_failed)),
+            summary=str(summary or "").strip()[:1000] or None,
+            planner_diagnostics=planner_diagnostics if isinstance(planner_diagnostics, dict) else None,
+            result_payload=result_payload if isinstance(result_payload, dict) else None,
+            created_by_user_id=created_by_user_id,
+            started_at=started_at or _utc_now_naive(),
+            completed_at=completed_at,
+        )
+        self.db.add(row)
+        await self.db.flush()
+        return row
+
+    async def get_run(self, *, tenant_id: int, trace_id: str) -> Optional[AIOrchestrationRun]:
+        return (
+            await self.db.execute(
+                select(AIOrchestrationRun).where(
+                    and_(
+                        AIOrchestrationRun.tenant_id == int(tenant_id),
+                        AIOrchestrationRun.trace_id == str(trace_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def update_run(
+        self,
+        *,
+        tenant_id: int,
+        trace_id: str,
+        status: Optional[str] = None,
+        planner_status: Optional[str] = None,
+        openclaw_status: Optional[str] = None,
+        worker_total: Optional[int] = None,
+        worker_completed: Optional[int] = None,
+        worker_failed: Optional[int] = None,
+        summary: Optional[str] = None,
+        planner_diagnostics: Optional[dict[str, Any]] = None,
+        result_payload: Optional[dict[str, Any]] = None,
+        completed_at: Optional[datetime] = None,
+    ) -> Optional[AIOrchestrationRun]:
+        row = await self.get_run(tenant_id=tenant_id, trace_id=trace_id)
+        if row is None:
+            return None
+        if status is not None:
+            row.status = str(status or "").strip()[:20]
+        if planner_status is not None:
+            row.planner_status = str(planner_status or "").strip()[:20] or None
+        if openclaw_status is not None:
+            row.openclaw_status = str(openclaw_status or "").strip()[:20] or None
+        if worker_total is not None:
+            row.worker_total = max(0, int(worker_total))
+        if worker_completed is not None:
+            row.worker_completed = max(0, int(worker_completed))
+        if worker_failed is not None:
+            row.worker_failed = max(0, int(worker_failed))
+        if summary is not None:
+            row.summary = str(summary or "").strip()[:1000] or None
+        if planner_diagnostics is not None:
+            row.planner_diagnostics = planner_diagnostics
+        if result_payload is not None:
+            row.result_payload = result_payload
+        if completed_at is not None:
+            row.completed_at = completed_at
+        return row
+
+    async def mark_stale_runs_interrupted(
+        self,
+        *,
+        older_than_minutes: int = 10,
+    ) -> int:
+        threshold = _utc_now_naive() - timedelta(minutes=max(1, int(older_than_minutes)))
+        rows = (
+            await self.db.execute(
+                select(AIOrchestrationRun).where(
+                    and_(
+                        AIOrchestrationRun.status.in_(("queued", "running")),
+                        AIOrchestrationRun.started_at.is_not(None),
+                        AIOrchestrationRun.started_at < threshold,
+                    )
+                )
+            )
+        ).scalars().all()
+        count = 0
+        for row in rows:
+            payload = row.result_payload if isinstance(row.result_payload, dict) else {}
+            payload["run_status"] = "interrupted"
+            payload["pending_workers"] = payload.get("pending_workers") or []
+            row.status = "interrupted"
+            row.completed_at = _utc_now_naive()
+            row.result_payload = payload
+            count += 1
+        return count
+
+    @staticmethod
+    def serialize_run(row: AIOrchestrationRun) -> dict[str, Any]:
+        payload = row.result_payload if isinstance(row.result_payload, dict) else {}
+        pending_workers = payload.get("pending_workers") if isinstance(payload.get("pending_workers"), list) else []
+        worker_total = int(row.worker_total or 0)
+        worker_completed = int(row.worker_completed or 0)
+        worker_failed = int(row.worker_failed or 0)
+        worker_pending = max(0, worker_total - worker_completed - worker_failed)
+        if not pending_workers and worker_pending > 0:
+            pending_workers = [f"pending_{idx}" for idx in range(1, worker_pending + 1)]
+        return {
+            "trace_id": str(row.trace_id or ""),
+            "run_type": str(row.run_type or ""),
+            "project_id": int(row.project_id),
+            "signal_id": int(row.signal_id) if row.signal_id is not None else None,
+            "response_mode": str(row.response_mode or ""),
+            "run_status": str(row.status or ""),
+            "planner_status": str(row.planner_status or ""),
+            "openclaw_status": str(row.openclaw_status or ""),
+            "worker_total": worker_total,
+            "worker_completed": worker_completed,
+            "worker_failed": worker_failed,
+            "worker_pending": worker_pending,
+            "pending_workers": pending_workers,
+            "summary": str(row.summary or payload.get("summary") or "").strip(),
+            "planner_diagnostics": row.planner_diagnostics if isinstance(row.planner_diagnostics, dict) else None,
+            "result_snapshot": payload,
+            "created_by_user_id": int(row.created_by_user_id) if row.created_by_user_id is not None else None,
+            "started_at": row.started_at.isoformat() if row.started_at else None,
+            "completed_at": row.completed_at.isoformat() if row.completed_at else None,
+        }
diff --git a/backend/app/services/ai_pmo_cadence_service.py b/backend/app/services/ai_pmo_cadence_service.py
new file mode 100644
index 00000000..c989c8b6
--- /dev/null
+++ b/backend/app/services/ai_pmo_cadence_service.py
@@ -0,0 +1,483 @@
+"""AI PMO cadence engine。"""
+
+from __future__ import annotations
+
+from datetime import datetime, timedelta, timezone
+from typing import Any, Optional
+
+from sqlalchemy import and_, func, or_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import Budget, Notification, PortfolioSnapshot, Project, ProjectDomain, ProjectSignalLink, Signal, SignalSourceRegistry, Task, UserProject, UserWorkload, WorkItem
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+class AIPMOCadenceService:
+    """日会 / 周会 / 风险例会 / 里程碑检查 / 项目群视角。"""
+
+    SUPPORTED_CADENCE_TYPES = {"daily_standup", "weekly_review", "risk_review", "milestone_check"}
+
+    def __init__(self, db: AsyncSession, tenant_id: int) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+
+    async def generate_snapshot(self, *, cadence_type: str, project_id: Optional[int] = None) -> dict[str, Any]:
+        normalized_type = str(cadence_type or "").strip().lower()
+        if normalized_type not in self.SUPPORTED_CADENCE_TYPES:
+            raise ValueError("unsupported cadence_type")
+        if normalized_type == "daily_standup":
+            payload = await self._build_daily_standup(project_id=project_id)
+        elif normalized_type == "weekly_review":
+            payload = await self._build_weekly_review(project_id=project_id)
+        elif normalized_type == "risk_review":
+            payload = await self._build_risk_review(project_id=project_id)
+        else:
+            payload = await self._build_milestone_check(project_id=project_id)
+
+        scope_ref = f"project:{int(project_id)}" if project_id is not None else f"tenant:{self.tenant_id}"
+        row = PortfolioSnapshot(
+            tenant_id=self.tenant_id,
+            scope_type=normalized_type[:20],
+            scope_ref=scope_ref[:64],
+            summary=payload.get("summary"),
+            payload=payload,
+        )
+        self.db.add(row)
+        await self.db.flush()
+        return {
+            "id": int(row.id),
+            "scope_type": row.scope_type,
+            "scope_ref": row.scope_ref,
+            "summary": row.summary,
+            "payload": row.payload,
+            "created_at": row.created_at.isoformat() if row.created_at else None,
+        }
+
+    async def get_latest_snapshot(self, *, cadence_type: str, project_id: Optional[int] = None) -> Optional[dict[str, Any]]:
+        normalized_type = str(cadence_type or "").strip().lower()
+        scope_ref = f"project:{int(project_id)}" if project_id is not None else f"tenant:{self.tenant_id}"
+        row = (
+            await self.db.execute(
+                select(PortfolioSnapshot).where(
+                    and_(
+                        PortfolioSnapshot.tenant_id == self.tenant_id,
+                        PortfolioSnapshot.scope_type == normalized_type,
+                        PortfolioSnapshot.scope_ref == scope_ref,
+                    )
+                ).order_by(PortfolioSnapshot.id.desc()).limit(1)
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return {
+            "id": int(row.id),
+            "scope_type": row.scope_type,
+            "scope_ref": row.scope_ref,
+            "summary": row.summary,
+            "payload": row.payload,
+            "created_at": row.created_at.isoformat() if row.created_at else None,
+        }
+
+    async def build_program_view(self) -> dict[str, Any]:
+        projects = (
+            await self.db.execute(
+                select(Project, ProjectDomain)
+                .outerjoin(ProjectDomain, and_(ProjectDomain.id == Project.domain_id, ProjectDomain.tenant_id == self.tenant_id))
+                .where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.status.in_(["active", "planning", "doing", "todo"]),
+                    )
+                )
+                .order_by(Project.id.asc())
+            )
+        ).all()
+
+        domain_groups: dict[str, dict[str, Any]] = {}
+        for project, domain in projects:
+            key = str((domain.slug if domain is not None else "unscoped") or "unscoped")
+            group = domain_groups.setdefault(
+                key,
+                {
+                    "program_key": key,
+                    "program_name": str((domain.name if domain is not None else "未分组项目") or "未分组项目"),
+                    "project_count": 0,
+                    "project_ids": [],
+                    "blocked_task_count": 0,
+                    "pending_manual_count": 0,
+                    "budget_usage_ratio_max": 0.0,
+                },
+            )
+            group["project_count"] += 1
+            group["project_ids"].append(int(project.id))
+
+        if not domain_groups:
+            return {"summary": {"program_count": 0, "active_project_count": 0}, "programs": []}
+
+        project_ids = [pid for group in domain_groups.values() for pid in group["project_ids"]]
+        task_rows = (
+            await self.db.execute(
+                select(Task.project_id, func.count(Task.id))
+                .where(
+                    and_(
+                        Task.tenant_id == self.tenant_id,
+                        Task.project_id.in_(project_ids),
+                        Task.status == "blocked",
+                    )
+                )
+                .group_by(Task.project_id)
+            )
+        ).all()
+        task_map = {int(project_id): int(count or 0) for project_id, count in task_rows}
+
+        work_rows = (
+            await self.db.execute(
+                select(WorkItem.project_id, func.count(WorkItem.id))
+                .where(
+                    and_(
+                        WorkItem.tenant_id == self.tenant_id,
+                        WorkItem.project_id.in_(project_ids),
+                        WorkItem.state == "waiting_manual",
+                    )
+                )
+                .group_by(WorkItem.project_id)
+            )
+        ).all()
+        work_map = {int(project_id): int(count or 0) for project_id, count in work_rows}
+
+        budget_rows = (
+            await self.db.execute(
+                select(Budget.project_id, Budget.budget_amount_usd, Budget.current_usage_usd)
+                .where(
+                    and_(
+                        Budget.tenant_id == self.tenant_id,
+                        Budget.project_id.in_(project_ids),
+                    )
+                )
+            )
+        ).all()
+        budget_ratio_map = {
+            int(project_id): (float(usage or 0.0) / float(amount or 1.0)) if float(amount or 0.0) > 0 else 0.0
+            for project_id, amount, usage in budget_rows
+            if project_id is not None
+        }
+
+        for group in domain_groups.values():
+            group["blocked_task_count"] = sum(task_map.get(pid, 0) for pid in group["project_ids"])
+            group["pending_manual_count"] = sum(work_map.get(pid, 0) for pid in group["project_ids"])
+            ratios = [budget_ratio_map.get(pid, 0.0) for pid in group["project_ids"]]
+            group["budget_usage_ratio_max"] = round(max(ratios, default=0.0), 4)
+
+        programs = sorted(domain_groups.values(), key=lambda item: (item["pending_manual_count"], item["blocked_task_count"], item["project_count"]), reverse=True)
+        return {
+            "summary": {
+                "program_count": len(programs),
+                "active_project_count": len(project_ids),
+                "pending_manual_total": sum(item["pending_manual_count"] for item in programs),
+                "blocked_task_total": sum(item["blocked_task_count"] for item in programs),
+            },
+            "programs": programs,
+        }
+
+    async def _build_daily_standup(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        now = _utc_now_naive()
+        since = now - timedelta(days=1)
+        work_items = await self._list_recent_work_items(since=since, project_id=project_id)
+        blocked_tasks = await self._list_blocked_tasks(project_id=project_id)
+        recent_signals = await self._list_recent_signals(since=since, project_id=project_id)
+        pending_manual_count = await self._count_work_items(state="waiting_manual", project_id=project_id)
+        return {
+            "summary": {
+                "cadence_type": "daily_standup",
+                "recent_work_item_count": len(work_items),
+                "pending_manual_count": pending_manual_count,
+                "blocked_task_count": len(blocked_tasks),
+                "recent_signal_count": len(recent_signals),
+            },
+            "work_items": work_items[:10],
+            "blocked_tasks": blocked_tasks[:10],
+            "recent_signals": recent_signals[:10],
+            "focus_actions": self._build_focus_actions(pending_manual_count=pending_manual_count, blocked_task_count=len(blocked_tasks)),
+        }
+
+    async def _build_weekly_review(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        now = _utc_now_naive()
+        since = now - timedelta(days=7)
+        weekly_rows = await self._list_recent_work_items(since=since, project_id=project_id)
+        budgets = await self._list_budget_hotspots(project_id=project_id)
+        capacity = await self._list_capacity_hotspots(project_id=project_id)
+        program_view = await self.build_program_view() if project_id is None else None
+        by_state: dict[str, int] = {}
+        for item in weekly_rows:
+            state = str(item.get("state") or "unknown")
+            by_state[state] = by_state.get(state, 0) + 1
+        return {
+            "summary": {
+                "cadence_type": "weekly_review",
+                "work_item_total": len(weekly_rows),
+                "by_state": by_state,
+                "budget_hotspot_count": len(budgets),
+                "capacity_hotspot_count": len(capacity),
+            },
+            "work_items": weekly_rows[:20],
+            "budget_hotspots": budgets[:10],
+            "capacity_hotspots": capacity[:10],
+            "program_view": program_view,
+        }
+
+    async def _build_risk_review(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        blocked_tasks = await self._list_blocked_tasks(project_id=project_id)
+        signal_sources = await self._list_pending_signal_sources(project_id=project_id)
+        notifications = await self._list_unread_risk_notifications(project_id=project_id)
+        return {
+            "summary": {
+                "cadence_type": "risk_review",
+                "blocked_task_count": len(blocked_tasks),
+                "pending_signal_source_count": len(signal_sources),
+                "unread_risk_notification_count": len(notifications),
+            },
+            "blocked_tasks": blocked_tasks[:20],
+            "pending_signal_sources": signal_sources[:20],
+            "risk_notifications": notifications[:20],
+        }
+
+    async def _build_milestone_check(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        overdue_tasks = await self._list_overdue_tasks(project_id=project_id)
+        work_items = await self._count_work_items(state="waiting_manual", project_id=project_id)
+        stmt = select(Project.id, Project.title, Project.status).where(Project.tenant_id == self.tenant_id)
+        if project_id is not None:
+            stmt = stmt.where(Project.id == int(project_id))
+        projects = (await self.db.execute(stmt)).all()
+        return {
+            "summary": {
+                "cadence_type": "milestone_check",
+                "project_count": len(projects),
+                "overdue_task_count": len(overdue_tasks),
+                "pending_manual_count": work_items,
+            },
+            "projects": [
+                {"project_id": int(pid), "title": str(title or ""), "status": str(status or "")}
+                for pid, title, status in projects
+            ],
+            "overdue_tasks": overdue_tasks[:20],
+        }
+
+    async def _list_recent_work_items(self, *, since: datetime, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(WorkItem).where(
+            and_(
+                WorkItem.tenant_id == self.tenant_id,
+                WorkItem.created_at >= since,
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(WorkItem.project_id == int(project_id))
+        rows = (await self.db.execute(stmt.order_by(WorkItem.created_at.desc()).limit(50))).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "state": row.state,
+                "intent": row.intent,
+                "project_id": row.project_id,
+                "signal_id": row.signal_id,
+                "created_at": row.created_at.isoformat() if row.created_at else None,
+            }
+            for row in rows
+        ]
+
+    async def _count_work_items(self, *, state: str, project_id: Optional[int]) -> int:
+        stmt = select(func.count(WorkItem.id)).where(
+            and_(
+                WorkItem.tenant_id == self.tenant_id,
+                WorkItem.state == state,
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(WorkItem.project_id == int(project_id))
+        return int((await self.db.execute(stmt)).scalar_one() or 0)
+
+    async def _list_blocked_tasks(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(Task).where(
+            and_(
+                Task.tenant_id == self.tenant_id,
+                Task.status == "blocked",
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(Task.project_id == int(project_id))
+        rows = (await self.db.execute(stmt.order_by(Task.updated_at.desc()).limit(50))).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "project_id": int(row.project_id),
+                "title": str(row.title or ""),
+                "priority": str(row.priority or ""),
+            }
+            for row in rows
+        ]
+
+    async def _list_overdue_tasks(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        now = _utc_now_naive()
+        stmt = select(Task).where(
+            and_(
+                Task.tenant_id == self.tenant_id,
+                Task.due_date.is_not(None),
+                Task.due_date < now,
+                Task.status.not_in(["done", "completed", "closed"]),
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(Task.project_id == int(project_id))
+        rows = (await self.db.execute(stmt.order_by(Task.due_date.asc()).limit(50))).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "project_id": int(row.project_id),
+                "title": str(row.title or ""),
+                "due_date": row.due_date.isoformat() if row.due_date else None,
+            }
+            for row in rows
+        ]
+
+    async def _list_recent_signals(self, *, since: datetime, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(Signal).where(
+            and_(
+                Signal.tenant_id == self.tenant_id,
+                Signal.created_at >= since,
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(
+                or_(
+                    Signal.matched_project_id == int(project_id),
+                    Signal.id.in_(
+                        select(ProjectSignalLink.signal_id).where(
+                            and_(
+                                ProjectSignalLink.tenant_id == self.tenant_id,
+                                ProjectSignalLink.project_id == int(project_id),
+                            )
+                        )
+                    ),
+                )
+            )
+        rows = (await self.db.execute(stmt.order_by(Signal.created_at.desc()).limit(50))).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "title": str(row.title or ""),
+                "status": str(row.status or ""),
+                "signal_type": str(row.signal_type or ""),
+            }
+            for row in rows
+        ]
+
+    async def _list_budget_hotspots(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(Budget).where(Budget.tenant_id == self.tenant_id)
+        if project_id is not None:
+            stmt = stmt.where(Budget.project_id == int(project_id))
+        rows = (await self.db.execute(stmt)).scalars().all()
+        hotspots = []
+        for row in rows:
+            amount = float(row.budget_amount_usd or 0.0)
+            usage = float(row.current_usage_usd or 0.0)
+            ratio = (usage / amount) if amount > 0 else 0.0
+            if ratio >= 0.8:
+                hotspots.append(
+                    {
+                        "budget_id": int(row.id),
+                        "project_id": row.project_id,
+                        "usage_ratio": round(ratio, 4),
+                        "hard_limit_enabled": bool(row.hard_limit_enabled),
+                    }
+                )
+        return sorted(hotspots, key=lambda item: item["usage_ratio"], reverse=True)
+
+    async def _list_capacity_hotspots(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(UserWorkload).where(
+            and_(
+                UserWorkload.tenant_id == self.tenant_id,
+                UserWorkload.workload_ratio >= 0.85,
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(
+                UserWorkload.user_id.in_(
+                    select(UserProject.user_id).where(
+                        and_(
+                            UserProject.tenant_id == self.tenant_id,
+                            UserProject.project_id == int(project_id),
+                        )
+                    )
+                )
+            )
+        rows = (await self.db.execute(stmt.limit(50))).scalars().all()
+        items = [
+            {
+                "user_id": int(row.user_id),
+                "workload_ratio": float(row.workload_ratio or 0.0),
+                "current_assigned_hours": float(row.current_assigned_hours or 0.0),
+            }
+            for row in rows
+        ]
+        return items
+
+    async def _list_pending_signal_sources(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(SignalSourceRegistry).where(
+            and_(
+                SignalSourceRegistry.tenant_id == self.tenant_id,
+                SignalSourceRegistry.onboarding_status == "pending_manual",
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(SignalSourceRegistry.project_id == int(project_id))
+        rows = (await self.db.execute(stmt.order_by(SignalSourceRegistry.updated_at.desc()).limit(50))).scalars().all()
+        return [
+            {
+                "registry_id": int(row.id),
+                "project_id": int(row.project_id),
+                "source_kind": str(row.source_kind or ""),
+                "source_name": str(row.source_name or ""),
+                "status": str(row.status or ""),
+                "onboarding_status": str(row.onboarding_status or ""),
+            }
+            for row in rows
+        ]
+
+    async def _list_unread_risk_notifications(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        stmt = select(Notification).where(
+            and_(
+                Notification.tenant_id == self.tenant_id,
+                Notification.is_read.is_(False),
+                Notification.kind.in_(["risk", "alert", "action_required"]),
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(
+                and_(
+                    Notification.entity_type == "project",
+                    Notification.entity_id == int(project_id),
+                )
+            )
+        rows = (await self.db.execute(stmt.order_by(Notification.updated_at.desc()).limit(50))).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "title": str(row.title or ""),
+                "status": str(row.status or ""),
+                "priority": str(row.priority or ""),
+            }
+            for row in rows
+        ]
+
+    @staticmethod
+    def _build_focus_actions(*, pending_manual_count: int, blocked_task_count: int) -> list[str]:
+        actions: list[str] = []
+        if pending_manual_count > 0:
+            actions.append("优先清理待人工确认队列，避免自动执行链路阻塞。")
+        if blocked_task_count > 0:
+            actions.append("优先处理 blocked tasks，降低关键依赖压力。")
+        if not actions:
+            actions.append("当前节奏平稳，继续按计划推进。")
+        return actions
diff --git a/backend/app/services/ai_pmo_domain_workflow.py b/backend/app/services/ai_pmo_domain_workflow.py
new file mode 100644
index 00000000..0cbb672d
--- /dev/null
+++ b/backend/app/services/ai_pmo_domain_workflow.py
@@ -0,0 +1,108 @@
+"""AI PMO domain workflow adapters。"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from .signal_source_orchestration_service import SignalSourceOrchestrationService
+
+
+class SignalSourceDomainWorkflow:
+    """把 signal source onboarding 归一成 domain workflow 结果。"""
+
+    def __init__(self, db: AsyncSession) -> None:
+        self.db = db
+
+    async def run(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        signal_id: int,
+        initiating_text: str,
+        requested_via: str,
+        operator_user_id: Optional[int],
+    ) -> dict[str, Any]:
+        raw_result = await SignalSourceOrchestrationService(self.db).orchestrate_from_signal(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            signal_id=signal_id,
+            initiating_text=initiating_text,
+            requested_via=requested_via,
+            operator_user_id=operator_user_id,
+            dry_run=False,
+            response_mode="fast",
+        )
+        normalized = self.normalize(raw_result)
+        normalized["raw_result"] = raw_result
+        return normalized
+
+    @classmethod
+    def normalize(cls, raw_result: dict[str, Any]) -> dict[str, Any]:
+        payload = raw_result if isinstance(raw_result, dict) else {}
+        signal_source = payload.get("signal_source") if isinstance(payload.get("signal_source"), dict) else {}
+        worker_executions = payload.get("worker_executions") if isinstance(payload.get("worker_executions"), list) else []
+        typed_actions = cls._extract_typed_actions(payload.get("openclaw"))
+        material_actions = [
+            row
+            for row in typed_actions
+            if str(row.get("operation_code") or "").strip() != "notification.create"
+            and str(row.get("status") or "").strip() in {"applied", "planned"}
+        ]
+        fallback_used = any(bool(item.get("used_fallback")) for item in worker_executions if isinstance(item, dict))
+
+        raw_status = str(payload.get("status") or "").strip().lower()
+        registry_status = str(signal_source.get("status") or "").strip().lower()
+        onboarding_status = str(signal_source.get("onboarding_status") or "").strip().lower()
+        openclaw_status = str((payload.get("openclaw") or {}).get("status") or "").strip().lower()
+
+        if raw_status == "failed" or openclaw_status == "failed":
+            final_status = "failed_safe"
+        elif onboarding_status != "completed":
+            final_status = "pending_manual"
+        elif not material_actions:
+            final_status = "pending_manual"
+        elif fallback_used:
+            final_status = "pending_manual"
+        elif raw_status in {"completed", "partial", "planned"}:
+            final_status = "completed"
+        else:
+            final_status = "pending_manual"
+
+        next_step = "等待人工确认" if final_status != "completed" else "继续自动执行后续任务"
+        return {
+            "status": final_status,
+            "registry_id": signal_source.get("registry_id"),
+            "registry_status": registry_status,
+            "onboarding_status": onboarding_status,
+            "openclaw_status": openclaw_status,
+            "typed_actions": typed_actions,
+            "material_action_count": len(material_actions),
+            "worker_executions": worker_executions,
+            "pending_workers": payload.get("pending_workers") if isinstance(payload.get("pending_workers"), list) else [],
+            "next_step": next_step,
+            "fallback_used": fallback_used,
+        }
+
+    @staticmethod
+    def _extract_typed_actions(openclaw_payload: Any) -> list[dict[str, Any]]:
+        if not isinstance(openclaw_payload, dict):
+            return []
+        result = openclaw_payload.get("result") if isinstance(openclaw_payload.get("result"), dict) else openclaw_payload
+        execution = result.get("execution") if isinstance(result, dict) and isinstance(result.get("execution"), dict) else {}
+        rows = execution.get("results") if isinstance(execution.get("results"), list) else []
+        actions: list[dict[str, Any]] = []
+        for row in rows:
+            if not isinstance(row, dict):
+                continue
+            actions.append(
+                {
+                    "operation_code": str(row.get("operation_code") or "").strip(),
+                    "status": str(row.get("status") or "").strip(),
+                    "entity_id": row.get("entity_id"),
+                    "payload": row.get("payload") if isinstance(row.get("payload"), dict) else {},
+                }
+            )
+        return actions
diff --git a/backend/app/services/ai_pmo_kernel_service.py b/backend/app/services/ai_pmo_kernel_service.py
new file mode 100644
index 00000000..2e4f22e1
--- /dev/null
+++ b/backend/app/services/ai_pmo_kernel_service.py
@@ -0,0 +1,2300 @@
+"""AI PMO 决策中枢内核服务。"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import re
+import time
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+from sqlalchemy import and_, func, or_, select
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.logging import get_logger
+from ..models import (
+    AIProcessingLog,
+    Budget,
+    DecisionRecord,
+    IngestEvent,
+    Notification,
+    PlanBundle,
+    PortfolioSnapshot,
+    Project,
+    ProjectDomain,
+    ProjectSignalLink,
+    ReplyReceipt,
+    Signal,
+    Task,
+    TruthNode,
+    User,
+    UserProject,
+    UserSkillBinding,
+    UserWorkload,
+    WorkItem,
+)
+from .adaptive_rag_service import AdaptiveRAGService
+from .ai_pmo_cadence_service import AIPMOCadenceService
+from .ai_steward import AIStewardService
+from .ai_pmo_review_service import AIPMOReviewService
+from .notification_service import NotificationService
+from .openclaw_commander_service import OpenClawCommanderService
+from .ai_pmo_domain_workflow import SignalSourceDomainWorkflow
+from .signal_source_orchestration_service import SignalSourceOrchestrationService
+
+logger = get_logger(__name__)
+
+INTENT_QUERY = "query.read_only"
+INTENT_PROJECT_CREATE = "project.create"
+INTENT_SIGNAL_CREATE = "signal.create"
+INTENT_SIGNAL_SOURCE_ONBOARD = "signal_source.onboard"
+INTENT_FOLLOWUP_CONFIRM = "followup.confirm"
+INTENT_FOLLOWUP_CONTINUE = "followup.continue"
+INTENT_IGNORE = "ignore"
+FOLLOWUP_TARGET_INTENTS = {
+    INTENT_PROJECT_CREATE,
+    INTENT_SIGNAL_CREATE,
+    INTENT_SIGNAL_SOURCE_ONBOARD,
+}
+
+STATE_RECEIVED = "received"
+STATE_MODELED = "modeled"
+STATE_DECIDED = "decided"
+STATE_PLANNED = "planned"
+STATE_EXECUTING = "executing"
+STATE_WAITING_MANUAL = "waiting_manual"
+STATE_SUCCEEDED = "succeeded"
+STATE_FAILED = "failed"
+STATE_IGNORED = "ignored"
+
+RECEIPT_PHASE_ACTION = "action_applied"
+RECEIPT_PHASE_PENDING = "pending_manual"
+RECEIPT_PHASE_ERROR = "error"
+
+QUESTION_HINTS = ("?", "？", "怎么", "为何", "为什么", "什么", "多少", "是否", "帮我看", "状态")
+SMALLTALK_EXACT = {
+    "收到",
+    "收到啦",
+    "好的",
+    "ok",
+    "okk",
+    "yes",
+    "no",
+    "谢谢",
+    "thanks",
+    "thx",
+    "在吗",
+    "在不在",
+}
+TRANSPORT_MODE_CORRECTION_HINTS = ("不是", "不再", "改为", "改成", "替换为", "取代", "现在都", "现在是")
+FOLLOWUP_CONFIRM_HINTS = ("确认归属", "确认项目", "确认到项目", "确认关联", "确认这条")
+FOLLOWUP_CONTINUE_HINTS = ("继续处理", "继续执行", "接着处理", "往下做", "继续推进")
+FOLLOWUP_WORK_ITEM_ID_RE = re.compile(r"(?:work[_ -]?item|工作项|队列项)\s*(?:id|ID)?\s*[:#：]?\s*(\d+)", re.IGNORECASE)
+SIGNAL_CREATE_HINTS = ("新增一个信号", "新增一条信号", "增加一个信号", "录入一个信号", "关注", "跟踪")
+PROJECT_CREATE_HINTS = ("创建项目", "新建项目", "建立项目", "做个项目", "立项", "开个项目")
+POLICY_GATE_RULES: tuple[tuple[str, tuple[str, ...]], ...] = (
+    ("delete_or_archive", ("删除项目", "删除信号", "归档项目", "归档", "archive", "清空项目")),
+    ("budget_overrun", ("超预算", "突破预算", "额外预算", "加预算")),
+    ("cross_project_staffing", ("跨项目抢人", "抽调", "抢占关键人力", "挪人")),
+    ("charter_rewrite", ("改写charter", "success criteria", "成功标准改", "范围重大变更")),
+    ("critical_path_change", ("关键路径", "里程碑改动", "影响多个项目", "跨项目依赖调整")),
+)
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+def _safe_int(value: Any) -> Optional[int]:
+    try:
+        if value is None or value == "":
+            return None
+        return int(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def _normalize_workload_ratio(value: Any) -> float:
+    try:
+        ratio = float(value or 0.0)
+    except (TypeError, ValueError):
+        return 0.0
+    if ratio <= 0:
+        return 0.0
+    return ratio / 100.0 if ratio > 1.0 else ratio
+
+
+@dataclass(slots=True)
+class CanonicalIntent:
+    name: str
+    confidence: float
+    project_id: Optional[int]
+    signal_title: str
+    signal_summary: str
+    metadata: dict[str, Any]
+
+
+class IntentResolver:
+    """把输入收口为 canonical intent。"""
+
+    @staticmethod
+    def resolve(*, text: str, project_id: Optional[int]) -> CanonicalIntent:
+        normalized = re.sub(r"\s+", "", str(text or "")).lower()
+        pretty = re.sub(r"\s+", " ", str(text or "")).strip()
+        if not pretty:
+            return CanonicalIntent(INTENT_IGNORE, 1.0, project_id, "", "", {"reason": "empty"})
+        if normalized in SMALLTALK_EXACT:
+            return CanonicalIntent(INTENT_IGNORE, 0.95, project_id, "", "", {"reason": "smalltalk"})
+        if any(token in pretty for token in FOLLOWUP_CONFIRM_HINTS):
+            return CanonicalIntent(INTENT_FOLLOWUP_CONFIRM, 0.82, project_id, "", "", {})
+        if any(token in pretty for token in FOLLOWUP_CONTINUE_HINTS):
+            return CanonicalIntent(INTENT_FOLLOWUP_CONTINUE, 0.82, project_id, "", "", {})
+        if SignalSourceOrchestrationService.looks_like_signal_source_request(pretty):
+            return CanonicalIntent(
+                INTENT_SIGNAL_SOURCE_ONBOARD,
+                0.92,
+                project_id,
+                IntentResolver._build_signal_title(pretty, default="新增信号源接入请求"),
+                pretty[:300],
+                {"source_request": True},
+            )
+        if any(token in pretty for token in PROJECT_CREATE_HINTS) or OpenClawCommanderService._extract_project_create_title(pretty):
+            return CanonicalIntent(
+                INTENT_PROJECT_CREATE,
+                0.9,
+                project_id,
+                IntentResolver._build_signal_title(pretty, default="新项目创建请求"),
+                pretty[:300],
+                {},
+            )
+        if any(token in pretty for token in QUESTION_HINTS):
+            return CanonicalIntent(
+                INTENT_QUERY,
+                0.88,
+                project_id,
+                "",
+                pretty[:300],
+                {},
+            )
+        if any(token in pretty for token in SIGNAL_CREATE_HINTS) or any(token in pretty for token in ("新增", "增加", "录入", "跟踪", "关注")):
+            return CanonicalIntent(
+                INTENT_SIGNAL_CREATE,
+                0.78,
+                project_id,
+                IntentResolver._build_signal_title(pretty, default="新信号"),
+                pretty[:300],
+                {},
+            )
+        return CanonicalIntent(
+            INTENT_SIGNAL_CREATE,
+            0.65,
+            project_id,
+            IntentResolver._build_signal_title(pretty, default="新信号"),
+            pretty[:300],
+            {"fallback": True},
+        )
+
+    @staticmethod
+    def _build_signal_title(text: str, *, default: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip(" ，,:：")
+        if not cleaned:
+            return default
+        for token in ("。", "；", ";", "!", "！", "?", "？"):
+            if token in cleaned:
+                cleaned = cleaned.split(token, 1)[0].strip()
+                break
+        return cleaned[:120] or default
+
+
+class StrategicBrain:
+    """组合级判断。"""
+
+    @staticmethod
+    def evaluate(*, intent: CanonicalIntent, world_model: dict[str, Any], message_text: str) -> dict[str, Any]:
+        project = world_model.get("project") if isinstance(world_model.get("project"), dict) else None
+        budget = world_model.get("budget") if isinstance(world_model.get("budget"), dict) else {}
+        capacity = world_model.get("capacity") if isinstance(world_model.get("capacity"), list) else []
+        risk_count = int((world_model.get("risk_summary") or {}).get("open_risk_count") or 0)
+        active_projects = int((world_model.get("portfolio") or {}).get("active_project_count") or 0)
+
+        budget_ratio = float(budget.get("usage_ratio") or 0.0)
+        overloaded_people = [row for row in capacity if _normalize_workload_ratio(row.get("workload_ratio")) >= 0.85]
+        risk_score = min(1.0, 0.18 + risk_count * 0.08 + budget_ratio * 0.25 + len(overloaded_people) * 0.1)
+        if intent.name == INTENT_PROJECT_CREATE:
+            risk_score = max(risk_score, 0.62)
+        if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD:
+            risk_score = max(risk_score, 0.58)
+
+        return {
+            "decision_level": "project" if project else "portfolio",
+            "decision_object_type": "project" if project else "portfolio",
+            "decision_object_id": _safe_int((project or {}).get("id")),
+            "risk_score": round(risk_score, 3),
+            "budget_impact": {
+                "budget_id": budget.get("budget_id"),
+                "usage_ratio": budget_ratio,
+                "hard_limit_enabled": bool(budget.get("hard_limit_enabled")),
+            },
+            "resource_impact": {
+                "overloaded_user_ids": [row.get("user_id") for row in overloaded_people],
+                "active_project_count": active_projects,
+            },
+            "rationale": (
+                f"intent={intent.name}; "
+                f"project_scope={'project' if project else 'portfolio'}; "
+                f"budget_ratio={budget_ratio:.2f}; "
+                f"open_risks={risk_count}; "
+                f"overloaded={len(overloaded_people)}"
+            ),
+            "message_excerpt": message_text[:240],
+        }
+
+
+class TacticalBrain:
+    """周/日执行计划。"""
+
+    @staticmethod
+    def plan(*, intent: CanonicalIntent, world_model: dict[str, Any], decision: dict[str, Any]) -> tuple[dict[str, Any], dict[str, Any]]:
+        project = world_model.get("project") if isinstance(world_model.get("project"), dict) else None
+        project_ref = f"项目#{project['id']}" if project and project.get("id") is not None else "项目群"
+        strategic_plan = {
+            "priority": "high" if float(decision.get("risk_score") or 0.0) >= 0.6 else "medium",
+            "objectives": [
+                f"围绕 {project_ref} 控制风险与节奏",
+                "优先保证关键路径和资源边界不被破坏",
+            ],
+            "signals": [
+                f"budget_usage_ratio={float((decision.get('budget_impact') or {}).get('usage_ratio') or 0.0):.2f}",
+                f"overloaded_users={len((decision.get('resource_impact') or {}).get('overloaded_user_ids') or [])}",
+            ],
+        }
+        execution_actions = []
+        if intent.name == INTENT_QUERY:
+            execution_actions.append({"operation_code": "query.answer", "mode": "read_only"})
+        elif intent.name == INTENT_PROJECT_CREATE:
+            execution_actions.append({"operation_code": "project.create", "mode": "write"})
+        elif intent.name == INTENT_SIGNAL_SOURCE_ONBOARD:
+            execution_actions.append({"operation_code": "signal.enter", "mode": "write"})
+            execution_actions.append({"operation_code": "signal_source.onboard", "mode": "write"})
+        elif intent.name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+            execution_actions.append({"operation_code": "work_item.resume", "mode": "write"})
+        elif intent.name == INTENT_SIGNAL_CREATE:
+            execution_actions.append({"operation_code": "signal.enter", "mode": "write"})
+
+        execution_plan = {
+            "intent": intent.name,
+            "project_id": intent.project_id,
+            "actions": execution_actions,
+            "summary": f"执行 {intent.name}，共 {len(execution_actions)} 个 typed actions",
+        }
+        return strategic_plan, execution_plan
+
+
+class SimulationEngine:
+    """执行前仿真。"""
+
+    BUDGET_COST_BY_OPERATION = {
+        "query.answer": 1.0,
+        "project.create": 120.0,
+        "signal.enter": 8.0,
+        "signal_source.onboard": 40.0,
+    }
+    HOURS_BY_OPERATION = {
+        "query.answer": 0.2,
+        "project.create": 12.0,
+        "signal.enter": 1.5,
+        "signal_source.onboard": 6.0,
+    }
+
+    @classmethod
+    def simulate(cls, *, intent: CanonicalIntent, world_model: dict[str, Any], execution_plan: dict[str, Any]) -> dict[str, Any]:
+        actions = execution_plan.get("actions") if isinstance(execution_plan.get("actions"), list) else []
+        operation_codes = [str(item.get("operation_code") or "").strip() for item in actions if isinstance(item, dict)]
+        estimated_budget_delta = sum(cls.BUDGET_COST_BY_OPERATION.get(code, 0.0) for code in operation_codes)
+        estimated_capacity_delta = sum(cls.HOURS_BY_OPERATION.get(code, 0.0) for code in operation_codes)
+
+        budget = world_model.get("budget") if isinstance(world_model.get("budget"), dict) else {}
+        budget_amount = float(budget.get("budget_amount_usd") or 0.0)
+        current_ratio = float(budget.get("usage_ratio") or 0.0)
+        projected_budget_ratio = current_ratio + (estimated_budget_delta / budget_amount if budget_amount > 0 else 0.0)
+
+        capacity = world_model.get("capacity") if isinstance(world_model.get("capacity"), list) else []
+        current_peak_ratio = max([_normalize_workload_ratio(item.get("workload_ratio")) for item in capacity], default=0.0)
+        project_member_count = max(1, int(world_model.get("member_count") or len(capacity) or 1))
+        projected_capacity_peak_ratio = current_peak_ratio + (estimated_capacity_delta / (40.0 * project_member_count))
+
+        task_summary = world_model.get("task_summary") if isinstance(world_model.get("task_summary"), dict) else {}
+        blocked_count = int((task_summary.get("by_status") or {}).get("blocked") or 0)
+        overdue_count = int(task_summary.get("overdue_count") or 0)
+        missing_skill_tag_ids = task_summary.get("uncovered_skill_tag_ids") if isinstance(task_summary.get("uncovered_skill_tag_ids"), list) else []
+
+        charter = world_model.get("charter") if isinstance(world_model.get("charter"), list) else []
+        dependency_pressure = "high" if blocked_count > 0 or overdue_count > 0 else ("medium" if not charter else "low")
+        projected_risk_score = min(
+            1.0,
+            0.12
+            + projected_budget_ratio * 0.24
+            + projected_capacity_peak_ratio * 0.28
+            + blocked_count * 0.08
+            + overdue_count * 0.05
+            + len(missing_skill_tag_ids) * 0.07,
+        )
+
+        return {
+            "intent": intent.name,
+            "operation_codes": operation_codes,
+            "estimated_budget_delta_usd": round(estimated_budget_delta, 2),
+            "projected_budget_usage_ratio": round(projected_budget_ratio, 4),
+            "estimated_capacity_delta_hours": round(estimated_capacity_delta, 2),
+            "projected_capacity_peak_ratio": round(projected_capacity_peak_ratio, 4),
+            "blocked_task_count": blocked_count,
+            "overdue_task_count": overdue_count,
+            "dependency_pressure": dependency_pressure,
+            "missing_skill_tag_ids": missing_skill_tag_ids[:20],
+            "missing_skill_count": len(missing_skill_tag_ids),
+            "projected_risk_score": round(projected_risk_score, 4),
+        }
+
+
+class ExecutionPolicyBrain:
+    """自动执行边界。"""
+
+    @staticmethod
+    def evaluate(
+        *,
+        message_text: str,
+        world_model: dict[str, Any],
+        intent: CanonicalIntent,
+        simulation: dict[str, Any],
+    ) -> dict[str, Any]:
+        normalized = re.sub(r"\s+", "", str(message_text or "")).lower()
+        for gate_code, tokens in POLICY_GATE_RULES:
+            if any(token.lower().replace(" ", "") in normalized for token in tokens):
+                return {
+                    "auto_execute": False,
+                    "policy_gate_hit": True,
+                    "policy_gate_code": gate_code,
+                    "approval_requirement": {
+                        "required": True,
+                        "reason": gate_code,
+                        "message": "命中硬策略边界，必须进入显式审批/人工确认。",
+                    },
+                }
+
+        if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD and intent.project_id is None:
+            return {
+                "auto_execute": False,
+                "policy_gate_hit": True,
+                "policy_gate_code": "missing_project_scope",
+                "approval_requirement": {
+                    "required": True,
+                    "reason": "missing_project_scope",
+                    "message": "信号源接入缺少项目归属，必须先确认项目。",
+                },
+            }
+
+        # 轻量级事实写入/只读查询不应被项目当前依赖或产能压力阻塞。
+        # 否则“AI 雷达”会在最需要先收集事实时失去自动化能力。
+        if intent.name in {INTENT_QUERY, INTENT_SIGNAL_CREATE}:
+            return {
+                "auto_execute": True,
+                "policy_gate_hit": False,
+                "policy_gate_code": None,
+                "approval_requirement": {"required": False},
+            }
+
+        budget = world_model.get("budget") if isinstance(world_model.get("budget"), dict) else {}
+        if bool(budget.get("hard_limit_enabled")) and float(budget.get("usage_ratio") or 0.0) >= 1.0:
+            return {
+                "auto_execute": False,
+                "policy_gate_hit": True,
+                "policy_gate_code": "budget_hard_limit",
+                "approval_requirement": {
+                    "required": True,
+                    "reason": "budget_hard_limit",
+                    "message": "项目预算已触达硬限制，自动执行被拦截。",
+                },
+            }
+
+        if float(simulation.get("projected_budget_usage_ratio") or 0.0) >= 1.0:
+            return {
+                "auto_execute": False,
+                "policy_gate_hit": True,
+                "policy_gate_code": "simulation_budget_limit",
+                "approval_requirement": {
+                    "required": True,
+                    "reason": "simulation_budget_limit",
+                    "message": "执行前仿真显示预算将触顶，自动执行被拦截。",
+                },
+            }
+
+        if float(simulation.get("projected_capacity_peak_ratio") or 0.0) >= 1.0:
+            return {
+                "auto_execute": False,
+                "policy_gate_hit": True,
+                "policy_gate_code": "simulation_capacity_limit",
+                "approval_requirement": {
+                    "required": True,
+                    "reason": "simulation_capacity_limit",
+                    "message": "执行前仿真显示产能会超载，自动执行被拦截。",
+                },
+            }
+
+        if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD and int(simulation.get("missing_skill_count") or 0) > 0:
+            return {
+                "auto_execute": False,
+                "policy_gate_hit": True,
+                "policy_gate_code": "simulation_skill_gap",
+                "approval_requirement": {
+                    "required": True,
+                    "reason": "simulation_skill_gap",
+                    "message": "执行前仿真显示当前项目存在技能缺口，信号源接入先进入待确认。",
+                },
+            }
+
+        if intent.name in {INTENT_SIGNAL_SOURCE_ONBOARD, INTENT_SIGNAL_CREATE} and str(simulation.get("dependency_pressure") or "") == "high":
+            return {
+                "auto_execute": False,
+                "policy_gate_hit": True,
+                "policy_gate_code": "simulation_dependency_pressure",
+                "approval_requirement": {
+                    "required": True,
+                    "reason": "simulation_dependency_pressure",
+                    "message": "执行前仿真显示关键依赖压力过高，当前先进入待人工确认。",
+                },
+            }
+
+        return {
+            "auto_execute": True,
+            "policy_gate_hit": False,
+            "policy_gate_code": None,
+            "approval_requirement": {"required": False},
+        }
+
+
+class DecisionEngine:
+    """汇总三脑输出。"""
+
+    def __init__(self) -> None:
+        self.strategic_brain = StrategicBrain()
+        self.tactical_brain = TacticalBrain()
+        self.simulation_engine = SimulationEngine()
+        self.execution_policy_brain = ExecutionPolicyBrain()
+
+    def decide(self, *, intent: CanonicalIntent, world_model: dict[str, Any], message_text: str) -> dict[str, Any]:
+        strategic = self.strategic_brain.evaluate(intent=intent, world_model=world_model, message_text=message_text)
+        strategic_plan, execution_plan = self.tactical_brain.plan(intent=intent, world_model=world_model, decision=strategic)
+        simulation = self.simulation_engine.simulate(intent=intent, world_model=world_model, execution_plan=execution_plan)
+        strategic_plan["simulation"] = simulation
+        execution_plan["simulation"] = simulation
+        policy = self.execution_policy_brain.evaluate(
+            message_text=message_text,
+            world_model=world_model,
+            intent=intent,
+            simulation=simulation,
+        )
+        strategic["risk_score"] = max(float(strategic.get("risk_score") or 0.0), float(simulation.get("projected_risk_score") or 0.0))
+        return {
+            "strategic": strategic,
+            "strategic_plan": strategic_plan,
+            "execution_plan": execution_plan,
+            "simulation": simulation,
+            "policy": policy,
+        }
+
+
+class ReplyRenderer:
+    """唯一回执出口。"""
+
+    @staticmethod
+    def render(
+        *,
+        work_state: str,
+        intent: CanonicalIntent,
+        decision_record: dict[str, Any],
+        execution_result: dict[str, Any],
+    ) -> dict[str, Any]:
+        if work_state == STATE_WAITING_MANUAL:
+            if intent.name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+                followup_text = str(execution_result.get("followup_reply") or "").strip()
+                if followup_text:
+                    return {
+                        "receipt_phase": RECEIPT_PHASE_PENDING,
+                        "template_key": "pending_manual",
+                        "payload": {"intent": intent.name, "reason": followup_text},
+                        "text": followup_text,
+                    }
+            gate_message = str(((decision_record.get("approval_requirement") or {}).get("message")) or "").strip()
+            if not gate_message:
+                gate_message = "已进入待人工确认队列。"
+            return {
+                "receipt_phase": RECEIPT_PHASE_PENDING,
+                "template_key": "pending_manual",
+                "payload": {"intent": intent.name, "reason": gate_message},
+                "text": gate_message,
+            }
+        if work_state == STATE_FAILED:
+            if intent.name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+                followup_text = str(execution_result.get("followup_reply") or "").strip()
+                if followup_text:
+                    return {
+                        "receipt_phase": RECEIPT_PHASE_ERROR,
+                        "template_key": "safe_error",
+                        "payload": {"intent": intent.name, "reason": followup_text},
+                        "text": followup_text,
+                    }
+            reason = str(execution_result.get("error_message") or "系统处理中出现异常，本次未执行。").strip()
+            return {
+                "receipt_phase": RECEIPT_PHASE_ERROR,
+                "template_key": "safe_error",
+                "payload": {"intent": intent.name, "reason": reason},
+                "text": reason,
+            }
+        if intent.name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+            followup_text = str(execution_result.get("followup_reply") or "已继续处理。").strip()
+            return {
+                "receipt_phase": RECEIPT_PHASE_ACTION,
+                "template_key": "action_succeeded",
+                "payload": {
+                    "intent": intent.name,
+                    "target_work_item_id": _safe_int(execution_result.get("target_work_item_id")),
+                    "target_state": execution_result.get("target_state"),
+                },
+                "text": followup_text,
+            }
+        if intent.name == INTENT_QUERY:
+            answer = str(execution_result.get("answer") or "当前暂无可用回答。").strip()
+            return {
+                "receipt_phase": RECEIPT_PHASE_ACTION,
+                "template_key": "action_succeeded",
+                "payload": {"intent": intent.name, "answer": answer},
+                "text": answer,
+            }
+        if intent.name == INTENT_PROJECT_CREATE:
+            project = execution_result.get("project") if isinstance(execution_result.get("project"), dict) else {}
+            project_id = _safe_int(project.get("id"))
+            project_title = str(project.get("title") or project.get("name") or "新项目").strip()
+            return {
+                "receipt_phase": RECEIPT_PHASE_ACTION,
+                "template_key": "action_succeeded",
+                "payload": {"intent": intent.name, "project_id": project_id, "project_title": project_title},
+                "text": f"已立项：项目#{project_id}《{project_title}》已创建完成。" if project_id else f"已立项：《{project_title}》已创建完成。",
+            }
+        if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD:
+            signal_id = _safe_int(execution_result.get("signal_id"))
+            workflow = execution_result.get("signal_source_workflow") if isinstance(execution_result.get("signal_source_workflow"), dict) else {}
+            status = str(workflow.get("status") or (execution_result.get("signal_source_orchestration") or {}).get("status") or "").strip()
+            if status == "pending_manual":
+                return {
+                    "receipt_phase": RECEIPT_PHASE_PENDING,
+                    "template_key": "pending_manual",
+                    "payload": {"intent": intent.name, "signal_id": signal_id, "status": status},
+                    "text": f"已记录信号源接入请求，当前为待人工确认。信号#{signal_id} 已入库。" if signal_id else "已记录信号源接入请求，当前为待人工确认。",
+                }
+            return {
+                "receipt_phase": RECEIPT_PHASE_ACTION,
+                "template_key": "action_succeeded",
+                "payload": {"intent": intent.name, "signal_id": signal_id, "status": status},
+                "text": f"已接收入库并启动信号源接入流程。信号#{signal_id}，状态：{status or 'running'}。",
+            }
+        signal_id = _safe_int(execution_result.get("signal_id"))
+        project_id = _safe_int(execution_result.get("project_id"))
+        if signal_id is not None and project_id is not None:
+            text = f"已入库：已为项目#{project_id}创建信号#{signal_id}。"
+        elif signal_id is not None:
+            text = f"已入库：已创建信号#{signal_id}。"
+        else:
+            text = "已收到并入库。"
+        return {
+            "receipt_phase": RECEIPT_PHASE_ACTION,
+            "template_key": "action_succeeded",
+            "payload": {"intent": intent.name, "signal_id": signal_id, "project_id": project_id},
+            "text": text,
+        }
+
+
+class TypedExecutor:
+    """执行层只做 typed action 落地。"""
+
+    def __init__(self, db: AsyncSession, tenant_id: int) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+        self.signal_source_workflow = SignalSourceDomainWorkflow(db)
+
+    async def execute(
+        self,
+        *,
+        intent: CanonicalIntent,
+        message_text: str,
+        sender: str,
+        operator_user_id: Optional[int],
+        source_system: str,
+        source_message_id: str,
+    ) -> dict[str, Any]:
+        if intent.name == INTENT_QUERY:
+            rag_service = AdaptiveRAGService(self.db, tenant_id=self.tenant_id)
+            rag_result = await rag_service.answer_question(
+                question=message_text,
+                asked_by=operator_user_id,
+                asked_via=source_system,
+                project_id=intent.project_id,
+                source_chat_id=source_message_id,
+                conversation_key=source_message_id,
+                action_mode="read",
+                sync_requirement=False,
+            )
+            return {
+                "status": "succeeded",
+                "answer": str(rag_result.get("answer") or "").strip(),
+                "rag_result": rag_result,
+                "project_id": intent.project_id,
+            }
+
+        if intent.name == INTENT_PROJECT_CREATE:
+            if operator_user_id is None:
+                return {
+                    "status": "failed_safe",
+                    "error_message": "未找到可用项目创建人，当前无法自动创建项目。",
+                }
+            steward = AIStewardService(self.db)
+            payload = await steward.create_project_from_natural_language(
+                input_text=message_text,
+                creator_id=int(operator_user_id),
+                tenant_id=self.tenant_id,
+            )
+            project = payload.get("project") if isinstance(payload.get("project"), dict) else {}
+            signal = payload.get("signal") if isinstance(payload.get("signal"), dict) else {}
+            return {
+                "status": "succeeded",
+                "project": project,
+                "signal": signal,
+                "project_id": _safe_int(project.get("id")),
+                "signal_id": _safe_int(signal.get("id")),
+                "tasks": payload.get("tasks") if isinstance(payload.get("tasks"), list) else [],
+                "raw_payload": payload,
+            }
+
+        if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD:
+            signal = await self._create_signal(
+                title=intent.signal_title or "新增信号源接入请求",
+                summary=intent.signal_summary or message_text,
+                content=message_text,
+                project_id=intent.project_id,
+                source="PMFlow",
+                signal_type="新需求",
+                recommended_action="enter_only",
+            )
+            domain_workflow = await self.signal_source_workflow.run(
+                tenant_id=self.tenant_id,
+                project_id=int(intent.project_id or 0),
+                signal_id=int(signal.id),
+                initiating_text=message_text,
+                requested_via=source_system,
+                operator_user_id=operator_user_id,
+            )
+            orchestration = domain_workflow.get("raw_result") if isinstance(domain_workflow.get("raw_result"), dict) else {}
+            return {
+                "status": "succeeded" if str(domain_workflow.get("status") or "").lower() == "completed" else (
+                    "pending_manual" if str(domain_workflow.get("status") or "").lower() == "pending_manual" else "failed_safe"
+                ),
+                "signal_id": int(signal.id),
+                "project_id": intent.project_id,
+                "signal_source_workflow": domain_workflow,
+                "signal_source_orchestration": orchestration,
+                "auto_tasks": self._extract_auto_tasks(orchestration.get("openclaw")),
+            }
+
+        if intent.name == INTENT_SIGNAL_CREATE:
+            signal = await self._create_signal(
+                title=intent.signal_title or "新信号",
+                summary=intent.signal_summary or message_text,
+                content=message_text,
+                project_id=intent.project_id,
+                source="PMFlow",
+                signal_type="新需求",
+                recommended_action="enter_only",
+            )
+            return {
+                "status": "succeeded",
+                "signal_id": int(signal.id),
+                "project_id": intent.project_id,
+                "auto_tasks": [],
+            }
+
+        return {
+            "status": "failed_safe",
+            "error_message": "当前意图尚未迁移到 AI PMO kernel，已保留给兼容链路处理。",
+        }
+
+    async def _create_signal(
+        self,
+        *,
+        title: str,
+        summary: str,
+        content: str,
+        project_id: Optional[int],
+        source: str,
+        signal_type: str,
+        recommended_action: str,
+    ) -> Signal:
+        signal = Signal(
+            tenant_id=self.tenant_id,
+            source=source,
+            title=(title or "新信号")[:120],
+            summary=(summary or content or "")[:300],
+            content=(content or summary or "")[:2000],
+            signal_type=signal_type,
+            reasoning=json.dumps(["由 AI PMO kernel 统一入库。"], ensure_ascii=False),
+            confidence=0.78,
+            recommended_action=recommended_action,
+            matched_project_id=int(project_id) if project_id is not None else None,
+            impact_score=0.7,
+            priority="medium",
+            processed=False,
+            tasks_generated=json.dumps([], ensure_ascii=False),
+            status="待处理",
+        )
+        self.db.add(signal)
+        await self.db.flush()
+        if project_id is not None:
+            exists = (
+                await self.db.execute(
+                    select(ProjectSignalLink.id).where(
+                        and_(
+                            ProjectSignalLink.tenant_id == self.tenant_id,
+                            ProjectSignalLink.project_id == int(project_id),
+                            ProjectSignalLink.signal_id == int(signal.id),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+            if exists is None:
+                self.db.add(
+                    ProjectSignalLink(
+                        tenant_id=self.tenant_id,
+                        project_id=int(project_id),
+                        signal_id=int(signal.id),
+                        link_type="related",
+                        relevance_score=1.0,
+                    )
+                )
+                await self.db.flush()
+        return signal
+
+    @staticmethod
+    def _extract_auto_tasks(openclaw_payload: Any) -> list[dict[str, Any]]:
+        if not isinstance(openclaw_payload, dict):
+            return []
+        result = openclaw_payload.get("result") if isinstance(openclaw_payload.get("result"), dict) else openclaw_payload
+        execution = result.get("execution") if isinstance(result, dict) and isinstance(result.get("execution"), dict) else {}
+        rows = execution.get("results") if isinstance(execution.get("results"), list) else []
+        items: list[dict[str, Any]] = []
+        for row in rows:
+            if not isinstance(row, dict):
+                continue
+            operation_code = str(row.get("operation_code") or "").strip()
+            if operation_code != "task.create":
+                continue
+            payload = row.get("payload") if isinstance(row.get("payload"), dict) else {}
+            items.append(
+                {
+                    "operation_code": operation_code,
+                    "status": str(row.get("status") or "").strip(),
+                    "entity_id": row.get("entity_id"),
+                    "title": payload.get("title"),
+                    "task_id": payload.get("task_id"),
+                }
+            )
+        return items
+
+
+class AIPMOKernelService:
+    """AI PMO kernel，作为新的统一事实链。"""
+
+    def __init__(self, db: AsyncSession, tenant_id: int = 1) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+        self.intent_resolver = IntentResolver()
+        self.decision_engine = DecisionEngine()
+        self.reply_renderer = ReplyRenderer()
+        self.executor = TypedExecutor(db, tenant_id=self.tenant_id)
+        self.review_service = AIPMOReviewService(db, tenant_id=self.tenant_id)
+        self.cadence_service = AIPMOCadenceService(db, tenant_id=self.tenant_id)
+
+    async def process_feishu_message(
+        self,
+        *,
+        group_chat_id: str,
+        message_content: str,
+        sender: str,
+        source_message_id: Optional[str],
+        source_sender_open_id: Optional[str] = None,
+        captured_at: Optional[datetime] = None,
+        chat_type: Optional[str] = None,
+    ) -> dict[str, Any]:
+        return await self.process_ingress(
+            source_system="feishu",
+            source_message_id=source_message_id,
+            message_text=message_content,
+            sender=sender,
+            channel_ref=group_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            captured_at=captured_at,
+            chat_type=chat_type,
+        )
+
+    async def process_ingress(
+        self,
+        *,
+        source_system: str,
+        source_message_id: Optional[str],
+        message_text: str,
+        sender: str,
+        channel_ref: Optional[str] = None,
+        source_sender_open_id: Optional[str] = None,
+        captured_at: Optional[datetime] = None,
+        chat_type: Optional[str] = None,
+        project_hint_id: Optional[int] = None,
+    ) -> dict[str, Any]:
+        started = time.perf_counter()
+        clean_text = self._normalize_text(message_text)
+        if not clean_text:
+            return {"handled": True, "status": "ignored", "classification": "empty", "direct_reply_required": False}
+        transport_mode_reply = self._build_transport_mode_reply(clean_text)
+        if transport_mode_reply:
+            return {
+                "handled": True,
+                "status": "answered",
+                "classification": "question",
+                "linked_project_ids": [int(project_hint_id)] if project_hint_id is not None else [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": None,
+                "results": [],
+                "intent": INTENT_QUERY,
+                "direct_reply_required": True,
+                "direct_reply_text": transport_mode_reply,
+                "reply_template_key": "transport_mode_ack",
+                "work_item_id": None,
+                "decision_level": None,
+                "review_summary": None,
+            }
+
+        synthetic_message_id = self._build_source_message_id(
+            source_system=source_system,
+            source_message_id=source_message_id,
+            channel_ref=channel_ref,
+            sender=sender,
+            message_text=clean_text,
+        )
+
+        ingest_event, created = await self._claim_ingress_event(
+            source_system=source_system,
+            source_message_id=synthetic_message_id,
+            payload={
+                "sender": sender,
+                "source_sender_open_id": source_sender_open_id,
+                "channel_ref": channel_ref,
+                "chat_type": chat_type,
+                "captured_at": captured_at.isoformat() if isinstance(captured_at, datetime) else None,
+                "message_text": clean_text,
+            },
+        )
+        if not created:
+            return {
+                "handled": True,
+                "status": "ignored",
+                "reason": "duplicate_filtered",
+                "classification": "duplicate",
+                "direct_reply_required": False,
+                "source_message_id": synthetic_message_id,
+            }
+
+        operator = await self._resolve_operator_user(sender=sender, sender_open_id=source_sender_open_id)
+        work_item = await self._create_work_item(
+            ingest_event=ingest_event,
+            source_system=source_system,
+            source_message_id=synthetic_message_id,
+            owner_user_id=_safe_int(getattr(operator, "id", None)),
+        )
+
+        try:
+            world_model = await self._build_world_model(
+                message_text=clean_text,
+                sender=sender,
+                project_hint_id=project_hint_id,
+            )
+            work_item.state = STATE_MODELED
+            work_item.context_snapshot = world_model
+            intent = self.intent_resolver.resolve(
+                text=clean_text,
+                project_id=_safe_int(world_model.get("resolved_project_id")),
+            )
+            if intent.name == INTENT_IGNORE:
+                work_item.state = STATE_IGNORED
+                work_item.intent = intent.name
+                ingest_event.status = "done"
+                ingest_event.processed_at = _utc_now_naive()
+                await self._write_processing_log(
+                    source_message_id=synthetic_message_id,
+                    work_item=work_item,
+                    message_text=clean_text,
+                    route="ai_pmo.ignore",
+                    operation_code=None,
+                    operation_result={"reason": intent.metadata.get("reason")},
+                    processing_status="success",
+                    duration_ms=int((time.perf_counter() - started) * 1000),
+                )
+                await self.db.commit()
+                return {
+                    "handled": True,
+                    "status": "ignored",
+                    "classification": "ignored",
+                    "direct_reply_required": False,
+                    "work_item_id": int(work_item.id),
+                    "ai_trace_id": f"ai-pmo-{int(work_item.id)}",
+                }
+
+            work_item.intent = intent.name
+            work_item.project_id = intent.project_id
+            decision_payload = self.decision_engine.decide(
+                intent=intent,
+                world_model=world_model,
+                message_text=clean_text,
+            )
+            work_item.state = STATE_DECIDED
+            work_item.decision_level = str((decision_payload["strategic"] or {}).get("decision_level") or "portfolio")
+
+            decision_row = await self._create_decision_record(
+                work_item=work_item,
+                world_model=world_model,
+                decision_payload=decision_payload,
+            )
+            await self._create_plan_bundle(
+                work_item_id=int(work_item.id),
+                plan_type="strategic",
+                summary="Strategic Brain 产出的项目群/项目策略。",
+                payload=decision_payload["strategic_plan"],
+            )
+            await self._create_plan_bundle(
+                work_item_id=int(work_item.id),
+                plan_type="execution",
+                summary="Execution Plan typed action bundle。",
+                payload=decision_payload["execution_plan"],
+            )
+
+            policy = decision_payload["policy"]
+            if not bool(policy.get("auto_execute")):
+                work_item.state = STATE_WAITING_MANUAL
+                work_item.action_bundle = decision_payload["execution_plan"]
+                work_item.execution_result = {
+                    "status": "waiting_manual",
+                    "reason": ((policy.get("approval_requirement") or {}).get("message") or "").strip(),
+                }
+                manual_issues = ["waiting_manual"]
+                policy_gate_code = str(policy.get("policy_gate_code") or "").strip()
+                if policy_gate_code:
+                    manual_issues.append(policy_gate_code)
+                work_item.review_summary = await self.review_service.review_work_item(
+                    work_item=work_item,
+                    intent_name=intent.name,
+                    execution_result=work_item.execution_result,
+                    world_model=world_model,
+                    verification={"verified": False, "issues": manual_issues},
+                    simulation=decision_payload["simulation"],
+                )
+                render = self.reply_renderer.render(
+                    work_state=STATE_WAITING_MANUAL,
+                    intent=intent,
+                    decision_record={
+                    "approval_requirement": policy.get("approval_requirement"),
+                    },
+                    execution_result={},
+                )
+                work_item.reply_template_key = str(render["template_key"])
+                await self._project_waiting_manual(
+                    work_item=work_item,
+                    message_text=clean_text,
+                    render=render,
+                    intent=intent,
+                )
+                await self._maybe_emit_risk_cadence_snapshot(
+                    project_id=_safe_int(work_item.project_id),
+                    review_summary=work_item.review_summary,
+                )
+                direct_reply_required = await self._claim_reply_receipt(
+                    work_item_id=int(work_item.id),
+                    source_system=source_system,
+                    source_message_id=synthetic_message_id,
+                    receipt_phase=str(render["receipt_phase"]),
+                    template_key=str(render["template_key"]),
+                    payload=render["payload"],
+                )
+                ingest_event.status = "done"
+                ingest_event.processed_at = _utc_now_naive()
+                await self._write_processing_log(
+                    source_message_id=synthetic_message_id,
+                    work_item=work_item,
+                    message_text=clean_text,
+                    route="ai_pmo.waiting_manual",
+                    operation_code=None,
+                    operation_result=work_item.execution_result,
+                    processing_status="success",
+                    duration_ms=int((time.perf_counter() - started) * 1000),
+                )
+                await self.db.commit()
+                return self._build_capture_result(
+                    work_item=work_item,
+                    intent=intent,
+                    execution_result=work_item.execution_result or {},
+                    render=render,
+                    direct_reply_required=direct_reply_required,
+                    status="captured",
+                    classification="command",
+                )
+
+            work_item.state = STATE_PLANNED
+            work_item.action_bundle = decision_payload["execution_plan"]
+            work_item.state = STATE_EXECUTING
+            if intent.name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+                execution_result, verification = await self._execute_followup_intent(
+                    work_item=work_item,
+                    intent=intent,
+                    message_text=clean_text,
+                    sender=sender,
+                    operator_user_id=_safe_int(getattr(operator, "id", None)),
+                    source_system=source_system,
+                    source_message_id=synthetic_message_id,
+                    channel_ref=channel_ref,
+                    source_sender_open_id=source_sender_open_id,
+                    world_model=world_model,
+                )
+            else:
+                execution_result = await self.executor.execute(
+                    intent=intent,
+                    message_text=clean_text,
+                    sender=sender,
+                    operator_user_id=_safe_int(getattr(operator, "id", None)),
+                    source_system=source_system,
+                    source_message_id=synthetic_message_id,
+                )
+                verification = await self._verify_execution(intent=intent, execution_result=execution_result, world_model=world_model)
+            work_item.execution_result = execution_result
+            if _safe_int(execution_result.get("project_id")) is not None:
+                work_item.project_id = _safe_int(execution_result.get("project_id"))
+            if _safe_int(execution_result.get("signal_id")) is not None:
+                work_item.signal_id = _safe_int(execution_result.get("signal_id"))
+            work_item.review_summary = await self.review_service.review_work_item(
+                work_item=work_item,
+                intent_name=intent.name,
+                execution_result=execution_result,
+                world_model=world_model,
+                verification=verification,
+                simulation=decision_payload["simulation"],
+            )
+            await self._maybe_emit_risk_cadence_snapshot(
+                project_id=_safe_int(work_item.project_id),
+                review_summary=work_item.review_summary,
+            )
+            decision_row.execution_result = execution_result
+            issues = verification.get("issues") if isinstance(verification.get("issues"), list) else []
+            if verification["verified"]:
+                work_item.state = STATE_SUCCEEDED
+                await self._project_success(intent=intent, execution_result=execution_result)
+            elif any(
+                issue in {"waiting_manual", "followup_ambiguous", "signal_source_pending_manual"}
+                for issue in issues
+            ) or str(execution_result.get("status") or "").strip().lower() == "pending_manual":
+                work_item.state = STATE_WAITING_MANUAL
+            else:
+                work_item.state = STATE_FAILED
+            decision_row.final_outcome = work_item.state
+
+            render = self.reply_renderer.render(
+                work_state=work_item.state,
+                intent=intent,
+                decision_record={
+                    "approval_requirement": decision_payload["policy"].get("approval_requirement"),
+                },
+                execution_result=execution_result,
+            )
+            work_item.reply_template_key = str(render["template_key"])
+            direct_reply_required = await self._claim_reply_receipt(
+                work_item_id=int(work_item.id),
+                source_system=source_system,
+                source_message_id=synthetic_message_id,
+                receipt_phase=str(render["receipt_phase"]),
+                template_key=str(render["template_key"]),
+                payload=render["payload"],
+            )
+            ingest_event.status = "done"
+            ingest_event.processed_at = _utc_now_naive()
+            await self._write_processing_log(
+                source_message_id=synthetic_message_id,
+                work_item=work_item,
+                message_text=clean_text,
+                route=f"ai_pmo.{intent.name}",
+                operation_code=self._primary_operation_code(intent.name),
+                operation_result=execution_result,
+                processing_status="success" if work_item.state in {STATE_SUCCEEDED, STATE_WAITING_MANUAL} else "failed",
+                duration_ms=int((time.perf_counter() - started) * 1000),
+            )
+            await self.db.commit()
+            return self._build_capture_result(
+                work_item=work_item,
+                intent=intent,
+                execution_result=execution_result,
+                render=render,
+                direct_reply_required=direct_reply_required,
+                status="answered" if intent.name in {INTENT_QUERY, INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE} and work_item.state == STATE_SUCCEEDED else "captured",
+                classification=(
+                    "question"
+                    if intent.name == INTENT_QUERY
+                    else ("signal" if intent.name == INTENT_SIGNAL_CREATE else "command")
+                ),
+            )
+        except Exception as exc:
+            await self.db.rollback()
+            logger.warning(
+                "ai_pmo_kernel_failed",
+                tenant_id=self.tenant_id,
+                source_system=source_system,
+                source_message_id=synthetic_message_id,
+                error=str(exc),
+            )
+            return {"handled": False, "reason": "kernel_exception", "error": str(exc)}
+
+    async def _claim_ingress_event(
+        self,
+        *,
+        source_system: str,
+        source_message_id: str,
+        payload: dict[str, Any],
+    ) -> tuple[IngestEvent, bool]:
+        channel = f"{source_system}_message"
+        existing = (
+            await self.db.execute(
+                select(IngestEvent).where(
+                    and_(
+                        IngestEvent.tenant_id == self.tenant_id,
+                        IngestEvent.channel == channel,
+                        IngestEvent.event_key == source_message_id,
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+        if existing is not None:
+            return existing, False
+        row = IngestEvent(
+            tenant_id=self.tenant_id,
+            channel=channel,
+            event_key=source_message_id,
+            status="processing",
+            payload=payload,
+            retry_count=0,
+            error_message=None,
+            processed_at=None,
+        )
+        self.db.add(row)
+        try:
+            await self.db.flush()
+        except IntegrityError:
+            await self.db.rollback()
+            existing = (
+                await self.db.execute(
+                    select(IngestEvent).where(
+                        and_(
+                            IngestEvent.tenant_id == self.tenant_id,
+                            IngestEvent.channel == channel,
+                            IngestEvent.event_key == source_message_id,
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one()
+            return existing, False
+        return row, True
+
+    async def _create_work_item(
+        self,
+        *,
+        ingest_event: IngestEvent,
+        source_system: str,
+        source_message_id: str,
+        owner_user_id: Optional[int],
+    ) -> WorkItem:
+        row = WorkItem(
+            tenant_id=self.tenant_id,
+            ingest_event_id=int(ingest_event.id),
+            source_system=source_system,
+            source_message_id=source_message_id,
+            state=STATE_RECEIVED,
+            owner_user_id=owner_user_id,
+            dedupe_key=f"{source_system}:{source_message_id}",
+        )
+        self.db.add(row)
+        await self.db.flush()
+        return row
+
+    async def _build_world_model(
+        self,
+        *,
+        message_text: str,
+        sender: str,
+        project_hint_id: Optional[int],
+    ) -> dict[str, Any]:
+        project = await self._resolve_project(message_text=message_text, project_hint_id=project_hint_id)
+        project_id = _safe_int(getattr(project, "id", None))
+        portfolio_summary = await self._build_portfolio_summary(project_id=project_id)
+        budget = await self._build_budget_summary(project_id=project_id)
+        capacity = await self._build_capacity_summary(project_id=project_id)
+        skill_graph = await self._build_skill_graph(project_id=project_id)
+        charter = await self._build_charter_snapshot(project_id=project_id)
+        task_summary = await self._build_task_summary(project_id=project_id)
+        risk_summary = await self._build_risk_summary(project_id=project_id)
+        domain = await self._build_project_domain_snapshot(project_id=project_id)
+        dependencies = self._build_dependency_summary(charter=charter, task_summary=task_summary)
+        recent_signals = await self._build_recent_signal_summary(project_id=project_id)
+
+        world_model = {
+            "sender": sender,
+            "resolved_project_id": project_id,
+            "project": self._serialize_project(project) if project is not None else None,
+            "project_domain": domain,
+            "portfolio": portfolio_summary,
+            "budget": budget,
+            "capacity": capacity,
+            "member_count": len(capacity),
+            "skill_graph": skill_graph,
+            "charter": charter,
+            "task_summary": task_summary,
+            "risk_summary": risk_summary,
+            "dependencies": dependencies,
+            "recent_signals": recent_signals,
+            "decision_context_snapshot": {
+                "captured_at": _utc_now_naive().isoformat(),
+                "project_id": project_id,
+                "budget": budget,
+                "capacity": capacity,
+                "skill_graph_size": len(skill_graph),
+                "dependency_pressure": dependencies.get("pressure"),
+            },
+        }
+        self.db.add(
+            PortfolioSnapshot(
+                tenant_id=self.tenant_id,
+                scope_type="tenant",
+                scope_ref=f"tenant:{self.tenant_id}",
+                summary=portfolio_summary,
+                payload=world_model,
+            )
+        )
+        await self.db.flush()
+        return world_model
+
+    async def _resolve_project(self, *, message_text: str, project_hint_id: Optional[int]) -> Optional[Project]:
+        if project_hint_id is not None:
+            return (
+                await self.db.execute(
+                    select(Project).where(
+                        and_(
+                            Project.tenant_id == self.tenant_id,
+                            Project.id == int(project_hint_id),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+
+        id_match = re.search(r"(?:项目(?:id|ID)?|project(?:id)?)\s*[:#：]?\s*(\d+)", str(message_text or ""))
+        if id_match:
+            project_id = int(id_match.group(1))
+            return (
+                await self.db.execute(
+                    select(Project).where(
+                        and_(
+                            Project.tenant_id == self.tenant_id,
+                            Project.id == project_id,
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+
+        projects = (
+            await self.db.execute(
+                select(Project).where(Project.tenant_id == self.tenant_id).order_by(Project.updated_at.desc(), Project.id.desc()).limit(50)
+            )
+        ).scalars().all()
+        normalized_message = re.sub(r"\s+", "", str(message_text or "")).lower()
+        for project in projects:
+            for candidate in (project.title, project.name):
+                normalized_candidate = re.sub(r"\s+", "", str(candidate or "")).lower()
+                if normalized_candidate and normalized_candidate in normalized_message:
+                    return project
+        return None
+
+    async def _build_portfolio_summary(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        active_project_count = (
+            await self.db.execute(
+                select(func.count(Project.id)).where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.status.in_(["active", "planning", "todo", "doing"]),
+                    )
+                )
+            )
+        ).scalar_one()
+        return {
+            "active_project_count": int(active_project_count or 0),
+            "focus_project_id": project_id,
+        }
+
+    async def _build_budget_summary(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        stmt = select(Budget).where(Budget.tenant_id == self.tenant_id)
+        if project_id is not None:
+            stmt = stmt.where(or_(Budget.project_id == int(project_id), Budget.project_id.is_(None)))
+        row = (await self.db.execute(stmt.order_by(Budget.project_id.desc().nulls_last(), Budget.id.desc()).limit(1))).scalars().first()
+        if row is None:
+            return {"budget_id": None, "usage_ratio": 0.0, "hard_limit_enabled": False}
+        amount = float(row.budget_amount_usd or 0.0)
+        usage = float(row.current_usage_usd or 0.0)
+        ratio = (usage / amount) if amount > 0 else 0.0
+        return {
+            "budget_id": int(row.id),
+            "budget_type": str(row.budget_type or ""),
+            "period": str(row.period or ""),
+            "budget_amount_usd": amount,
+            "current_usage_usd": usage,
+            "usage_ratio": round(ratio, 4),
+            "hard_limit_enabled": bool(row.hard_limit_enabled),
+        }
+
+    async def _build_capacity_summary(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        user_ids: list[int] = []
+        if project_id is not None:
+            rows = (
+                await self.db.execute(
+                    select(UserProject.user_id).where(
+                        and_(
+                            UserProject.tenant_id == self.tenant_id,
+                            UserProject.project_id == int(project_id),
+                        )
+                    )
+                )
+            ).scalars().all()
+            user_ids = [int(uid) for uid in rows]
+        if not user_ids and project_id is not None:
+            owner_id = (
+                await self.db.execute(
+                    select(Project.owner_id).where(
+                        and_(
+                            Project.tenant_id == self.tenant_id,
+                            Project.id == int(project_id),
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if owner_id is not None:
+                user_ids = [int(owner_id)]
+        if not user_ids:
+            return []
+        rows = (
+            await self.db.execute(
+                select(UserWorkload).where(
+                    and_(
+                        UserWorkload.tenant_id == self.tenant_id,
+                        UserWorkload.user_id.in_(user_ids),
+                    )
+                )
+            )
+        ).scalars().all()
+        return [
+            {
+                "user_id": int(row.user_id),
+                "current_assigned_hours": float(row.current_assigned_hours or 0.0),
+                "max_available_hours_per_week": float(row.max_available_hours_per_week or 0.0),
+                "workload_ratio": float(row.workload_ratio or 0.0),
+            }
+            for row in rows
+        ]
+
+    async def _build_skill_graph(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        user_ids: list[int] = []
+        if project_id is not None:
+            rows = (
+                await self.db.execute(
+                    select(UserProject.user_id).where(
+                        and_(
+                            UserProject.tenant_id == self.tenant_id,
+                            UserProject.project_id == int(project_id),
+                        )
+                    )
+                )
+            ).scalars().all()
+            user_ids = [int(uid) for uid in rows]
+        if not user_ids:
+            return []
+        rows = (
+            await self.db.execute(
+                select(UserSkillBinding).where(
+                    and_(
+                        UserSkillBinding.tenant_id == self.tenant_id,
+                        UserSkillBinding.user_id.in_(user_ids),
+                    )
+                ).limit(50)
+            )
+        ).scalars().all()
+        return [
+            {
+                "user_id": int(row.user_id),
+                "skill_tag_id": int(row.skill_tag_id),
+                "proficiency_level": str(row.proficiency_level or ""),
+                "quality_score": float(row.quality_score or 0.0),
+                "efficiency_score": float(row.efficiency_score or 0.0),
+            }
+            for row in rows
+        ]
+
+    async def _build_project_domain_snapshot(self, *, project_id: Optional[int]) -> Optional[dict[str, Any]]:
+        if project_id is None:
+            return None
+        domain_row = (
+            await self.db.execute(
+                select(ProjectDomain)
+                .join(Project, Project.domain_id == ProjectDomain.id)
+                .where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.id == int(project_id),
+                        ProjectDomain.tenant_id == self.tenant_id,
+                    )
+                )
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+        if domain_row is None:
+            return None
+        return {
+            "id": int(domain_row.id),
+            "name": str(domain_row.name or ""),
+            "slug": str(domain_row.slug or ""),
+            "description": str(domain_row.description or ""),
+            "owner_id": domain_row.owner_id,
+        }
+
+    async def _build_charter_snapshot(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        if project_id is None:
+            return []
+        rows = (
+            await self.db.execute(
+                select(TruthNode).where(
+                    and_(
+                        TruthNode.tenant_id == self.tenant_id,
+                        TruthNode.project_id == int(project_id),
+                        TruthNode.node_type == "charter",
+                    )
+                ).order_by(TruthNode.updated_at.desc(), TruthNode.id.desc()).limit(3)
+            )
+        ).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "title": str(row.title or ""),
+                "status": str(row.status or ""),
+                "content": row.content if isinstance(row.content, dict) else {},
+            }
+            for row in rows
+        ]
+
+    async def _build_task_summary(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        if project_id is None:
+            return {"total": 0, "by_status": {}, "overdue_count": 0, "uncovered_skill_tag_ids": [], "blocked_tasks": []}
+        tasks = (
+            await self.db.execute(
+                select(Task).where(and_(Task.tenant_id == self.tenant_id, Task.project_id == int(project_id)))
+            )
+        ).scalars().all()
+        by_status: dict[str, int] = {}
+        required_skill_tag_ids: set[int] = set()
+        blocked_tasks: list[dict[str, Any]] = []
+        overdue_count = 0
+        now = _utc_now_naive()
+        for row in tasks:
+            status = str(row.status or "unknown")
+            by_status[status] = by_status.get(status, 0) + 1
+            for skill_id in row.required_skill_tag_ids or []:
+                if skill_id is not None:
+                    required_skill_tag_ids.add(int(skill_id))
+            if status == "blocked":
+                blocked_tasks.append({"id": int(row.id), "title": str(row.title or "")[:120]})
+            if row.due_date and str(row.status or "").lower() not in {"done", "completed", "closed"}:
+                due_date = row.due_date.replace(tzinfo=None) if row.due_date.tzinfo else row.due_date
+                if due_date < now:
+                    overdue_count += 1
+        skill_graph = await self._build_skill_graph(project_id=project_id)
+        available_skill_ids = {int(item["skill_tag_id"]) for item in skill_graph if item.get("skill_tag_id") is not None}
+        uncovered_skill_ids = sorted(skill_id for skill_id in required_skill_tag_ids if skill_id not in available_skill_ids)
+        return {
+            "total": sum(by_status.values()),
+            "by_status": by_status,
+            "overdue_count": overdue_count,
+            "required_skill_tag_ids": sorted(required_skill_tag_ids),
+            "uncovered_skill_tag_ids": uncovered_skill_ids,
+            "blocked_tasks": blocked_tasks[:8],
+        }
+
+    async def _build_recent_signal_summary(self, *, project_id: Optional[int]) -> list[dict[str, Any]]:
+        if project_id is None:
+            return []
+        rows = (
+            await self.db.execute(
+                select(Signal).where(
+                    and_(
+                        Signal.tenant_id == self.tenant_id,
+                        or_(
+                            Signal.matched_project_id == int(project_id),
+                            Signal.id.in_(
+                                select(ProjectSignalLink.signal_id).where(
+                                    and_(
+                                        ProjectSignalLink.tenant_id == self.tenant_id,
+                                        ProjectSignalLink.project_id == int(project_id),
+                                    )
+                                )
+                            ),
+                        ),
+                    )
+                ).order_by(Signal.updated_at.desc(), Signal.id.desc()).limit(5)
+            )
+        ).scalars().all()
+        return [
+            {
+                "id": int(row.id),
+                "title": str(row.title or ""),
+                "status": str(row.status or ""),
+                "signal_type": str(row.signal_type or ""),
+            }
+            for row in rows
+        ]
+
+    async def _build_risk_summary(self, *, project_id: Optional[int]) -> dict[str, Any]:
+        stmt = select(func.count(Notification.id)).where(
+            and_(
+                Notification.tenant_id == self.tenant_id,
+                Notification.is_read.is_(False),
+                Notification.kind.in_(["risk", "alert", "action_required"]),
+            )
+        )
+        if project_id is not None:
+            stmt = stmt.where(or_(Notification.entity_id == int(project_id), Notification.payload.is_(None), Notification.entity_type == "project"))
+        open_risk_count = (await self.db.execute(stmt)).scalar_one()
+        return {"open_risk_count": int(open_risk_count or 0)}
+
+    @staticmethod
+    def _build_dependency_summary(*, charter: list[dict[str, Any]], task_summary: dict[str, Any]) -> dict[str, Any]:
+        blocked_tasks = task_summary.get("blocked_tasks") if isinstance(task_summary.get("blocked_tasks"), list) else []
+        overdue_count = int(task_summary.get("overdue_count") or 0)
+        pressure = "high" if blocked_tasks or overdue_count > 0 else ("medium" if not charter else "low")
+        return {
+            "pressure": pressure,
+            "blocked_tasks": blocked_tasks[:5],
+            "overdue_count": overdue_count,
+            "charter_ready": bool(charter),
+        }
+
+    async def _create_decision_record(
+        self,
+        *,
+        work_item: WorkItem,
+        world_model: dict[str, Any],
+        decision_payload: dict[str, Any],
+    ) -> DecisionRecord:
+        strategic = decision_payload["strategic"]
+        policy = decision_payload["policy"]
+        row = DecisionRecord(
+            tenant_id=self.tenant_id,
+            work_item_id=int(work_item.id),
+            decision_object_type=str(strategic.get("decision_object_type") or "portfolio"),
+            decision_object_id=_safe_int(strategic.get("decision_object_id")),
+            decision_level=str(strategic.get("decision_level") or "portfolio"),
+            context_snapshot=world_model,
+            rationale=str(strategic.get("rationale") or ""),
+            confidence=0.82,
+            risk_score=float(strategic.get("risk_score") or 0.0),
+            budget_impact=strategic.get("budget_impact"),
+            resource_impact=strategic.get("resource_impact"),
+            simulation_result=decision_payload.get("simulation"),
+            policy_gate_hit=bool(policy.get("policy_gate_hit")),
+            policy_gate_code=(str(policy.get("policy_gate_code") or "")[:80] or None),
+            action_bundle=decision_payload["execution_plan"],
+            approval_requirement=policy.get("approval_requirement"),
+            final_outcome=STATE_DECIDED,
+        )
+        self.db.add(row)
+        await self.db.flush()
+        return row
+
+    async def _create_plan_bundle(
+        self,
+        *,
+        work_item_id: int,
+        plan_type: str,
+        summary: str,
+        payload: dict[str, Any],
+    ) -> PlanBundle:
+        row = PlanBundle(
+            tenant_id=self.tenant_id,
+            work_item_id=int(work_item_id),
+            plan_type=plan_type[:20],
+            summary=summary[:2000],
+            payload=payload,
+        )
+        self.db.add(row)
+        await self.db.flush()
+        return row
+
+    async def _claim_reply_receipt(
+        self,
+        *,
+        work_item_id: int,
+        source_system: str,
+        source_message_id: str,
+        receipt_phase: str,
+        template_key: str,
+        payload: dict[str, Any],
+    ) -> bool:
+        row = ReplyReceipt(
+            tenant_id=self.tenant_id,
+            work_item_id=int(work_item_id),
+            source_system=source_system[:32],
+            source_message_id=source_message_id[:128],
+            receipt_phase=receipt_phase[:32],
+            template_key=template_key[:64],
+            payload=payload,
+            status="ready",
+        )
+        self.db.add(row)
+        try:
+            await self.db.flush()
+            return True
+        except IntegrityError:
+            await self.db.rollback()
+            existing_ingest = (
+                await self.db.execute(
+                    select(IngestEvent).where(
+                        and_(
+                            IngestEvent.tenant_id == self.tenant_id,
+                            IngestEvent.channel == f"{source_system}_message",
+                            IngestEvent.event_key == source_message_id,
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one()
+            existing_work_item = (
+                await self.db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == self.tenant_id,
+                            WorkItem.ingest_event_id == int(existing_ingest.id),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+            if existing_work_item is not None:
+                self.db.add(existing_work_item)
+            return False
+
+    async def _project_waiting_manual(
+        self,
+        *,
+        work_item: WorkItem,
+        message_text: str,
+        render: dict[str, Any],
+        intent: CanonicalIntent,
+    ) -> None:
+        await NotificationService(self.db).create_notifications(
+            tenant_id=self.tenant_id,
+            user_ids=await NotificationService(self.db)._resolve_recipients(tenant_id=self.tenant_id),
+            title="AI PMO 待人工确认",
+            content=str(render.get("text") or message_text)[:4000],
+            category="decision",
+            source="ai_pmo",
+            entity_type="work_item",
+            entity_id=int(work_item.id),
+            target_url=f"/inbox?workItemId={int(work_item.id)}",
+            kind="action_required",
+            priority="high",
+            status="pending_manual",
+            recommended_action="view_detail",
+            dedupe_key=f"ai-pmo-waiting:{int(work_item.id)}",
+            payload={
+                "work_item_id": int(work_item.id),
+                "intent": intent.name,
+                "project_id": intent.project_id,
+                "message_text": message_text[:500],
+            },
+            commit=False,
+        )
+
+    async def _project_success(self, *, intent: CanonicalIntent, execution_result: dict[str, Any]) -> None:
+        if intent.name == INTENT_SIGNAL_CREATE and _safe_int(execution_result.get("signal_id")) is not None:
+            await NotificationService(self.db).notify_signal_created(
+                tenant_id=self.tenant_id,
+                signal_id=int(execution_result["signal_id"]),
+                signal_content=str(execution_result.get("summary") or execution_result.get("signal_id") or ""),
+                source="ai_pmo",
+                target_url=(
+                    f"/projects/{int(execution_result['project_id'])}/board"
+                    if _safe_int(execution_result.get("project_id")) is not None
+                    else None
+                ),
+            )
+
+    async def _execute_followup_intent(
+        self,
+        *,
+        work_item: WorkItem,
+        intent: CanonicalIntent,
+        message_text: str,
+        sender: str,
+        operator_user_id: Optional[int],
+        source_system: str,
+        source_message_id: str,
+        channel_ref: Optional[str],
+        source_sender_open_id: Optional[str],
+        world_model: dict[str, Any],
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        explicit_target_id = self._extract_followup_work_item_id(message_text)
+        candidates = await self._load_recent_followup_candidates(
+            current_work_item_id=int(work_item.id),
+            source_system=source_system,
+            channel_ref=channel_ref,
+            sender=sender,
+            source_sender_open_id=source_sender_open_id,
+            explicit_target_id=explicit_target_id,
+        )
+        actionable = [
+            (target, ingest)
+            for target, ingest in candidates
+            if str(target.state or "") == STATE_WAITING_MANUAL
+            and str(target.intent or "") in FOLLOWUP_TARGET_INTENTS
+        ]
+        if explicit_target_id is not None and not actionable:
+            matched = next((target for target, _ in candidates if int(target.id) == int(explicit_target_id)), None)
+            if matched is not None:
+                reply_text = f"工作项#{int(matched.id)} 当前状态为 {matched.state}，无需继续处理。"
+                return {
+                    "status": "succeeded",
+                    "followup_reply": reply_text,
+                    "target_work_item_id": int(matched.id),
+                    "target_state": matched.state,
+                    "target_intent": matched.intent,
+                    "project_id": _safe_int(matched.project_id),
+                    "signal_id": _safe_int(matched.signal_id),
+                }, {"verified": True, "issues": []}
+
+        if not actionable:
+            reply_text = "当前群最近没有可继续处理的待确认工作项。"
+            return {
+                "status": "pending_manual",
+                "followup_reply": reply_text,
+                "candidate_work_item_ids": [],
+            }, {"verified": False, "issues": ["waiting_manual"]}
+
+        if len(actionable) > 1:
+            reply_text = self._build_followup_candidate_clarification([target for target, _ in actionable])
+            return {
+                "status": "pending_manual",
+                "followup_reply": reply_text,
+                "candidate_work_item_ids": [int(target.id) for target, _ in actionable[:5]],
+            }, {"verified": False, "issues": ["waiting_manual", "followup_ambiguous"]}
+
+        target_work_item, target_ingest = actionable[0]
+        return await self._resume_followup_target(
+            target_work_item=target_work_item,
+            target_ingest=target_ingest,
+            followup_intent=intent,
+            followup_sender=sender,
+            operator_user_id=operator_user_id,
+            source_system=source_system,
+            source_message_id=source_message_id,
+            world_model=world_model,
+        )
+
+    async def _load_recent_followup_candidates(
+        self,
+        *,
+        current_work_item_id: int,
+        source_system: str,
+        channel_ref: Optional[str],
+        sender: str,
+        source_sender_open_id: Optional[str],
+        explicit_target_id: Optional[int],
+        limit: int = 12,
+    ) -> list[tuple[WorkItem, IngestEvent]]:
+        rows = (
+            await self.db.execute(
+                select(WorkItem, IngestEvent)
+                .join(IngestEvent, IngestEvent.id == WorkItem.ingest_event_id)
+                .where(
+                    and_(
+                        WorkItem.tenant_id == self.tenant_id,
+                        WorkItem.id != int(current_work_item_id),
+                    )
+                )
+                .order_by(WorkItem.id.desc())
+                .limit(max(limit * 3, 18))
+            )
+        ).all()
+        selected: list[tuple[WorkItem, IngestEvent]] = []
+        normalized_channel = str(channel_ref or "").strip()
+        normalized_sender = str(sender or "").strip()
+        normalized_open_id = str(source_sender_open_id or "").strip()
+        for target_work_item, ingest in rows:
+            if explicit_target_id is not None and int(target_work_item.id) == int(explicit_target_id):
+                selected.append((target_work_item, ingest))
+                continue
+            payload = ingest.payload if isinstance(ingest.payload, dict) else {}
+            same_channel = normalized_channel and str(payload.get("channel_ref") or "").strip() == normalized_channel
+            same_open_id = normalized_open_id and str(payload.get("source_sender_open_id") or "").strip() == normalized_open_id
+            same_sender = normalized_sender and str(payload.get("sender") or "").strip() == normalized_sender
+            if not (same_channel or same_open_id or same_sender):
+                continue
+            selected.append((target_work_item, ingest))
+            if len(selected) >= limit:
+                break
+        return selected
+
+    async def _resume_followup_target(
+        self,
+        *,
+        target_work_item: WorkItem,
+        target_ingest: IngestEvent,
+        followup_intent: CanonicalIntent,
+        followup_sender: str,
+        operator_user_id: Optional[int],
+        source_system: str,
+        source_message_id: str,
+        world_model: dict[str, Any],
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        payload = target_ingest.payload if isinstance(target_ingest.payload, dict) else {}
+        original_message = str(payload.get("message_text") or "").strip()
+        override_project_id = _safe_int(world_model.get("resolved_project_id"))
+        target_project_id = _safe_int(target_work_item.project_id) or override_project_id
+        rebuilt_intent = self.intent_resolver.resolve(text=original_message, project_id=target_project_id)
+        target_intent_name = str(target_work_item.intent or rebuilt_intent.name or "").strip()
+        target_intent = CanonicalIntent(
+            name=target_intent_name or rebuilt_intent.name,
+            confidence=float(rebuilt_intent.confidence or 0.88),
+            project_id=target_project_id,
+            signal_title=rebuilt_intent.signal_title,
+            signal_summary=rebuilt_intent.signal_summary,
+            metadata=dict(rebuilt_intent.metadata),
+        )
+        if target_intent.name not in FOLLOWUP_TARGET_INTENTS:
+            reply_text = f"工作项#{int(target_work_item.id)} 当前类型为 {target_intent.name or 'unknown'}，暂不支持继续处理。"
+            return {
+                "status": "failed_safe",
+                "followup_reply": reply_text,
+                "target_work_item_id": int(target_work_item.id),
+                "target_state": target_work_item.state,
+                "target_intent": target_intent.name,
+            }, {"verified": False, "issues": ["unsupported_intent"]}
+        if target_intent.name in {INTENT_SIGNAL_CREATE, INTENT_SIGNAL_SOURCE_ONBOARD} and target_project_id is None:
+            reply_text = "未执行：目标工作项仍缺少项目归属，请补充项目 ID 后再继续处理。"
+            return {
+                "status": "pending_manual",
+                "followup_reply": reply_text,
+                "target_work_item_id": int(target_work_item.id),
+                "target_state": STATE_WAITING_MANUAL,
+                "target_intent": target_intent.name,
+            }, {"verified": False, "issues": ["waiting_manual"]}
+
+        target_work_item.project_id = target_project_id
+        target_work_item.state = STATE_EXECUTING
+        target_execution_result = await self.executor.execute(
+            intent=target_intent,
+            message_text=original_message,
+            sender=followup_sender,
+            operator_user_id=operator_user_id,
+            source_system=source_system,
+            source_message_id=f"{source_message_id}:resume:{int(target_work_item.id)}",
+        )
+        target_work_item.execution_result = target_execution_result
+        if _safe_int(target_execution_result.get("signal_id")) is not None:
+            target_work_item.signal_id = _safe_int(target_execution_result.get("signal_id"))
+
+        target_verification = await self._verify_execution(
+            intent=target_intent,
+            execution_result=target_execution_result,
+            world_model=target_work_item.context_snapshot if isinstance(target_work_item.context_snapshot, dict) else world_model,
+        )
+        target_work_item.review_summary = await self.review_service.review_work_item(
+            work_item=target_work_item,
+            intent_name=target_intent.name,
+            execution_result=target_execution_result,
+            world_model=target_work_item.context_snapshot if isinstance(target_work_item.context_snapshot, dict) else world_model,
+            verification=target_verification,
+            simulation={
+                "followup_resumed": True,
+                "target_work_item_id": int(target_work_item.id),
+                "followup_intent": followup_intent.name,
+            },
+        )
+        await self._maybe_emit_risk_cadence_snapshot(
+            project_id=_safe_int(target_work_item.project_id),
+            review_summary=target_work_item.review_summary,
+        )
+        if target_verification.get("verified"):
+            target_work_item.state = STATE_SUCCEEDED
+            await self._project_success(intent=target_intent, execution_result=target_execution_result)
+        else:
+            target_issues = target_verification.get("issues") if isinstance(target_verification.get("issues"), list) else []
+            if any(issue in {"pending_manual", "signal_source_pending_manual"} for issue in target_issues):
+                target_work_item.state = STATE_WAITING_MANUAL
+            else:
+                target_work_item.state = STATE_FAILED
+
+        target_decision = (
+            await self.db.execute(
+                select(DecisionRecord)
+                .where(
+                    and_(
+                        DecisionRecord.tenant_id == self.tenant_id,
+                        DecisionRecord.work_item_id == int(target_work_item.id),
+                    )
+                )
+                .order_by(DecisionRecord.id.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+        if target_decision is not None:
+            target_decision.execution_result = target_execution_result
+            target_decision.final_outcome = target_work_item.state
+
+        target_render = self.reply_renderer.render(
+            work_state=target_work_item.state,
+            intent=target_intent,
+            decision_record={},
+            execution_result=target_execution_result,
+        )
+        reply_text = str(target_render.get("text") or "").strip()
+        if target_work_item.state == STATE_SUCCEEDED:
+            reply_text = f"已继续处理：{reply_text}"
+        elif not reply_text:
+            reply_text = "未执行：继续处理后仍未形成可验证终态。"
+        project_id = _safe_int(target_execution_result.get("project_id")) or _safe_int(target_work_item.project_id)
+        signal_id = _safe_int(target_execution_result.get("signal_id")) or _safe_int(target_work_item.signal_id)
+        status = (
+            "succeeded"
+            if target_work_item.state == STATE_SUCCEEDED
+            else ("pending_manual" if target_work_item.state == STATE_WAITING_MANUAL else "failed_safe")
+        )
+        current_verification = {"verified": target_work_item.state == STATE_SUCCEEDED, "issues": []}
+        if target_work_item.state == STATE_WAITING_MANUAL:
+            current_verification = {"verified": False, "issues": ["waiting_manual"]}
+        elif target_work_item.state == STATE_FAILED:
+            current_verification = {"verified": False, "issues": ["executor_failed"]}
+        return {
+            "status": status,
+            "followup_reply": reply_text,
+            "target_work_item_id": int(target_work_item.id),
+            "target_state": target_work_item.state,
+            "target_intent": target_intent.name,
+            "project_id": project_id,
+            "signal_id": signal_id,
+            "auto_tasks": target_execution_result.get("auto_tasks") if isinstance(target_execution_result.get("auto_tasks"), list) else [],
+        }, current_verification
+
+    async def _verify_execution(
+        self,
+        *,
+        intent: CanonicalIntent,
+        execution_result: dict[str, Any],
+        world_model: dict[str, Any],
+    ) -> dict[str, Any]:
+        if str(execution_result.get("status") or "").lower() == "failed_safe":
+            return {"verified": False, "issues": ["executor_failed"]}
+        if str(execution_result.get("status") or "").lower() == "pending_manual":
+            return {"verified": False, "issues": ["pending_manual"]}
+        if intent.name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+            return {
+                "verified": bool(
+                    _safe_int(execution_result.get("target_work_item_id")) is not None
+                    and str(execution_result.get("target_state") or "").strip() == STATE_SUCCEEDED
+                ),
+                "issues": [] if str(execution_result.get("target_state") or "").strip() == STATE_SUCCEEDED else ["waiting_manual"],
+            }
+        if intent.name == INTENT_QUERY:
+            return {"verified": bool(str(execution_result.get("answer") or "").strip()), "issues": []}
+        if intent.name == INTENT_PROJECT_CREATE:
+            return {"verified": _safe_int(execution_result.get("project_id")) is not None, "issues": []}
+        if intent.name in {INTENT_SIGNAL_CREATE, INTENT_SIGNAL_SOURCE_ONBOARD}:
+            signal_id = _safe_int(execution_result.get("signal_id"))
+            if signal_id is None:
+                return {"verified": False, "issues": ["signal_missing"]}
+            if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD:
+                workflow = execution_result.get("signal_source_workflow") if isinstance(execution_result.get("signal_source_workflow"), dict) else {}
+                if str(workflow.get("status") or "").lower() != "completed":
+                    return {"verified": False, "issues": ["signal_source_pending_manual"]}
+            if intent.project_id is None:
+                return {"verified": True, "issues": []}
+            link_exists = (
+                await self.db.execute(
+                    select(ProjectSignalLink.id).where(
+                        and_(
+                            ProjectSignalLink.tenant_id == self.tenant_id,
+                            ProjectSignalLink.project_id == int(intent.project_id),
+                            ProjectSignalLink.signal_id == int(signal_id),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+            return {"verified": link_exists is not None, "issues": [] if link_exists is not None else ["project_signal_link_missing"]}
+        return {"verified": False, "issues": ["unsupported_intent"]}
+
+    @staticmethod
+    def _build_review_summary(
+        *,
+        intent: CanonicalIntent,
+        verification: dict[str, Any],
+        world_model: dict[str, Any],
+    ) -> dict[str, Any]:
+        return {
+            "verified": bool(verification.get("verified")),
+            "issues": verification.get("issues") if isinstance(verification.get("issues"), list) else [],
+            "next_step": "进入下一轮调度" if verification.get("verified") else "等待人工复核",
+            "risk_score_after": float((world_model.get("risk_summary") or {}).get("open_risk_count") or 0.0),
+            "intent": intent.name,
+        }
+
+    async def _maybe_emit_risk_cadence_snapshot(
+        self,
+        *,
+        project_id: Optional[int],
+        review_summary: dict[str, Any],
+    ) -> None:
+        if not isinstance(review_summary, dict):
+            return
+        risk_after = float(review_summary.get("risk_score_after") or 0.0)
+        divergences = review_summary.get("divergences") if isinstance(review_summary.get("divergences"), list) else []
+        if risk_after < 0.7 and not divergences:
+            return
+        await self.cadence_service.generate_snapshot(
+            cadence_type="risk_review",
+            project_id=int(project_id) if project_id is not None else None,
+        )
+
+    async def _resolve_operator_user(self, *, sender: str, sender_open_id: Optional[str]) -> Optional[User]:
+        cleaned_sender = str(sender or "").strip()
+        if cleaned_sender:
+            user = (
+                await self.db.execute(
+                    select(User).where(
+                        and_(
+                            User.tenant_id == self.tenant_id,
+                            User.is_active.is_(True),
+                            or_(User.name == cleaned_sender, User.username == cleaned_sender),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+            if user is not None:
+                return user
+        preferred = (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.is_active.is_(True),
+                        User.status == "active",
+                        User.role.in_(["BOSS", "ADMIN", "PM"]),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+            )
+        ).scalar_one_or_none()
+        if preferred is not None:
+            return preferred
+        return (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.is_active.is_(True),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+            )
+        ).scalar_one_or_none()
+
+    async def _write_processing_log(
+        self,
+        *,
+        source_message_id: str,
+        work_item: WorkItem,
+        message_text: str,
+        route: str,
+        operation_code: Optional[str],
+        operation_result: Any,
+        processing_status: str,
+        duration_ms: int,
+    ) -> None:
+        self.db.add(
+            AIProcessingLog(
+                tenant_id=self.tenant_id,
+                feishu_message_id=source_message_id if work_item.source_system == "feishu" else None,
+                feishu_item_id=None,
+                raw_message=message_text[:8000],
+                ai_raw_output=json.dumps(
+                    {
+                        "work_item_id": int(work_item.id),
+                        "intent": work_item.intent,
+                        "state": work_item.state,
+                        "reply_template_key": work_item.reply_template_key,
+                    },
+                    ensure_ascii=False,
+                ),
+                project_score=None,
+                signal_confidence=None,
+                level="ai_pmo",
+                normalized_text=message_text[:8000],
+                route=route[:80],
+                intent=(str(work_item.intent or "")[:40] or None),
+                executor="ai_pmo_kernel",
+                operation_code=(operation_code[:80] if operation_code else None),
+                operation_result=operation_result if isinstance(operation_result, dict) else {"result": operation_result},
+                reply_source=(str(work_item.reply_template_key or "")[:40] or None),
+                routing_decision=route[:64],
+                routing_reason=f"AI PMO kernel route={route}",
+                reasoning_bullets=[
+                    f"work_item_id={int(work_item.id)}",
+                    f"state={work_item.state}",
+                    f"intent={work_item.intent}",
+                ],
+                processing_status=processing_status[:20],
+                error_message=None,
+                duration_ms=int(duration_ms),
+            )
+        )
+        await self.db.flush()
+
+    def _build_capture_result(
+        self,
+        *,
+        work_item: WorkItem,
+        intent: CanonicalIntent,
+        execution_result: dict[str, Any],
+        render: dict[str, Any],
+        direct_reply_required: bool,
+        status: str,
+        classification: str,
+    ) -> dict[str, Any]:
+        linked_project_ids = [int(work_item.project_id)] if work_item.project_id is not None else []
+        result = {
+            "handled": True,
+            "status": status,
+            "classification": classification,
+            "linked_project_ids": linked_project_ids,
+            "signal_id": _safe_int(execution_result.get("signal_id")) or _safe_int(work_item.signal_id),
+            "created_project_id": _safe_int(execution_result.get("project_id")) if intent.name == INTENT_PROJECT_CREATE else None,
+            "auto_tasks": execution_result.get("auto_tasks") if isinstance(execution_result.get("auto_tasks"), list) else execution_result.get("tasks", []),
+            "ai_trace_id": f"ai-pmo-{int(work_item.id)}",
+            "results": [
+                {
+                    "work_item_id": int(work_item.id),
+                    "state": work_item.state,
+                    "intent": intent.name,
+                    "route_status": "queued_for_decision" if work_item.state == STATE_WAITING_MANUAL else "auto_done",
+                }
+            ],
+            "intent": intent.name,
+            "direct_reply_required": bool(direct_reply_required),
+            "direct_reply_text": str(render.get("text") or "").strip(),
+            "reply_template_key": str(render.get("template_key") or "").strip(),
+            "work_item_id": int(work_item.id),
+            "decision_level": work_item.decision_level,
+            "review_summary": work_item.review_summary,
+        }
+        if intent.name == INTENT_QUERY and isinstance(execution_result.get("rag_result"), dict):
+            result["rag_result"] = execution_result["rag_result"]
+        if intent.name == INTENT_PROJECT_CREATE:
+            raw_payload = execution_result.get("raw_payload") if isinstance(execution_result.get("raw_payload"), dict) else {}
+            result["project"] = raw_payload.get("project")
+        if intent.name == INTENT_SIGNAL_SOURCE_ONBOARD:
+            result["signal_source_workflow"] = execution_result.get("signal_source_workflow")
+            result["signal_source_orchestration"] = execution_result.get("signal_source_orchestration")
+            result["ai_execution"] = execution_result.get("signal_source_orchestration")
+        return result
+
+    @staticmethod
+    def _primary_operation_code(intent_name: str) -> Optional[str]:
+        if intent_name == INTENT_QUERY:
+            return "query.answer"
+        if intent_name == INTENT_PROJECT_CREATE:
+            return "project.create"
+        if intent_name == INTENT_SIGNAL_CREATE:
+            return "signal.enter"
+        if intent_name == INTENT_SIGNAL_SOURCE_ONBOARD:
+            return "signal_source.onboard"
+        if intent_name in {INTENT_FOLLOWUP_CONFIRM, INTENT_FOLLOWUP_CONTINUE}:
+            return "work_item.resume"
+        return None
+
+    @staticmethod
+    def _extract_followup_work_item_id(message: str) -> Optional[int]:
+        match = FOLLOWUP_WORK_ITEM_ID_RE.search(str(message or ""))
+        return int(match.group(1)) if match else None
+
+    @staticmethod
+    def _build_followup_candidate_clarification(items: list[WorkItem]) -> str:
+        if not items:
+            return "当前没有可继续处理的待确认工作项。"
+        labels = [f"工作项#{int(item.id)}({str(item.intent or '').strip() or 'unknown'})" for item in items[:3]]
+        return f"当前存在多个待确认工作项：{'、'.join(labels)}。请补充工作项 ID 后再继续处理。"
+
+    @staticmethod
+    def _build_source_message_id(
+        *,
+        source_system: str,
+        source_message_id: Optional[str],
+        channel_ref: Optional[str],
+        sender: str,
+        message_text: str,
+    ) -> str:
+        explicit = str(source_message_id or "").strip()
+        if explicit:
+            return explicit[:128]
+        basis = f"{source_system}|{channel_ref or ''}|{sender}|{message_text}"
+        digest = hashlib.sha1(basis.encode("utf-8"), usedforsecurity=False).hexdigest()
+        return f"{source_system}-{digest[:32]}"
+
+    @staticmethod
+    def _normalize_text(text: str) -> str:
+        cleaned = re.sub(r"</?at[^>]*>", " ", str(text or ""), flags=re.IGNORECASE)
+        cleaned = re.sub(r"^@PMFlow\s*", "", cleaned, flags=re.IGNORECASE)
+        cleaned = re.sub(r"\s+", " ", cleaned).strip()
+        return cleaned[:4000]
+
+    @staticmethod
+    def _build_transport_mode_reply(message_text: str) -> Optional[str]:
+        normalized = re.sub(r"\s+", "", str(message_text or "")).lower()
+        has_long_connection = "长连接" in normalized or "longconnection" in normalized
+        has_webhook = "webhook" in normalized
+        if not (has_long_connection and has_webhook):
+            return None
+        if not any(token in normalized for token in TRANSPORT_MODE_CORRECTION_HINTS):
+            return None
+        return "已按长连接模式处理飞书接入，当前不会走 webhook；这条说明不会再被当成信号源类型变更。"
+
+    @staticmethod
+    def _serialize_project(project: Optional[Project]) -> Optional[dict[str, Any]]:
+        if project is None:
+            return None
+        return {
+            "id": int(project.id),
+            "name": str(project.name or ""),
+            "title": str(project.title or project.name or ""),
+            "description": str(project.description or ""),
+            "status": str(project.status or ""),
+            "owner_id": _safe_int(project.owner_id),
+        }
diff --git a/backend/app/services/ai_pmo_review_service.py b/backend/app/services/ai_pmo_review_service.py
new file mode 100644
index 00000000..40ca3396
--- /dev/null
+++ b/backend/app/services/ai_pmo_review_service.py
@@ -0,0 +1,189 @@
+"""AI PMO review engine。"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from sqlalchemy import and_, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import PortfolioSnapshot, ProjectSignalLink, ReplyReceipt, SignalSourceRegistry, WorkItem
+
+
+def _safe_int(value: Any) -> int | None:
+    try:
+        if value is None or value == "":
+            return None
+        return int(value)
+    except (TypeError, ValueError):
+        return None
+
+
+class AIPMOReviewService:
+    """复盘层：结果核验、偏差检测、风险再评分、规则建议。"""
+
+    def __init__(self, db: AsyncSession, tenant_id: int) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+
+    async def review_work_item(
+        self,
+        *,
+        work_item: WorkItem,
+        intent_name: str,
+        execution_result: dict[str, Any],
+        world_model: dict[str, Any],
+        verification: dict[str, Any],
+        simulation: dict[str, Any],
+    ) -> dict[str, Any]:
+        divergences: list[dict[str, Any]] = []
+        rule_updates: list[dict[str, Any]] = []
+
+        signal_id = _safe_int(execution_result.get("signal_id")) or _safe_int(work_item.signal_id)
+        project_id = _safe_int(work_item.project_id)
+        if signal_id is not None and project_id is not None:
+            link_exists = (
+                await self.db.execute(
+                    select(ProjectSignalLink.id).where(
+                        and_(
+                            ProjectSignalLink.tenant_id == self.tenant_id,
+                            ProjectSignalLink.project_id == int(project_id),
+                            ProjectSignalLink.signal_id == int(signal_id),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+            if link_exists is None:
+                divergences.append(
+                    {
+                        "code": "project_signal_link_missing",
+                        "severity": "high",
+                        "detail": f"signal#{signal_id} 未关联 project#{project_id}",
+                    }
+                )
+                rule_updates.append(
+                    {
+                        "rule_key": "signal_link_persistence_guard",
+                        "reason": "信号已写入但 link 缺失，后续查询会出现空白板。",
+                    }
+                )
+
+        workflow = execution_result.get("signal_source_workflow") if isinstance(execution_result.get("signal_source_workflow"), dict) else {}
+        raw_orchestration = execution_result.get("signal_source_orchestration") if isinstance(execution_result.get("signal_source_orchestration"), dict) else {}
+        registry_id = _safe_int(workflow.get("registry_id"))
+        if registry_id is not None:
+            registry = (
+                await self.db.execute(
+                    select(SignalSourceRegistry).where(
+                        and_(
+                            SignalSourceRegistry.tenant_id == self.tenant_id,
+                            SignalSourceRegistry.id == int(registry_id),
+                        )
+                    ).limit(1)
+                )
+            ).scalar_one_or_none()
+            if registry is not None:
+                normalized_status = str(workflow.get("status") or "").strip().lower()
+                registry_onboarding = str(registry.onboarding_status or "").strip().lower()
+                if normalized_status == "pending_manual" and registry_onboarding == "completed":
+                    divergences.append(
+                        {
+                            "code": "registry_workflow_divergence",
+                            "severity": "high",
+                            "detail": f"registry#{registry_id} 已 completed，但 domain workflow 为 pending_manual",
+                        }
+                    )
+                    rule_updates.append(
+                        {
+                            "rule_key": "signal_source_finalize_single_source",
+                            "reason": "registry 与 workflow 终态不一致，必须收口到 domain workflow。",
+                        }
+                    )
+        if workflow and str(workflow.get("status") or "").strip().lower() != "completed":
+            rule_updates.append(
+                {
+                    "rule_key": "onboarding_requires_material_actions",
+                    "reason": "signal source onboarding 只有 material actions + 非 fallback 才能标记 completed。",
+                }
+            )
+
+        duplicate_reply_groups = (
+            await self.db.execute(
+                select(ReplyReceipt.receipt_phase, func.count(ReplyReceipt.id))
+                .where(
+                    and_(
+                        ReplyReceipt.tenant_id == self.tenant_id,
+                        ReplyReceipt.work_item_id == int(work_item.id),
+                    )
+                )
+                .group_by(ReplyReceipt.receipt_phase)
+            )
+        ).all()
+        duplicate_phases = [str(phase) for phase, count in duplicate_reply_groups if int(count or 0) > 1]
+        if duplicate_phases:
+            divergences.append(
+                {
+                    "code": "duplicate_reply_receipts",
+                    "severity": "medium",
+                    "detail": ",".join(duplicate_phases),
+                }
+            )
+            rule_updates.append(
+                {
+                    "rule_key": "reply_receipt_single_phase",
+                    "reason": "同一 source_message 不得生成多条同 phase 回执。",
+                }
+            )
+
+        issues = verification.get("issues") if isinstance(verification.get("issues"), list) else []
+        manual_issue_codes = {"pending_manual", "signal_source_pending_manual", "waiting_manual"}
+        base_score = 0.92 if verification.get("verified") else (0.56 if any(issue in manual_issue_codes for issue in issues) else 0.32)
+        fallback_penalty = 0.15 if bool(workflow.get("fallback_used")) else 0.0
+        divergence_penalty = min(0.45, len(divergences) * 0.12)
+        agent_quality_score = max(0.0, round(base_score - fallback_penalty - divergence_penalty, 4))
+
+        risk_before = float((world_model.get("risk_summary") or {}).get("open_risk_count") or 0.0)
+        projected_risk = float(simulation.get("projected_risk_score") or 0.0)
+        risk_after = min(1.0, projected_risk + len(divergences) * 0.08 + (0.1 if not verification.get("verified") else 0.0))
+
+        next_actions: list[str] = []
+        if divergences:
+            next_actions.append("优先修正 projection / registry / reply 收口偏差。")
+        if not verification.get("verified"):
+            next_actions.append("当前工作项未形成可验证终态，进入风险例会或人工确认。")
+        if risk_after >= 0.7:
+            next_actions.append("风险重评分偏高，建议纳入风险例会升级。")
+        if not next_actions:
+            next_actions.append("复盘通过，进入下一轮节奏调度。")
+
+        quality_grade = "A" if agent_quality_score >= 0.85 else ("B" if agent_quality_score >= 0.7 else ("C" if agent_quality_score >= 0.5 else "D"))
+        review_payload = {
+            "verified": bool(verification.get("verified")),
+            "issues": issues,
+            "divergences": divergences,
+            "risk_score_before": round(risk_before, 4),
+            "risk_score_after": round(risk_after, 4),
+            "agent_quality_score": agent_quality_score,
+            "agent_quality_grade": quality_grade,
+            "rule_updates": rule_updates,
+            "next_actions": next_actions,
+            "simulation_summary": simulation,
+            "workflow_summary": workflow or raw_orchestration,
+        }
+        self.db.add(
+            PortfolioSnapshot(
+                tenant_id=self.tenant_id,
+                scope_type="review_cycle",
+                scope_ref=f"work_item:{int(work_item.id)}",
+                summary={
+                    "work_item_id": int(work_item.id),
+                    "intent": intent_name,
+                    "verified": bool(verification.get("verified")),
+                    "agent_quality_score": agent_quality_score,
+                    "risk_score_after": round(risk_after, 4),
+                },
+                payload=review_payload,
+            )
+        )
+        await self.db.flush()
+        return review_payload
diff --git a/backend/app/services/ai_service.py b/backend/app/services/ai_service.py
new file mode 100644
index 00000000..299aab13
--- /dev/null
+++ b/backend/app/services/ai_service.py
@@ -0,0 +1,590 @@
+"""
+PMFlow AI 核心能力封装
+
+技术栈：LangChain 0.3+、OpenAI SDK、Redis
+- 封装 GPT-4o mini 调用（流式/非流式）
+- Prompt 集中在 config/ai_prompts
+- AI 调用日志写入 ai_call_logs
+- Redis 结果缓存（1 小时 TTL）
+- 指数退避重试
+"""
+
+import hashlib
+import json
+import os
+import time
+from datetime import datetime
+from typing import Any, AsyncGenerator, Optional, Union
+from sqlalchemy import and_, select
+
+# LangChain & OpenAI
+try:
+    from langchain_openai import ChatOpenAI
+    from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
+    from langchain_core.language_models.chat_models import BaseChatModel
+    LANGCHAIN_AVAILABLE = True
+except ImportError:
+    LANGCHAIN_AVAILABLE = False
+    BaseChatModel = object
+
+# Redis
+try:
+    import redis.asyncio as redis
+    REDIS_AVAILABLE = True
+except ImportError:
+    REDIS_AVAILABLE = False
+
+# Token 定价 (GPT-4o-mini, 2024)
+INPUT_PRICE_PER_1K = 0.00015
+OUTPUT_PRICE_PER_1K = 0.0006
+CACHE_TTL_SECONDS = 3600  # 1 小时
+PROMPT_PROFILE_CACHE_TTL_SECONDS = 60
+
+from ..config.ai_prompts import build_default_role_prompt_bundle
+from ..database import AsyncSessionLocal
+from ..models import Integration
+
+
+def _t2s(v: Union[str, int]) -> str:
+    """统一转为 str"""
+    return str(v) if v is not None else ""
+
+
+def _hash_input(fn: str, *args, **kwargs) -> str:
+    """生成缓存 key"""
+    payload = json.dumps({"fn": fn, "args": args, "kwargs": kwargs}, sort_keys=True, default=str)
+    return hashlib.sha256(payload.encode()).hexdigest()
+
+
+def _estimate_cost(prompt_tokens: int, completion_tokens: int) -> float:
+    return (prompt_tokens * INPUT_PRICE_PER_1K + completion_tokens * OUTPUT_PRICE_PER_1K) / 1000
+
+
+def _tenant_to_int(tenant_id: Union[str, int]) -> int:
+    try:
+        return int(tenant_id)
+    except (TypeError, ValueError):
+        return 1
+
+
+class AIService:
+    """PMFlow AI 服务"""
+
+    def __init__(
+        self,
+        api_key: Optional[str] = None,
+        model: str = "gpt-4o-mini",
+        redis_url: Optional[str] = None,
+        max_retries: int = 3,
+    ):
+        self.api_key = api_key or os.getenv("OPENAI_API_KEY", "")
+        self.model = model
+        self.redis_url = redis_url or os.getenv("REDIS_URL")
+        self.max_retries = max_retries
+        self._llm: Optional[BaseChatModel] = None
+        self._redis: Optional[Any] = None
+        self._prompt_profile_cache: dict[int, tuple[float, tuple[str, str]]] = {}
+
+    def _get_llm(self) -> BaseChatModel:
+        if not LANGCHAIN_AVAILABLE:
+            raise RuntimeError("请安装 langchain-openai: pip install langchain-openai langchain-core")
+        if not self.api_key:
+            raise RuntimeError("请设置 OPENAI_API_KEY 环境变量")
+        if self._llm is None:
+            self._llm = ChatOpenAI(
+                model=self.model,
+                api_key=self.api_key,
+                temperature=0.3,
+            )
+        return self._llm
+
+    async def _get_redis(self):
+        if not REDIS_AVAILABLE or not self.redis_url:
+            return None
+        if self._redis is None:
+            self._redis = redis.from_url(self.redis_url, decode_responses=True)
+        return self._redis
+
+    async def _cache_get(self, key: str) -> Optional[str]:
+        r = await self._get_redis()
+        if r is None:
+            return None
+        try:
+            return await r.get(key)
+        except Exception:
+            return None
+
+    async def _cache_set(self, key: str, value: str, ttl: int = CACHE_TTL_SECONDS):
+        r = await self._get_redis()
+        if r is None:
+            return
+        try:
+            await r.setex(key, ttl, value)
+        except Exception:
+            pass
+
+    async def _resolve_active_system_prompt(
+        self,
+        *,
+        tenant_id: Union[str, int],
+        db=None,
+    ) -> tuple[str, str]:
+        tenant_int = _tenant_to_int(tenant_id)
+        now = time.time()
+        cached = self._prompt_profile_cache.get(tenant_int)
+        if cached and cached[0] > now:
+            return cached[1]
+
+        role_name = ""
+        system_prompt = ""
+        try:
+            async def _load_profile(session) -> tuple[str, str]:
+                row = (
+                    await session.execute(
+                        select(Integration).where(
+                            and_(
+                                Integration.tenant_id == int(tenant_int),
+                                Integration.type == "ai_prompt_profiles",
+                            )
+                        )
+                    )
+                ).scalar_one_or_none()
+                bundle = row.config if row is not None and isinstance(row.config, dict) else {}
+                active_key = str(bundle.get("active_role_key") or "").strip()
+                roles = bundle.get("roles") if isinstance(bundle.get("roles"), list) else []
+                selected = None
+                for item in roles:
+                    if not isinstance(item, dict):
+                        continue
+                    if str(item.get("role_key") or "").strip() == active_key:
+                        selected = item
+                        break
+                if selected is None:
+                    for item in roles:
+                        if isinstance(item, dict) and bool(item.get("enabled", True)):
+                            selected = item
+                            break
+                if not isinstance(selected, dict):
+                    return ("", "")
+                return (
+                    str(selected.get("role_name") or "").strip(),
+                    str(selected.get("system_prompt") or "").strip(),
+                )
+
+            if db is not None:
+                role_name, system_prompt = await _load_profile(db)
+            else:
+                async with AsyncSessionLocal() as session:
+                    role_name, system_prompt = await _load_profile(session)
+        except Exception:
+            role_name = ""
+            system_prompt = ""
+
+        if len(system_prompt) < 20:
+            fallback = build_default_role_prompt_bundle()
+            active_key = str(fallback.get("active_role_key") or "").strip()
+            roles = fallback.get("roles") if isinstance(fallback.get("roles"), list) else []
+            selected = None
+            for row in roles:
+                if not isinstance(row, dict):
+                    continue
+                if str(row.get("role_key") or "").strip() == active_key:
+                    selected = row
+                    break
+            if selected is None and roles and isinstance(roles[0], dict):
+                selected = roles[0]
+            if isinstance(selected, dict):
+                role_name = str(selected.get("role_name") or role_name or "默认角色").strip()
+                system_prompt = str(selected.get("system_prompt") or "").strip()
+
+        result = (role_name, system_prompt)
+        self._prompt_profile_cache[tenant_int] = (now + PROMPT_PROFILE_CACHE_TTL_SECONDS, result)
+        return result
+
+    async def _log_call(
+        self,
+        db,
+        tenant_id: str,
+        function_name: str,
+        input_hash: str,
+        status: str,
+        prompt_tokens: int = 0,
+        completion_tokens: int = 0,
+        latency_ms: Optional[int] = None,
+        error_message: Optional[str] = None,
+    ):
+        try:
+            from ..models import AICallLog
+
+            total = prompt_tokens + completion_tokens
+            cost = _estimate_cost(prompt_tokens, completion_tokens)
+            log = AICallLog(
+                tenant_id=_t2s(tenant_id),
+                function_name=function_name,
+                input_hash=input_hash,
+                status=status,
+                prompt_tokens=prompt_tokens,
+                completion_tokens=completion_tokens,
+                total_tokens=total,
+                cost_usd=cost,
+                latency_ms=latency_ms,
+                error_message=error_message,
+            )
+            db.add(log)
+            await db.commit()
+        except Exception:
+            pass
+
+    async def _invoke_with_retry(
+        self,
+        prompt: str,
+        db=None,
+        tenant_id: str = "default",
+        function_name: str = "unknown",
+        use_cache: bool = True,
+        cache_key: Optional[str] = None,
+        project_id: Optional[int] = None,
+    ) -> str:
+        """带缓存、重试、日志的 LLM 调用"""
+        input_hash = cache_key or _hash_input(function_name, prompt)
+
+        if use_cache and db is None:
+            cached = await self._cache_get(input_hash)
+            if cached:
+                return cached
+
+        # 预算前置检查（仅在有 db 会话时启用）
+        if db is not None:
+            try:
+                from .budget_enforcement_service import BudgetEnforcementService
+
+                tenant_int = _tenant_to_int(tenant_id)
+                # 预估：按 prompt 长度粗略估算，保守给定 completion 预算
+                estimated_prompt_tokens = max(len(prompt) // 4, 1)
+                estimated_completion_tokens = 1200
+                estimated_cost = _estimate_cost(estimated_prompt_tokens, estimated_completion_tokens)
+
+                budget_service = BudgetEnforcementService(db)
+                budget_check = await budget_service.check_budget_before_ai_call(
+                    tenant_id=tenant_int,
+                    estimated_cost_usd=estimated_cost,
+                    project_id=project_id,
+                )
+                if not budget_check.get("allowed", True):
+                    raise RuntimeError("预算超限，AI 调用被阻止（hard limit）")
+            except RuntimeError:
+                raise
+            except Exception:
+                # 预算系统异常不阻断主流程，避免影响现有可用性
+                pass
+
+        llm = self._get_llm()
+        _, system_prompt = await self._resolve_active_system_prompt(
+            tenant_id=tenant_id,
+            db=db,
+        )
+        last_err = None
+        for attempt in range(self.max_retries):
+            try:
+                t0 = time.perf_counter()
+                messages = [HumanMessage(content=prompt)]
+                if system_prompt:
+                    messages = [SystemMessage(content=system_prompt), *messages]
+                resp = await llm.ainvoke(messages)
+                latency_ms = int((time.perf_counter() - t0) * 1000)
+                text = resp.content if hasattr(resp, "content") else str(resp)
+                pt = getattr(resp, "response_metadata", None) or {}
+                usage = (pt or {}).get("usage", {}) or (getattr(resp, "response_metadata", None) or {}).get("usage", {})
+                prompt_tokens = usage.get("input_tokens", 0) or usage.get("prompt_tokens", 0)
+                completion_tokens = usage.get("output_tokens", 0) or usage.get("completion_tokens", 0)
+
+                if db:
+                    await self._log_call(
+                        db, tenant_id, function_name, input_hash, "success",
+                        prompt_tokens, completion_tokens, latency_ms,
+                    )
+                    # 按实际 token 成本累计预算
+                    try:
+                        from .budget_enforcement_service import BudgetEnforcementService
+
+                        actual_cost = _estimate_cost(prompt_tokens, completion_tokens)
+                        await BudgetEnforcementService(db).update_budget_usage_after_ai_call(
+                            tenant_id=_tenant_to_int(tenant_id),
+                            actual_cost_usd=actual_cost,
+                            project_id=project_id,
+                        )
+                    except Exception:
+                        pass
+                if use_cache:
+                    await self._cache_set(input_hash, text)
+                return text
+            except Exception as e:
+                last_err = e
+                wait = 2 ** attempt
+                time.sleep(wait)
+
+        if db:
+            await self._log_call(
+                db, tenant_id, function_name, input_hash, "failed",
+                error_message=str(last_err),
+            )
+        raise last_err
+
+    def _parse_json(self, text: str) -> dict:
+        text = text.strip()
+        for p in ("```json", "```"):
+            if text.startswith(p):
+                text = text[len(p):]
+            if text.endswith("```"):
+                text = text[:-3]
+        return json.loads(text.strip())
+
+    # ========================================================================
+    # 核心函数
+    # ========================================================================
+
+    async def generate_charter_from_signal(
+        self,
+        signal_content: str,
+        tenant_id: Union[str, int] = "default",
+        db=None,
+    ) -> dict:
+        """从市场/模型信号生成需求宪章草案"""
+        from ..config.ai_prompts import PROMPT_GENERATE_CHARTER
+
+        tid = _t2s(tenant_id)
+        prompt = PROMPT_GENERATE_CHARTER.format(signal_content=signal_content)
+        text = await self._invoke_with_retry(
+            prompt, db=db, tenant_id=tid, function_name="generate_charter_from_signal",
+        )
+        return self._parse_json(text)
+
+    async def decompose_tasks_from_charter(
+        self,
+        charter_content: str,
+        project_id: Union[str, int],
+        tenant_id: Union[str, int] = "default",
+        db=None,
+    ) -> dict:
+        """从宪章拆解任务，自动识别所需技能标签，返回带技能要求的任务列表"""
+        from ..config.ai_prompts import PROMPT_DECOMPOSE_TASKS
+
+        tid = _t2s(tenant_id)
+        prompt = PROMPT_DECOMPOSE_TASKS.format(
+            charter_content=charter_content,
+            project_id=_t2s(project_id),
+        )
+        text = await self._invoke_with_retry(
+            prompt, db=db, tenant_id=tid, function_name="decompose_tasks_from_charter",
+        )
+        return self._parse_json(text)
+
+    async def recommend_assignees_for_task(
+        self,
+        task_content: str,
+        required_skill_tag_ids: list,
+        project_id: Union[str, int],
+        tenant_id: Union[str, int] = "default",
+        db=None,
+        candidates_json: Optional[str] = None,
+    ) -> dict:
+        """基于人员能力图谱、负载、熟练度，推荐 Top3 人员"""
+        from ..config.ai_prompts import PROMPT_RECOMMEND_ASSIGNEES
+
+        tid = _t2s(tenant_id)
+        candidates = candidates_json or "[]"
+        prompt = PROMPT_RECOMMEND_ASSIGNEES.format(
+            task_content=task_content,
+            required_skill_tag_ids=json.dumps(required_skill_tag_ids),
+            candidates_json=candidates,
+        )
+        text = await self._invoke_with_retry(
+            prompt, db=db, tenant_id=tid, function_name="recommend_assignees_for_task",
+        )
+        return self._parse_json(text)
+
+    async def update_user_skill_scores(
+        self,
+        task_id: Union[str, int],
+        tenant_id: Union[str, int] = "default",
+        db=None,
+        quality_score: float = 0.0,
+        efficiency_score: float = 0.0,
+        assignee_ids: Optional[list] = None,
+        skill_tag_ids: Optional[list] = None,
+    ) -> dict:
+        """任务完成后，根据质量/效率评分更新相关人员技能分"""
+        from sqlalchemy import select
+        from ..models import UserSkillBinding
+        from uuid import UUID
+
+        def to_uuid(v):
+            if v is None:
+                return None
+            if isinstance(v, UUID):
+                return v
+            s = str(v)
+            try:
+                return UUID(s) if len(s) == 36 else None
+            except ValueError:
+                return None
+
+        if db is None:
+            return {"updated_count": 0, "updates": []}
+
+        tid = _t2s(tenant_id)
+        updates = []
+        if assignee_ids and skill_tag_ids:
+            for uid in assignee_ids:
+                for sid in skill_tag_ids:
+                    uid_u = to_uuid(uid)
+                    sid_u = to_uuid(sid)
+                    if not uid_u or not sid_u:
+                        continue
+                    try:
+                        r = await db.execute(
+                            select(UserSkillBinding).where(
+                                UserSkillBinding.tenant_id == tid,
+                                UserSkillBinding.user_id == uid_u,
+                                UserSkillBinding.skill_tag_id == sid_u,
+                            )
+                        )
+                        b = r.scalar_one_or_none()
+                        if b:
+                            b.quality_score = min(100, max(0, (b.quality_score or 0) + quality_score))
+                            b.efficiency_score = min(100, max(0, (b.efficiency_score or 0) + efficiency_score))
+                            b.last_used_at = datetime.utcnow()
+                            await db.commit()
+                            updates.append({
+                                "user_id": str(uid),
+                                "skill_tag_id": str(sid),
+                                "quality_score_delta": quality_score,
+                                "efficiency_score_delta": efficiency_score,
+                                "last_used_at": b.last_used_at.isoformat(),
+                            })
+                    except Exception:
+                        pass
+
+        return {"updated_count": len(updates), "updates": updates}
+
+    async def calculate_alignment_score(
+        self,
+        user_id: Union[str, int],
+        truth_node_id: Union[str, int],
+        tenant_id: Union[str, int] = "default",
+        db=None,
+        truth_content: str = "",
+        user_context: str = "",
+    ) -> dict:
+        """计算用户对真理节点的认知对齐度（0-100）"""
+        from ..config.ai_prompts import PROMPT_ALIGNMENT_SCORE
+        from sqlalchemy import select
+        from ..models import TruthNode
+
+        tid = _t2s(tenant_id)
+        if db and not truth_content:
+            r = await db.execute(
+                select(TruthNode).where(
+                    TruthNode.id == truth_node_id,
+                    TruthNode.tenant_id == tid,
+                )
+            )
+            tn = r.scalar_one_or_none()
+            if tn:
+                truth_content = tn.content or tn.title or ""
+
+        prompt = PROMPT_ALIGNMENT_SCORE.format(
+            user_id=_t2s(user_id),
+            truth_node_id=_t2s(truth_node_id),
+            truth_content=truth_content or "(无内容)",
+            user_context=user_context or "(无)",
+        )
+        text = await self._invoke_with_retry(
+            prompt, db=db, tenant_id=tid, function_name="calculate_alignment_score",
+        )
+        data = self._parse_json(text)
+        data["user_id"] = _t2s(user_id)
+        data["truth_node_id"] = _t2s(truth_node_id)
+        return data
+
+    async def detect_cognitive_conflicts(
+        self,
+        truth_node_id: Union[str, int],
+        tenant_id: Union[str, int] = "default",
+        db=None,
+        truth_content: str = "",
+        user_viewpoints: str = "",
+    ) -> dict:
+        """识别真理节点下的隐性认知冲突"""
+        from ..config.ai_prompts import PROMPT_COGNITIVE_CONFLICTS
+        from sqlalchemy import select
+        from ..models import TruthNode
+
+        tid = _t2s(tenant_id)
+        if db and not truth_content:
+            r = await db.execute(
+                select(TruthNode).where(
+                    TruthNode.id == truth_node_id,
+                    TruthNode.tenant_id == tid,
+                )
+            )
+            tn = r.scalar_one_or_none()
+            if tn:
+                truth_content = tn.content or tn.title or ""
+
+        prompt = PROMPT_COGNITIVE_CONFLICTS.format(
+            truth_node_id=_t2s(truth_node_id),
+            truth_content=truth_content or "(无内容)",
+            user_viewpoints=user_viewpoints or "(无)",
+        )
+        text = await self._invoke_with_retry(
+            prompt, db=db, tenant_id=tid, function_name="detect_cognitive_conflicts",
+        )
+        data = self._parse_json(text)
+        data["truth_node_id"] = _t2s(truth_node_id)
+        data["total_count"] = len(data.get("conflicts", []))
+        return data
+
+    async def evaluate_model_requirement(
+        self,
+        charter_content: str,
+        model_version_id: Union[str, int],
+        tenant_id: Union[str, int] = "default",
+        db=None,
+        model_info: str = "",
+    ) -> dict:
+        """评估需求与模型的匹配度"""
+        from ..config.ai_prompts import PROMPT_EVALUATE_MODEL_REQUIREMENT
+
+        tid = _t2s(tenant_id)
+        prompt = PROMPT_EVALUATE_MODEL_REQUIREMENT.format(
+            charter_content=charter_content,
+            model_version_id=_t2s(model_version_id),
+            model_info=model_info or "(未提供)",
+        )
+        text = await self._invoke_with_retry(
+            prompt, db=db, tenant_id=tid, function_name="evaluate_model_requirement",
+        )
+        return self._parse_json(text)
+
+    # ========================================================================
+    # 流式调用
+    # ========================================================================
+
+    async def stream_chat(self, prompt: str) -> AsyncGenerator[str, None]:
+        """流式生成"""
+        llm = self._get_llm()
+        async for chunk in llm.astream([HumanMessage(content=prompt)]):
+            if hasattr(chunk, "content") and chunk.content:
+                yield chunk.content
+
+
+# 单例
+_ai_service: Optional[AIService] = None
+
+
+def get_ai_service() -> AIService:
+    global _ai_service
+    if _ai_service is None:
+        _ai_service = AIService()
+    return _ai_service
diff --git a/backend/app/services/ai_steward.py b/backend/app/services/ai_steward.py
new file mode 100644
index 00000000..939eb742
--- /dev/null
+++ b/backend/app/services/ai_steward.py
@@ -0,0 +1,656 @@
+"""AI Steward Service - 主动式 AI 管家核心服务。"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+from typing import Any, Dict, List
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..core.logging import get_logger
+from ..database import get_db
+from ..models import CognitiveConflict, Project, ProjectSignalLink, Signal, TruthNode
+from ..utils.project_members import ensure_project_default_members
+from ..utils.sequence import ensure_project_id_sequence
+from .adaptive_rag_service import AdaptiveRAGService
+from .llm import LLMService
+from .task_persistence import (
+    build_task_create_draft,
+    normalize_skill_names,
+    persist_task,
+    resolve_skill_tag_ids_by_names,
+)
+
+logger = get_logger(__name__)
+PROJECT_CREATION_DRAFT_TIMEOUT_SECONDS = 8.0
+
+
+class AIStewardService:
+    """主动式 AI 管家服务。"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+        self.llm = LLMService()
+
+    async def create_project_from_natural_language(
+        self,
+        input_text: str,
+        creator_id: int,
+        tenant_id: int = 1,
+    ) -> Dict[str, Any]:
+        """零配置项目创建，统一走 draft -> normalize -> persist 三段式。"""
+        analysis = await self._analyze_project_input(input_text, tenant_id)
+        draft = await self._generate_project_creation_draft(
+            input_text=input_text,
+            analysis=analysis,
+            tenant_id=tenant_id,
+        )
+        normalized = await self._normalize_project_creation_draft(
+            draft=draft,
+            input_text=input_text,
+            creator_id=creator_id,
+            tenant_id=tenant_id,
+        )
+        payload = await self._persist_project_creation_draft(
+            normalized=normalized,
+            input_text=input_text,
+            creator_id=creator_id,
+            tenant_id=tenant_id,
+        )
+
+        await self.db.commit()
+
+        try:
+            charter = payload["charter"]
+            project = payload["project"]
+            asyncio.create_task(
+                AdaptiveRAGService.upsert_pmflow_record_async(
+                    tenant_id=tenant_id,
+                    record_type="charter",
+                    record_id=int(charter["id"]),
+                    project_id=int(project["id"]),
+                    title=str(charter.get("title") or "项目宪章"),
+                    content=json.dumps(charter.get("content") or {}, ensure_ascii=False),
+                )
+            )
+            for task in payload.get("tasks", []):
+                task_id = int(str(task.get("id") or "0") or 0)
+                if task_id <= 0:
+                    continue
+                asyncio.create_task(
+                    AdaptiveRAGService.upsert_pmflow_record_async(
+                        tenant_id=tenant_id,
+                        record_type="task",
+                        record_id=task_id,
+                        project_id=int(project["id"]),
+                        title=str(task.get("title") or "任务"),
+                        content=str(task.get("description") or ""),
+                    )
+                )
+        except Exception:
+            pass
+
+        return payload
+
+    async def _generate_project_creation_draft(
+        self,
+        *,
+        input_text: str,
+        analysis: Dict[str, Any],
+        tenant_id: int,
+    ) -> dict[str, Any]:
+        signal_draft, charter_draft, task_drafts = await asyncio.gather(
+            self._generate_competitive_signal_draft(analysis=analysis, tenant_id=tenant_id),
+            self._generate_charter_draft(analysis=analysis, tenant_id=tenant_id),
+            self._generate_task_chain_draft(analysis=analysis, tenant_id=tenant_id),
+        )
+        return {
+            "analysis": analysis,
+            "project": {
+                "name": analysis.get("project_name") or input_text[:20] or "未命名项目",
+                "title": analysis.get("project_title") or input_text[:50] or "未命名项目",
+                "description": input_text,
+                "assigned_role": analysis.get("primary_role") or "backend",
+                "domain_id": analysis.get("domain_id"),
+            },
+            "signal": signal_draft,
+            "charter": charter_draft,
+            "tasks": task_drafts,
+            "risks": self._build_risk_alert_drafts(analysis=analysis),
+        }
+
+    async def _generate_project_creation_text(
+        self,
+        *,
+        prompt: str,
+        tenant_id: int,
+        function_name: str,
+        timeout_seconds: float = PROJECT_CREATION_DRAFT_TIMEOUT_SECONDS,
+    ) -> str:
+        try:
+            return await asyncio.wait_for(
+                self.llm.generate(
+                    prompt,
+                    db=self.db,
+                    tenant_id=tenant_id,
+                    function_name=function_name,
+                ),
+                timeout=timeout_seconds,
+            )
+        except Exception as exc:
+            logger.warning(
+                "ai_steward_project_creation_step_fallback",
+                tenant_id=int(tenant_id),
+                function_name=function_name,
+                error=str(exc),
+            )
+            return ""
+
+    async def _normalize_project_creation_draft(
+        self,
+        *,
+        draft: dict[str, Any],
+        input_text: str,
+        creator_id: int,
+        tenant_id: int,
+    ) -> dict[str, Any]:
+        analysis = draft.get("analysis") if isinstance(draft.get("analysis"), dict) else {}
+        project_draft = draft.get("project") if isinstance(draft.get("project"), dict) else {}
+        charter_draft = draft.get("charter") if isinstance(draft.get("charter"), dict) else {}
+        signal_draft = draft.get("signal") if isinstance(draft.get("signal"), dict) else {}
+        risk_drafts = draft.get("risks") if isinstance(draft.get("risks"), list) else []
+
+        normalized_task_drafts = []
+        task_audit: list[dict[str, Any]] = []
+        for raw_task in (draft.get("tasks") if isinstance(draft.get("tasks"), list) else [])[:12]:
+            if not isinstance(raw_task, dict):
+                continue
+            raw_required_skills = normalize_skill_names(raw_task.get("required_skills"))
+            resolution = await resolve_skill_tag_ids_by_names(
+                self.db,
+                tenant_id=tenant_id,
+                skill_names=raw_required_skills,
+            )
+            validation_error = None
+            try:
+                task_draft = build_task_create_draft(
+                    title=str(raw_task.get("title") or ""),
+                    description=str(raw_task.get("description") or "").strip() or None,
+                    required_skill_tag_ids=resolution.matched_ids,
+                    priority=raw_task.get("priority"),
+                    status="todo",
+                    estimated_hours=raw_task.get("estimated_hours"),
+                    due_date=raw_task.get("due_date"),
+                )
+            except ValueError as exc:
+                validation_error = str(exc)
+                task_draft = build_task_create_draft(
+                    title=str(raw_task.get("title") or ""),
+                    description=str(raw_task.get("description") or "").strip() or None,
+                    required_skill_tag_ids=resolution.matched_ids,
+                    priority=raw_task.get("priority"),
+                    status="todo",
+                    estimated_hours=None,
+                    due_date=None,
+                )
+            normalized_task_drafts.append(task_draft)
+            task_audit.append(
+                {
+                    "raw_task_draft": raw_task,
+                    "normalized_title": task_draft.title,
+                    "required_skills_input": raw_required_skills,
+                    "required_skill_tag_ids": resolution.matched_ids,
+                    "skill_resolution": resolution.matched_items,
+                    "dropped_required_skills": resolution.unmatched_names,
+                    "validation_error": validation_error,
+                }
+            )
+
+        normalized_project = {
+            "name": str(project_draft.get("name") or "未命名项目").strip()[:50] or "未命名项目",
+            "title": str(project_draft.get("title") or input_text[:50] or "未命名项目").strip()[:50] or "未命名项目",
+            "description": str(project_draft.get("description") or input_text).strip()[:4000] or input_text[:4000],
+            "assigned_role": str(project_draft.get("assigned_role") or "backend").strip()[:50] or "backend",
+            "domain_id": analysis.get("domain_id"),
+        }
+        normalized_charter_content = charter_draft.get("content") if isinstance(charter_draft.get("content"), dict) else {
+            "rationale": str(charter_draft.get("content") or ""),
+            "success_criteria": charter_draft.get("success_criteria") or [],
+            "constraints": charter_draft.get("constraints") or [],
+            "signal_ids": [],
+        }
+        normalized_signal = {
+            "content": str(signal_draft.get("content") or "").strip()[:2000] or f"竞品分析：{normalized_project['title']}",
+            "impact_score": float(signal_draft.get("impact_score") or 0.8),
+            "priority": str(signal_draft.get("priority") or "high").strip().lower() or "high",
+            "strategic_analysis": signal_draft.get("strategic_analysis") if isinstance(signal_draft.get("strategic_analysis"), dict) else {},
+        }
+        normalized_risks = [risk for risk in risk_drafts if isinstance(risk, dict)]
+
+        return {
+            "analysis": analysis,
+            "project": normalized_project,
+            "signal": normalized_signal,
+            "charter": {
+                "title": str(charter_draft.get("title") or f"{normalized_project['name']}项目宪章").strip()[:200] or "项目宪章",
+                "content": normalized_charter_content,
+            },
+            "tasks": normalized_task_drafts,
+            "task_audit": task_audit,
+            "risks": normalized_risks,
+            "creator_id": int(creator_id),
+        }
+
+    async def _persist_project_creation_draft(
+        self,
+        *,
+        normalized: dict[str, Any],
+        input_text: str,
+        creator_id: int,
+        tenant_id: int,
+    ) -> dict[str, Any]:
+        project_data = normalized["project"]
+        await ensure_project_id_sequence(self.db)
+        project = Project(
+            name=project_data["name"],
+            title=project_data["title"],
+            description=project_data["description"],
+            owner_id=int(creator_id),
+            status="active",
+            domain_id=project_data.get("domain_id"),
+            assigned_role=project_data["assigned_role"],
+            signal_count=0,
+            tenant_id=int(tenant_id),
+        )
+        self.db.add(project)
+        await self.db.flush()
+        await ensure_project_default_members(
+            self.db,
+            tenant_id=int(tenant_id),
+            project_id=int(project.id),
+            owner_id=int(creator_id),
+            assigned_role=project.assigned_role,
+        )
+
+        signal_data = normalized["signal"]
+        signal = Signal(
+            source="AI-Steward",
+            content=signal_data["content"],
+            impact_score=signal_data["impact_score"],
+            priority=signal_data["priority"],
+            processed=True,
+            strategic_analysis=json.dumps(signal_data.get("strategic_analysis") or {}, ensure_ascii=False),
+            tasks_generated=json.dumps([]),
+            tenant_id=int(tenant_id),
+        )
+        self.db.add(signal)
+        await self.db.flush()
+
+        self.db.add(
+            ProjectSignalLink(
+                tenant_id=int(tenant_id),
+                project_id=int(project.id),
+                signal_id=int(signal.id),
+                link_type="competitive_analysis",
+                relevance_score=0.9,
+            )
+        )
+
+        charter_data = normalized["charter"]
+        charter = TruthNode(
+            node_type="charter",
+            project_id=int(project.id),
+            created_by=int(creator_id),
+            title=charter_data["title"],
+            content=charter_data["content"],
+            version=1,
+            status="active",
+            tenant_id=int(tenant_id),
+        )
+        self.db.add(charter)
+        await self.db.flush()
+
+        persisted_tasks: list[dict[str, Any]] = []
+        persisted_task_ids: list[int] = []
+        for task_draft in normalized["tasks"]:
+            task = await persist_task(
+                self.db,
+                tenant_id=int(tenant_id),
+                truth_node_id=int(charter.id),
+                project_id=int(project.id),
+                draft=task_draft,
+                actor_user_id=int(creator_id),
+                action="task.create",
+                remark="ai_steward.project_from_text",
+                commit=False,
+                refresh_workload=False,
+            )
+            persisted_task_ids.append(int(task.id))
+            persisted_tasks.append(
+                {
+                    "id": str(task.id),
+                    "title": task.title,
+                    "description": task.description or "",
+                    "estimated_hours": float(task.estimated_hours or 0),
+                    "priority": task.priority,
+                    "status": task.status,
+                }
+            )
+
+        risks_payload: list[dict[str, Any]] = []
+        for risk in normalized["risks"]:
+            description = str(risk.get("description") or "").strip()
+            conflict_type = str(risk.get("type") or "risk").strip() or "risk"
+            if not description:
+                continue
+            self.db.add(
+                CognitiveConflict(
+                    truth_node_id=int(charter.id),
+                    conflict_type=conflict_type,
+                    description=description,
+                    involved_user_ids=[],
+                    status="active",
+                    tenant_id=int(tenant_id),
+                )
+            )
+            risks_payload.append(
+                {
+                    "type": conflict_type,
+                    "description": description,
+                    "threshold_hours": risk.get("threshold_hours"),
+                    "threshold": risk.get("threshold"),
+                    "severity": risk.get("severity"),
+                }
+            )
+
+        logger.info(
+            "ai_steward_project_from_text_persisted",
+            tenant_id=int(tenant_id),
+            creator_id=int(creator_id),
+            project_id=int(project.id),
+            signal_id=int(signal.id),
+            charter_id=int(charter.id),
+            persisted_task_ids=persisted_task_ids,
+            task_audit=normalized["task_audit"],
+            input_text=input_text[:500],
+        )
+
+        return {
+            "project": {
+                "id": str(project.id),
+                "name": project.name,
+                "title": project.title,
+                "description": project.description,
+                "status": project.status,
+            },
+            "signal": {
+                "id": str(signal.id),
+                "content": signal.content,
+                "impact_score": signal.impact_score,
+            },
+            "charter": {
+                "id": str(charter.id),
+                "title": charter.title,
+                "content": charter.content,
+            },
+            "tasks": persisted_tasks,
+            "risks": risks_payload,
+            "analysis": normalized["analysis"],
+        }
+
+    async def _analyze_project_input(self, input_text: str, tenant_id: int) -> Dict[str, Any]:
+        """使用 AI 分析项目输入，提取关键信息。"""
+        prompt = f"""分析以下项目需求，提取关键信息并以 JSON 格式返回：
+
+用户输入：{input_text}
+
+请提取以下信息：
+{{
+    "project_name": "项目名称（简短）",
+    "project_title": "项目标题",
+    "tech_stack": ["技术栈列表"],
+    "target_audience": "目标用户",
+    "timeline": "时间周期",
+    "primary_role": "主要角色类型（ai_engineer/backend/frontend）",
+    "domain_id": 1,
+    "key_features": ["核心功能列表"],
+    "competitors": ["竞品列表"],
+    "success_metrics": ["成功指标"]
+}}
+
+只返回 JSON，不要其他内容。"""
+
+        response = await self._generate_project_creation_text(
+            prompt=prompt,
+            tenant_id=tenant_id,
+            function_name="project_from_text_analysis",
+        )
+        try:
+            response = response.strip()
+            if response.startswith("```"):
+                response = response.split("```")[1]
+                if response.startswith("json"):
+                    response = response[4:]
+            analysis = json.loads(response.strip())
+        except Exception:
+            analysis = {
+                "project_name": input_text[:20],
+                "project_title": input_text[:50],
+                "tech_stack": ["GPT-4o-mini"],
+                "target_audience": "通用",
+                "timeline": "3周",
+                "primary_role": "ai_engineer",
+                "domain_id": 1,
+                "key_features": [],
+                "competitors": [],
+                "success_metrics": [],
+            }
+        return analysis
+
+    async def _generate_competitive_signal_draft(
+        self,
+        *,
+        analysis: Dict[str, Any],
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        competitors = analysis.get("competitors", [])
+        if not competitors:
+            competitors = ["行业同类产品"]
+
+        prompt = f"""基于以下项目信息，生成竞品分析信号：
+
+项目：{analysis.get('project_title')}
+技术栈：{', '.join(analysis.get('tech_stack', []))}
+竞品：{', '.join(competitors)}
+
+请生成：
+1. 竞品现状分析
+2. 差异化机会
+3. 技术风险评估
+
+以 JSON 格式返回：
+{{
+    "content": "信号内容（200字以内）",
+    "impact_score": 0.9,
+    "priority": "high",
+    "strategic_analysis": {{"is_strategic": true, "threat_level": "MEDIUM"}}
+}}"""
+
+        response = await self._generate_project_creation_text(
+            prompt=prompt,
+            tenant_id=tenant_id,
+            function_name="project_from_text_signal",
+        )
+        try:
+            response = response.strip()
+            if response.startswith("```"):
+                response = response.split("```")[1]
+                if response.startswith("json"):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except Exception:
+            return {
+                "content": f"竞品分析：{', '.join(competitors)} 的技术现状和差异化机会",
+                "impact_score": 0.8,
+                "priority": "high",
+                "strategic_analysis": {"is_strategic": True},
+            }
+
+    async def _generate_charter_draft(
+        self,
+        *,
+        analysis: Dict[str, Any],
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        prompt = f"""基于以下项目信息，生成项目宪章：
+
+项目：{analysis.get('project_title')}
+描述：{analysis.get('project_title')}
+技术栈：{', '.join(analysis.get('tech_stack', []))}
+核心功能：{', '.join(analysis.get('key_features', []))}
+成功指标：{', '.join(analysis.get('success_metrics', []))}
+
+请生成包含以下内容的宪章：
+1. 项目目标（明确、可衡量）
+2. 成功标准（具体指标）
+3. 技术约束
+4. 时间节点
+
+以 JSON 格式返回：
+{{
+    "title": "宪章标题",
+    "content": "宪章完整内容（包含上述4个方面）",
+    "success_criteria": ["标准1", "标准2"],
+    "constraints": ["约束1", "约束2"]
+}}"""
+
+        response = await self._generate_project_creation_text(
+            prompt=prompt,
+            tenant_id=tenant_id,
+            function_name="project_from_text_charter",
+        )
+        try:
+            response = response.strip()
+            if response.startswith("```"):
+                response = response.split("```")[1]
+                if response.startswith("json"):
+                    response = response[4:]
+            charter_data = json.loads(response.strip())
+        except Exception:
+            charter_data = {
+                "title": f"{analysis.get('project_name')}项目宪章",
+                "content": f"项目目标：{analysis.get('project_title')}\n成功标准：按时上线",
+                "success_criteria": ["按时交付", "功能完整"],
+                "constraints": ["时间限制"],
+            }
+
+        raw_content = charter_data.get("content", "")
+        if isinstance(raw_content, dict):
+            content = dict(raw_content)
+        else:
+            content = {
+                "rationale": str(raw_content),
+                "success_criteria": charter_data.get("success_criteria", []),
+                "constraints": charter_data.get("constraints", []),
+                "signal_ids": [],
+            }
+        return {
+            "title": charter_data.get("title", "项目宪章"),
+            "content": content,
+        }
+
+    async def _generate_task_chain_draft(
+        self,
+        *,
+        analysis: Dict[str, Any],
+        tenant_id: int,
+    ) -> List[Dict[str, Any]]:
+        prompt = f"""基于以下信息，生成完整的任务链：
+
+项目：{analysis.get('project_title')}
+技术栈：{', '.join(analysis.get('tech_stack', []))}
+时间周期：{analysis.get('timeline', '3周')}
+
+请生成：
+1. 任务列表（按优先级排序）
+2. 每个任务包含：名称、类型、预估工时、依赖关系、技能要求
+
+以 JSON 格式返回：
+{{
+    "tasks": [
+        {{
+            "title": "任务名称",
+            "type": "backend/frontend/ai_engineer",
+            "estimated_hours": 8,
+            "description": "任务描述",
+            "priority": "high/medium/low",
+            "required_skills": ["技能1", "技能2"]
+        }}
+    ]
+}}"""
+
+        response = await self._generate_project_creation_text(
+            prompt=prompt,
+            tenant_id=tenant_id,
+            function_name="project_from_text_task_chain",
+        )
+        try:
+            response = response.strip()
+            if response.startswith("```"):
+                response = response.split("```")[1]
+                if response.startswith("json"):
+                    response = response[4:]
+            tasks_data = json.loads(response.strip())
+            tasks_list = tasks_data.get("tasks", [])
+            if isinstance(tasks_list, list) and tasks_list:
+                return tasks_list
+        except Exception:
+            pass
+        return [
+            {
+                "title": "技术方案设计",
+                "type": analysis.get("primary_role", "backend"),
+                "estimated_hours": 16,
+                "description": "完成项目技术方案设计",
+                "priority": "high",
+                "required_skills": ["系统设计"],
+            },
+            {
+                "title": "核心功能开发",
+                "type": analysis.get("primary_role", "backend"),
+                "estimated_hours": 40,
+                "description": "完成核心功能开发",
+                "priority": "high",
+                "required_skills": ["开发"],
+            },
+        ]
+
+    @staticmethod
+    def _build_risk_alert_drafts(*, analysis: Dict[str, Any]) -> List[Dict[str, Any]]:
+        estimated_hours = analysis.get("estimated_hours", 40)
+        return [
+            {
+                "type": "schedule_delay",
+                "description": "任务超时预警",
+                "threshold_hours": estimated_hours * 1.2,
+                "severity": "medium",
+            },
+            {
+                "type": "quality_drop",
+                "description": "质量下降预警",
+                "threshold": 0.85,
+                "severity": "high",
+            },
+            {
+                "type": "skill_mismatch",
+                "description": "技能匹配度不足",
+                "threshold": 0.6,
+                "severity": "medium",
+            },
+        ]
+
+
+def get_ai_steward_service(db: AsyncSession) -> AIStewardService:
+    """获取 AI Steward 服务实例。"""
+    return AIStewardService(db)
diff --git a/backend/app/services/alignment_orchestrator.py b/backend/app/services/alignment_orchestrator.py
new file mode 100644
index 00000000..e69de29b
diff --git a/backend/app/services/analyzer.py b/backend/app/services/analyzer.py
new file mode 100644
index 00000000..9dd9ae10
--- /dev/null
+++ b/backend/app/services/analyzer.py
@@ -0,0 +1,12 @@
+"""信号影响评分分析（轻量实现）。"""
+
+
+def analyze_impact(content: str) -> float:
+    text = (content or "").strip()
+    if not text:
+        return 0.1
+    if len(text) > 300:
+        return 0.8
+    if len(text) > 120:
+        return 0.6
+    return 0.4
diff --git a/backend/app/services/auditor.py b/backend/app/services/auditor.py
new file mode 100644
index 00000000..09203e8e
--- /dev/null
+++ b/backend/app/services/auditor.py
@@ -0,0 +1,635 @@
+"""
+Neural Sync - Enhanced Code Auditor Service
+
+增强版代码审计服务 - 支持严格模式的自动合并
+"""
+
+import asyncio
+import re
+import logging
+from typing import Dict, List, Optional, Any, Tuple
+from pydantic import BaseModel, Field
+from datetime import datetime
+from enum import Enum
+
+logger = logging.getLogger(__name__)
+
+
+class AuditSeverity(str, Enum):
+    """审计问题严重程度"""
+    CRITICAL = "critical"  # 阻止合并
+    HIGH = "high"        # 大幅扣分
+    MEDIUM = "medium"    # 中度扣分
+    LOW = "low"          # 轻微扣分
+    INFO = "info"        # 仅提示
+
+
+class AuditIssue(BaseModel):
+    """审计问题"""
+    severity: AuditSeverity
+    category: str  # security, performance, style, logic, etc.
+    file_path: str
+    line_number: Optional[int] = None
+    message: str
+    suggestion: Optional[str] = None
+    score_penalty: float = 0.0
+
+
+class AuditResult(BaseModel):
+    """审计结果"""
+    signal_id: Optional[str] = None
+    commit_sha: Optional[str] = None
+    task_name: Optional[str] = None
+    passed: bool
+    score: float = Field(default=0.0, ge=0.0, le=100.0)
+    issues: List[AuditIssue] = Field(default_factory=list)
+    summary: str = ""
+    details: Dict[str, Any] = Field(default_factory=dict)
+    timestamp: float = Field(default_factory=lambda: datetime.utcnow().timestamp())
+
+
+class SecurityIssue(BaseModel):
+    """安全问题"""
+    type: str  # sql_injection, xss, csrf, etc.
+    severity: AuditSeverity
+    file_path: str
+    line_number: int
+    code_snippet: str
+    description: str
+    fix_suggestion: str
+
+
+class CodeAuditor:
+    """
+    增强版代码审计器
+
+    特性:
+    - 安全漏洞扫描
+    - 代码质量评估
+    - 性能问题检测
+    - 最佳实践检查
+    - 严格评分机制（98+ 分阈值）
+    """
+
+    # 审计权重
+    WEIGHTS = {
+        "security": 30,      # 安全性最重要
+        "functionality": 25, # 功能正确性
+        "performance": 15,   # 性能优化
+        "style": 10,         # 代码风格
+        "documentation": 10, # 文档完整性
+        "testing": 10,       # 测试覆盖
+    }
+
+    # 安全模式库
+    SECURITY_PATTERNS = {
+        "sql_injection": [
+            r'execute\s*\(\s*["\'].*?\+.*?["\']',  # 字符串拼接 SQL
+            r'cursor\.execute\s*\(\s*["\'].*?%s.*?["\']',  # %s 格式化
+            r'\$\{.*\}.*?(?:SELECT|INSERT|UPDATE|DELETE)',  # 模板字符串 SQL
+        ],
+        "xss": [
+            r'innerHTML\s*=\s*.*?\+.*?(?:request|input|form)',
+            r'eval\s*\(',
+            r'document\.write\s*\(',
+        ],
+        "hardcoded_secrets": [
+            r'(?:password|secret|api_key|token)\s*=\s*["\'][^"\']+["\']',
+            r'Bearer\s+[A-Za-z0-9]{20,}',
+        ],
+        "command_injection": [
+            r'subprocess\.call\s*\(\s*["\'].*?\+.*?["\']',
+            r'os\.system\s*\(\s*["\'].*?\+.*?["\']',
+        ],
+        "path_traversal": [
+            r'open\s*\(\s*["\'].*?\.\.\/.*?["\']',
+            r'Path\s*\(\s*["\'].*?\.\.\/.*?["\']',
+        ],
+    }
+
+    # 性能问题模式
+    PERFORMANCE_PATTERNS = {
+        "n_plus_1": [
+            r'for\s+\w+\s+in\s+range\s*\(\s*len\s*\(',
+            r'\[[^\]]+\]\s*\.append\s*\(.*?\)\s*for\s+\w+\s+in\s+',
+        ],
+        "inefficient_string_concat": [
+            r'["\'][^"\']+["\']\s*\+.*?\+',
+        ],
+        "unused_import": [],
+        "large_file": None,  # 文件行数检查
+    }
+
+    # 代码风格检查
+    STYLE_PATTERNS = {
+        "line_too_long": 100,  # 最大行长度
+        "function_too_long": 50,  # 最大函数行数
+        "class_too_long": 200,    # 最大类行数
+    }
+
+    def __init__(self, llm_service):
+        self.llm = llm_service
+
+    async def audit_code_change(
+        self,
+        task_name: str,
+        commit_message: str,
+        code_diff: str,
+        signal_id: Optional[str] = None,
+        commit_sha: Optional[str] = None,
+        file_contexts: Optional[Dict[str, str]] = None,
+    ) -> AuditResult:
+        """
+        审计代码变更
+
+        Args:
+            task_name: 任务名称
+            commit_message: 提交消息
+            code_diff: 代码 diff
+            signal_id: 关联的信号 ID
+            commit_sha: 提交 SHA
+            file_contexts: 文件上下文（完整文件内容）
+
+        Returns:
+            审计结果
+        """
+        issues = []
+
+        try:
+            # 解析 diff 获取变更的文件
+            changed_files = self._parse_diff(code_diff)
+
+            # 对每个变更的文件进行审计
+            for file_path, file_changes in changed_files.items():
+                # 获取文件内容（如果提供了上下文）
+                full_content = file_contexts.get(file_path) if file_contexts else None
+
+                # 安全审计
+                security_issues = await self._audit_security(
+                    file_path, file_changes, full_content
+                )
+                issues.extend(security_issues)
+
+                # 性能审计
+                perf_issues = self._audit_performance(
+                    file_path, file_changes, full_content
+                )
+                issues.extend(perf_issues)
+
+                # 代码风格审计
+                style_issues = self._audit_style(
+                    file_path, file_changes, full_content
+                )
+                issues.extend(style_issues)
+
+                # 功能正确性审计（使用 AI）
+                functional_issues = await self._audit_functionality(
+                    task_name, file_path, file_changes, full_content
+                )
+                issues.extend(functional_issues)
+
+            # AI 深度审计
+            ai_issues = await self._ai_deep_audit(
+                task_name, commit_message, code_diff, changed_files
+            )
+            issues.extend(ai_issues)
+
+            # 计算总分
+            score = self._calculate_score(issues)
+
+            # 生成摘要
+            summary = self._generate_summary(issues, score)
+
+            # 判断是否通过（严格模式：98分）
+            passed = score >= 98.0 and not any(
+                i.severity == AuditSeverity.CRITICAL for i in issues
+            )
+
+            return AuditResult(
+                signal_id=signal_id,
+                commit_sha=commit_sha,
+                task_name=task_name,
+                passed=passed,
+                score=score,
+                issues=issues,
+                summary=summary,
+                details={
+                    "changed_files_count": len(changed_files),
+                    "issues_by_severity": self._count_issues_by_severity(issues),
+                    "issues_by_category": self._count_issues_by_category(issues),
+                    "changed_files": list(changed_files.keys()),
+                },
+            )
+
+        except Exception as e:
+            logger.error(f"Audit failed: {e}", exc_info=True)
+            return AuditResult(
+                signal_id=signal_id,
+                commit_sha=commit_sha,
+                task_name=task_name,
+                passed=False,
+                score=0.0,
+                issues=[
+                    AuditIssue(
+                        severity=AuditSeverity.CRITICAL,
+                        category="system",
+                        file_path="unknown",
+                        message=f"审计过程出错: {str(e)}",
+                        score_penalty=100.0,
+                    )
+                ],
+                summary=f"审计失败: {str(e)}",
+            )
+
+    def _parse_diff(self, diff: str) -> Dict[str, str]:
+        """解析 unified diff 格式"""
+        files = {}
+        current_file = None
+        current_content = []
+
+        for line in diff.split("\n"):
+            if line.startswith("diff --git"):
+                # 新文件开始
+                if current_file and current_content:
+                    files[current_file] = "\n".join(current_content)
+
+                # 提取文件路径
+                parts = line.split()
+                if len(parts) >= 4:
+                    current_file = parts[3].lstrip("b/")
+                else:
+                    current_file = parts[2].lstrip("b/")
+                current_content = []
+
+            elif line.startswith("@@") or line.startswith("---") or line.startswith("+++"):
+                continue
+            elif current_file:
+                current_content.append(line)
+
+        # 保存最后一个文件
+        if current_file and current_content:
+            files[current_file] = "\n".join(current_content)
+
+        return files
+
+    async def _audit_security(
+        self, file_path: str, file_changes: str, full_content: Optional[str]
+    ) -> List[AuditIssue]:
+        """安全审计"""
+        issues = []
+        content = full_content or file_changes
+
+        # 检查常见安全问题
+        for vuln_type, patterns in self.SECURITY_PATTERNS.items():
+            for pattern in patterns:
+                matches = re.finditer(pattern, content, re.IGNORECASE | re.MULTILINE)
+                for match in matches:
+                    line_num = content[:match.start()].count("\n") + 1
+
+                    # 计算扣分
+                    if vuln_type in ["sql_injection", "xss", "command_injection"]:
+                        severity = AuditSeverity.CRITICAL
+                        penalty = 20.0
+                    elif vuln_type == "hardcoded_secrets":
+                        severity = AuditSeverity.HIGH
+                        penalty = 10.0
+                    elif vuln_type == "path_traversal":
+                        severity = AuditSeverity.HIGH
+                        penalty = 15.0
+                    else:
+                        severity = AuditSeverity.MEDIUM
+                        penalty = 5.0
+
+                    issues.append(
+                        AuditIssue(
+                            severity=severity,
+                            category="security",
+                            file_path=file_path,
+                            line_number=line_num,
+                            message=f"潜在{vuln_type}漏洞",
+                            suggestion=f"使用参数化查询/输入验证来避免{vuln_type}",
+                            score_penalty=penalty,
+                        )
+                    )
+
+        return issues
+
+    def _audit_performance(
+        self, file_path: str, file_changes: str, full_content: Optional[str]
+    ) -> List[AuditIssue]:
+        """性能审计"""
+        issues = []
+        content = full_content or file_changes
+
+        # 检查 N+1 问题
+        for pattern in self.PERFORMANCE_PATTERNS.get("n_plus_1", []):
+            matches = re.finditer(pattern, content, re.MULTILINE)
+            for match in matches:
+                line_num = content[:match.start()].count("\n") + 1
+                issues.append(
+                    AuditIssue(
+                        severity=AuditSeverity.MEDIUM,
+                        category="performance",
+                        file_path=file_path,
+                        line_number=line_num,
+                        message="可能的 N+1 查询问题",
+                        suggestion="考虑使用 bulk_create 或 select_related 优化",
+                        score_penalty=5.0,
+                    )
+                )
+
+        # 检查低效字符串拼接
+        for pattern in self.PERFORMANCE_PATTERNS.get("inefficient_string_concat", []):
+            matches = re.finditer(pattern, content)
+            for match in matches:
+                line_num = content[:match.start()].count("\n") + 1
+                issues.append(
+                    AuditIssue(
+                        severity=AuditSeverity.LOW,
+                        category="performance",
+                        file_path=file_path,
+                        line_number=line_num,
+                        message="低效的字符串拼接",
+                        suggestion="使用 f-string 或 join()",
+                        score_penalty=2.0,
+                    )
+                )
+
+        # 检查文件大小
+        if full_content:
+            line_count = len(full_content.split("\n"))
+            if line_count > 500:
+                issues.append(
+                    AuditIssue(
+                        severity=AuditSeverity.MEDIUM,
+                        category="performance",
+                        file_path=file_path,
+                        message=f"文件过大 ({line_count} 行)",
+                        suggestion="考虑拆分为更小的模块",
+                        score_penalty=5.0,
+                    )
+                )
+
+        return issues
+
+    def _audit_style(
+        self, file_path: str, file_changes: str, full_content: Optional[str]
+    ) -> List[AuditIssue]:
+        """代码风格审计"""
+        issues = []
+        content = full_content or file_changes
+
+        # 检查行长度
+        max_length = self.STYLE_PATTERNS["line_too_long"]
+        for i, line in enumerate(content.split("\n"), 1):
+            if len(line) > max_length:
+                issues.append(
+                    AuditIssue(
+                        severity=AuditSeverity.INFO,
+                        category="style",
+                        file_path=file_path,
+                        line_number=i,
+                        message=f"行过长 ({len(line)} > {max_length})",
+                        suggestion="拆分为多行",
+                        score_penalty=1.0,
+                    )
+                )
+
+        return issues
+
+    async def _audit_functionality(
+        self, task_name: str, file_path: str, file_changes: str, full_content: Optional[str]
+    ) -> List[AuditIssue]:
+        """功能正确性审计（使用 AI）"""
+        issues = []
+
+        # 构建 AI 审计 prompt
+        prompt = f"""作为一位资深代码审查员，请审计以下代码变更：
+
+任务需求：{task_name}
+
+文件：{file_path}
+
+代码变更：
+```diff
+{file_changes}
+```
+
+{f'完整文件内容：\n```{self._detect_language(file_path)}\n{full_content}\n```' if full_content else ''}
+
+请从以下维度进行审查：
+1. 功能正确性：代码是否正确实现了需求
+2. 逻辑完整性：是否考虑了边界情况
+3. 错误处理：是否有适当的异常处理
+4. 资源管理：是否有资源泄漏风险
+
+返回格式（JSON）：
+```json
+{{
+  "issues": [
+    {{
+      "severity": "critical|high|medium|low|info",
+      "line_number": 123,
+      "message": "问题描述",
+      "suggestion": "改进建议",
+      "score_penalty": 10.0
+    }}
+  ]
+}}
+```
+"""
+
+        try:
+            # 调用 LLM 进行深度审计
+            response = await self.llm._call_llm_chat(
+                messages=[
+                    {"role": "system", "content": "你是一位严格但公平的代码审查专家。"},
+                    {"role": "user", "content": prompt}
+                ],
+                temperature=0.2,
+            )
+
+            # 解析 AI 响应
+            import json
+            json_match = re.search(r'```json\s*([\s\S]*?)\s*```', response)
+            if json_match:
+                ai_result = json.loads(json_match.group(1))
+                for issue_data in ai_result.get("issues", []):
+                    # 限制 AI 生成的问题数量，避免过度敏感
+                    if len(issues) < 10:  # 最多 10 个 AI 发现的问题
+                        severity_map = {
+                            "critical": AuditSeverity.CRITICAL,
+                            "high": AuditSeverity.HIGH,
+                            "medium": AuditSeverity.MEDIUM,
+                            "low": AuditSeverity.LOW,
+                            "info": AuditSeverity.INFO,
+                        }
+                        issues.append(
+                            AuditIssue(
+                                severity=severity_map.get(issue_data.get("severity", "info"), AuditSeverity.INFO),
+                                category="functionality",
+                                file_path=file_path,
+                                line_number=issue_data.get("line_number"),
+                                message=issue_data.get("message", ""),
+                                suggestion=issue_data.get("suggestion", ""),
+                                score_penalty=issue_data.get("score_penalty", 5.0),
+                            )
+                        )
+        except Exception as e:
+            logger.warning(f"AI functionality audit failed: {e}")
+
+        return issues
+
+    async def _ai_deep_audit(
+        self, task_name: str, commit_message: str, code_diff: str, changed_files: Dict[str, str]
+    ) -> List[AuditIssue]:
+        """AI 深度审计 - 跨文件分析"""
+        issues = []
+
+        # 限制文件数量，避免 prompt 过长
+        files_to_audit = list(changed_files.keys())[:5]
+
+        prompt = f"""作为一位资深架构师，请从系统层面审查以下代码变更：
+
+任务：{task_name}
+提交消息：{commit_message}
+
+变更的文件：
+{', '.join(files_to_audit)}
+
+完整 diff：
+```diff
+{code_diff[:2000]}  # 限制长度
+```
+
+请检查：
+1. 架构一致性：变更是否符合项目架构
+2. 依赖关系：是否引入了不兼容的依赖
+3. 配置变更：是否需要更新配置文件
+4. 迁移影响：是否需要数据迁移
+5. 测试覆盖：是否需要添加测试
+
+返回格式（JSON）：
+```json
+{{
+  "issues": [
+    {{
+      "severity": "critical|high|medium|low",
+      "category": "architecture|dependencies|configuration|migration|testing",
+      "message": "问题描述",
+      "suggestion": "建议",
+      "score_penalty": 5.0
+    }}
+  ]
+}}
+```
+"""
+
+        try:
+            response = await self.llm._call_llm_chat(
+                messages=[
+                    {"role": "system", "content": "你是一位经验丰富的软件架构师。"},
+                    {"role": "user", "content": prompt}
+                ],
+                temperature=0.2,
+            )
+
+            import json
+            json_match = re.search(r'```json\s*([\s\S]*?)\s*```', response)
+            if json_match:
+                ai_result = json.loads(json_match.group(1))
+                for issue_data in ai_result.get("issues", []):
+                    if len(issues) < 5:  # 限制数量
+                        severity_map = {
+                            "critical": AuditSeverity.CRITICAL,
+                            "high": AuditSeverity.HIGH,
+                            "medium": AuditSeverity.MEDIUM,
+                            "low": AuditSeverity.LOW,
+                        }
+                        issues.append(
+                            AuditIssue(
+                                severity=severity_map.get(issue_data.get("severity", "low"), AuditSeverity.LOW),
+                                category=issue_data.get("category", "architecture"),
+                                file_path="multiple",
+                                message=issue_data.get("message", ""),
+                                suggestion=issue_data.get("suggestion", ""),
+                                score_penalty=issue_data.get("score_penalty", 3.0),
+                            )
+                        )
+        except Exception as e:
+            logger.warning(f"AI deep audit failed: {e}")
+
+        return issues
+
+    def _calculate_score(self, issues: List[AuditIssue]) -> float:
+        """计算审计分数"""
+        # 初始满分
+        score = 100.0
+
+        # 按权重扣分
+        for issue in issues:
+            weight = self.WEIGHTS.get(issue.category, 10)
+            penalty = issue.score_penalty * (weight / 100)
+            score -= penalty
+
+        return max(0.0, score)
+
+    def _generate_summary(self, issues: List[AuditIssue], score: float) -> str:
+        """生成审计摘要"""
+        if not issues:
+            return f"完美！代码审计通过，得分: {score:.1f}/100"
+
+        # 按严重程度统计
+        critical = sum(1 for i in issues if i.severity == AuditSeverity.CRITICAL)
+        high = sum(1 for i in issues if i.severity == AuditSeverity.HIGH)
+        medium = sum(1 for i in issues if i.severity == AuditSeverity.MEDIUM)
+
+        summary_parts = [f"审计得分: {score:.1f}/100"]
+        if critical > 0:
+            summary_parts.append(f"{critical} 个严重问题")
+        if high > 0:
+            summary_parts.append(f"{high} 个高优先级问题")
+        if medium > 0:
+            summary_parts.append(f"{medium} 个中等问题")
+
+        return " | ".join(summary_parts)
+
+    def _count_issues_by_severity(self, issues: List[AuditIssue]) -> Dict[str, int]:
+        """按严重程度统计问题"""
+        counts = {s.value: 0 for s in AuditSeverity}
+        for issue in issues:
+            counts[issue.severity.value] += 1
+        return counts
+
+    def _count_issues_by_category(self, issues: List[AuditIssue]) -> Dict[str, int]:
+        """按类别统计问题"""
+        counts = {}
+        for issue in issues:
+            counts[issue.category] = counts.get(issue.category, 0) + 1
+        return counts
+
+    def _detect_language(self, file_path: str) -> str:
+        """检测文件语言"""
+        ext = file_path.rsplit(".", 1)[-1].lower()
+        lang_map = {
+            "py": "python",
+            "ts": "typescript",
+            "tsx": "typescript",
+            "js": "javascript",
+            "jsx": "javascript",
+            "vue": "vue",
+            "sql": "sql",
+        }
+        return lang_map.get(ext, "text")
+
+
+# 全局审计器实例
+_auditor = None
+
+
+def get_auditor(llm_service) -> CodeAuditor:
+    """获取代码审计器实例"""
+    global _auditor
+    if _auditor is None:
+        _auditor = CodeAuditor(llm_service)
+    return _auditor
diff --git a/backend/app/services/auth.py b/backend/app/services/auth.py
new file mode 100644
index 00000000..fe4a0492
--- /dev/null
+++ b/backend/app/services/auth.py
@@ -0,0 +1,218 @@
+"""
+Authentication & Authorization Service
+
+Provides password hashing, JWT token management, and permission checking.
+"""
+
+from datetime import datetime, timedelta
+from typing import Optional
+import hashlib
+import hmac
+import os
+
+# Try to import passlib, fallback to simple hashing
+try:
+    from passlib.context import CryptContext
+    from jose import JWTError, jwt
+    HAS_PASSLIB = True
+except ImportError:
+    HAS_PASSLIB = False
+    # Fallback JWT implementation
+    try:
+        import jwt as jwt_lib
+        jwt = jwt_lib
+        JWTError = Exception
+    except ImportError:
+        jwt = None
+        JWTError = Exception
+
+# JWT Configuration
+SECRET_KEY = os.getenv("SECRET_KEY", "neural-sync-secret-key-change-in-production")
+ALGORITHM = "HS256"
+ACCESS_TOKEN_EXPIRE_HOURS = 24
+MAX_BCRYPT_PASSWORD_BYTES = 72
+
+# Password hashing context (if passlib available)
+if HAS_PASSLIB:
+    pwd_context = CryptContext(schemes=["bcrypt"], deprecated="auto")
+
+
+class AuthService:
+    """Service for authentication operations"""
+
+    @staticmethod
+    def password_utf8_bytes(password: str) -> int:
+        return len(str(password or "").encode("utf-8"))
+
+    @staticmethod
+    def validate_password(password: str) -> None:
+        if len(str(password or "")) < 6:
+            raise ValueError("密码至少需要 6 个字符")
+        if AuthService.password_utf8_bytes(password) > MAX_BCRYPT_PASSWORD_BYTES:
+            raise ValueError("密码最长支持 72 个 UTF-8 字节，请缩短后重试")
+
+    @staticmethod
+    def verify_password(plain_password: str, hashed_password: str) -> bool:
+        """Verify a plain password against a hashed password"""
+        if HAS_PASSLIB:
+            try:
+                # 优先走 bcrypt（新数据）
+                if pwd_context.verify(plain_password, hashed_password):
+                    return True
+            except Exception:
+                # 旧库可能是历史 SHA256 哈希，回退兼容
+                pass
+
+            return hmac.compare_digest(
+                AuthService._simple_hash(plain_password),
+                hashed_password,
+            )
+
+        # Simple SHA256 fallback for development
+        return hmac.compare_digest(
+            AuthService._simple_hash(plain_password),
+            hashed_password
+        )
+
+    @staticmethod
+    def hash_password(password: str) -> str:
+        """Hash a password for storage"""
+        AuthService.validate_password(password)
+        if HAS_PASSLIB:
+            try:
+                return pwd_context.hash(password)
+            except ValueError as exc:
+                if "72 bytes" in str(exc).lower() and AuthService.password_utf8_bytes(password) > MAX_BCRYPT_PASSWORD_BYTES:
+                    raise ValueError("密码最长支持 72 个 UTF-8 字节，请缩短后重试") from exc
+                return AuthService._simple_hash(password)
+            except Exception:
+                return AuthService._simple_hash(password)
+        else:
+            # Simple SHA256 fallback for development
+            return AuthService._simple_hash(password)
+
+    @staticmethod
+    def _simple_hash(password: str) -> str:
+        """Simple SHA256 hash with salt for development use"""
+        salt = SECRET_KEY[:16].encode() if SECRET_KEY else b'neural-sync-salt'
+        return hashlib.sha256(password.encode() + salt).hexdigest()
+
+    @staticmethod
+    def create_access_token(data: dict, expires_delta: Optional[timedelta] = None) -> str:
+        """Create a JWT access token"""
+        if jwt is None:
+            # Fallback: return a simple token (not recommended for production)
+            import base64
+            import json
+            token_data = {
+                **data,
+                "exp": (datetime.utcnow() + (expires_delta or timedelta(hours=ACCESS_TOKEN_EXPIRE_HOURS))).timestamp()
+            }
+            return base64.b64encode(json.dumps(token_data).encode()).decode()
+
+        to_encode = data.copy()
+        if expires_delta:
+            expire = datetime.utcnow() + expires_delta
+        else:
+            expire = datetime.utcnow() + timedelta(hours=ACCESS_TOKEN_EXPIRE_HOURS)
+        to_encode.update({"exp": expire})
+        return jwt.encode(to_encode, SECRET_KEY, algorithm=ALGORITHM)
+
+    @staticmethod
+    def verify_token(token: str) -> Optional[str]:
+        """
+        Verify a JWT token and return the username
+
+        Returns None if token is invalid or expired
+        """
+        if jwt is None:
+            # Fallback: decode simple token
+            import base64
+            import json
+            try:
+                token_data = json.loads(base64.b64decode(token).decode())
+                if token_data.get("exp", 0) > datetime.utcnow().timestamp():
+                    return token_data.get("sub")
+                return None
+            except:
+                return None
+
+        try:
+            payload = jwt.decode(token, SECRET_KEY, algorithms=[ALGORITHM])
+            username: str = payload.get("sub")
+            return username
+        except JWTError:
+            return None
+
+
+# Permission matrix for role-based access control
+PERMISSION_MATRIX = {
+    "BOSS": [
+        "view_all_signals",
+        "assign_tasks",
+        "manage_users",
+        "delete_users",
+        "manage_domains",
+        "view_projects",
+        "view_context",
+        "view_settings",
+        "manage_staff",
+    ],
+    "ADMIN": [
+        "view_all_signals",
+        "assign_tasks",
+        "manage_domains",
+        "view_projects",
+        "view_context",
+        "view_settings",
+        "manage_staff",
+    ],
+    "AI": [
+        "view_assigned_signals",
+        "create_experiments",
+        "view_context",
+    ],
+    "FE": [
+        "view_assigned_signals",
+        "view_assigned_projects",
+    ],
+    "BE": [
+        "view_assigned_signals",
+        "view_assigned_projects",
+    ],
+}
+
+
+def has_permission(user_role: str, permission: str) -> bool:
+    """Check if a role has a specific permission"""
+    roles = parse_role_values(user_role)
+    return any(permission in PERMISSION_MATRIX.get(role, []) for role in roles)
+
+
+def has_any_role(user_role: str, allowed_roles: list) -> bool:
+    """Check if user_role is in the list of allowed roles"""
+    allowed = {str(role).upper() for role in allowed_roles}
+    return any(role in allowed for role in parse_role_values(user_role))
+
+
+def parse_role_values(user_role: str | None) -> list[str]:
+    """Parse single/multi role string into uppercase role tokens."""
+    raw = str(user_role or "")
+    if not raw:
+        return []
+    values: list[str] = []
+    for token in raw.split(","):
+        role = token.strip().upper()
+        if role and role not in values:
+            values.append(role)
+    return values
+
+
+def can_manage_users(user_role: str) -> bool:
+    """Check if user can manage other users (BOSS only)"""
+    return has_any_role(user_role, ["BOSS"])
+
+
+def can_assign_tasks(user_role: str) -> bool:
+    """Check if user can assign tasks to others"""
+    return has_any_role(user_role, ["BOSS", "ADMIN"])
diff --git a/backend/app/services/auto_merger.py b/backend/app/services/auto_merger.py
new file mode 100644
index 00000000..df92096a
--- /dev/null
+++ b/backend/app/services/auto_merger.py
@@ -0,0 +1,523 @@
+"""
+Neural Sync - Auto Merge Service
+
+自动合并服务 - 严格模式
+- 审计评分 >= 98 分
+- 全量测试通过
+- 二次人工确认
+- 熔断机制
+"""
+
+import asyncio
+import logging
+from typing import Optional, Dict, List, Any, Tuple
+from pydantic import BaseModel, Field
+from datetime import datetime, timedelta
+from enum import Enum
+from ..events import Event, EventType
+
+logger = logging.getLogger(__name__)
+
+
+class MergeStatus(str, Enum):
+    """合并状态"""
+    PENDING = "pending"
+    WAITING_CONFIRMATION = "waiting_confirmation"
+    APPROVED = "approved"
+    REJECTED = "rejected"
+    FAILED = "failed"
+    COOLDOWN = "cooldown"
+
+
+class MergeRequest(BaseModel):
+    """合并请求"""
+    signal_id: str
+    commit_sha: str
+    task_name: str
+    audit_score: float = Field(ge=0.0, le=100.0)
+    test_results: Dict[str, Any]
+    requester: Optional[str] = None
+
+
+class MergeResponse(BaseModel):
+    """合并响应"""
+    status: MergeStatus
+    success: bool
+    merge_sha: Optional[str] = None
+    reason: Optional[str] = None
+    requires_human_confirmation: bool = False
+    confirmation_token: Optional[str] = None
+    cooldown_remaining: Optional[float] = None
+    details: Dict[str, Any] = Field(default_factory=dict)
+
+
+class ConfirmationRequest(BaseModel):
+    """人工确认请求"""
+    token: str
+    merge_request: MergeRequest
+    created_at: float
+    expires_at: float
+
+
+class AutoMergeService:
+    """
+    自动合并服务 - 严格模式
+
+    条件：
+    1. 审计评分 >= 98 分
+    2. 全量测试通过
+    3. 二次人工确认
+    4. 不在熔断期内
+
+    特性：
+    - 熔断机制（连续失败后暂停）
+    - 人工确认流程
+    - 回滚支持
+    - 事件通知
+    """
+
+    # 严格模式阈值
+    AUDIT_SCORE_THRESHOLD = 98.0
+    HUMAN_CONFIRMATION_REQUIRED = True
+    CONFIRMATION_TIMEOUT = 300  # 5 minutes
+
+    # 熔断机制
+    MAX_CONSECUTIVE_FAILURES = 3
+    FAILURE_COOLDOWN_SECONDS = 3600  # 1 hour
+
+    def __init__(self, git_service=None, plane_service=None, lark_service=None, event_bus=None):
+        self.git = git_service
+        self.plane = plane_service
+        self.lark = lark_service
+        self.event_bus = event_bus
+
+        # 状态跟踪
+        self._failure_count = 0
+        self._last_failure_time: Optional[datetime] = None
+        self._successful_merges = 0
+
+        # 待确认的请求
+        self._pending_confirmations: Dict[str, ConfirmationRequest] = {}
+
+    async def try_auto_merge(self, request: MergeRequest) -> MergeResponse:
+        """
+        尝试自动合并
+        """
+        # 1. 检查熔断状态
+        if self._is_in_cooldown():
+            return MergeResponse(
+                status=MergeStatus.COOLDOWN,
+                success=False,
+                reason=f"系统处于熔断期，请在 {self._get_cooldown_remaining()}s 后重试",
+                requires_human_confirmation=True,
+                cooldown_remaining=self._get_cooldown_remaining(),
+            )
+
+        # 2. 验证审计评分
+        if request.audit_score < self.AUDIT_SCORE_THRESHOLD:
+            return MergeResponse(
+                status=MergeStatus.REJECTED,
+                success=False,
+                reason=f"审计评分 {request.audit_score} 低于阈值 {self.AUDIT_SCORE_THRESHOLD}",
+                requires_human_confirmation=True,
+            )
+
+        # 3. 验证测试结果
+        if not self._all_tests_passed(request.test_results):
+            return MergeResponse(
+                status=MergeStatus.REJECTED,
+                success=False,
+                reason="测试未全部通过",
+                requires_human_confirmation=True,
+                details={
+                    "test_results": request.test_results,
+                    "failed_tests": self._get_failed_tests(request.test_results),
+                },
+            )
+
+        # 4. 创建人工确认请求
+        if self.HUMAN_CONFIRMATION_REQUIRED:
+            confirmation_token = self._generate_confirmation_token()
+            expires_at = datetime.utcnow().timestamp() + self.CONFIRMATION_TIMEOUT
+
+            confirmation = ConfirmationRequest(
+                token=confirmation_token,
+                merge_request=request,
+                created_at=datetime.utcnow().timestamp(),
+                expires_at=expires_at,
+            )
+            self._pending_confirmations[confirmation_token] = confirmation
+
+            # 发送确认请求事件
+            await self._notify_confirmation_request(confirmation)
+
+            return MergeResponse(
+                status=MergeStatus.WAITING_CONFIRMATION,
+                success=False,
+                reason="等待人工二次确认",
+                requires_human_confirmation=True,
+                confirmation_token=confirmation_token,
+                details={
+                    "expires_at": expires_at,
+                    "timeout_seconds": self.CONFIRMATION_TIMEOUT,
+                },
+            )
+
+        # 5. 直接执行自动合并（不需要确认）
+        return await self._execute_merge(request)
+
+    async def confirm_merge(self, confirmation_token: str, requester: str) -> MergeResponse:
+        """
+        人工确认合并
+        """
+        confirmation = self._pending_confirmations.get(confirmation_token)
+        if not confirmation:
+            return MergeResponse(
+                status=MergeStatus.REJECTED,
+                success=False,
+                reason="确认令牌无效或已过期",
+            )
+
+        # 检查是否过期
+        if datetime.utcnow().timestamp() > confirmation.expires_at:
+            del self._pending_confirmations[confirmation_token]
+            return MergeResponse(
+                status=MergeStatus.REJECTED,
+                success=False,
+                reason="确认请求已过期",
+            )
+
+        # 执行合并
+        result = await self._execute_merge(confirmation.merge_request)
+
+        # 清理确认请求
+        if result.success:
+            del self._pending_confirmations[confirmation_token]
+
+        return result
+
+    async def reject_merge(self, confirmation_token: str, reason: str) -> MergeResponse:
+        """
+        拒绝合并
+        """
+        confirmation = self._pending_confirmations.get(confirmation_token)
+        if not confirmation:
+            return MergeResponse(
+                status=MergeStatus.REJECTED,
+                success=False,
+                reason="确认令牌无效",
+            )
+
+        # 清理确认请求
+        del self._pending_confirmations[confirmation_token]
+
+        return MergeResponse(
+            status=MergeStatus.REJECTED,
+            success=False,
+            reason=f"合并被拒绝: {reason}",
+        )
+
+    async def _execute_merge(self, request: MergeRequest) -> MergeResponse:
+        """
+        执行实际的合并操作
+        """
+        try:
+            # 1. 执行 Git 合并
+            merge_sha = await self._perform_git_merge(request.commit_sha)
+
+            # 2. 推送到远程
+            await self._push_to_remote()
+
+            # 3. 更新 Plane 任务状态
+            if self.plane:
+                await self._update_plane_status(request.signal_id, merge_sha)
+
+            # 4. 发送飞书通知
+            if self.lark:
+                await self._send_lark_notification(request, merge_sha)
+
+            # 5. 发送成功事件
+            await self._publish_merge_success_event(request, merge_sha)
+
+            # 重置失败计数
+            self._reset_failure_count()
+            self._successful_merges += 1
+
+            return MergeResponse(
+                status=MergeStatus.APPROVED,
+                success=True,
+                merge_sha=merge_sha,
+                reason="自动合并成功",
+                details={
+                    "successful_merges": self._successful_merges,
+                },
+            )
+
+        except Exception as e:
+            logger.error(f"Merge failed: {e}", exc_info=True)
+            self._record_failure()
+
+            await self._publish_merge_failed_event(request, str(e))
+
+            return MergeResponse(
+                status=MergeStatus.FAILED,
+                success=False,
+                reason=f"合并失败: {str(e)}",
+                requires_human_confirmation=True,
+            )
+
+    async def _perform_git_merge(self, commit_sha: str) -> str:
+        """执行 Git 合并"""
+        # 如果有 git 服务，使用它；否则使用 subprocess
+        if self.git:
+            return await self.git.merge(commit_sha)
+
+        # 使用 subprocess 执行 git 命令
+        import subprocess
+
+        # 合并提交（不创建合并提交，直接快进）
+        result = subprocess.run(
+            ["git", "merge", "--ff", commit_sha],
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+
+        if result.returncode != 0:
+            # 尝试 cherry-pick
+            result = subprocess.run(
+                ["git", "cherry-pick", commit_sha],
+                capture_output=True,
+                text=True,
+                timeout=30,
+            )
+            if result.returncode != 0:
+                raise Exception(f"Git merge/cherry-pick failed: {result.stderr}")
+
+        # 获取合并后的 HEAD
+        result = subprocess.run(
+            ["git", "rev-parse", "HEAD"],
+            capture_output=True,
+            text=True,
+            check=True,
+        )
+        return result.stdout.strip()
+
+    async def _push_to_remote(self):
+        """推送到远程仓库"""
+        import subprocess
+
+        result = subprocess.run(
+            ["git", "push"],
+            capture_output=True,
+            text=True,
+            timeout=60,
+        )
+
+        if result.returncode != 0:
+            raise Exception(f"Git push failed: {result.stderr}")
+
+    async def _update_plane_status(self, signal_id: str, merge_sha: str):
+        """更新 Plane 任务状态"""
+        if not self.plane:
+            return
+
+        try:
+            # 这里应该根据 signal_id 找到对应的 Plane issue
+            # 简化实现：假设任务 ID 已知
+            await self.plane.update_issue_status(
+                signal_id,  # 或提取的任务 ID
+                status="Done"
+            )
+        except Exception as e:
+            logger.warning(f"Failed to update Plane status: {e}")
+
+    async def _send_lark_notification(self, request: MergeRequest, merge_sha: str):
+        """发送飞书通知"""
+        if not self.lark:
+            return
+
+        try:
+            await self.lark.send_merge_notification(
+                task_name=request.task_name,
+                merge_sha=merge_sha,
+                audit_score=request.audit_score,
+            )
+        except Exception as e:
+            logger.warning(f"Failed to send Lark notification: {e}")
+
+    async def _notify_confirmation_request(self, confirmation: ConfirmationRequest):
+        """通知确认请求"""
+        if self.event_bus:
+            from .events import EventType
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.AUTO_MERGE_REQUESTED,
+                    data={
+                        "token": confirmation.token,
+                        "task_name": confirmation.merge_request.task_name,
+                        "commit_sha": confirmation.merge_request.commit_sha,
+                        "audit_score": confirmation.merge_request.audit_score,
+                        "expires_at": confirmation.expires_at,
+                    },
+                )
+            )
+
+    async def _publish_merge_success_event(self, request: MergeRequest, merge_sha: str):
+        """发布合并成功事件"""
+        if self.event_bus:
+            from .events import EventType
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.AUTO_MERGE_SUCCESS,
+                    data={
+                        "signal_id": request.signal_id,
+                        "commit_sha": request.commit_sha,
+                        "merge_sha": merge_sha,
+                        "task_name": request.task_name,
+                        "audit_score": request.audit_score,
+                    },
+                )
+            )
+
+    async def _publish_merge_failed_event(self, request: MergeRequest, error: str):
+        """发布合并失败事件"""
+        if self.event_bus:
+            from .events import EventType
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.AUTO_MERGE_FAILED,
+                    data={
+                        "signal_id": request.signal_id,
+                        "commit_sha": request.commit_sha,
+                        "task_name": request.task_name,
+                        "error": error,
+                    },
+                )
+            )
+
+    def _is_in_cooldown(self) -> bool:
+        """检查是否在熔断期"""
+        if self._last_failure_time is None:
+            return False
+
+        elapsed = (datetime.utcnow() - self._last_failure_time).total_seconds()
+        return elapsed < self.FAILURE_COOLDOWN_SECONDS
+
+    def _get_cooldown_remaining(self) -> float:
+        """获取熔断剩余时间"""
+        if self._last_failure_time is None:
+            return 0.0
+
+        elapsed = (datetime.utcnow() - self._last_failure_time).total_seconds()
+        return max(0, self.FAILURE_COOLDOWN_SECONDS - elapsed)
+
+    def _all_tests_passed(self, test_results: Dict[str, Any]) -> bool:
+        """检查所有测试是否通过"""
+        if not test_results:
+            return True  # 没有测试结果视为通过
+
+        # 检查是否有失败的测试
+        for test_name, result in test_results.items():
+            if isinstance(result, dict):
+                if result.get("status") == "failed":
+                    return False
+                if result.get("passed") is False:
+                    return False
+            elif isinstance(result, bool):
+                if not result:
+                    return False
+
+        return True
+
+    def _get_failed_tests(self, test_results: Dict[str, Any]) -> List[str]:
+        """获取失败的测试列表"""
+        failed = []
+        for test_name, result in test_results.items():
+            if isinstance(result, dict) and result.get("status") == "failed":
+                failed.append(test_name)
+            elif isinstance(result, bool) and not result:
+                failed.append(test_name)
+        return failed
+
+    def _generate_confirmation_token(self) -> str:
+        """生成确认令牌"""
+        import uuid
+        return str(uuid.uuid4())
+
+    def _record_failure(self):
+        """记录失败"""
+        self._failure_count += 1
+        self._last_failure_time = datetime.utcnow()
+
+        # 检查是否需要进入熔断期
+        if self._failure_count >= self.MAX_CONSECUTIVE_FAILURES:
+            logger.warning(
+                f"[AutoMerge] {self._failure_count} consecutive failures. "
+                f"Entering cooldown for {self.FAILURE_COOLDOWN_SECONDS}s"
+            )
+            if self.event_bus:
+                import asyncio
+                asyncio.create_task(
+                    self.event_bus.publish(
+                        Event(
+                            type=EventType.AUTO_MERGE_COOLDOWN,
+                            data={
+                                "failure_count": self._failure_count,
+                                "cooldown_seconds": self.FAILURE_COOLDOWN_SECONDS,
+                            },
+                        )
+                    )
+                )
+
+    def _reset_failure_count(self):
+        """重置失败计数"""
+        self._failure_count = 0
+        self._last_failure_time = None
+
+    async def get_pending_confirmations(self) -> List[ConfirmationRequest]:
+        """获取待确认的请求"""
+        # 清理过期的确认请求
+        now = datetime.utcnow().timestamp()
+        expired = [
+            token for token, conf in self._pending_confirmations.items()
+            if now > conf.expires_at
+        ]
+        for token in expired:
+            del self._pending_confirmations[token]
+
+        return list(self._pending_confirmations.values())
+
+    def get_status(self) -> Dict[str, Any]:
+        """获取服务状态"""
+        return {
+            "audit_threshold": self.AUDIT_SCORE_THRESHOLD,
+            "human_confirmation_required": self.HUMAN_CONFIRMATION_REQUIRED,
+            "failure_count": self._failure_count,
+            "successful_merges": self._successful_merges,
+            "in_cooldown": self._is_in_cooldown(),
+            "cooldown_remaining": self._get_cooldown_remaining() if self._is_in_cooldown() else 0,
+            "pending_confirmations": len(self._pending_confirmations),
+        }
+
+
+# 全局服务实例
+_auto_merger = None
+
+
+def get_auto_merger(
+    git_service=None,
+    plane_service=None,
+    lark_service=None,
+    event_bus=None,
+) -> AutoMergeService:
+    """获取自动合并服务实例"""
+    global _auto_merger
+    if _auto_merger is None:
+        _auto_merger = AutoMergeService(
+            git_service=git_service,
+            plane_service=plane_service,
+            lark_service=lark_service,
+            event_bus=event_bus,
+        )
+    return _auto_merger
diff --git a/backend/app/services/blocker_detection.py b/backend/app/services/blocker_detection.py
new file mode 100644
index 00000000..19eda2ca
--- /dev/null
+++ b/backend/app/services/blocker_detection.py
@@ -0,0 +1,423 @@
+"""
+Proactive Blocker Detection Service - 主动式阻塞检测服务
+
+核心功能：
+1. 扫描代码注释中的阻塞关键词
+2. 自动识别常见的开发阻塞模式
+3. AI 生成解决方案建议
+4. 追踪阻塞解决进度
+
+支持的阻塞模式：
+- "等待 X 接口完成"
+- "这个 API 达不到效果"
+- "由于 Y 的限制，暂时跳过"
+- "TODO: 等待 Z 确定"
+"""
+import asyncio
+import json
+import re
+from datetime import datetime, timezone, timedelta
+from typing import Any, Dict, List, Optional
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from sqlalchemy import text
+
+from ..models import (
+    User, Project, Task, Signal,
+    CognitiveConflict, TruthNode
+)
+from .llm import LLMService
+
+
+# 阻塞模式正则表达式
+BLOCKER_PATTERNS = [
+    r'等待.*接口',
+    r'等待.*API',
+    r'等待.*确定',
+    r'等待.*完成',
+    r'blocked by',
+    r'blocked due to',
+    r'can\'t proceed until',
+    r'暂时跳过',
+    r'TODO.*等待',
+    r'FIXME.*等待',
+    r'HACK.*临时',
+    r'达不到效果',
+    r'无法.*因为',
+    r'pending.*approval',
+    r'waiting for',
+    r'blocked on',
+]
+
+
+class BlockerDetectionService:
+    """主动式阻塞检测服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+        self.llm = LLMService()
+
+    async def scan_code_comment(
+        self,
+        code_content: str,
+        file_path: str,
+        line_number: int,
+        project_id: int,
+        user_id: int,
+        tenant_id: int = 1
+    ) -> Optional[Dict[str, Any]]:
+        """
+        扫描代码注释，检测阻塞
+
+        Args:
+            code_content: 代码内容
+            file_path: 文件路径
+            line_number: 行号
+            project_id: 项目ID
+            user_id: 用户ID
+            tenant_id: 租户ID
+
+        Returns:
+            如果检测到阻塞，返回阻塞信息；否则返回 None
+        """
+        # 检查是否匹配阻塞模式
+        detected_pattern = None
+        for pattern in BLOCKER_PATTERNS:
+            if re.search(pattern, code_content, re.IGNORECASE):
+                detected_pattern = pattern
+                break
+
+        if not detected_pattern:
+            return None
+
+        # AI 分析阻塞严重程度和生成解决方案
+        analysis = await self._analyze_blocker_with_ai(
+            code_content=code_content,
+            file_path=file_path,
+            pattern=detected_pattern,
+            tenant_id=tenant_id,
+        )
+
+        # 创建阻塞记录（复用 CognitiveConflict 模型）
+        from sqlalchemy import text
+        next_conflict_id = (await self.db.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM cognitive_conflicts"))).scalar() or 1
+
+        blocker = CognitiveConflict(
+            id=next_conflict_id,
+            truth_node_id=None,  # 阻塞可能不关联特定宪章
+            conflict_type='code_blocker',
+            description=f"代码阻塞检测：\n\n文件：{file_path}:{line_number}\n代码：{code_content[:200]}\n\nAI 分析：{analysis.get('problem_analysis', '')}",
+            involved_user_ids=json.dumps([user_id]),
+            status='detected',
+            tenant_id=tenant_id
+        )
+        self.db.add(blocker)
+        await self.db.commit()
+
+        # 自动生成任务建议
+        suggested_tasks = await self._generate_blocker_resolution_tasks(
+            blocker_id=blocker.id,
+            analysis=analysis,
+            project_id=project_id,
+            tenant_id=tenant_id
+        )
+
+        return {
+            'blocker_id': blocker.id,
+            'severity': analysis.get('severity', 'medium'),
+            'problem_analysis': analysis.get('problem_analysis', ''),
+            'suggested_solutions': analysis.get('suggested_solutions', []),
+            'suggested_tasks': suggested_tasks
+        }
+
+    async def detect_task_blockers(
+        self,
+        tenant_id: int = 1
+    ) -> List[Dict[str, Any]]:
+        """
+        检测所有任务中的阻塞
+
+        扫描所有进行中的任务，检测是否存在阻塞
+        """
+        from sqlalchemy import text
+
+        # 获取所有进行中的任务
+        result = await self.db.execute(
+            text("""
+                SELECT t.id, t.title, t.description, t.status, t.project_id
+                FROM tasks t
+                WHERE t.status IN ('pending', 'in_progress')
+                AND t.tenant_id = :tenant_id
+                AND (t.updated_at < datetime('now', '-3 days')
+                     OR t.updated_at < datetime('now', '-3 days', 'localtime'))
+            """),
+            {"tenant_id": tenant_id}
+        )
+
+        tasks = result.fetchall()
+
+        detected_blockers = []
+
+        for task in tasks:
+            task_id, title, description, status, project_id = task
+
+            # AI 分析任务是否可能被阻塞
+            analysis = await self._analyze_task_for_blockers(
+                task_id=task_id,
+                title=title,
+                description=description,
+                tenant_id=tenant_id,
+            )
+
+            if analysis.get('is_blocked', False):
+                # 创建阻塞记录
+                next_conflict_id = (await self.db.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM cognitive_conflicts"))).scalar() or 1
+
+                blocker = CognitiveConflict(
+                    id=next_conflict_id,
+                    truth_node_id=None,
+                    conflict_type='task_blocker',
+                    description=f"任务阻塞检测：\n\n任务：{title}\nID: {task_id}\n\nAI 分析：{analysis.get('reason', '')}",
+                    involved_user_ids=json.dumps([]),
+                    status='detected',
+                    tenant_id=tenant_id
+                )
+                self.db.add(blocker)
+                await self.db.commit()
+
+                detected_blockers.append({
+                    'blocker_id': blocker.id,
+                    'task_id': task_id,
+                    'task_title': title,
+                    'severity': analysis.get('severity', 'medium'),
+                    'reason': analysis.get('reason', ''),
+                    'suggested_actions': analysis.get('suggested_actions', [])
+                })
+
+        return detected_blockers
+
+    async def resolve_blocker(
+        self,
+        blocker_id: int,
+        resolution_note: str,
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        标记阻塞为已解决
+
+        Args:
+            blocker_id: 阻塞ID
+            resolution_note: 解决方案说明
+            tenant_id: 租户ID
+        """
+        result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.id == blocker_id,
+                CognitiveConflict.tenant_id == tenant_id
+            )
+        )
+        blocker = result.scalar_one_or_none()
+
+        if not blocker:
+            raise ValueError(f"阻塞 {blocker_id} 不存在")
+
+        blocker.status = 'resolved'
+        blocker.resolved_at = datetime.now(timezone.utc)
+        blocker.description += f"\n\n解决方案：{resolution_note}"
+
+        await self.db.commit()
+
+        return {
+            'blocker_id': blocker_id,
+            'status': 'resolved',
+            'resolved_at': blocker.resolved_at.isoformat()
+        }
+
+    async def get_active_blockers(
+        self,
+        tenant_id: int = 1
+    ) -> List[Dict[str, Any]]:
+        """
+        获取所有活跃的阻塞
+
+        Returns:
+            阻塞列表，包含严重程度、描述等信息
+        """
+        result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.tenant_id == tenant_id,
+                CognitiveConflict.conflict_type.in_(['code_blocker', 'task_blocker']),
+                CognitiveConflict.status == 'detected'
+            )
+        )
+        blockers = result.scalars().all()
+
+        return [
+            {
+                'id': b.id,
+                'type': b.conflict_type,
+                'description': b.description,
+                'severity': self._extract_severity(b.description),
+                'detected_at': b.created_at.isoformat() if hasattr(b, 'created_at') else None,
+                'involved_users': json.loads(b.involved_user_ids) if b.involved_user_ids else []
+            }
+            for b in blockers
+        ]
+
+    async def _analyze_blocker_with_ai(
+        self,
+        code_content: str,
+        file_path: str,
+        pattern: str,
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        """AI 分析代码阻塞"""
+        prompt = f"""分析以下代码注释中提到的阻塞问题：
+
+文件：{file_path}
+匹配到的模式：{pattern}
+代码内容：
+{code_content}
+
+请分析：
+1. 阻塞的根本原因是什么？
+2. 严重程度（high/medium/low）？
+3. 可能的解决方案有哪些？
+4. 是否需要创建新任务来跟踪？
+
+以 JSON 格式返回：
+{{
+    "problem_analysis": "问题分析",
+    "severity": "high/medium/low",
+    "suggested_solutions": ["方案1", "方案2"],
+    "requires_new_task": boolean,
+    "suggested_task_title": "建议的任务标题"
+}}"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except:
+            return {
+                'problem_analysis': f'检测到可能的阻塞：{pattern}',
+                'severity': 'medium',
+                'suggested_solutions': ['请团队负责人评估'],
+                'requires_new_task': False
+            }
+
+    async def _analyze_task_for_blockers(
+        self,
+        task_id: int,
+        title: str,
+        description: str,
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        """AI 分析任务是否被阻塞"""
+        prompt = f"""分析以下任务是否可能被阻塞：
+
+任务标题：{title}
+任务描述：{description}
+
+请判断：
+1. 任务是否被阻塞？
+2. 如果阻塞，原因是什么？
+3. 严重程度（high/medium/low）？
+4. 建议的下一步行动？
+
+以 JSON 格式返回：
+{{
+    "is_blocked": boolean,
+    "reason": "阻塞原因（如果阻塞）",
+    "severity": "high/medium/low",
+    "suggested_actions": ["行动1", "行动2"]
+}}
+
+如果任务没有被阻塞，is_blocked 设为 false。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except:
+            return {'is_blocked': False}
+
+    async def _generate_blocker_resolution_tasks(
+        self,
+        blocker_id: int,
+        analysis: Dict[str, Any],
+        project_id: int,
+        tenant_id: int
+    ) -> List[Dict[str, Any]]:
+        """生成阻塞解决任务"""
+        if not analysis.get('requires_new_task', False):
+            return []
+
+        from sqlalchemy import text
+
+        # 获取下一个任务ID
+        next_task_id = (await self.db.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM tasks"))).scalar() or 1
+
+        task_title = analysis.get('suggested_task_title', f'解决阻塞 #{blocker_id}')
+        task_description = f"""自动生成的阻塞解决任务
+
+阻塞原因：
+{analysis.get('problem_analysis', '')}
+
+建议的解决方案：
+{chr(10).join(analysis.get('suggested_solutions', []))}"""
+
+        task = Task(
+            id=next_task_id,
+            truth_node_id=None,  # 不关联特定宪章
+            project_id=project_id,
+            assignee_id=None,
+            title=task_title,
+            description=task_description,
+            required_skill_tag_ids=[],
+            status='pending',
+            priority='HIGH' if analysis.get('severity') == 'high' else 'MEDIUM',
+            estimated_hours=4,
+            actual_hours=0,
+            quality_score=0,
+            tenant_id=tenant_id
+        )
+        self.db.add(task)
+        await self.db.commit()
+
+        return [{
+            'task_id': task.id,
+            'title': task.title,
+            'priority': task.priority
+        }]
+
+    def _extract_severity(self, description: str) -> str:
+        """从描述中提取严重程度"""
+        desc_lower = description.lower()
+        if 'high' in desc_lower or '严重' in desc_lower or 'critical' in desc_lower:
+            return 'high'
+        elif 'low' in desc_lower or '轻微' in desc_lower:
+            return 'low'
+        return 'medium'
+
+
+# 全局服务实例
+def get_blocker_detection_service(db: AsyncSession) -> BlockerDetectionService:
+    """获取阻塞检测服务实例"""
+    return BlockerDetectionService(db)
diff --git a/backend/app/services/budget_enforcement_service.py b/backend/app/services/budget_enforcement_service.py
new file mode 100644
index 00000000..f6bf65ce
--- /dev/null
+++ b/backend/app/services/budget_enforcement_service.py
@@ -0,0 +1,345 @@
+"""预算强制执行服务
+
+实现：
+- AI 调用前预算检查
+- 超支硬限制执行
+- 预算使用量累计
+- 80% 阈值预警与超支告警
+"""
+
+from __future__ import annotations
+
+from datetime import datetime, timedelta
+from typing import Any, Optional
+
+from sqlalchemy import and_, desc, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import Budget, BudgetAlert
+
+
+class BudgetEnforcementService:
+    """预算强制执行服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def check_budget_before_ai_call(
+        self,
+        tenant_id: int,
+        estimated_cost_usd: float,
+        project_id: Optional[int] = None,
+    ) -> dict[str, Any]:
+        """AI 调用前预算检查"""
+        budget = await self._get_active_budget(tenant_id=tenant_id, project_id=project_id)
+        if not budget:
+            return {
+                "allowed": True,
+                "reason": "no_budget_configured",
+                "budget_checked": False,
+            }
+
+        now = datetime.utcnow()
+        await self._ensure_current_period(budget, now)
+
+        current_usage = float(budget.current_usage_usd or 0.0)
+        budget_amount = float(budget.budget_amount_usd or 0.0)
+        if budget_amount <= 0:
+            return {
+                "allowed": True,
+                "reason": "invalid_budget_amount",
+                "budget_checked": True,
+            }
+
+        if budget.status == "suspended":
+            return {
+                "allowed": False,
+                "reason": "budget_suspended",
+                "budget_checked": True,
+                "budget_id": budget.id,
+                "current_usage_usd": round(current_usage, 4),
+                "budget_amount_usd": round(budget_amount, 4),
+            }
+
+        projected_usage = current_usage + max(estimated_cost_usd, 0.0)
+        usage_percent = (projected_usage / budget_amount) * 100
+
+        # 阈值预警（80% 默认）
+        if usage_percent >= float(budget.alert_threshold_percent or 80.0):
+            await self.generate_budget_alert(
+                budget=budget,
+                alert_type="threshold_reached",
+                usage_usd=projected_usage,
+                usage_percent=usage_percent,
+                message=(
+                    f"预算使用率达到 {usage_percent:.2f}% "
+                    f"(阈值 {float(budget.alert_threshold_percent or 80.0):.2f}%)"
+                ),
+                action_taken=None,
+            )
+
+        # 超支硬拦截
+        if (
+            projected_usage > budget_amount
+            and bool(budget.hard_limit_enabled)
+            and budget.status != "overridden"
+        ):
+            await self.enforce_spending_limit(
+                budget=budget,
+                usage_usd=projected_usage,
+                usage_percent=usage_percent,
+            )
+            return {
+                "allowed": False,
+                "reason": "budget_exceeded_hard_limit",
+                "budget_checked": True,
+                "budget_id": budget.id,
+                "current_usage_usd": round(current_usage, 4),
+                "projected_usage_usd": round(projected_usage, 4),
+                "budget_amount_usd": round(budget_amount, 4),
+                "usage_percent": round(usage_percent, 2),
+            }
+
+        return {
+            "allowed": True,
+            "reason": "within_budget",
+            "budget_checked": True,
+            "budget_id": budget.id,
+            "current_usage_usd": round(current_usage, 4),
+            "projected_usage_usd": round(projected_usage, 4),
+            "budget_amount_usd": round(budget_amount, 4),
+            "usage_percent": round(usage_percent, 2),
+        }
+
+    async def update_budget_usage_after_ai_call(
+        self,
+        tenant_id: int,
+        actual_cost_usd: float,
+        project_id: Optional[int] = None,
+    ) -> dict[str, Any]:
+        """AI 调用成功后累计预算使用量"""
+        budget = await self._get_active_budget(tenant_id=tenant_id, project_id=project_id)
+        if not budget:
+            return {"updated": False, "reason": "no_budget_configured"}
+
+        now = datetime.utcnow()
+        await self._ensure_current_period(budget, now)
+
+        cost = max(float(actual_cost_usd or 0.0), 0.0)
+        budget.current_usage_usd = float(budget.current_usage_usd or 0.0) + cost
+
+        budget_amount = float(budget.budget_amount_usd or 0.0)
+        usage_percent = (float(budget.current_usage_usd) / budget_amount * 100) if budget_amount > 0 else 0.0
+
+        if usage_percent >= float(budget.alert_threshold_percent or 80.0):
+            await self.generate_budget_alert(
+                budget=budget,
+                alert_type="threshold_reached",
+                usage_usd=float(budget.current_usage_usd),
+                usage_percent=usage_percent,
+                message=(
+                    f"预算使用率达到 {usage_percent:.2f}% "
+                    f"(阈值 {float(budget.alert_threshold_percent or 80.0):.2f}%)"
+                ),
+                action_taken=None,
+            )
+
+        if budget_amount > 0 and float(budget.current_usage_usd) > budget_amount and bool(budget.hard_limit_enabled):
+            await self.enforce_spending_limit(
+                budget=budget,
+                usage_usd=float(budget.current_usage_usd),
+                usage_percent=usage_percent,
+            )
+
+        await self.db.commit()
+        await self.db.refresh(budget)
+
+        return {
+            "updated": True,
+            "budget_id": budget.id,
+            "current_usage_usd": float(budget.current_usage_usd),
+            "usage_percent": round(usage_percent, 2),
+            "status": budget.status,
+        }
+
+    async def enforce_spending_limit(
+        self,
+        budget: Budget,
+        usage_usd: float,
+        usage_percent: float,
+    ) -> None:
+        """执行超支硬性限制：标记预算为 suspended 并记录告警"""
+        budget.status = "suspended"
+
+        await self.generate_budget_alert(
+            budget=budget,
+            alert_type="exceeded",
+            usage_usd=usage_usd,
+            usage_percent=usage_percent,
+            message=f"预算已超支，当前使用率 {usage_percent:.2f}%",
+            action_taken="mark_exceeded",
+        )
+        await self.generate_budget_alert(
+            budget=budget,
+            alert_type="suspended",
+            usage_usd=usage_usd,
+            usage_percent=usage_percent,
+            message="预算超支，已暂停新的 AI 调用",
+            action_taken="suspend_ai_calls",
+        )
+        await self.db.commit()
+
+    async def generate_budget_alert(
+        self,
+        budget: Budget,
+        alert_type: str,
+        usage_usd: float,
+        usage_percent: float,
+        message: str,
+        action_taken: Optional[str],
+    ) -> BudgetAlert:
+        """生成预算预警（同类 active 预警存在则复用更新）"""
+        result = await self.db.execute(
+            select(BudgetAlert).where(
+                and_(
+                    BudgetAlert.budget_id == budget.id,
+                    BudgetAlert.alert_type == alert_type,
+                    BudgetAlert.status == "active",
+                )
+            ).order_by(desc(BudgetAlert.created_at))
+        )
+        existing = result.scalar_one_or_none()
+        if existing:
+            existing.usage_usd = usage_usd
+            existing.usage_percent = usage_percent
+            existing.message = message
+            existing.action_taken = action_taken
+            await self.db.commit()
+            await self.db.refresh(existing)
+            return existing
+
+        alert = BudgetAlert(
+            tenant_id=budget.tenant_id,
+            budget_id=budget.id,
+            alert_type=alert_type,
+            usage_usd=usage_usd,
+            usage_percent=usage_percent,
+            message=message,
+            is_read=False,
+            action_taken=action_taken,
+            status="active",
+        )
+        self.db.add(alert)
+        await self.db.commit()
+        await self.db.refresh(alert)
+        return alert
+
+    async def override_budget_limit(
+        self,
+        budget_id: int,
+        tenant_id: int,
+        operator_user_id: int,
+        note: str = "",
+        duration_hours: int = 24,
+    ) -> dict[str, Any]:
+        """BOSS/ADMIN 超支解锁"""
+        result = await self.db.execute(
+            select(Budget).where(
+                and_(
+                    Budget.id == budget_id,
+                    Budget.tenant_id == tenant_id,
+                )
+            )
+        )
+        budget = result.scalar_one_or_none()
+        if not budget:
+            raise ValueError(f"预算 {budget_id} 不存在")
+
+        now = datetime.utcnow()
+        expires_at = now + timedelta(hours=max(duration_hours, 1))
+        history = dict(budget.override_history or {})
+        records = list(history.get("records", []))
+        records.append(
+            {
+                "operator_user_id": operator_user_id,
+                "note": note,
+                "overridden_at": now.isoformat(),
+                "expires_at": expires_at.isoformat(),
+            }
+        )
+        history["records"] = records
+        budget.override_history = history
+        budget.status = "overridden"
+
+        await self.db.commit()
+        await self.db.refresh(budget)
+        return {
+            "budget_id": budget.id,
+            "status": budget.status,
+            "expires_at": expires_at.isoformat(),
+            "override_count": len(records),
+        }
+
+    async def _get_active_budget(self, tenant_id: int, project_id: Optional[int]) -> Optional[Budget]:
+        """优先读取项目预算，否则回退租户预算"""
+        if project_id is not None:
+            result = await self.db.execute(
+                select(Budget).where(
+                    and_(
+                        Budget.tenant_id == tenant_id,
+                        Budget.budget_type == "project",
+                        Budget.project_id == project_id,
+                    )
+                ).order_by(desc(Budget.created_at))
+            )
+            project_budget = result.scalar_one_or_none()
+            if project_budget:
+                return project_budget
+
+        result = await self.db.execute(
+            select(Budget).where(
+                and_(
+                    Budget.tenant_id == tenant_id,
+                    Budget.budget_type == "tenant",
+                )
+            ).order_by(desc(Budget.created_at))
+        )
+        return result.scalar_one_or_none()
+
+    async def _ensure_current_period(self, budget: Budget, now: datetime) -> None:
+        """若预算周期已过，自动切换到新周期并清零使用量"""
+        start = budget.current_period_start
+        end = budget.current_period_end
+        if start is None or end is None or now >= end:
+            next_start, next_end = self._compute_period_window(now, budget.period)
+            budget.current_period_start = next_start
+            budget.current_period_end = next_end
+            budget.current_usage_usd = 0.0
+            # 周期切换后恢复 active，允许重新消费
+            budget.status = "active"
+            await self.db.commit()
+
+    def _compute_period_window(self, now: datetime, period: str) -> tuple[datetime, datetime]:
+        period = (period or "monthly").lower()
+        if period == "yearly":
+            start = datetime(now.year, 1, 1)
+            end = datetime(now.year + 1, 1, 1)
+            return start, end
+        if period == "quarterly":
+            quarter = (now.month - 1) // 3
+            start_month = quarter * 3 + 1
+            start = datetime(now.year, start_month, 1)
+            if start_month == 10:
+                end = datetime(now.year + 1, 1, 1)
+            else:
+                end = datetime(now.year, start_month + 3, 1)
+            return start, end
+
+        # default monthly
+        start = datetime(now.year, now.month, 1)
+        if now.month == 12:
+            end = datetime(now.year + 1, 1, 1)
+        else:
+            end = datetime(now.year, now.month + 1, 1)
+        return start, end
+
diff --git a/backend/app/services/codegen.py b/backend/app/services/codegen.py
new file mode 100644
index 00000000..eea6bad0
--- /dev/null
+++ b/backend/app/services/codegen.py
@@ -0,0 +1,552 @@
+"""
+Neural Sync - Code Draft Generation Service
+
+AI 代码草案生成服务 - 激进策略
+AI 尝试生成所有代码（包括复杂逻辑），但提供详细的 review points
+"""
+
+import asyncio
+import os
+import re
+import subprocess
+import logging
+from typing import List, Optional, Dict, Any
+from pydantic import BaseModel, Field
+from datetime import datetime
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+class CodeDraftRequest(BaseModel):
+    """代码草案生成请求"""
+    signal_id: str
+    signal_content: str
+    strategic_analysis: Dict[str, Any]
+    target_files: List[str]
+    project_root: str = "."
+
+
+class FileInfo(BaseModel):
+    """文件信息"""
+    path: str
+    language: str
+    exists: bool
+    content: Optional[str] = None
+    lines_count: int = 0
+
+
+class GeneratedFile(BaseModel):
+    """生成的文件"""
+    path: str
+    content: str
+    language: str
+    diff: str
+    is_new: bool
+    review_points: List[str] = []
+
+
+class CodeDraftResponse(BaseModel):
+    """代码草案生成响应"""
+    success: bool
+    files: List[GeneratedFile]
+    summary: str
+    confidence: float = Field(default=0.0, ge=0.0, le=1.0)
+    warnings: List[str] = Field(default_factory=list)
+    review_points: List[str] = Field(default_factory=list)
+    estimated_complexity: str = "low"
+    generation_time: float = 0.0
+    metadata: Dict[str, Any] = Field(default_factory=dict)
+
+
+class CodeDraftService:
+    """
+    AI 代码草案生成服务 - 激进策略
+
+    特性:
+    - 支持全栈代码生成
+    - 读取现有代码上下文
+    - 生成 unified diff
+    - 提取人工 review 关键点
+    - 深度代码验证
+    """
+
+    def __init__(self, llm_service):
+        self.llm = llm_service
+        self._language_map = {
+            ".py": "python",
+            ".ts": "typescript",
+            ".tsx": "typescript",
+            ".js": "javascript",
+            ".jsx": "javascript",
+            ".vue": "vue",
+            ".sql": "sql",
+            ".sh": "bash",
+            ".yaml": "yaml",
+            ".yml": "yaml",
+            ".json": "json",
+        }
+
+    async def generate_draft(self, request: CodeDraftRequest) -> CodeDraftResponse:
+        """
+        生成代码草案（激进策略）
+        """
+        start_time = datetime.utcnow().timestamp()
+
+        try:
+            # 1. 读取目标文件
+            file_contexts = await self._read_target_files(request)
+
+            # 2. 分析代码库结构
+            repo_context = await self._analyze_repo_structure(request)
+
+            # 3. 构建激进策略的生成 prompt
+            prompt = self._build_aggressive_prompt(request, file_contexts, repo_context)
+
+            # 4. 调用 LLM 生成代码
+            llm_response = await self._call_llm_for_code(prompt)
+
+            # 5. 解析和验证生成的代码
+            generated_files = await self._parse_and_validate(
+                llm_response, file_contexts
+            )
+
+            # 6. 生成 unified diff
+            for file in generated_files:
+                file.diff = self._generate_diff(file_contexts, file)
+
+            # 7. 提取 review points
+            review_points = await self._extract_review_points(
+                generated_files, file_contexts, request
+            )
+
+            # 8. 计算置信度和复杂度
+            confidence = self._calculate_confidence(generated_files, file_contexts)
+            complexity = self._estimate_complexity(generated_files)
+
+            # 9. 生成摘要
+            summary = self._generate_summary(generated_files, request)
+
+            # 10. 收集警告
+            warnings = self._collect_warnings(generated_files, file_contexts)
+
+            generation_time = datetime.utcnow().timestamp() - start_time
+
+            return CodeDraftResponse(
+                success=True,
+                files=generated_files,
+                summary=summary,
+                confidence=confidence,
+                warnings=warnings,
+                review_points=review_points,
+                estimated_complexity=complexity,
+                generation_time=generation_time,
+                metadata={
+                    "signal_id": request.signal_id,
+                    "files_count": len(generated_files),
+                    "new_files": sum(1 for f in generated_files if f.is_new),
+                    "modified_files": sum(1 for f in generated_files if not f.is_new),
+                },
+            )
+
+        except Exception as e:
+            logger.error(f"Code generation failed: {e}", exc_info=True)
+            return CodeDraftResponse(
+                success=False,
+                files=[],
+                summary=f"代码生成失败: {str(e)}",
+                confidence=0.0,
+                warnings=[f"生成过程出错: {str(e)}"],
+                review_points=[],
+                generation_time=datetime.utcnow().timestamp() - start_time,
+            )
+
+    async def _read_target_files(
+        self, request: CodeDraftRequest
+    ) -> Dict[str, FileInfo]:
+        """读取目标文件内容"""
+        file_contexts = {}
+
+        for file_path in request.target_files:
+            full_path = os.path.join(request.project_root, file_path)
+
+            info = FileInfo(
+                path=file_path,
+                language=self._detect_language(file_path),
+                exists=os.path.exists(full_path),
+            )
+
+            if info.exists:
+                try:
+                    with open(full_path, "r", encoding="utf-8") as f:
+                        info.content = f.read()
+                        info.lines_count = len(info.content.split("\n"))
+                except Exception as e:
+                    logger.warning(f"Failed to read {file_path}: {e}")
+                    info.content = None
+
+            file_contexts[file_path] = info
+
+        return file_contexts
+
+    def _detect_language(self, file_path: str) -> str:
+        """检测文件语言"""
+        for ext, lang in self._language_map.items():
+            if file_path.endswith(ext):
+                return lang
+        return "text"
+
+    async def _analyze_repo_structure(self, request: CodeDraftRequest) -> Dict[str, Any]:
+        """分析代码库结构"""
+        repo_context = {
+            "existing_modules": [],
+            "imports_available": [],
+            "patterns_used": [],
+        }
+
+        try:
+            # 扫描 Python 文件的导入
+            py_files = list(
+                Path(request.project_root).rglob("*.py")
+            )
+            for py_file in py_files[:20]:  # 限制扫描数量
+                try:
+                    content = py_file.read_text(encoding="utf-8", errors="ignore")
+                    # 提取导入
+                    imports = re.findall(r"^import (\w+)|^from (\w+) import", content, re.MULTILINE)
+                    for imp in imports:
+                        module = imp[0] or imp[1]
+                        if module and module not in repo_context["imports_available"]:
+                            repo_context["imports_available"].append(module)
+                except:
+                    pass
+
+        except Exception as e:
+            logger.warning(f"Repo structure analysis failed: {e}")
+
+        return repo_context
+
+    def _build_aggressive_prompt(
+        self,
+        request: CodeDraftRequest,
+        file_contexts: Dict[str, FileInfo],
+        repo_context: Dict[str, Any],
+    ) -> str:
+        """构建激进的代码生成 prompt"""
+        prompt = f"""You are a senior full-stack engineer. Generate COMPLETE, PRODUCTION-READY code based on the following requirements.
+
+## Signal (Requirement)
+**Content**: {request.signal_content}
+
+## Strategic Analysis
+```json
+{self._format_json(request.strategic_analysis)}
+```
+
+## Available Imports/Modules
+{', '.join(repo_context['imports_available'][:20])}
+
+## Existing Code Context
+"""
+
+        for file_path, info in file_contexts.items():
+            prompt += f"\n### {file_path} ({info.language})\n"
+            if info.exists and info.content:
+                # 只显示关键部分（导入、类定义、函数签名）
+                lines = info.content.split("\n")
+                key_lines = []
+                for line in lines:
+                    stripped = line.strip()
+                    # 保留导入、类/函数定义、装饰器
+                    if any(
+                        stripped.startswith(p)
+                        for p in ["import ", "from ", "class ", "def ", "@", "async def "]
+                    ):
+                        key_lines.append(line)
+                    elif len(key_lines) < 30:  # 保留前30行作为上下文
+                        key_lines.append(line)
+
+                prompt += "```" + info.language + "\n"
+                prompt += "\n".join(key_lines)
+                if len(lines) > 30:
+                    prompt += f"\n... ({len(lines) - 30} more lines)\n"
+                prompt += "```\n"
+            else:
+                prompt += "*(New file to be created)*\n"
+
+        prompt += """
+## Code Generation Requirements
+1. **Complete Implementation**: Generate FULL, WORKING code including complex business logic
+2. **Error Handling**: Proper exception handling and edge cases
+3. **Performance**: Consider performance implications
+4. **Type Safety**: Use proper type hints
+5. **Comments**: Add concise comments for complex logic
+6. **Testing**: Consider testability
+
+## Output Format
+Return a JSON response with this EXACT structure:
+```json
+{
+  "files": [
+    {
+      "path": "relative/path/to/file.ext",
+      "content": "full file content here",
+      "is_new": true/false,
+      "review_points": ["list of key points needing human review"]
+    }
+  ],
+  "summary": "Brief description of changes",
+  "warnings": ["any potential issues or risks"]
+}
+```
+
+IMPORTANT:
+- Generate complete, runnable code
+- Handle edge cases and errors
+- Follow existing code patterns in the project
+- Highlight complex logic in review_points
+"""
+
+        return prompt
+
+    async def _call_llm_for_code(self, prompt: str) -> Dict[str, Any]:
+        """调用 LLM 生成代码"""
+        try:
+            # 使用 LLM 服务的聊天接口
+            response = await self.llm._call_llm_chat(
+                messages=[
+                    {"role": "system", "content": "You are an expert software engineer."},
+                    {"role": "user", "content": prompt}
+                ],
+                temperature=0.3,  # 降低温度以获得更稳定的代码
+            )
+
+            # 解析 JSON 响应
+            json_match = re.search(r'```json\s*([\s\S]*?)\s*```', response)
+            if json_match:
+                import json
+                return json.loads(json_match.group(1))
+
+            # 尝试直接解析 JSON
+            import json
+            return json.loads(response)
+
+        except Exception as e:
+            logger.error(f"LLM call or parsing failed: {e}")
+            # 返回错误响应
+            return {
+                "files": [],
+                "summary": f"LLM 生成失败: {str(e)}",
+                "warnings": [str(e)]
+            }
+
+    async def _parse_and_validate(
+        self, llm_response: Dict[str, Any], file_contexts: Dict[str, FileInfo]
+    ) -> List[GeneratedFile]:
+        """解析和验证生成的代码"""
+        generated_files = []
+
+        for file_data in llm_response.get("files", []):
+            file_path = file_data.get("path", "")
+            content = file_data.get("content", "")
+            review_points = file_data.get("review_points", [])
+
+            file_info = file_contexts.get(file_path)
+            if not file_info:
+                file_info = FileInfo(
+                    path=file_path,
+                    language=self._detect_language(file_path),
+                    exists=False,
+                )
+
+            # 基础验证
+            validation_warnings = self._validate_code_syntax(content, file_info.language)
+            if validation_warnings:
+                review_points.extend(validation_warnings)
+
+            generated_files.append(
+                GeneratedFile(
+                    path=file_path,
+                    content=content,
+                    language=file_info.language,
+                    diff="",  # 稍后生成
+                    is_new=not file_info.exists,
+                    review_points=review_points,
+                )
+            )
+
+        return generated_files
+
+    def _validate_code_syntax(self, content: str, language: str) -> List[str]:
+        """验证代码语法"""
+        warnings = []
+
+        try:
+            if language == "python":
+                import ast
+                ast.parse(content)
+            elif language in ["typescript", "javascript"]:
+                # 简单的括号匹配检查
+                if content.count("(") != content.count(")"):
+                    warnings.append("括号不匹配")
+                if content.count("{") != content.count("}"):
+                    warnings.append("花括号不匹配")
+        except SyntaxError as e:
+            warnings.append(f"语法错误: {str(e)}")
+
+        return warnings
+
+    def _generate_diff(
+        self, file_contexts: Dict[str, FileInfo], file: GeneratedFile
+    ) -> str:
+        """生成 unified diff"""
+        if file.is_new:
+            return f"--- /dev/null\n+++ {file.path}\n@@ -0,0 +1,{file.content.count(chr(10))} @@\n" + "\n".join(
+                f"+{line}" for line in file.content.split("\n")
+            )
+
+        # 对于修改的文件
+        old_content = file_contexts.get(file.path)?.content or ""
+        old_lines = old_content.split("\n")
+        new_lines = file.content.split("\n")
+
+        # 简化的 diff 生成（实际应该使用 difflib）
+        diff_lines = [f"--- {file.path}", f"+++ {file.path}"]
+
+        # 找出差异的行范围
+        changed_ranges = self._find_changed_ranges(old_lines, new_lines)
+
+        for start_old, start_new, length in changed_ranges:
+            diff_lines.append(f"@@ -{start_old + 1},{length} +{start_new + 1},{length} @@")
+            for i in range(length):
+                old_line = old_lines[start_old + i] if start_old + i < len(old_lines) else ""
+                new_line = new_lines[start_new + i] if start_new + i < len(new_lines) else ""
+
+                if old_line == new_line:
+                    diff_lines.append(f" {new_line}")
+                elif new_line and not old_line:
+                    diff_lines.append(f"+{new_line}")
+                elif old_line and not new_line:
+                    diff_lines.append(f"-{old_line}")
+                else:
+                    diff_lines.append(f"-{old_line}")
+                    diff_lines.append(f"+{new_line}")
+
+        return "\n".join(diff_lines)
+
+    def _find_changed_ranges(
+        self, old_lines: List[str], new_lines: List[str]
+    ) -> List[tuple]:
+        """找出变化的代码行范围"""
+        # 简化实现：假设整个文件是变化的
+        return [(0, 0, max(len(old_lines), len(new_lines)))]
+
+    async def _extract_review_points(
+        self, generated_files: List[GeneratedFile],
+        file_contexts: Dict[str, FileInfo],
+        request: CodeDraftRequest,
+    ) -> List[str]:
+        """提取需要人工 review 的关键点"""
+        review_points = []
+
+        # 收集所有文件的 review points
+        for file in generated_files:
+            review_points.extend(file.review_points)
+
+        # 添加通用 review points
+        review_points.extend([
+            "验证生成的代码是否符合项目的编码规范",
+            "检查是否有安全漏洞（SQL注入、XSS等）",
+            "确认复杂业务逻辑的正确性",
+            "验证错误处理是否完整",
+            "检查是否有性能优化空间",
+        ])
+
+        # 根据战略分析添加特定 review points
+        strategic_tags = request.strategic_analysis.get("decision_proposal", {}).get("tags", [])
+        if "security" in strategic_tags:
+            review_points.append("重点关注安全相关代码")
+        if "performance" in strategic_tags:
+            review_points.append("验证性能指标是否达标")
+
+        return list(set(review_points))  # 去重
+
+    def _calculate_confidence(
+        self, generated_files: List[GeneratedFile], file_contexts: Dict[str, FileInfo]
+    ) -> float:
+        """计算生成置信度"""
+        if not generated_files:
+            return 0.0
+
+        confidence = 0.7  # 基础置信度
+
+        # 根据文件数量调整
+        confidence -= min(len(generated_files) * 0.05, 0.2)
+
+        # 根据警告数量调整
+        total_warnings = sum(len(f.review_points) for f in generated_files)
+        confidence -= min(total_warnings * 0.02, 0.3)
+
+        return max(0.0, min(1.0, confidence))
+
+    def _estimate_complexity(self, generated_files: List[GeneratedFile]) -> str:
+        """估算代码复杂度"""
+        total_lines = sum(len(f.content.split("\n")) for f in generated_files)
+
+        if total_lines < 100:
+            return "low"
+        elif total_lines < 300:
+            return "medium"
+        else:
+            return "high"
+
+    def _generate_summary(
+        self, generated_files: List[GeneratedFile], request: CodeDraftRequest
+    ) -> str:
+        """生成变更摘要"""
+        new_files = [f for f in generated_files if f.is_new]
+        modified_files = [f for f in generated_files if not f.is_new]
+
+        summary_parts = [
+            f"为信号 '{request.signal_content[:50]}...' 生成代码草案",
+            f"新建文件: {len(new_files)} 个",
+            f"修改文件: {len(modified_files)} 个",
+        ]
+
+        if modified_files:
+            summary_parts.append(f"主要变更: {modified_files[0].path}")
+
+        return " | ".join(summary_parts)
+
+    def _collect_warnings(
+        self, generated_files: List[GeneratedFile], file_contexts: Dict[str, FileInfo]
+    ) -> List[str]:
+        """收集警告信息"""
+        warnings = []
+
+        for file in generated_files:
+            if not file.content.strip():
+                warnings.append(f"{file.path}: 文件内容为空")
+
+            # 检查是否有 TODO/FIXME
+            if "TODO" in file.content or "FIXME" in file.content:
+                warnings.append(f"{file.path}: 包含未完成的标记")
+
+        return warnings
+
+    def _format_json(self, obj: Any) -> str:
+        """格式化 JSON"""
+        import json
+        return json.dumps(obj, ensure_ascii=False, indent=2)
+
+
+# 全局服务实例
+_codegen_service = None
+
+
+def get_codegen_service(llm_service) -> CodeDraftService:
+    """获取代码生成服务实例"""
+    global _codegen_service
+    if _codegen_service is None:
+        _codegen_service = CodeDraftService(llm_service)
+    return _codegen_service
diff --git a/backend/app/services/cognitive_conflict.py b/backend/app/services/cognitive_conflict.py
new file mode 100644
index 00000000..26475a88
--- /dev/null
+++ b/backend/app/services/cognitive_conflict.py
@@ -0,0 +1,473 @@
+"""
+Cognitive Conflict Detection & Resolution Service - 认知冲突检测与解决服务
+
+核心功能：
+1. 自动检测团队隐性认知差异
+2. 识别定义不一致（如"留存"vs"DAU"）
+3. 检测目标冲突（产品说X，开发实现Y）
+4. 主动推送对齐提醒
+5. 异步 1v1 认知对齐
+"""
+import asyncio
+import json
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from sqlalchemy import text
+
+from ..models import (
+    User, Project, TruthNode, Task,
+    CognitiveConflict, ProjectDomain
+)
+from .llm import LLMService
+
+
+class CognitiveConflictService:
+    """认知冲突检测与解决服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+        self.llm = LLMService()
+
+    async def detect_charter_conflicts(
+        self,
+        project_id: int,
+        tenant_id: int = 1
+    ) -> List[Dict[str, Any]]:
+        """
+        检测项目宪章中的潜在冲突
+
+        扫描：
+        1. 宪章中的成功指标定义是否清晰
+        2. 是否有模糊的目标描述
+        3. 是否有相互冲突的约束
+        """
+        # 获取项目宪章
+        result = await self.db.execute(
+            select(TruthNode).where(
+                TruthNode.project_id == project_id,
+                TruthNode.node_type == 'charter',
+                TruthNode.tenant_id == tenant_id
+            )
+        )
+        charter = result.scalar_one_or_none()
+
+        if not charter:
+            return []
+
+        # AI 分析宪章内容
+        conflicts = await self._analyze_charter_for_conflicts(
+            charter.content,
+            tenant_id=tenant_id,
+        )
+
+        # 保存检测到的冲突
+        for conflict_data in conflicts:
+            await self._create_conflict_record(
+                truth_node_id=charter.id,
+                conflict_type=conflict_data['type'],
+                description=conflict_data['description'],
+                severity=conflict_data['severity'],
+                suggested_resolution=conflict_data['suggested_resolution'],
+                tenant_id=tenant_id
+            )
+
+        return conflicts
+
+    async def detect_task_charter_mismatch(
+        self,
+        task: Task,
+        charter_content: str,
+        tenant_id: int = 1
+    ) -> Optional[Dict[str, Any]]:
+        """
+        检测任务与宪章的一致性
+
+        检查：
+        1. 任务描述是否与宪章目标一致
+        2. 技术选型是否符合约束
+        3. 是否偏离项目方向
+        """
+        prompt = f"""检查以下任务是否符合项目宪章：
+
+宪章内容：
+{charter_content}
+
+任务信息：
+标题：{task.title}
+描述：{task.description}
+优先级：{task.priority}
+预计工时：{task.estimated_hours} 小时
+
+请判断：
+1. 任务是否与宪章目标一致？
+2. 是否有潜在的偏离？
+3. 是否存在隐性冲突？
+
+以 JSON 格式返回：
+{{
+    "is_aligned": boolean,
+    "conflict_type": "目标不一致/技术约束冲突/优先级错配",
+    "description": "冲突描述",
+    "severity": "high/medium/low",
+    "suggestion": "建议调整"
+}}
+
+如果任务与宪章一致，is_aligned 设为 true，conflict_type 为 null。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                project_id=task.project_id if getattr(task, "project_id", None) is not None else None,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            result = json.loads(response.strip())
+
+            if not result.get('is_aligned', True):
+                # 创建冲突记录
+                await self._create_conflict_record(
+                    truth_node_id=task.truth_node_id,
+                    conflict_type=result.get('conflict_type', 'unknown'),
+                    description=result.get('description', ''),
+                    severity=result.get('severity', 'medium'),
+                    suggested_resolution=result.get('suggestion', ''),
+                    related_task_id=task.id,
+                    tenant_id=tenant_id
+                )
+                return result
+
+            return None
+
+        except Exception as e:
+            print(f"[CognitiveConflict] Task mismatch detection failed: {e}")
+            return None
+
+    async def detect_user_definition_conflicts(
+        self,
+        project_id: int,
+        tenant_id: int = 1
+    ) -> List[Dict[str, Any]]:
+        """
+        检测用户定义冲突
+
+        例如：
+        - 产品在宪章中写"用户留存提升10%"
+        - 开发在实现"DAU提升10%"
+        - 双方都没有争论，但目标不一致
+        """
+        # 获取项目的所有宪章和任务
+        charters_result = await self.db.execute(
+            select(TruthNode).where(
+                TruthNode.project_id == project_id,
+                TruthNode.node_type == 'charter',
+                TruthNode.tenant_id == tenant_id
+            )
+        )
+        charters = charters_result.scalars().all()
+
+        tasks_result = await self.db.execute(
+            select(Task).where(
+                Task.project_id == project_id,
+                Task.tenant_id == tenant_id
+            )
+        )
+        tasks = tasks_result.scalars().all()
+
+        # 构建所有定义的上下文
+        definitions = []
+
+        for charter in charters:
+            definitions.append({
+                'source_type': 'charter',
+                'source_id': charter.id,
+                'content': charter.content
+            })
+
+        for task in tasks:
+            definitions.append({
+                'source_type': 'task',
+                'source_id': task.id,
+                'content': f"{task.title}: {task.description}"
+            })
+
+        if not definitions:
+            return []
+
+        # AI 检测定义冲突
+        prompt = f"""分析以下项目定义中是否存在隐性认知冲突：
+
+{json.dumps(definitions, ensure_ascii=False, indent=2)}
+
+请查找：
+1. 相同概念的不同表述（如"留存"vs"DAU"）
+2. 相矛盾的目标（如"降低成本"vs"增加功能"）
+3. 不一致的指标定义
+
+以 JSON 格式返回：
+[
+    {{
+        "conflict_type": "定义不一致",
+        "description": "具体描述冲突",
+        "involved_sources": ["charter:1", "task:2"],
+        "severity": "high/medium/low",
+        "suggested_alignment": "建议的对齐方案"
+    }}
+]
+
+如果没有发现冲突，返回空数组。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                project_id=project_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            conflicts = json.loads(response.strip())
+
+            # 保存检测到的冲突
+            for conflict in conflicts:
+                await self._create_conflict_record(
+                    truth_node_id=int(conflict['involved_sources'][0].split(':')[1]),
+                    conflict_type=conflict['conflict_type'],
+                    description=conflict['description'],
+                    severity=conflict['severity'],
+                    suggested_resolution=conflict.get('suggested_alignment', ''),
+                    tenant_id=tenant_id
+                )
+
+            return conflicts
+
+        except Exception as e:
+            print(f"[CognitiveConflict] Definition conflict detection failed: {e}")
+            return []
+
+    async def propose_alignment_resolution(
+        self,
+        conflict_id: int,
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        提议认知冲突解决方案
+
+        为冲突双方生成对齐方案：
+        1. 分析冲突根因
+        2. 提出妥协方案
+        3. 生成可执行的对齐计划
+        """
+        # 获取冲突记录
+        result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.id == conflict_id,
+                CognitiveConflict.tenant_id == tenant_id
+            )
+        )
+        conflict = result.scalar_one_or_none()
+
+        if not conflict:
+            raise ValueError(f"Conflict {conflict_id} not found")
+
+        # AI 生成解决方案
+        prompt = f"""为以下认知冲突提供解决方案：
+
+冲突类型：{conflict.conflict_type}
+描述：{conflict.description}
+
+请提供：
+1. 冲突根因分析
+2. 双方都能接受的妥协方案
+3. 具体的对齐步骤
+
+以 JSON 格式返回：
+{{
+    "root_cause": "根本原因",
+    "resolution_plan": ["步骤1", "步骤2"],
+    "compromise": "妥协方案",
+    "expected_outcome": "预期结果"
+}}"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            resolution = json.loads(response.strip())
+
+            # 更新冲突状态
+            conflict.status = 'resolution_proposed'
+            conflict.description += f"\n\nAI 建议方案：{resolution.get('compromise', '')}"
+
+            await self.db.commit()
+
+            return resolution
+
+        except Exception as e:
+            print(f"[CognitiveConflict] Resolution proposal failed: {e}")
+            return {}
+
+    async def trigger_async_alignment(
+        self,
+        conflict_id: int,
+        user1_id: int,
+        user2_id: int,
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        触发异步 1v1 认知对齐
+
+        为冲突双方创建独立的对齐任务：
+        1. 发送对齐邀请
+        2. 提供背景信息
+        3. 收集双方观点
+        4. AI 生成共识方案
+        """
+        # 获取用户信息
+        result1 = await self.db.execute(
+            select(User).where(User.id == user1_id)
+        )
+        user1 = result1.scalar_one_or_none()
+
+        result2 = await self.db.execute(
+            select(User).where(User.id == user2_id)
+        )
+        user2 = result2.scalar_one_or_none()
+
+        if not user1 or not user2:
+            raise ValueError("Users not found")
+
+        # 获取冲突详情
+        result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.id == conflict_id,
+                CognitiveConflict.tenant_id == tenant_id
+            )
+        )
+        conflict = result.scalar_one_or_none()
+
+        # TODO: 实际实现需要：
+        # 1. 发送飞书消息/邮件通知双方
+        # 2. 创建对齐任务
+        # 3. 记录对齐过程
+
+        return {
+            'alignment_session_id': f"align_{conflict_id}",
+            'user1': user1.username,
+            'user2': user2.username,
+            'conflict': conflict.conflict_type,
+            'status': 'initiated'
+        }
+
+    async def _create_conflict_record(
+        self,
+        truth_node_id: int,
+        conflict_type: str,
+        description: str,
+        severity: str,
+        suggested_resolution: str = "",
+        related_task_id: Optional[int] = None,
+        tenant_id: int = 1
+    ):
+        """创建认知冲突记录"""
+        from sqlalchemy import text
+        next_conflict_id = (await self.db.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM cognitive_conflicts"))).scalar() or 1
+
+        conflict = CognitiveConflict(
+            id=next_conflict_id,
+            truth_node_id=truth_node_id,
+            conflict_type=conflict_type,
+            description=description,
+            involved_user_ids=json.dumps([]),
+            status='detected',
+            tenant_id=tenant_id
+        )
+        self.db.add(conflict)
+        await self.db.commit()
+
+    async def _analyze_charter_for_conflicts(
+        self,
+        charter_content: str,
+        tenant_id: int,
+    ) -> List[Dict[str, Any]]:
+        """AI 分析宪章内容中的冲突"""
+        prompt = f"""分析以下项目宪章中是否存在潜在冲突或问题：
+
+{charter_content}
+
+请检查：
+1. 目标是否模糊或相互矛盾？
+2. 约束条件是否现实？
+3. 成功指标是否可衡量？
+4. 是否遗漏重要方面？
+
+以 JSON 格式返回：
+[
+    {{
+        "type": "目标模糊/约束冲突/指标不可测量",
+        "description": "具体问题描述",
+        "severity": "high/medium/low",
+        "suggested_resolution": "建议的改进方案"
+    }}
+"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except:
+            return []
+
+    async def get_detected_conflicts(
+        self,
+        tenant_id: int = 1
+    ) -> List[Dict[str, Any]]:
+        """获取所有检测到的冲突"""
+        result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.tenant_id == tenant_id,
+                CognitiveConflict.status == 'detected'
+            )
+        )
+        conflicts = result.scalars().all()
+
+        return [
+            {
+                'id': str(c.id),
+                'type': c.conflict_type,
+                'description': c.description,
+                'severity': 'medium' if 'high' not in c.description.lower() else 'high',
+                'truth_node_id': str(c.truth_node_id)
+            }
+            for c in conflicts
+        ]
+
+
+# 全局服务实例
+def get_cognitive_conflict_service(db: AsyncSession) -> CognitiveConflictService:
+    """获取认知冲突服务实例"""
+    return CognitiveConflictService(db)
diff --git a/backend/app/services/context_cache.py b/backend/app/services/context_cache.py
new file mode 100644
index 00000000..ead979b0
--- /dev/null
+++ b/backend/app/services/context_cache.py
@@ -0,0 +1,1155 @@
+"""AI Steward 上下文缓存服务（Redis）与失效钩子。"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Optional
+
+from sqlalchemy import and_, case, event, func, inspect, or_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import Session
+
+from ..models import Project, ProjectSignalLink, Signal, Task, User, UserProject, UserWorkload
+
+try:
+    import redis.asyncio as redis_async
+    from redis import Redis as RedisSync
+
+    REDIS_AVAILABLE = True
+except Exception:  # pragma: no cover
+    REDIS_AVAILABLE = False
+    redis_async = None
+    RedisSync = None
+
+
+SYSTEM_TTL_SECONDS = 30 * 60
+USER_TTL_SECONDS = 5 * 60
+SESSION_TTL_SECONDS = 60 * 60
+PROJECT_LIST_TTL_SECONDS = 10 * 60
+SIGNAL_PENDING_TTL_SECONDS = 5 * 60
+SESSION_MAX_MESSAGES = 20
+DONE_TASK_STATUSES = {"done", "completed", "closed", "resolved", "archived"}
+PENDING_SIGNAL_STATUSES = {"pending", "todo", "open", "待处理"}
+
+
+def _safe_int(value: Any, fallback: int = 0) -> int:
+    try:
+        return int(value)
+    except Exception:
+        return fallback
+
+
+def _json_loads(raw: Optional[str], default: Any) -> Any:
+    if not raw:
+        return default
+    try:
+        return json.loads(raw)
+    except Exception:
+        return default
+
+
+class ContextCacheService:
+    def __init__(self, redis_url: Optional[str] = None) -> None:
+        self.redis_url = redis_url or os.getenv("REDIS_URL")
+        self._redis: Optional[Any] = None
+
+    @staticmethod
+    def _system_key(tenant_id: int) -> str:
+        return f"pmflow:context:system:{tenant_id}"
+
+    @staticmethod
+    def _user_key(tenant_id: int, user_id: int) -> str:
+        return f"pmflow:context:user:{tenant_id}:{user_id}"
+
+    @staticmethod
+    def _session_key(tenant_id: int, session_id: int) -> str:
+        return f"pmflow:qa:session:{tenant_id}:{session_id}"
+
+    @staticmethod
+    def _project_list_key(tenant_id: int) -> str:
+        return f"pmflow:project:list:{tenant_id}"
+
+    @staticmethod
+    def _signal_pending_key(tenant_id: int) -> str:
+        return f"pmflow:signal:pending_count:{tenant_id}"
+
+    async def _get_redis(self):
+        if not REDIS_AVAILABLE or not self.redis_url:
+            return None
+        if self._redis is None:
+            self._redis = redis_async.from_url(self.redis_url, decode_responses=True)
+        return self._redis
+
+    async def _get_json(self, key: str) -> Optional[dict[str, Any]]:
+        redis = await self._get_redis()
+        if redis is None:
+            return None
+        try:
+            raw = await redis.get(key)
+        except Exception:
+            return None
+        if not raw:
+            return None
+        payload = _json_loads(raw, None)
+        return payload if isinstance(payload, dict) else None
+
+    async def _set_json(self, key: str, value: dict[str, Any], ttl: int) -> None:
+        redis = await self._get_redis()
+        if redis is None:
+            return
+        try:
+            await redis.setex(key, ttl, json.dumps(value, ensure_ascii=False))
+        except Exception:
+            pass
+
+    @staticmethod
+    def _is_system_snapshot_complete(payload: Optional[dict[str, Any]]) -> bool:
+        if not isinstance(payload, dict):
+            return False
+        return isinstance(payload.get("assistant_profile"), dict) and isinstance(payload.get("platform_profile"), dict)
+
+    async def get_system_snapshot(self, *, db: AsyncSession, tenant_id: int) -> dict[str, Any]:
+        tid = _safe_int(tenant_id, 1)
+        key = self._system_key(tid)
+        cached = await self._get_json(key)
+        if self._is_system_snapshot_complete(cached):
+            return cached
+
+        payload = await self._build_system_snapshot(db=db, tenant_id=tid)
+        await self._set_json(key, payload, SYSTEM_TTL_SECONDS)
+        return payload
+
+    async def get_user_snapshot(
+        self,
+        *,
+        db: AsyncSession,
+        tenant_id: int,
+        user_id: Optional[int],
+    ) -> dict[str, Any]:
+        if user_id is None:
+            return {}
+
+        tid = _safe_int(tenant_id, 1)
+        uid = _safe_int(user_id, 0)
+        if uid <= 0:
+            return {}
+
+        key = self._user_key(tid, uid)
+        cached = await self._get_json(key)
+        if cached is not None:
+            return cached
+
+        payload = await self._build_user_snapshot(db=db, tenant_id=tid, user_id=uid)
+        await self._set_json(key, payload, USER_TTL_SECONDS)
+        return payload
+
+    async def _build_system_snapshot(self, *, db: AsyncSession, tenant_id: int) -> dict[str, Any]:
+        projects = await self._get_project_list_cached(db=db, tenant_id=tenant_id)
+        team = await self._build_team_members_snapshot(db=db, tenant_id=tenant_id)
+        pending = await self._get_pending_signal_cached(db=db, tenant_id=tenant_id)
+        assistant_profile = self._build_assistant_profile()
+        platform_profile = self._build_platform_profile()
+        return {
+            "tenant_id": tenant_id,
+            "assistant_profile": assistant_profile,
+            "platform_profile": platform_profile,
+            "projects": projects.get("items", []),
+            "team_members": team,
+            "pending_signals": pending,
+            "snapshot_at": datetime.now(timezone.utc).isoformat(),
+        }
+
+    @staticmethod
+    def _build_assistant_profile() -> dict[str, Any]:
+        return {
+            "name": "PMFlow AI 管家",
+            "identity": "PMFlow 内置的项目管理与技术协同助手",
+            "responsibilities": [
+                "识别需求并路由到项目/信号",
+                "结合结构化数据与知识库回答问题",
+                "辅助任务拆解、分配与优先级判断",
+                "输出基于证据的系统状态与风险提示",
+            ],
+            "principles": [
+                "只基于可验证上下文回答",
+                "证据不足时明确说明不足字段",
+                "优先使用系统快照中的实时结构化数据",
+            ],
+        }
+
+    @staticmethod
+    def _count_python_files(target: Path) -> int:
+        if not target.exists() or not target.is_dir():
+            return 0
+        return sum(1 for p in target.rglob("*.py") if p.is_file())
+
+    @staticmethod
+    def _estimate_api_routes(api_dir: Path) -> int:
+        if not api_dir.exists() or not api_dir.is_dir():
+            return 0
+        markers = (
+            "@router.get(",
+            "@router.post(",
+            "@router.put(",
+            "@router.patch(",
+            "@router.delete(",
+            "@router.websocket(",
+        )
+        route_count = 0
+        for py_file in api_dir.rglob("*.py"):
+            if not py_file.is_file() or py_file.name.startswith("__"):
+                continue
+            try:
+                text = py_file.read_text(encoding="utf-8", errors="ignore")
+            except Exception:
+                continue
+            for marker in markers:
+                route_count += text.count(marker)
+        return route_count
+
+    @staticmethod
+    def _detect_database_stack() -> str:
+        url = str(os.getenv("DATABASE_URL") or "").lower()
+        if "postgresql" in url:
+            return "PostgreSQL (pgvector-ready)"
+        if "mysql" in url:
+            return "MySQL"
+        if "sqlite" in url:
+            return "SQLite"
+        return "Unknown"
+
+    @staticmethod
+    def _default_feature_catalog_items() -> list[dict[str, Any]]:
+        return [
+            {
+                "route": "/workloads",
+                "name": "团队负载",
+                "summary": "查看团队成员负载、瓶颈与调配建议。",
+                "page_component": "WorkloadsPage",
+                "page_file": "frontend/src/pages/WorkloadsPage.tsx",
+                "apis": ["/workloads/", "/workloads/ai-optimize", "/workloads/ai-optimize/apply"],
+                "aliases": ["workloads", "负载", "工作量", "capacity", "团队负载"],
+            },
+            {
+                "route": "/projects",
+                "name": "项目列表",
+                "summary": "管理项目清单、状态、入口跳转与总览信息。",
+                "page_component": "ProjectsPage",
+                "page_file": "frontend/src/pages/ProjectsPage.tsx",
+                "apis": ["/projects/", "/projects/?summary=1"],
+                "aliases": ["projects", "项目", "项目列表", "project list"],
+            },
+            {
+                "route": "/ai-steward",
+                "name": "AI 管家",
+                "summary": "AI 任务协同与治理中心，包含简报、冲突检测、问答历史。",
+                "page_component": "AIStewardPage",
+                "page_file": "frontend/src/pages/AIStewardPage.tsx",
+                "apis": ["/ai-steward/", "/ai-steward/rag/ask", "/rag/qa-sessions", "/rag/qa-records"],
+                "aliases": ["ai steward", "ai管家", "问答", "rag", "智能问答"],
+            },
+            {
+                "route": "/ai-config",
+                "name": "智能配置",
+                "summary": "配置飞书/企微/GitHub/知识库与模型参数，决定 AI 数据源与行为。",
+                "page_component": "AIConfigPage",
+                "page_file": "frontend/src/pages/AIConfigPage.tsx",
+                "apis": [
+                    "/ai-steward/config/suggest",
+                    "/feishu/config",
+                    "/wecom/config",
+                    "/rag/knowledge-base/docs",
+                    "/rag/knowledge-base/upload",
+                    "/rag/knowledge-base/urls",
+                ],
+                "aliases": [
+                    "智能配置",
+                    "ai配置",
+                    "配置中心",
+                    "integration",
+                    "集成配置",
+                    "knowledge base",
+                ],
+            },
+            {
+                "route": "/skills",
+                "name": "能力管理",
+                "summary": "管理能力标签、人员能力画像、任务匹配与推荐。",
+                "page_component": "SkillsPage",
+                "page_file": "frontend/src/pages/SkillsPage.tsx",
+                "apis": ["/skills/tags", "/skills/recommend", "/skills/users/{id}/bindings"],
+                "aliases": ["skills", "能力", "技能", "ability", "能力管理"],
+            },
+            {
+                "route": "/users",
+                "name": "用户管理",
+                "summary": "管理用户档案、角色、状态与权限入口。",
+                "page_component": "UsersPage",
+                "page_file": "frontend/src/pages/UsersPage.tsx",
+                "apis": ["/users/", "/users/{id}", "/users/bulk-delete", "/users/bulk-assign"],
+                "aliases": ["users", "成员管理", "用户", "人员管理"],
+            },
+        ]
+
+    @staticmethod
+    def _extract_router_import_map(router_text: str) -> dict[str, str]:
+        mapping: dict[str, str] = {}
+        for component, page_rel in re.findall(
+            r'import\s+([A-Za-z_][A-Za-z0-9_]*)\s+from\s+"@/pages/([^"]+)";',
+            str(router_text or ""),
+        ):
+            rel = str(page_rel).strip()
+            if not rel:
+                continue
+            if not rel.endswith(".tsx") and not rel.endswith(".ts"):
+                rel = f"{rel}.tsx"
+            mapping[str(component).strip()] = f"frontend/src/pages/{rel}"
+        return mapping
+
+    @staticmethod
+    def _extract_router_entries(router_text: str) -> list[tuple[str, str]]:
+        entries: list[tuple[str, str]] = []
+        for route, component in re.findall(
+            r'\{\s*path:\s*"([^"]+)"\s*,\s*element:\s*<([A-Za-z_][A-Za-z0-9_]*)\s*/?>\s*\}',
+            str(router_text or ""),
+        ):
+            clean_route = str(route or "").strip()
+            if not clean_route:
+                continue
+            if not clean_route.startswith("/"):
+                clean_route = f"/{clean_route}"
+            entries.append((clean_route, str(component).strip()))
+        return entries
+
+    @staticmethod
+    def _extract_sidebar_labels(sidebar_text: str) -> dict[str, str]:
+        labels: dict[str, str] = {}
+        for route, label in re.findall(
+            r'\{\s*to:\s*"([^"]+)"\s*,\s*label:\s*"([^"]+)"',
+            str(sidebar_text or ""),
+        ):
+            clean_route = str(route or "").strip()
+            if clean_route and not clean_route.startswith("/"):
+                clean_route = f"/{clean_route}"
+            if clean_route:
+                labels[clean_route] = str(label or "").strip()
+        return labels
+
+    @staticmethod
+    def _route_feature_summary(route: str) -> tuple[str, list[str], list[str]]:
+        route_map: dict[str, tuple[str, list[str], list[str]]] = {
+            "/workloads": (
+                "查看团队成员负载、瓶颈与调配建议。",
+                ["/workloads/", "/workloads/ai-optimize", "/workloads/ai-optimize/apply"],
+                ["workloads", "负载", "工作量", "capacity"],
+            ),
+            "/projects": (
+                "管理项目清单、状态、入口跳转与总览信息。",
+                ["/projects/", "/projects/?summary=1"],
+                ["projects", "项目", "项目列表", "project list"],
+            ),
+            "/ai-steward": (
+                "AI 任务协同与治理中心，包含简报、冲突检测、问答历史。",
+                ["/ai-steward/", "/ai-steward/rag/ask", "/rag/qa-sessions", "/rag/qa-records"],
+                ["ai steward", "ai管家", "问答", "rag", "智能问答"],
+            ),
+            "/ai-config": (
+                "配置飞书/企微/GitHub/知识库与模型参数，决定 AI 数据源与行为。",
+                [
+                    "/ai-steward/config/suggest",
+                    "/feishu/config",
+                    "/wecom/config",
+                    "/rag/knowledge-base/docs",
+                    "/rag/knowledge-base/upload",
+                    "/rag/knowledge-base/urls",
+                ],
+                ["智能配置", "ai配置", "配置中心", "integration", "集成配置", "knowledge base"],
+            ),
+            "/skills": (
+                "管理能力标签、人员能力画像、任务匹配与推荐。",
+                ["/skills/tags", "/skills/recommend", "/skills/users/{id}/bindings"],
+                ["skills", "能力", "技能", "ability", "能力管理"],
+            ),
+            "/users": (
+                "管理用户档案、角色、状态与权限入口。",
+                ["/users/", "/users/{id}", "/users/bulk-delete", "/users/bulk-assign"],
+                ["users", "成员管理", "用户", "人员管理"],
+            ),
+        }
+        if route in route_map:
+            return route_map[route]
+        return (
+            "系统功能页面入口。",
+            [],
+            [route.strip("/"), route.replace("/", "_").strip("_")],
+        )
+
+    @staticmethod
+    def _normalize_aliases(*groups: list[str]) -> list[str]:
+        merged: list[str] = []
+        for group in groups:
+            for item in group:
+                alias = str(item or "").strip().lower()
+                if not alias:
+                    continue
+                if alias not in merged:
+                    merged.append(alias)
+        return merged
+
+    def _build_feature_catalog(
+        self,
+        *,
+        backend_root: Path,
+    ) -> dict[str, Any]:
+        workspace_root = backend_root.parent
+        frontend_src = workspace_root / "frontend" / "src"
+        router_file = frontend_src / "router" / "index.tsx"
+        sidebar_file = frontend_src / "components" / "layout" / "Sidebar.tsx"
+
+        if not frontend_src.exists():
+            fallback_items = self._default_feature_catalog_items()
+            return {
+                "items": fallback_items,
+                "count": len(fallback_items),
+                "source": "fallback",
+            }
+
+        router_text = ""
+        sidebar_text = ""
+        try:
+            router_text = router_file.read_text(encoding="utf-8", errors="ignore")
+        except Exception:
+            router_text = ""
+        try:
+            sidebar_text = sidebar_file.read_text(encoding="utf-8", errors="ignore")
+        except Exception:
+            sidebar_text = ""
+
+        import_map = self._extract_router_import_map(router_text)
+        router_entries = self._extract_router_entries(router_text)
+        sidebar_labels = self._extract_sidebar_labels(sidebar_text)
+
+        merged: dict[str, dict[str, Any]] = {}
+        for route, component in router_entries:
+            page_file = import_map.get(component) or ""
+            default_name = component.replace("Page", "")
+            name = sidebar_labels.get(route) or default_name
+            summary, apis, aliases = self._route_feature_summary(route)
+            route_aliases = [route.strip("/"), route.replace("/", "_").strip("_"), component]
+            merged[route] = {
+                "route": route,
+                "name": name,
+                "summary": summary,
+                "page_component": component,
+                "page_file": page_file,
+                "apis": apis,
+                "aliases": self._normalize_aliases(aliases, route_aliases, [name]),
+            }
+
+        for route, label in sidebar_labels.items():
+            if route not in merged:
+                summary, apis, aliases = self._route_feature_summary(route)
+                merged[route] = {
+                    "route": route,
+                    "name": label,
+                    "summary": summary,
+                    "page_component": "",
+                    "page_file": "",
+                    "apis": apis,
+                    "aliases": self._normalize_aliases(aliases, [route.strip("/"), label]),
+                }
+            else:
+                merged[route]["name"] = label
+                merged[route]["aliases"] = self._normalize_aliases(
+                    list(merged[route].get("aliases") or []),
+                    [label],
+                )
+
+        items = sorted(merged.values(), key=lambda x: str(x.get("route") or ""))
+        if not items:
+            items = self._default_feature_catalog_items()
+            source = "fallback"
+        else:
+            source = "frontend_scan"
+        return {
+            "items": items,
+            "count": len(items),
+            "source": source,
+        }
+
+    def _build_platform_profile(self) -> dict[str, Any]:
+        backend_root = Path(__file__).resolve().parents[2]
+        workspace_root = backend_root.parent
+        app_root = backend_root / "app"
+        api_dir = app_root / "api"
+        services_dir = app_root / "services"
+        models_dir = app_root / "models"
+        schemas_dir = app_root / "schemas"
+        dependencies_dir = app_root / "dependencies"
+        core_dir = app_root / "core"
+        config_dir = app_root / "config"
+
+        module_counts = {
+            "api": self._count_python_files(api_dir),
+            "services": self._count_python_files(services_dir),
+            "models": self._count_python_files(models_dir),
+            "schemas": self._count_python_files(schemas_dir),
+            "dependencies": self._count_python_files(dependencies_dir),
+            "core": self._count_python_files(core_dir),
+            "config": self._count_python_files(config_dir),
+        }
+        app_python_files = self._count_python_files(app_root)
+        route_count = self._estimate_api_routes(api_dir)
+        running_in_docker = Path("/.dockerenv").exists()
+        redis_enabled = bool(str(os.getenv("REDIS_URL") or "").strip())
+        feature_catalog = self._build_feature_catalog(backend_root=backend_root)
+        frontend_src = workspace_root / "frontend" / "src"
+        frontend_ts_files = 0
+        if frontend_src.exists():
+            frontend_ts_files = sum(
+                1
+                for p in frontend_src.rglob("*")
+                if p.is_file() and p.suffix.lower() in {".ts", ".tsx", ".js", ".jsx"}
+            )
+
+        return {
+            "runtime": "docker" if running_in_docker else "local",
+            "backend": {
+                "framework": "FastAPI",
+                "orm": "SQLAlchemy 2.0",
+                "python": "3.11+",
+                "module_counts": module_counts,
+                "api_route_count_estimate": route_count,
+            },
+            "infrastructure": {
+                "database": self._detect_database_stack(),
+                "cache": "Redis" if redis_enabled else "Disabled/Unknown",
+            },
+            "frontend": {
+                "framework": "React 18 + TypeScript + Vite",
+                "source_root": str(frontend_src),
+                "route_count": int(feature_catalog.get("count") or 0),
+                "source_file_count": int(frontend_ts_files),
+            },
+            "feature_catalog": feature_catalog.get("items") if isinstance(feature_catalog.get("items"), list) else [],
+            "source_scan": {
+                "backend_root": str(backend_root),
+                "app_python_files": app_python_files,
+                "feature_catalog_source": str(feature_catalog.get("source") or "unknown"),
+                "scanned_at": datetime.now(timezone.utc).isoformat(),
+            },
+        }
+
+    async def _get_project_list_cached(self, *, db: AsyncSession, tenant_id: int) -> dict[str, Any]:
+        key = self._project_list_key(tenant_id)
+        cached = await self._get_json(key)
+        if cached is not None:
+            return cached
+
+        project_rows = (
+            await db.execute(
+                select(
+                    Project.id,
+                    Project.name,
+                    Project.title,
+                    Project.status,
+                    Project.owner_id,
+                    User.name.label("owner_name"),
+                )
+                .outerjoin(
+                    User,
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id == Project.owner_id,
+                    ),
+                )
+                .where(Project.tenant_id == tenant_id)
+                .order_by(Project.updated_at.desc(), Project.id.desc())
+            )
+        ).all()
+
+        project_ids = [int(row.id) for row in project_rows]
+        signal_count_map: dict[int, int] = {}
+        if project_ids:
+            signal_rows = (
+                await db.execute(
+                    select(ProjectSignalLink.project_id, func.count(ProjectSignalLink.id))
+                    .where(
+                        and_(
+                            ProjectSignalLink.tenant_id == tenant_id,
+                            ProjectSignalLink.project_id.in_(project_ids),
+                        )
+                    )
+                    .group_by(ProjectSignalLink.project_id)
+                )
+            ).all()
+            signal_count_map = {int(pid): int(cnt or 0) for pid, cnt in signal_rows}
+
+        task_count_map: dict[int, int] = {}
+        task_done_map: dict[int, int] = {}
+        if project_ids:
+            done_case = case((func.lower(Task.status).in_(DONE_TASK_STATUSES), 1), else_=0)
+            task_rows = (
+                await db.execute(
+                    select(
+                        Task.project_id,
+                        func.count(Task.id).label("task_total"),
+                        func.sum(done_case).label("done_count"),
+                    )
+                    .where(
+                        and_(
+                            Task.tenant_id == tenant_id,
+                            Task.project_id.in_(project_ids),
+                        )
+                    )
+                    .group_by(Task.project_id)
+                )
+            ).all()
+            task_count_map = {int(pid): int(total or 0) for pid, total, _ in task_rows}
+            task_done_map = {int(pid): int(done or 0) for pid, _, done in task_rows}
+
+        items: list[dict[str, Any]] = []
+        for row in project_rows:
+            pid = int(row.id)
+            task_total = int(task_count_map.get(pid, 0))
+            done_total = int(task_done_map.get(pid, 0))
+            completion_rate = round((done_total / task_total) * 100, 2) if task_total > 0 else 0.0
+            title = str(row.title or row.name or f"Project#{pid}")
+            items.append(
+                {
+                    "id": pid,
+                    "name": title,
+                    "status": str(row.status or "unknown"),
+                    "current_stage": str(row.status or "unknown"),
+                    "owner_id": _safe_int(row.owner_id, 0) or None,
+                    "owner_name": str(row.owner_name or "未知"),
+                    "signal_count": int(signal_count_map.get(pid, 0)),
+                    "task_count": task_total,
+                    "completion_rate": completion_rate,
+                }
+            )
+
+        payload = {
+            "items": items,
+            "count": len(items),
+            "snapshot_at": datetime.now(timezone.utc).isoformat(),
+        }
+        await self._set_json(key, payload, PROJECT_LIST_TTL_SECONDS)
+        return payload
+
+    async def _build_team_members_snapshot(self, *, db: AsyncSession, tenant_id: int) -> list[dict[str, Any]]:
+        users = (
+            await db.execute(
+                select(User.id, User.name, User.role, User.workload).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.is_active.is_(True),
+                    )
+                )
+            )
+        ).all()
+        if not users:
+            return []
+
+        user_ids = [int(u.id) for u in users]
+        workload_rows = (
+            await db.execute(
+                select(UserWorkload.user_id, UserWorkload.workload_ratio).where(
+                    and_(
+                        UserWorkload.tenant_id == tenant_id,
+                        UserWorkload.user_id.in_(user_ids),
+                    )
+                )
+            )
+        ).all()
+        workload_map = {int(uid): float(ratio or 0.0) for uid, ratio in workload_rows}
+
+        project_rows = (
+            await db.execute(
+                select(UserProject.user_id, UserProject.project_id).where(
+                    and_(
+                        UserProject.tenant_id == tenant_id,
+                        UserProject.user_id.in_(user_ids),
+                    )
+                )
+            )
+        ).all()
+        projects_map: dict[int, list[int]] = {}
+        for uid, pid in project_rows:
+            projects_map.setdefault(int(uid), []).append(int(pid))
+
+        result: list[dict[str, Any]] = []
+        for row in users:
+            uid = int(row.id)
+            ratio = workload_map.get(uid)
+            if ratio is None:
+                ratio = round(float(row.workload or 0) / 100.0, 2)
+            result.append(
+                {
+                    "id": uid,
+                    "name": str(row.name or f"User#{uid}"),
+                    "role": str(row.role or "unknown"),
+                    "workload_ratio": float(ratio),
+                    "assigned_projects": projects_map.get(uid, []),
+                }
+            )
+        return result
+
+    async def _get_pending_signal_cached(self, *, db: AsyncSession, tenant_id: int) -> dict[str, Any]:
+        key = self._signal_pending_key(tenant_id)
+        cached = await self._get_json(key)
+        if cached is not None:
+            return cached
+
+        pending_filter = or_(
+            Signal.processed.is_(False),
+            func.lower(Signal.status).in_(PENDING_SIGNAL_STATUSES),
+        )
+        p0_case = case(
+            (
+                func.lower(Signal.priority).in_({"critical", "p0", "urgent", "highest"}),
+                1,
+            ),
+            else_=0,
+        )
+        p1_case = case((func.lower(Signal.priority).in_({"high", "p1"}), 1), else_=0)
+        p2_case = case((func.lower(Signal.priority).in_({"medium", "p2", "normal"}), 1), else_=0)
+
+        row = (
+            await db.execute(
+                select(
+                    func.count(Signal.id),
+                    func.sum(p0_case),
+                    func.sum(p1_case),
+                    func.sum(p2_case),
+                ).where(
+                    and_(
+                        Signal.tenant_id == tenant_id,
+                        pending_filter,
+                    )
+                )
+            )
+        ).first()
+        total, p0_count, p1_count, p2_count = row or (0, 0, 0, 0)
+
+        payload = {
+            "total": int(total or 0),
+            "p0_count": int(p0_count or 0),
+            "p1_count": int(p1_count or 0),
+            "p2_count": int(p2_count or 0),
+            "snapshot_at": datetime.now(timezone.utc).isoformat(),
+        }
+        await self._set_json(key, payload, SIGNAL_PENDING_TTL_SECONDS)
+        return payload
+
+    async def _build_user_snapshot(
+        self,
+        *,
+        db: AsyncSession,
+        tenant_id: int,
+        user_id: int,
+    ) -> dict[str, Any]:
+        owned_rows = (
+            await db.execute(
+                select(Project.id).where(
+                    and_(
+                        Project.tenant_id == tenant_id,
+                        Project.owner_id == user_id,
+                    )
+                )
+            )
+        ).scalars().all()
+        member_rows = (
+            await db.execute(
+                select(UserProject.project_id).where(
+                    and_(
+                        UserProject.tenant_id == tenant_id,
+                        UserProject.user_id == user_id,
+                    )
+                )
+            )
+        ).scalars().all()
+        project_ids = sorted({_safe_int(pid, 0) for pid in [*owned_rows, *member_rows] if _safe_int(pid, 0) > 0})
+
+        project_rows: list[Any] = []
+        if project_ids:
+            project_rows = (
+                await db.execute(
+                    select(Project.id, Project.name, Project.title, Project.status).where(
+                        and_(
+                            Project.tenant_id == tenant_id,
+                            Project.id.in_(project_ids),
+                        )
+                    )
+                )
+            ).all()
+
+        open_tasks = (
+            await db.execute(
+                select(Task.id, Task.project_id, Task.title, Task.status, Task.due_date).where(
+                    and_(
+                        Task.tenant_id == tenant_id,
+                        Task.assignee_id == user_id,
+                        func.lower(Task.status).notin_(DONE_TASK_STATUSES),
+                    )
+                )
+            )
+        ).all()
+
+        pending_signals: list[dict[str, Any]] = []
+        if project_ids:
+            pending_rows = (
+                await db.execute(
+                    select(Signal.id, Signal.title, Signal.priority, Signal.status, Signal.matched_project_id).where(
+                        and_(
+                            Signal.tenant_id == tenant_id,
+                            Signal.matched_project_id.in_(project_ids),
+                            or_(
+                                Signal.processed.is_(False),
+                                func.lower(Signal.status).in_(PENDING_SIGNAL_STATUSES),
+                            ),
+                        )
+                    ).limit(60)
+                )
+            ).all()
+            pending_signals = [
+                {
+                    "id": int(row.id),
+                    "title": str(row.title or ""),
+                    "priority": str(row.priority or ""),
+                    "status": str(row.status or ""),
+                    "project_id": _safe_int(row.matched_project_id, 0) or None,
+                }
+                for row in pending_rows
+            ]
+
+        workload_row = (
+            await db.execute(
+                select(UserWorkload.workload_ratio).where(
+                    and_(
+                        UserWorkload.tenant_id == tenant_id,
+                        UserWorkload.user_id == user_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if workload_row is None:
+            fallback = (
+                await db.execute(
+                    select(User.workload).where(
+                        and_(
+                            User.tenant_id == tenant_id,
+                            User.id == user_id,
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            workload_ratio = round(float(fallback or 0) / 100.0, 2)
+        else:
+            workload_ratio = float(workload_row or 0.0)
+
+        return {
+            "user_id": user_id,
+            "projects": [
+                {
+                    "id": int(row.id),
+                    "name": str(row.title or row.name or f"Project#{int(row.id)}"),
+                    "status": str(row.status or "unknown"),
+                }
+                for row in project_rows
+            ],
+            "assigned_tasks": [
+                {
+                    "id": int(row.id),
+                    "project_id": int(row.project_id),
+                    "title": str(row.title or ""),
+                    "status": str(row.status or ""),
+                    "due_date": row.due_date.isoformat() if row.due_date else None,
+                }
+                for row in open_tasks
+            ],
+            "pending_signals": pending_signals,
+            "workload_ratio": workload_ratio,
+            "snapshot_at": datetime.now(timezone.utc).isoformat(),
+        }
+
+    async def get_session_messages(self, *, tenant_id: int, session_id: Optional[int]) -> list[dict[str, str]]:
+        if session_id is None:
+            return []
+        sid = _safe_int(session_id, 0)
+        if sid <= 0:
+            return []
+
+        redis = await self._get_redis()
+        if redis is None:
+            return []
+        key = self._session_key(_safe_int(tenant_id, 1), sid)
+        try:
+            raw_rows = await redis.lrange(key, -SESSION_MAX_MESSAGES, -1)
+        except Exception:
+            return []
+        messages: list[dict[str, str]] = []
+        for raw in raw_rows:
+            row = _json_loads(raw, None)
+            if not isinstance(row, dict):
+                continue
+            role = str(row.get("role") or "").strip().lower()
+            content = str(row.get("content") or "").strip()
+            if role in {"user", "assistant"} and content:
+                messages.append({"role": role, "content": content})
+        return messages
+
+    async def append_session_messages(
+        self,
+        *,
+        tenant_id: int,
+        session_id: Optional[int],
+        question: str,
+        answer: str,
+    ) -> None:
+        if session_id is None:
+            return
+        sid = _safe_int(session_id, 0)
+        if sid <= 0:
+            return
+
+        redis = await self._get_redis()
+        if redis is None:
+            return
+
+        key = self._session_key(_safe_int(tenant_id, 1), sid)
+        user_payload = json.dumps({"role": "user", "content": str(question or "")[:3000]}, ensure_ascii=False)
+        assistant_payload = json.dumps({"role": "assistant", "content": str(answer or "")[:4000]}, ensure_ascii=False)
+        try:
+            pipe = redis.pipeline()
+            pipe.rpush(key, user_payload, assistant_payload)
+            pipe.ltrim(key, -SESSION_MAX_MESSAGES, -1)
+            pipe.expire(key, SESSION_TTL_SECONDS)
+            await pipe.execute()
+        except Exception:
+            return
+
+    async def invalidate_system_context(self, *, tenant_id: int) -> None:
+        redis = await self._get_redis()
+        if redis is None:
+            return
+        tid = _safe_int(tenant_id, 1)
+        keys = [
+            self._system_key(tid),
+            self._project_list_key(tid),
+            self._signal_pending_key(tid),
+        ]
+        try:
+            await redis.delete(*keys)
+        except Exception:
+            pass
+
+    async def invalidate_user_context(self, *, tenant_id: int, user_id: int) -> None:
+        redis = await self._get_redis()
+        if redis is None:
+            return
+        uid = _safe_int(user_id, 0)
+        if uid <= 0:
+            return
+        try:
+            await redis.delete(self._user_key(_safe_int(tenant_id, 1), uid))
+        except Exception:
+            pass
+
+    @staticmethod
+    def messages_to_turns(messages: list[dict[str, str]]) -> list[dict[str, str]]:
+        turns: list[dict[str, str]] = []
+        pending_question: Optional[str] = None
+        for row in messages:
+            role = str(row.get("role") or "").lower()
+            content = str(row.get("content") or "")
+            if role == "user":
+                if pending_question:
+                    turns.append({"question": pending_question, "answer": ""})
+                pending_question = content
+                continue
+            if role == "assistant":
+                if pending_question is None:
+                    continue
+                turns.append({"question": pending_question, "answer": content})
+                pending_question = None
+        if pending_question:
+            turns.append({"question": pending_question, "answer": ""})
+        return turns
+
+
+class ContextCacheInvalidator:
+    def __init__(self, redis_url: Optional[str] = None) -> None:
+        self.redis_url = redis_url or os.getenv("REDIS_URL")
+        self._redis: Optional[Any] = None
+
+    def _get_redis(self):
+        if not REDIS_AVAILABLE or not self.redis_url or RedisSync is None:
+            return None
+        if self._redis is None:
+            self._redis = RedisSync.from_url(
+                self.redis_url,
+                decode_responses=True,
+                socket_connect_timeout=1,
+                socket_timeout=1,
+            )
+        return self._redis
+
+    @staticmethod
+    def _system_key(tenant_id: int) -> str:
+        return ContextCacheService._system_key(_safe_int(tenant_id, 1))
+
+    @staticmethod
+    def _project_list_key(tenant_id: int) -> str:
+        return ContextCacheService._project_list_key(_safe_int(tenant_id, 1))
+
+    @staticmethod
+    def _signal_pending_key(tenant_id: int) -> str:
+        return ContextCacheService._signal_pending_key(_safe_int(tenant_id, 1))
+
+    @staticmethod
+    def _user_key(tenant_id: int, user_id: int) -> str:
+        return ContextCacheService._user_key(_safe_int(tenant_id, 1), _safe_int(user_id, 0))
+
+    def invalidate(
+        self,
+        *,
+        tenant_id: int,
+        invalidate_system: bool,
+        invalidate_project_list: bool,
+        invalidate_signal_pending: bool,
+        user_ids: Optional[set[int]] = None,
+    ) -> None:
+        redis = self._get_redis()
+        if redis is None:
+            return
+
+        tid = _safe_int(tenant_id, 1)
+        keys: list[str] = []
+        if invalidate_system:
+            keys.append(self._system_key(tid))
+        if invalidate_project_list:
+            keys.append(self._project_list_key(tid))
+        if invalidate_signal_pending:
+            keys.append(self._signal_pending_key(tid))
+        if user_ids:
+            for uid in user_ids:
+                if _safe_int(uid, 0) > 0:
+                    keys.append(self._user_key(tid, uid))
+        if not keys:
+            return
+        try:
+            redis.delete(*keys)
+        except Exception:
+            return
+
+
+def register_context_cache_invalidation_hooks() -> None:
+    if getattr(register_context_cache_invalidation_hooks, "_registered", False):
+        return
+
+    invalidator = ContextCacheInvalidator()
+
+    @event.listens_for(Session, "after_flush")
+    def _collect_cache_invalidation(session: Session, _flush_context) -> None:  # pragma: no cover
+        payload = session.info.get("_pmflow_cache_invalidation")
+        if not isinstance(payload, dict):
+            payload = {
+                "tenant_ids": set(),
+                "invalidate_system": False,
+                "invalidate_project_list": False,
+                "invalidate_signal_pending": False,
+                "user_ids": set(),
+            }
+            session.info["_pmflow_cache_invalidation"] = payload
+
+        objects = set(session.new).union(session.dirty).union(session.deleted)
+        for obj in objects:
+            tenant_id = _safe_int(getattr(obj, "tenant_id", 0), 0)
+            if tenant_id > 0:
+                payload["tenant_ids"].add(tenant_id)
+
+            if isinstance(obj, Project):
+                payload["invalidate_system"] = True
+                payload["invalidate_project_list"] = True
+                continue
+
+            if isinstance(obj, ProjectSignalLink):
+                payload["invalidate_system"] = True
+                payload["invalidate_project_list"] = True
+                continue
+
+            if isinstance(obj, Signal):
+                payload["invalidate_system"] = True
+                payload["invalidate_signal_pending"] = True
+                continue
+
+            if isinstance(obj, Task):
+                payload["invalidate_system"] = True
+                state = inspect(obj)
+                assignee_hist = state.attrs.assignee_id.history if hasattr(state.attrs, "assignee_id") else None
+                if assignee_hist is not None:
+                    for old_value in assignee_hist.deleted:
+                        uid = _safe_int(old_value, 0)
+                        if uid > 0:
+                            payload["user_ids"].add(uid)
+                    for new_value in assignee_hist.added:
+                        uid = _safe_int(new_value, 0)
+                        if uid > 0:
+                            payload["user_ids"].add(uid)
+                uid = _safe_int(getattr(obj, "assignee_id", 0), 0)
+                if uid > 0:
+                    payload["user_ids"].add(uid)
+                continue
+
+            if isinstance(obj, UserProject):
+                payload["invalidate_system"] = True
+                payload["invalidate_project_list"] = True
+                uid = _safe_int(getattr(obj, "user_id", 0), 0)
+                if uid > 0:
+                    payload["user_ids"].add(uid)
+                continue
+
+            if isinstance(obj, User):
+                payload["invalidate_system"] = True
+                uid = _safe_int(getattr(obj, "id", 0), 0)
+                if uid > 0:
+                    payload["user_ids"].add(uid)
+                continue
+
+            if isinstance(obj, UserWorkload):
+                payload["invalidate_system"] = True
+                uid = _safe_int(getattr(obj, "user_id", 0), 0)
+                if uid > 0:
+                    payload["user_ids"].add(uid)
+                continue
+
+    @event.listens_for(Session, "after_commit")
+    def _flush_cache_invalidation(session: Session) -> None:  # pragma: no cover
+        payload = session.info.pop("_pmflow_cache_invalidation", None)
+        if not isinstance(payload, dict):
+            return
+        tenant_ids = payload.get("tenant_ids") or set()
+        if not tenant_ids:
+            tenant_ids = {_safe_int(os.getenv("DEFAULT_TENANT_ID", "1"), 1)}
+        for tenant_id in tenant_ids:
+            invalidator.invalidate(
+                tenant_id=_safe_int(tenant_id, 1),
+                invalidate_system=bool(payload.get("invalidate_system")),
+                invalidate_project_list=bool(payload.get("invalidate_project_list")),
+                invalidate_signal_pending=bool(payload.get("invalidate_signal_pending")),
+                user_ids=set(payload.get("user_ids") or set()),
+            )
+
+    @event.listens_for(Session, "after_rollback")
+    def _clear_cache_invalidation(session: Session) -> None:  # pragma: no cover
+        session.info.pop("_pmflow_cache_invalidation", None)
+
+    setattr(register_context_cache_invalidation_hooks, "_registered", True)
+
+
+context_cache_service = ContextCacheService()
diff --git a/backend/app/services/cost_optimization_service.py b/backend/app/services/cost_optimization_service.py
new file mode 100644
index 00000000..47ddc402
--- /dev/null
+++ b/backend/app/services/cost_optimization_service.py
@@ -0,0 +1,183 @@
+"""成本优化服务（P1）"""
+
+from typing import Any, Dict, Optional
+
+from sqlalchemy import and_, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import AICallLog, CostTracking, ModelRequirementBinding, ModelVersion, TruthNode
+
+
+class CostOptimizationService:
+    """面向宪章维度的成本优化建议。"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def analyze_cost_optimization(
+        self,
+        truth_node_id: int,
+        tenant_id: int,
+        period: str = "weekly",
+        project_id: Optional[int] = None,
+    ) -> Dict[str, Any]:
+        if project_id is not None:
+            await self.validate_truth_node_in_project(
+                truth_node_id=truth_node_id,
+                tenant_id=tenant_id,
+                project_id=project_id,
+            )
+        tracking = await self._get_tracking(truth_node_id, tenant_id, period)
+        if not tracking:
+            return {
+                "truth_node_id": truth_node_id,
+                "period": period,
+                "monthly_cost": 0.0,
+                "recommendations": [],
+                "total_potential_savings_percent": 0,
+            }
+
+        monthly_cost = float(tracking.actual_cost_usd or 0.0) * self._period_multiplier(period)
+        recs = []
+        recs.extend(await self._recommend_model_switch(truth_node_id, tenant_id))
+        recs.extend(await self._recommend_cache_and_batching(truth_node_id, tenant_id))
+
+        total_pct = sum(int(item.get("estimated_savings_percent", 0)) for item in recs)
+        total_pct = min(total_pct, 60)
+
+        result = {
+            "truth_node_id": truth_node_id,
+            "period": period,
+            "monthly_cost": round(monthly_cost, 2),
+            "recommendations": recs,
+            "total_potential_savings_percent": total_pct,
+            "estimated_monthly_savings_usd": round(monthly_cost * total_pct / 100, 2),
+        }
+        tracking.optimization_suggestions = result
+        await self.db.commit()
+        return result
+
+    async def validate_truth_node_in_project(
+        self, truth_node_id: int, tenant_id: int, project_id: int
+    ) -> TruthNode:
+        result = await self.db.execute(
+            select(TruthNode).where(
+                and_(
+                    TruthNode.id == truth_node_id,
+                    TruthNode.tenant_id == tenant_id,
+                )
+            )
+        )
+        truth_node = result.scalar_one_or_none()
+        if not truth_node:
+            raise ValueError(f"宪章 {truth_node_id} 不存在")
+        if truth_node.project_id != project_id:
+            raise ValueError(
+                f"宪章 {truth_node_id} 不属于项目 {project_id}，拒绝跨项目分析"
+            )
+        return truth_node
+
+    async def _get_tracking(
+        self, truth_node_id: int, tenant_id: int, period: str
+    ) -> Optional[CostTracking]:
+        result = await self.db.execute(
+            select(CostTracking).where(
+                and_(
+                    CostTracking.truth_node_id == truth_node_id,
+                    CostTracking.tenant_id == tenant_id,
+                    CostTracking.period == period,
+                )
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def _recommend_model_switch(
+        self, truth_node_id: int, tenant_id: int
+    ) -> list[Dict[str, Any]]:
+        result = await self.db.execute(
+            select(ModelRequirementBinding).where(
+                and_(
+                    ModelRequirementBinding.truth_node_id == truth_node_id,
+                    ModelRequirementBinding.tenant_id == tenant_id,
+                )
+            )
+        )
+        bindings = result.scalars().all()
+        if not bindings:
+            return []
+
+        current = bindings[0]
+        model = await self._get_model(current.model_version_id)
+        if not model or not model.cost_per_1k_tokens:
+            return []
+
+        cheaper = await self._find_cheaper_compatible_model(float(model.cost_per_1k_tokens))
+        if not cheaper:
+            return []
+
+        return [
+            {
+                "type": "model_switch",
+                "priority": "high",
+                "description": f"可从 {model.name} 切换到 {cheaper.name}（更低 token 成本）",
+                "estimated_savings_percent": 25,
+            }
+        ]
+
+    async def _recommend_cache_and_batching(
+        self, truth_node_id: int, tenant_id: int
+    ) -> list[Dict[str, Any]]:
+        # 当前日志没有强关联 truth_node 字段，按租户统计调用频次近似分析。
+        call_count_res = await self.db.execute(
+            select(func.count(AICallLog.id)).where(AICallLog.tenant_id == str(tenant_id))
+        )
+        call_count = int(call_count_res.scalar() or 0)
+
+        recs: list[Dict[str, Any]] = [
+            {
+                "type": "cache",
+                "priority": "high",
+                "description": "启用/提升相似请求缓存命中率（去重 prompt + TTL 分层）",
+                "estimated_savings_percent": 15,
+            }
+        ]
+        if call_count > 200:
+            recs.append(
+                {
+                    "type": "batching",
+                    "priority": "medium",
+                    "description": "批量聚合低优先级推理请求，减少并发和重复上下文开销",
+                    "estimated_savings_percent": 10,
+                }
+            )
+        return recs
+
+    async def _get_model(self, model_version_id: int) -> Optional[ModelVersion]:
+        result = await self.db.execute(
+            select(ModelVersion).where(ModelVersion.id == model_version_id)
+        )
+        return result.scalar_one_or_none()
+
+    async def _find_cheaper_compatible_model(
+        self, current_cost_per_1k_tokens: float
+    ) -> Optional[ModelVersion]:
+        result = await self.db.execute(
+            select(ModelVersion)
+            .where(ModelVersion.is_active == True)  # noqa: E712
+            .order_by(ModelVersion.cost_per_1k_tokens.asc())
+        )
+        candidates = result.scalars().all()
+        for m in candidates:
+            cost = float(m.cost_per_1k_tokens or 0.0)
+            if cost > 0 and cost < current_cost_per_1k_tokens:
+                return m
+        return None
+
+    def _period_multiplier(self, period: str) -> int:
+        if period == "daily":
+            return 30
+        if period == "weekly":
+            return 4
+        if period == "monthly":
+            return 1
+        return 1
diff --git a/backend/app/services/daily_briefing.py b/backend/app/services/daily_briefing.py
new file mode 100644
index 00000000..aa80e2d3
--- /dev/null
+++ b/backend/app/services/daily_briefing.py
@@ -0,0 +1,480 @@
+"""
+Daily Progress Briefing Service - 每日进度简报服务
+
+核心功能：
+1. 为每个角色生成个性化简报
+2. 智能过滤无关信息
+3. 根据角色优先级排序内容
+4. 替代传统日报/周报
+
+个性化规则：
+- AI工程师：关注模型更新、技术信号、实验结果
+- 后端开发：关注API变更、任务阻塞、依赖更新
+- 前端开发：关注UI需求、设计变更、接口更新
+- BOSS/ADMIN：关注整体进度、风险、决策点
+"""
+import asyncio
+import json
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from sqlalchemy import text, func, and_
+
+from ..models import (
+    User, Project, Task, Signal, CognitiveConflict,
+    TruthNode, UserRole, ProjectSignalLink
+)
+from .llm import LLMService
+
+
+def _parse_involved_user_ids(raw_value: Any) -> List[int]:
+    """兼容 involved_user_ids 的 list / json-string / 单值字符串。"""
+    if raw_value is None:
+        return []
+    data = raw_value
+    if isinstance(data, (bytes, bytearray)):
+        data = data.decode("utf-8", errors="ignore")
+    if isinstance(data, str):
+        text = data.strip()
+        if not text:
+            return []
+        try:
+            data = json.loads(text)
+        except (TypeError, ValueError, json.JSONDecodeError):
+            data = [v.strip() for v in text.split(",") if v.strip()]
+    if not isinstance(data, list):
+        data = [data]
+
+    parsed: List[int] = []
+    for item in data:
+        if item is None or isinstance(item, bool):
+            continue
+        try:
+            parsed.append(int(item))
+        except (TypeError, ValueError):
+            continue
+    return parsed
+
+
+class DailyBriefingService:
+    """每日进度简报服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+        self.llm = LLMService()
+
+    async def generate_personalized_briefing(
+        self,
+        user_id: int,
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        生成个性化每日简报
+
+        简报内容根据用户角色自动过滤：
+        - 只显示相关项目的内容
+        - 按优先级排序
+        - 突出风险和需要行动的事项
+        """
+        # 获取用户信息
+        user_result = await self.db.execute(
+            select(User).where(User.id == user_id)
+        )
+        user = user_result.scalar_one_or_none()
+        if not user:
+            raise ValueError(f"User {user_id} not found")
+
+        # 计算时间范围
+        now = datetime.now(timezone.utc)
+        yesterday = now - timedelta(days=1)
+
+        # 根据角色收集相关信息
+        relevant_data = await self._collect_relevant_data(
+            user=user,
+            since=yesterday,
+            tenant_id=tenant_id
+        )
+
+        # AI 生成个性化摘要
+        summary = await self._generate_ai_summary(
+            user=user,
+            data=relevant_data,
+            tenant_id=tenant_id,
+        )
+
+        return {
+            'user_id': user_id,
+            'user_role': self._role_code(user.role),
+            'date': now.strftime('%Y-%m-%d'),
+            'summary': summary,
+            'highlights': relevant_data.get('highlights', []),
+            'your_tasks': relevant_data.get('your_tasks', []),
+            'risks': relevant_data.get('risks', []),
+            'signals': relevant_data.get('signals', []),
+            'conflicts': relevant_data.get('conflicts', []),
+            'action_items': relevant_data.get('action_items', []),
+            'team_updates': relevant_data.get('team_updates', [])
+        }
+
+    async def _collect_relevant_data(
+        self,
+        user: User,
+        since: datetime,
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """收集与用户相关的数据"""
+
+        data = {
+            'highlights': [],
+            'your_tasks': [],
+            'risks': [],
+            'signals': [],
+            'conflicts': [],
+            'action_items': [],
+            'team_updates': []
+        }
+
+        # 获取用户相关的项目
+        user_projects = await self._get_user_projects(user.id, tenant_id)
+
+        if not user_projects:
+            return data
+
+        project_ids = [p.id for p in user_projects]
+        role_code = self._role_code(user.role)
+
+        # 1. 获取用户的任务更新
+        task_result = await self.db.execute(
+            select(Task, Project).join(
+                Project,
+                and_(Project.id == Task.project_id, Project.tenant_id == tenant_id),
+            ).where(
+                Task.tenant_id == tenant_id,
+                Task.project_id.in_(project_ids),
+                Task.updated_at >= since
+            )
+        )
+        task_rows = task_result.all()
+
+        for task, project in task_rows:
+            # 我的任务
+            if task.assignee_id == user.id:
+                data['your_tasks'].append({
+                    'id': task.id,
+                    'title': task.title,
+                    'status': task.status,
+                    'priority': task.priority,
+                    'project': project.title,
+                    'updated_at': task.updated_at.isoformat() if hasattr(task, 'updated_at') else None
+                })
+
+                # 高优先级或阻塞的任务加入 highlights
+                if task.priority == 'HIGH' or task.status == 'blocked':
+                    data['highlights'].append({
+                        'type': 'task',
+                        'priority': 'high',
+                        'content': f'任务"{task.title}"状态: {task.status}',
+                        'project': project.title
+                    })
+
+        # 2. 获取风险（阻塞的任务）
+        for task, project in task_rows:
+            if task.status == 'blocked':
+                data['risks'].append({
+                    'type': 'task_blocked',
+                    'task': task.title,
+                    'project': project.title,
+                    'severity': 'high'
+                })
+
+        # 3. 获取新信号（根据角色过滤）
+        signal_result = await self.db.execute(
+            select(Signal).join(
+                ProjectSignalLink,
+                ProjectSignalLink.signal_id == Signal.id
+            ).where(
+                Signal.tenant_id == tenant_id,
+                ProjectSignalLink.project_id.in_(project_ids),
+                Signal.created_at >= since,
+                Signal.impact_score >= 0.7  # 只显示高影响信号
+            ).order_by(
+                Signal.impact_score.desc()
+            ).limit(10)
+        )
+        signals = signal_result.scalars().all()
+
+        for signal in signals:
+            # 根据角色过滤信号内容
+            if self._is_signal_relevant_to_role(signal, role_code):
+                data['signals'].append({
+                    'id': signal.id,
+                    'source': signal.source,
+                    'content': signal.content[:200],
+                    'impact_score': signal.impact_score,
+                    'priority': signal.priority
+                })
+
+                # 高影响信号加入 highlights
+                if signal.impact_score >= 0.9:
+                    data['highlights'].append({
+                        'type': 'signal',
+                        'priority': 'high',
+                        'content': f'高影响信号: {signal.content[:100]}',
+                        'source': signal.source
+                    })
+
+        # 4. 获取认知冲突
+        conflict_result = await self.db.execute(
+            select(CognitiveConflict).join(
+                TruthNode,
+                TruthNode.id == CognitiveConflict.truth_node_id
+            ).where(
+                CognitiveConflict.tenant_id == tenant_id,
+                TruthNode.project_id.in_(project_ids),
+                CognitiveConflict.status == 'active',
+                CognitiveConflict.created_at >= since
+            )
+        )
+        conflicts = conflict_result.scalars().all()
+
+        for conflict in conflicts:
+            involved_users = _parse_involved_user_ids(conflict.involved_user_ids)
+            # 如果用户参与了冲突，或是管理员
+            if user.id in involved_users or role_code in [UserRole.ADMIN.value, UserRole.BOSS.value]:
+                data['conflicts'].append({
+                    'id': conflict.id,
+                    'type': conflict.conflict_type,
+                    'description': conflict.description[:200],
+                    'severity': 'high' if 'critical' in conflict.description.lower() else 'medium'
+                })
+
+                # 冲突加入 highlights
+                data['highlights'].append({
+                    'type': 'conflict',
+                    'priority': 'high',
+                    'content': f'认知冲突: {conflict.conflict_type}',
+                    'project_id': conflict.truth_node_id
+                })
+
+        # 5. 生成行动项
+        if data['your_tasks']:
+            for task in data['your_tasks'][:3]:
+                if task['status'] in ['pending', 'in_progress']:
+                    data['action_items'].append({
+                        'type': 'task',
+                        'content': f'继续推进任务: {task["title"]}',
+                        'priority': task['priority']
+                    })
+
+        if data['conflicts']:
+            data['action_items'].append({
+                'type': 'conflict',
+                'content': f'解决 {len(data["conflicts"])} 个认知冲突',
+                'priority': 'high'
+            })
+
+        # 6. 团队更新（其他用户的相关任务）
+        team_task_result = await self.db.execute(
+            select(Task, Project).join(
+                Project,
+                and_(Project.id == Task.project_id, Project.tenant_id == tenant_id),
+            ).where(
+                Task.tenant_id == tenant_id,
+                Task.project_id.in_(project_ids),
+                Task.assignee_id != user.id,
+                Task.updated_at >= since
+            )
+        )
+        team_tasks = team_task_result.all()
+
+        for task, project in team_tasks[:5]:
+            data['team_updates'].append({
+                'assignee_id': task.assignee_id,
+                'task': task.title,
+                'status': task.status,
+                'project': project.title
+            })
+
+        return data
+
+    async def _get_user_projects(
+        self,
+        user_id: int,
+        tenant_id: int
+    ) -> List[Project]:
+        """获取用户相关的项目"""
+        # 获取用户参与的所有项目
+        result = await self.db.execute(
+            select(Project).where(
+                Project.tenant_id == tenant_id,
+                Project.status == 'active'
+            )
+        )
+        return result.scalars().all()
+
+    def _is_signal_relevant_to_role(
+        self,
+        signal: Signal,
+        user_role: Any
+    ) -> bool:
+        """判断信号是否与用户角色相关"""
+        content = signal.content.lower()
+        role_code = self._role_code(user_role)
+        _ = signal.source.lower()
+
+        # AI工程师关注的内容
+        if role_code == UserRole.AI.value:
+            ai_keywords = ['model', 'llm', 'gpt', 'embedding', 'vector', '训练', '推理', 'ollama', 'langchain']
+            if any(kw in content for kw in ai_keywords):
+                return True
+
+        # 后端开发关注的内容
+        elif role_code in {UserRole.BE.value, "BACKEND"}:
+            backend_keywords = ['api', 'backend', 'server', 'database', 'sql', '后端', '接口', '服务']
+            if any(kw in content for kw in backend_keywords):
+                return True
+
+        # 前端开发关注的内容
+        elif role_code in {UserRole.FE.value, "FRONTEND"}:
+            frontend_keywords = ['ui', 'frontend', 'react', 'vue', 'component', '前端', '界面', '交互']
+            if any(kw in content for kw in frontend_keywords):
+                return True
+
+        # 管理员看到所有高影响信号
+        elif role_code in {UserRole.ADMIN.value, UserRole.BOSS.value}:
+            return signal.impact_score >= 0.8
+
+        # 默认返回中等影响以上的信号
+        return signal.impact_score >= 0.7
+
+    async def _generate_ai_summary(
+        self,
+        user: User,
+        data: Dict[str, Any],
+        tenant_id: int,
+    ) -> str:
+        """AI 生成个性化摘要"""
+        role_code = self._role_code(user.role)
+        prompt = f"""为以下用户生成每日工作摘要：
+
+用户角色：{role_code}
+用户名：{user.username}
+
+今日概况：
+- 你的任务数：{len(data.get('your_tasks', []))}
+- 新信号数：{len(data.get('signals', []))}
+- 活跃冲突数：{len(data.get('conflicts', []))}
+- 风险数：{len(data.get('risks', []))}
+
+重点事项：
+{json.dumps(data.get('highlights', []), ensure_ascii=False, indent=2)[:500]}
+
+请生成一段简洁的每日摘要（100字以内），突出最重要的信息和行动项。"""
+
+        try:
+            summary = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            return summary.strip()
+        except:
+            # 降级：生成简单摘要
+            parts = []
+            if data.get('your_tasks'):
+                parts.append(f"你有 {len(data['your_tasks'])} 个任务需要关注")
+            if data.get('risks'):
+                parts.append(f"注意 {len(data['risks'])} 个风险点")
+            if data.get('conflicts'):
+                parts.append(f"需要解决 {len(data['conflicts'])} 个认知冲突")
+            return "；".join(parts) if parts else "今日无重要更新"
+
+    @staticmethod
+    def _role_code(raw_role: Any) -> str:
+        if raw_role is None:
+            return ""
+        value = getattr(raw_role, "value", raw_role)
+        return str(value or "").strip().upper()
+
+    async def generate_team_briefing(
+        self,
+        project_id: int,
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        生成团队简报（给管理者）
+
+        包含：
+        - 整体进度
+        - 每个成员的工作量
+        - 风险和阻塞
+        - 需要决策的事项
+        """
+        # 获取项目信息
+        project_result = await self.db.execute(
+            select(Project).where(
+                Project.id == project_id,
+                Project.tenant_id == tenant_id
+            )
+        )
+        project = project_result.scalar_one_or_none()
+        if not project:
+            raise ValueError(f"Project {project_id} not found")
+
+        # 获取项目所有任务
+        task_result = await self.db.execute(
+            select(Task).where(
+                Task.project_id == project_id,
+                Task.tenant_id == tenant_id
+            )
+        )
+        tasks = task_result.scalars().all()
+
+        # 统计任务状态
+        status_counts = {}
+        for task in tasks:
+            status_counts[task.status] = status_counts.get(task.status, 0) + 1
+
+        # 获取阻塞的任务
+        blocked_tasks = [t for t in tasks if t.status == 'blocked']
+
+        # 获取项目相关信号
+        signal_result = await self.db.execute(
+            select(Signal).join(
+                ProjectSignalLink,
+                ProjectSignalLink.signal_id == Signal.id
+            ).where(
+                Signal.tenant_id == tenant_id,
+                ProjectSignalLink.project_id == project_id
+            ).order_by(
+                Signal.created_at.desc()
+            ).limit(10)
+        )
+        signals = signal_result.scalars().all()
+
+        return {
+            'project_id': project_id,
+            'project_name': project.title,
+            'generated_at': datetime.now(timezone.utc).isoformat(),
+            'task_summary': {
+                'total': len(tasks),
+                'by_status': status_counts,
+                'completion_rate': status_counts.get('done', 0) / len(tasks) if tasks else 0
+            },
+            'blocked_tasks': [
+                {'id': t.id, 'title': t.title, 'assignee_id': t.assignee_id}
+                for t in blocked_tasks
+            ],
+            'recent_signals': [
+                {'id': s.id, 'content': s.content[:100], 'impact_score': s.impact_score}
+                for s in signals[:5]
+            ],
+            'needs_attention': len(blocked_tasks) > 0
+        }
+
+
+# 全局服务实例
+def get_daily_briefing_service(db: AsyncSession) -> DailyBriefingService:
+    """获取每日简报服务实例"""
+    return DailyBriefingService(db)
diff --git a/backend/app/services/domain.py b/backend/app/services/domain.py
new file mode 100644
index 00000000..02f324de
--- /dev/null
+++ b/backend/app/services/domain.py
@@ -0,0 +1,448 @@
+"""
+Neural Sync 2.0 - Project Domain Service
+
+项目域服务 - 管理动态项目域
+每个域有独立的信号源、知识库和配置
+"""
+
+import asyncio
+import logging
+from typing import Optional, Dict, List, Any
+from pydantic import BaseModel, Field
+from datetime import datetime
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+logger = logging.getLogger(__name__)
+
+
+class DomainConfig(BaseModel):
+    """项目域配置"""
+    signal_sources: List[str] = Field(default_factory=lambda: ["hackernews"])
+    ai_config: Dict[str, Any] = Field(default_factory=dict)
+    audit_threshold: float = 98.0
+    auto_merge_enabled: bool = False
+    webhook_urls: Dict[str, str] = Field(default_factory=dict)
+
+
+class ProjectDomainCreate(BaseModel):
+    """创建项目域请求"""
+    name: str
+    slug: str
+    description: Optional[str] = None
+    config: Optional[DomainConfig] = None
+    is_default: bool = False
+
+
+class ProjectDomainUpdate(BaseModel):
+    """更新项目域请求"""
+    name: Optional[str] = None
+    description: Optional[str] = None
+    config: Optional[DomainConfig] = None
+    is_active: Optional[bool] = None
+    is_default: Optional[bool] = None
+
+
+class ProjectDomainResponse(BaseModel):
+    """项目域响应"""
+    id: str
+    name: str
+    slug: str
+    description: Optional[str]
+    config: DomainConfig
+    is_active: bool
+    is_default: bool
+    signal_count: int
+    context_node_count: int
+    experiment_count: int
+    created_at: datetime
+    updated_at: datetime
+    last_synced_at: Optional[datetime]
+
+
+class DomainSyncProgress(BaseModel):
+    """域同步进度"""
+    domain_id: str
+    phase: str  # connecting, loading_knowledge, syncing_signals, complete
+    progress: float  # 0.0 - 1.0
+    message: str
+
+
+class ProjectDomainService:
+    """
+    项目域服务
+
+    功能：
+    - CRUD 操作
+    - 域切换和知识同步
+    - 统计信息更新
+    - 默认域管理
+    """
+
+    def __init__(self, db_session: AsyncSession, event_bus=None):
+        self.db = db_session
+        self.event_bus = event_bus
+        self._active_domain_id: Optional[str] = None
+
+    async def create_domain(self, request: ProjectDomainCreate) -> ProjectDomainResponse:
+        """创建新项目域"""
+        from ..models import ProjectDomain
+        import uuid
+
+        # 检查 slug 唯一性
+        existing = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.slug == request.slug)
+        )
+        if existing.scalar_one_or_none():
+            raise ValueError(f"Slug '{request.slug}' already exists")
+
+        # 如果设置为默认域，取消其他域的默认状态
+        if request.is_default:
+            await self._clear_default_domains()
+
+        # 序列化配置
+        config_dict = request.config.model_dump() if request.config else {}
+
+        domain = ProjectDomain(
+            name=request.name,
+            slug=request.slug,
+            description=request.description,
+            config=config_dict,
+            is_default=request.is_default,
+        )
+
+        self.db.add(domain)
+        await self.db.commit()
+        await self.db.refresh(domain)
+
+        # 发布域创建事件
+        if self.event_bus:
+            from ..events import EventType, Event
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.DOMAIN_CREATED,
+                    data={
+                        "domain_id": str(domain.id),
+                        "name": domain.name,
+                        "slug": domain.slug,
+                    },
+                )
+            )
+
+        return self._to_response(domain)
+
+    async def get_domain(self, domain_id: str) -> Optional[ProjectDomainResponse]:
+        """获取单个域"""
+        from ..models import ProjectDomain
+        import uuid
+
+        try:
+            domain_uuid = uuid.UUID(domain_id)
+        except ValueError:
+            return None
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.id == domain_uuid)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            return None
+
+        return self._to_response(domain)
+
+    async def get_domain_by_slug(self, slug: str) -> Optional[ProjectDomainResponse]:
+        """通过 slug 获取域"""
+        from ..models import ProjectDomain
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.slug == slug)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            return None
+
+        return self._to_response(domain)
+
+    async def list_domains(
+        self, include_inactive: bool = False
+    ) -> List[ProjectDomainResponse]:
+        """列出所有域"""
+        from ..models import ProjectDomain
+
+        query = select(ProjectDomain)
+        if not include_inactive:
+            query = query.where(ProjectDomain.is_active == True)
+
+        query = query.order_by(ProjectDomain.is_default.desc(), ProjectDomain.created_at.asc())
+
+        result = await self.db.execute(query)
+        domains = result.scalars().all()
+
+        return [self._to_response(d) for d in domains]
+
+    async def update_domain(
+        self, domain_id: str, request: ProjectDomainUpdate
+    ) -> Optional[ProjectDomainResponse]:
+        """更新域"""
+        from ..models import ProjectDomain
+        import uuid
+
+        try:
+            domain_uuid = uuid.UUID(domain_id)
+        except ValueError:
+            return None
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.id == domain_uuid)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            return None
+
+        # 更新字段
+        if request.name is not None:
+            domain.name = request.name
+        if request.description is not None:
+            domain.description = request.description
+        if request.config is not None:
+            domain.config = request.config.model_dump()
+        if request.is_active is not None:
+            domain.is_active = request.is_active
+        if request.is_default is not None and request.is_default != domain.is_default:
+            if request.is_default:
+                await self._clear_default_domains()
+            domain.is_default = request.is_default
+
+        domain.updated_at = datetime.utcnow()
+
+        await self.db.commit()
+        await self.db.refresh(domain)
+
+        # 发布域更新事件
+        if self.event_bus:
+            from ..events import EventType, Event
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.DOMAIN_UPDATED,
+                    data={
+                        "domain_id": str(domain.id),
+                        "name": domain.name,
+                        "slug": domain.slug,
+                    },
+                )
+            )
+
+        return self._to_response(domain)
+
+    async def delete_domain(self, domain_id: str) -> bool:
+        """删除域"""
+        from ..models import ProjectDomain
+        import uuid
+
+        try:
+            domain_uuid = uuid.UUID(domain_id)
+        except ValueError:
+            return False
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.id == domain_uuid)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            return False
+
+        # 不允许删除默认域
+        if domain.is_default:
+            raise ValueError("Cannot delete default domain")
+
+        await self.db.delete(domain)
+        await self.db.commit()
+
+        # 发布域删除事件
+        if self.event_bus:
+            from ..events import EventType, Event
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.DOMAIN_DELETED,
+                    data={
+                        "domain_id": str(domain.id),
+                        "name": domain.name,
+                        "slug": domain.slug,
+                    },
+                )
+            )
+
+        return True
+
+    async def switch_domain(self, domain_id: str) -> DomainSyncProgress:
+        """
+        切换活动域
+
+        流程：
+        1. 验证域存在且激活
+        2. 开始知识同步（1.5s 动画）
+        3. 更新活动域 ID
+        4. 返回同步进度
+        """
+        from ..models import ProjectDomain
+        import uuid
+
+        try:
+            domain_uuid = uuid.UUID(domain_id)
+        except ValueError:
+            raise ValueError("Invalid domain ID")
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.id == domain_uuid)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            raise ValueError("Domain not found")
+
+        if not domain.is_active:
+            raise ValueError("Domain is not active")
+
+        # 更新最后同步时间
+        domain.last_synced_at = datetime.utcnow()
+        await self.db.commit()
+
+        # 设置活动域
+        self._active_domain_id = str(domain.id)
+
+        # 发布域切换事件
+        if self.event_bus:
+            from ..events import EventType, Event
+            await self.event_bus.publish(
+                Event(
+                    type=EventType.DOMAIN_SWITCHED,
+                    data={
+                        "domain_id": str(domain.id),
+                        "name": domain.name,
+                        "slug": domain.slug,
+                    },
+                )
+            )
+
+        # 模拟同步进度（1.5s）
+        return DomainSyncProgress(
+            domain_id=str(domain.id),
+            phase="complete",
+            progress=1.0,
+            message=f"已切换到 {domain.name}",
+        )
+
+    async def get_default_domain(self) -> Optional[ProjectDomainResponse]:
+        """获取默认域"""
+        from ..models import ProjectDomain
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.is_default == True)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            # 如果没有默认域，返回第一个激活域
+            result = await self.db.execute(
+                select(ProjectDomain)
+                .where(ProjectDomain.is_active == True)
+                .order_by(ProjectDomain.created_at.asc())
+                .limit(1)
+            )
+            domain = result.scalar_one_or_none()
+
+        if not domain:
+            return None
+
+        return self._to_response(domain)
+
+    async def update_domain_stats(self, domain_id: str) -> None:
+        """更新域统计信息"""
+        from ..models import ProjectDomain, Signal, ContextNode, Experiment
+        import uuid
+
+        try:
+            domain_uuid = uuid.UUID(domain_id)
+        except ValueError:
+            return
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.id == domain_uuid)
+        )
+        domain = result.scalar_one_or_none()
+
+        if not domain:
+            return
+
+        # TODO: 需要添加域关联到 Signal, ContextNode, Experiment
+        # 当前使用全局计数
+        signal_count = await self.db.execute(select(Signal))
+        domain.signal_count = len(signal_count.scalars().all())
+
+        context_count = await self.db.execute(select(ContextNode))
+        domain.context_node_count = len(context_count.scalars().all())
+
+        experiment_count = await self.db.execute(select(Experiment))
+        domain.experiment_count = len(experiment_count.scalars().all())
+
+        await self.db.commit()
+
+    async def get_active_domain_id(self) -> Optional[str]:
+        """获取当前活动域 ID"""
+        if self._active_domain_id:
+            return self._active_domain_id
+
+        # 如果没有设置，返回默认域
+        default = await self.get_default_domain()
+        if default:
+            self._active_domain_id = default.id
+            return default.id
+
+        return None
+
+    async def _clear_default_domains(self) -> None:
+        """清除所有域的默认状态"""
+        from ..models import ProjectDomain
+
+        result = await self.db.execute(
+            select(ProjectDomain).where(ProjectDomain.is_default == True)
+        )
+        domains = result.scalars().all()
+
+        for domain in domains:
+            domain.is_default = False
+
+        await self.db.commit()
+
+    def _to_response(self, domain) -> ProjectDomainResponse:
+        """转换为响应模型"""
+        config = DomainConfig(**domain.config) if domain.config else DomainConfig()
+
+        return ProjectDomainResponse(
+            id=str(domain.id),
+            name=domain.name,
+            slug=domain.slug,
+            description=domain.description,
+            config=config,
+            is_active=domain.is_active,
+            is_default=domain.is_default,
+            signal_count=domain.signal_count,
+            context_node_count=domain.context_node_count,
+            experiment_count=domain.experiment_count,
+            created_at=domain.created_at,
+            updated_at=domain.updated_at,
+            last_synced_at=getattr(domain, 'last_synced_at', None),
+        )
+
+
+# 全局服务实例
+_domain_service = None
+
+
+def get_domain_service(db_session: AsyncSession, event_bus=None) -> ProjectDomainService:
+    """获取项目域服务实例"""
+    return ProjectDomainService(db_session, event_bus)
diff --git a/backend/app/services/event_workflow.py b/backend/app/services/event_workflow.py
new file mode 100644
index 00000000..01596669
--- /dev/null
+++ b/backend/app/services/event_workflow.py
@@ -0,0 +1,614 @@
+"""
+Event-Driven Workflow Engine - 事件驱动工作流引擎
+
+核心功能：
+1. 替代固定 Sprint 迭代，实现实时事件驱动迭代
+2. 市场变化可在10分钟内反映到项目优先级
+3. 自动触发任务重排、依赖更新
+4. 推送实时通知给相关角色
+
+工作流事件：
+- 高影响信号到达 → 自动重新评估项目优先级
+- 阻塞检测 → 触发替代方案生成
+- 认知冲突 → 触发对齐流程
+- 技术栈变更 → 自动更新任务依赖
+- 竞品动态 → 触发战略调整
+"""
+import asyncio
+import json
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional, Callable
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from sqlalchemy import text, update
+
+from ..models import (
+    User, Project, Task, Signal, CognitiveConflict,
+    TruthNode, ProjectSignalLink, UserRole
+)
+from .llm import LLMService
+from ..events import EventType, Event, get_event_bus
+
+
+class WorkflowEventType:
+    """工作流事件类型"""
+    HIGH_IMPACT_SIGNAL = "high_impact_signal"
+    BLOCKER_DETECTED = "blocker_detected"
+    COGNITIVE_CONFLICT = "cognitive_conflict"
+    TECH_STACK_CHANGE = "tech_stack_change"
+    COMPETITOR_UPDATE = "competitor_update"
+    TASK_COMPLETED = "task_completed"
+    TASK_BLOCKED = "task_blocked"
+    PROJECT_RISK = "project_risk"
+
+
+class EventDrivenWorkflowEngine:
+    """事件驱动工作流引擎"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+        self.llm = LLMService()
+        self.event_bus = get_event_bus()
+
+        # 注册工作流处理器
+        self._handlers = {
+            WorkflowEventType.HIGH_IMPACT_SIGNAL: self._handle_high_impact_signal,
+            WorkflowEventType.BLOCKER_DETECTED: self._handle_blocker_detected,
+            WorkflowEventType.COGNITIVE_CONFLICT: self._handle_cognitive_conflict,
+            WorkflowEventType.TECH_STACK_CHANGE: self._handle_tech_stack_change,
+            WorkflowEventType.COMPETITOR_UPDATE: self._handle_competitor_update,
+            WorkflowEventType.TASK_COMPLETED: self._handle_task_completed,
+            WorkflowEventType.TASK_BLOCKED: self._handle_task_blocked,
+            WorkflowEventType.PROJECT_RISK: self._handle_project_risk,
+        }
+
+    async def trigger_workflow(
+        self,
+        event_type: str,
+        event_data: Dict[str, Any],
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        触发工作流
+
+        Args:
+            event_type: 事件类型
+            event_data: 事件数据
+            tenant_id: 租户ID
+
+        Returns:
+            处理结果
+        """
+        handler = self._handlers.get(event_type)
+        if not handler:
+            return {'status': 'ignored', 'reason': 'Unknown event type'}
+
+        try:
+            result = await handler(event_data, tenant_id)
+
+            # 发布到事件总线
+            await self.event_bus.publish(
+                EventType.WORKFLOW_TRIGGERED,
+                {
+                    'workflow_type': event_type,
+                    'result': result,
+                    'timestamp': datetime.now(timezone.utc).isoformat()
+                }
+            )
+
+            return {
+                'status': 'processed',
+                'event_type': event_type,
+                'result': result
+            }
+
+        except Exception as e:
+            return {
+                'status': 'error',
+                'event_type': event_type,
+                'error': str(e)
+            }
+
+    async def _handle_high_impact_signal(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理高影响信号
+
+        动作：
+        1. 评估对关联项目的影响
+        2. 调整项目优先级
+        3. 创建响应任务
+        4. 通知相关角色
+        """
+        signal_id = event_data.get('signal_id')
+        impact_threshold = 0.85
+
+        # 获取信号
+        signal_result = await self.db.execute(
+            select(Signal).where(
+                Signal.id == signal_id,
+                Signal.tenant_id == tenant_id
+            )
+        )
+        signal = signal_result.scalar_one_or_none()
+        if not signal:
+            raise ValueError(f"Signal {signal_id} not found")
+
+        # 获取关联的项目
+        link_result = await self.db.execute(
+            select(ProjectSignalLink).where(
+                ProjectSignalLink.signal_id == signal_id,
+                ProjectSignalLink.tenant_id == tenant_id
+            )
+        )
+        links = link_result.scalars().all()
+
+        affected_projects = []
+        created_tasks = []
+
+        for link in links:
+            project_result = await self.db.execute(
+                select(Project).where(Project.id == link.project_id)
+            )
+            project = project_result.scalar_one_or_none()
+            if not project:
+                continue
+
+            # AI 评估影响
+            impact = await self._assess_signal_impact_on_project(signal, project)
+
+            if impact.get('should_adjust_priority', False):
+                # 调整项目优先级
+                old_priority = project.priority if hasattr(project, 'priority') else 'medium'
+                new_priority = impact.get('suggested_priority', 'high')
+
+                if hasattr(project, 'priority'):
+                    project.priority = new_priority
+                    await self.db.commit()
+
+                affected_projects.append({
+                    'project_id': project.id,
+                    'project_name': project.title,
+                    'old_priority': old_priority,
+                    'new_priority': new_priority,
+                    'reason': impact.get('reason', '')
+                })
+
+                # 创建响应任务
+                if impact.get('needs_response_task', False):
+                    task = await self._create_response_task(
+                        project=project,
+                        signal=signal,
+                        impact=impact,
+                        tenant_id=tenant_id
+                    )
+                    created_tasks.append({
+                        'task_id': task.id,
+                        'title': task.title,
+                        'priority': task.priority
+                    })
+
+        return {
+            'affected_projects': affected_projects,
+            'created_tasks': created_tasks,
+            'message': f'处理了 {len(affected_projects)} 个受影响的项目'
+        }
+
+    async def _handle_blocker_detected(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理阻塞检测
+
+        动作：
+        1. AI 生成替代方案
+        2. 创建绕过任务
+        3. 通知相关人员
+        """
+        blocker_id = event_data.get('blocker_id')
+        task_id = event_data.get('task_id')
+
+        # 获取阻塞信息
+        conflict_result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.id == blocker_id,
+                CognitiveConflict.tenant_id == tenant_id
+            )
+        )
+        blocker = conflict_result.scalar_one_or_none()
+
+        if not blocker:
+            raise ValueError(f"Blocker {blocker_id} not found")
+
+        # AI 生成解决方案
+        prompt = f"""为以下阻塞生成绕过方案：
+
+阻塞描述：{blocker.description}
+
+请提供：
+1. 问题根本原因分析
+2. 3个可行的绕过方案
+3. 每个方案的预估工作量
+4. 推荐的最优方案
+
+以 JSON 格式返回。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            # 解析 AI 响应（简化处理）
+            return {
+                'blocker_id': blocker_id,
+                'alternatives_generated': True,
+                'ai_response': response
+            }
+        except:
+            return {
+                'blocker_id': blocker_id,
+                'alternatives_generated': False,
+                'message': 'AI 方案生成失败'
+            }
+
+    async def _handle_cognitive_conflict(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理认知冲突
+
+        动作：
+        1. 识别涉及的用户
+        2. 发送对齐邀请
+        3. 调度异步对齐会话
+        """
+        conflict_id = event_data.get('conflict_id')
+
+        conflict_result = await self.db.execute(
+            select(CognitiveConflict).where(
+                CognitiveConflict.id == conflict_id,
+                CognitiveConflict.tenant_id == tenant_id
+            )
+        )
+        conflict = conflict_result.scalar_one_or_none()
+
+        if not conflict:
+            raise ValueError(f"Conflict {conflict_id} not found")
+
+        # 获取涉及的用户
+        involved_user_ids = json.loads(conflict.involved_user_ids) if conflict.involved_user_ids else []
+
+        return {
+            'conflict_id': conflict_id,
+            'involved_users': involved_user_ids,
+            'alignment_session_created': len(involved_user_ids) >= 2,
+            'message': '对齐邀请已发送'
+        }
+
+    async def _handle_tech_stack_change(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理技术栈变更
+
+        动作：
+        1. 识别受影响的任务
+        2. 更新任务描述和依赖
+        3. 重新估算工时
+        """
+        project_id = event_data.get('project_id')
+        new_tech = event_data.get('new_tech', '')
+
+        # 获取项目的所有任务
+        task_result = await self.db.execute(
+            select(Task).where(
+                Task.project_id == project_id,
+                Task.tenant_id == tenant_id,
+                Task.status.in_(['pending', 'in_progress'])
+            )
+        )
+        tasks = task_result.scalars().all()
+
+        affected_tasks = []
+        for task in tasks:
+            # AI 判断任务是否受影响
+            is_affected = await self._is_task_affected_by_tech_change(task, new_tech)
+            if is_affected:
+                affected_tasks.append({
+                    'task_id': task.id,
+                    'task_title': task.title,
+                    'impact': 'high'
+                })
+
+        return {
+            'project_id': project_id,
+            'new_tech': new_tech,
+            'affected_tasks': affected_tasks,
+            'message': f'识别出 {len(affected_tasks)} 个受影响的任务'
+        }
+
+    async def _handle_competitor_update(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理竞品动态
+
+        动作：
+        1. 评估战略影响
+        2. 调整项目方向
+        3. 创建应对任务
+        """
+        signal_content = event_data.get('signal_content', '')
+        competitor_name = event_data.get('competitor_name', '竞品')
+
+        # AI 分析影响
+        prompt = f"""分析以下竞品动态的战略影响：
+
+竞品：{competitor_name}
+动态：{signal_content}
+
+请评估：
+1. 对我们的威胁等级（high/medium/low）
+2. 建议的应对措施
+3. 是否需要调整项目方向
+
+以 JSON 格式返回。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            return {
+                'competitor': competitor_name,
+                'threat_assessment': response,
+                'strategic_adjustment_recommended': True
+            }
+        except:
+            return {
+                'competitor': competitor_name,
+                'threat_assessment': 'medium',
+                'message': 'AI 评估失败'
+            }
+
+    async def _handle_task_completed(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理任务完成
+
+        动作：
+        1. 检查依赖任务，解锁可开始的
+        2. 更新项目进度
+        3. 触发后续任务
+        """
+        task_id = event_data.get('task_id')
+
+        # 获取任务
+        task_result = await self.db.execute(
+            select(Task).where(
+                Task.id == task_id,
+                Task.tenant_id == tenant_id
+            )
+        )
+        task = task_result.scalar_one_or_none()
+
+        if not task:
+            raise ValueError(f"Task {task_id} not found")
+
+        # 查找依赖此任务的其他任务
+        # 这里简化处理：查找同项目的待办任务
+        dependent_result = await self.db.execute(
+            select(Task).where(
+                Task.project_id == task.project_id,
+                Task.tenant_id == tenant_id,
+                Task.status == 'pending',
+                Task.id != task_id
+            ).limit(5)
+        )
+        dependents = dependent_result.scalars().all()
+
+        return {
+            'completed_task_id': task_id,
+            'project_id': task.project_id,
+            'unlocked_tasks': [
+                {'id': t.id, 'title': t.title}
+                for t in dependents
+            ],
+            'message': f'解锁了 {len(dependents)} 个待办任务'
+        }
+
+    async def _handle_task_blocked(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理任务阻塞
+
+        动作：
+        1. 标记任务状态
+        2. 触发阻塞检测工作流
+        3. 通知项目负责人
+        """
+        task_id = event_data.get('task_id')
+        blocker_reason = event_data.get('reason', 'Unknown')
+
+        # 更新任务状态
+        await self.db.execute(
+            update(Task)
+            .where(Task.id == task_id, Task.tenant_id == tenant_id)
+            .values(status='blocked')
+        )
+        await self.db.commit()
+
+        return {
+            'task_id': task_id,
+            'blocker_reason': blocker_reason,
+            'status': 'blocked',
+            'workflow_triggered': True
+        }
+
+    async def _handle_project_risk(
+        self,
+        event_data: Dict[str, Any],
+        tenant_id: int
+    ) -> Dict[str, Any]:
+        """
+        处理项目风险
+
+        动作：
+        1. 评估风险等级
+        2. 创建缓解任务
+        3. 通知管理层
+        """
+        project_id = event_data.get('project_id')
+        risk_type = event_data.get('risk_type', 'general')
+        risk_description = event_data.get('description', '')
+
+        # 获取项目
+        project_result = await self.db.execute(
+            select(Project).where(
+                Project.id == project_id,
+                Project.tenant_id == tenant_id
+            )
+        )
+        project = project_result.scalar_one_or_none()
+
+        if not project:
+            raise ValueError(f"Project {project_id} not found")
+
+        return {
+            'project_id': project_id,
+            'project_name': project.title if project else 'Unknown',
+            'risk_type': risk_type,
+            'risk_level': 'high',
+            'mitigation_required': True,
+            'message': '风险已记录，等待处理'
+        }
+
+    # ==================== Helper Methods ====================
+
+    async def _assess_signal_impact_on_project(
+        self,
+        signal: Signal,
+        project: Project
+    ) -> Dict[str, Any]:
+        """AI 评估信号对项目的影响"""
+        prompt = f"""评估以下信号对项目的影响：
+
+项目：{project.title}
+描述：{project.description}
+
+信号：
+来源：{signal.source}
+内容：{signal.content}
+影响分数：{signal.impact_score}
+
+请判断：
+1. 是否需要调整项目优先级？
+2. 建议的优先级是什么？
+3. 是否需要创建响应任务？
+4. 原因是什么？
+
+以 JSON 格式返回：
+{{
+    "should_adjust_priority": boolean,
+    "suggested_priority": "high/medium/low",
+    "needs_response_task": boolean,
+    "reason": "原因说明"
+}}"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=signal.tenant_id if getattr(signal, "tenant_id", None) is not None else None,
+                project_id=project.id if project else None,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except:
+            return {
+                'should_adjust_priority': signal.impact_score > 0.9,
+                'suggested_priority': 'high' if signal.impact_score > 0.9 else 'medium',
+                'needs_response_task': signal.impact_score > 0.85,
+                'reason': f'信号影响分数 {signal.impact_score}'
+            }
+
+    async def _create_response_task(
+        self,
+        project: Project,
+        signal: Signal,
+        impact: Dict[str, Any],
+        tenant_id: int
+    ):
+        """创建响应任务"""
+        from sqlalchemy import text
+
+        next_task_id = (await self.db.execute(text("SELECT COALESCE(MAX(id),0)+1 FROM tasks"))).scalar() or 1
+
+        task = Task(
+            id=next_task_id,
+            truth_node_id=None,
+            project_id=project.id,
+            assignee_id=None,
+            title=f'响应信号: {signal.content[:50]}',
+            description=f"""自动生成的响应任务
+
+触发信号：
+来源：{signal.source}
+内容：{signal.content}
+
+影响评估：{impact.get('reason', '')}
+
+建议行动：
+请根据信号内容制定具体的响应方案。""",
+            required_skill_tag_ids=[],
+            status='pending',
+            priority='HIGH',
+            estimated_hours=8,
+            actual_hours=0,
+            quality_score=0,
+            tenant_id=tenant_id
+        )
+        self.db.add(task)
+        await self.db.commit()
+        await self.db.refresh(task)
+        return task
+
+    async def _is_task_affected_by_tech_change(
+        self,
+        task: Task,
+        new_tech: str
+    ) -> bool:
+        """判断任务是否受技术变更影响"""
+        # 简化处理：检查任务描述中是否包含相关技术关键词
+        task_keywords = task.title.lower() + ' ' + task.description.lower()
+        # 这里可以更复杂，使用 AI 判断
+        return new_tech.lower() in task_keywords
+
+
+# 全局服务实例
+def get_event_workflow_engine(db: AsyncSession) -> EventDrivenWorkflowEngine:
+    """获取事件驱动工作流引擎实例"""
+    return EventDrivenWorkflowEngine(db)
diff --git a/backend/app/services/experiment_sync_service.py b/backend/app/services/experiment_sync_service.py
new file mode 100644
index 00000000..0e2a13f0
--- /dev/null
+++ b/backend/app/services/experiment_sync_service.py
@@ -0,0 +1,300 @@
+"""实验结果同步服务（P1）"""
+
+from datetime import datetime, timezone
+from typing import Any, Dict, Optional, Set
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import Experiment, ExperimentTruthNodeSync, TruthNode
+
+
+class ExperimentSyncService:
+    """实验结果 -> 宪章同步，支持自动应用、人工审核与回滚。"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def auto_sync_experiment_result(
+        self,
+        experiment_id: int,
+        truth_node_id: int,
+        tenant_id: int,
+        project_id: Optional[int] = None,
+        sync_type: str = "result_sync",
+        auto_apply_threshold: float = 0.8,
+    ) -> Dict[str, Any]:
+        experiment = await self._get_experiment(experiment_id, tenant_id)
+        truth_node = await self._get_truth_node(truth_node_id, tenant_id)
+        if not experiment or not truth_node:
+            raise ValueError("实验或宪章不存在")
+        if project_id is not None:
+            await self.validate_experiment_truth_same_project(
+                experiment=experiment,
+                truth_node=truth_node,
+                tenant_id=tenant_id,
+                project_id=project_id,
+            )
+
+        confidence = self._calc_confidence(experiment.current_score)
+        sync_data = self._build_sync_data(experiment, truth_node, sync_type)
+        auto_apply = confidence >= auto_apply_threshold
+
+        record = ExperimentTruthNodeSync(
+            tenant_id=tenant_id,
+            experiment_id=experiment.id,
+            experiment_external_id=str(experiment.id),
+            truth_node_id=truth_node.id,
+            sync_type=sync_type,
+            sync_data=sync_data,
+            confidence_score=confidence,
+            auto_applied=auto_apply,
+            apply_status="pending",
+        )
+
+        if auto_apply:
+            await self._apply_sync_to_truth_node(record, truth_node)
+            record.apply_status = "applied"
+
+        self.db.add(record)
+        await self.db.commit()
+        await self.db.refresh(record)
+
+        return {
+            "sync_id": record.id,
+            "confidence_score": record.confidence_score,
+            "auto_applied": record.auto_applied,
+            "apply_status": record.apply_status,
+            "sync_data": record.sync_data,
+        }
+
+    async def validate_experiment_truth_same_project(
+        self,
+        experiment: Experiment,
+        truth_node: TruthNode,
+        tenant_id: int,
+        project_id: int,
+    ) -> None:
+        """交叉校验实验与宪章均属于指定项目。
+
+        注意：Experiment 当前无 project_id 字段，使用 linked_context_nodes 与项目宪章集合交叉验证。
+        """
+        if truth_node.project_id != project_id:
+            raise ValueError(
+                f"宪章 {truth_node.id} 不属于项目 {project_id}，拒绝跨项目同步"
+            )
+
+        # 强约束优先：Experiment.project_id 命中则直接使用。
+        if experiment.project_id is not None:
+            if int(experiment.project_id) != int(project_id):
+                raise ValueError(
+                    f"实验 {experiment.id} 不属于项目 {project_id}，拒绝跨项目同步"
+                )
+            return
+
+        project_truth_nodes = await self._get_project_truth_node_ids(project_id, tenant_id)
+        if not project_truth_nodes:
+            raise ValueError(f"项目 {project_id} 下无可用宪章，无法校验实验归属")
+
+        linked = experiment.linked_context_nodes or []
+        linked_ids = self._normalize_ids(linked)
+        if not linked_ids:
+            raise ValueError(
+                "实验未关联任何上下文节点（linked_context_nodes 为空），无法通过项目归属校验"
+            )
+
+        if linked_ids.isdisjoint(project_truth_nodes):
+            raise ValueError(
+                f"实验 {experiment.id} 未关联项目 {project_id} 下宪章，拒绝跨项目同步"
+            )
+
+        # 迁移期兼容：可推断归属时补写 project_id，减少后续重复推断。
+        experiment.project_id = project_id
+
+    async def _get_project_truth_node_ids(self, project_id: int, tenant_id: int) -> Set[int]:
+        result = await self.db.execute(
+            select(TruthNode.id).where(
+                and_(
+                    TruthNode.tenant_id == tenant_id,
+                    TruthNode.project_id == project_id,
+                )
+            )
+        )
+        return {int(row[0]) for row in result.all()}
+
+    def _normalize_ids(self, raw: Any) -> Set[int]:
+        ids: Set[int] = set()
+        if isinstance(raw, list):
+            for item in raw:
+                try:
+                    ids.add(int(item))
+                except (TypeError, ValueError):
+                    continue
+        return ids
+
+    async def review_sync(
+        self,
+        sync_id: int,
+        tenant_id: int,
+        reviewer_id: int,
+        action: str,
+        review_notes: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        record = await self._get_sync_record(sync_id, tenant_id)
+        if not record:
+            raise ValueError("同步记录不存在")
+        if record.apply_status == "applied" and action == "approve":
+            return {"sync_id": record.id, "apply_status": record.apply_status}
+
+        record.reviewed_by = reviewer_id
+        record.reviewed_at = datetime.now(timezone.utc)
+        record.review_notes = review_notes
+
+        if action == "approve":
+            truth_node = await self._get_truth_node(record.truth_node_id, tenant_id)
+            if truth_node:
+                await self._apply_sync_to_truth_node(record, truth_node)
+            record.apply_status = "applied"
+        else:
+            record.apply_status = "rejected"
+
+        await self.db.commit()
+        await self.db.refresh(record)
+        return {
+            "sync_id": record.id,
+            "apply_status": record.apply_status,
+            "reviewed_at": record.reviewed_at.isoformat() if record.reviewed_at else None,
+        }
+
+    async def rollback_sync(
+        self,
+        sync_id: int,
+        tenant_id: int,
+        reviewer_id: int,
+        rollback_notes: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        record = await self._get_sync_record(sync_id, tenant_id)
+        if not record:
+            raise ValueError("同步记录不存在")
+        if record.apply_status != "applied":
+            raise ValueError("仅已应用记录支持回滚")
+
+        truth_node = await self._get_truth_node(record.truth_node_id, tenant_id)
+        if not truth_node:
+            raise ValueError("关联宪章不存在")
+
+        snapshot = (record.sync_data or {}).get("previous_truth_content")
+        if snapshot is None:
+            raise ValueError("缺少回滚快照，无法回滚")
+
+        truth_node.content = snapshot
+        truth_node.version = int(truth_node.version or 1) + 1
+
+        record.apply_status = "requires_review"
+        record.reviewed_by = reviewer_id
+        record.reviewed_at = datetime.now(timezone.utc)
+        record.review_notes = rollback_notes or "manual rollback"
+
+        await self.db.commit()
+        await self.db.refresh(record)
+        return {
+            "sync_id": record.id,
+            "apply_status": record.apply_status,
+            "rolled_back": True,
+        }
+
+    async def list_sync_records(
+        self,
+        tenant_id: int,
+        truth_node_id: Optional[int] = None,
+        limit: int = 50,
+    ) -> Dict[str, Any]:
+        query = select(ExperimentTruthNodeSync).where(
+            ExperimentTruthNodeSync.tenant_id == tenant_id
+        )
+        if truth_node_id is not None:
+            query = query.where(ExperimentTruthNodeSync.truth_node_id == truth_node_id)
+        query = query.order_by(ExperimentTruthNodeSync.created_at.desc()).limit(limit)
+
+        result = await self.db.execute(query)
+        items = result.scalars().all()
+        return {
+            "items": [
+                {
+                    "id": i.id,
+                    "experiment_id": i.experiment_id,
+                    "truth_node_id": i.truth_node_id,
+                    "sync_type": i.sync_type,
+                    "confidence_score": i.confidence_score,
+                    "auto_applied": i.auto_applied,
+                    "apply_status": i.apply_status,
+                    "created_at": i.created_at.isoformat() if i.created_at else None,
+                }
+                for i in items
+            ]
+        }
+
+    async def _get_experiment(self, experiment_id: int, tenant_id: int) -> Optional[Experiment]:
+        result = await self.db.execute(
+            select(Experiment).where(
+                and_(Experiment.id == experiment_id, Experiment.tenant_id == tenant_id)
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def _get_truth_node(self, truth_node_id: int, tenant_id: int) -> Optional[TruthNode]:
+        result = await self.db.execute(
+            select(TruthNode).where(
+                and_(TruthNode.id == truth_node_id, TruthNode.tenant_id == tenant_id)
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def _get_sync_record(
+        self, sync_id: int, tenant_id: int
+    ) -> Optional[ExperimentTruthNodeSync]:
+        result = await self.db.execute(
+            select(ExperimentTruthNodeSync).where(
+                and_(ExperimentTruthNodeSync.id == sync_id, ExperimentTruthNodeSync.tenant_id == tenant_id)
+            )
+        )
+        return result.scalar_one_or_none()
+
+    def _calc_confidence(self, current_score: Optional[float]) -> float:
+        score = float(current_score or 0.0)
+        if score > 1:
+            score /= 100.0
+        return round(max(0.0, min(1.0, score)), 2)
+
+    def _build_sync_data(
+        self, experiment: Experiment, truth_node: TruthNode, sync_type: str
+    ) -> Dict[str, Any]:
+        return {
+            "sync_type": sync_type,
+            "experiment_summary": {
+                "id": experiment.id,
+                "title": experiment.title,
+                "status": experiment.status,
+                "target_metric": experiment.target_metric,
+                "current_score": experiment.current_score,
+            },
+            "previous_truth_content": truth_node.content,
+        }
+
+    async def _apply_sync_to_truth_node(
+        self, record: ExperimentTruthNodeSync, truth_node: TruthNode
+    ) -> None:
+        content = truth_node.content if isinstance(truth_node.content, dict) else {}
+        if not isinstance(content, dict):
+            content = {"raw": str(truth_node.content)}
+        sync_section = content.get("experiment_sync", {})
+        sync_section["last_sync"] = {
+            "sync_id": record.id,
+            "sync_type": record.sync_type,
+            "confidence_score": record.confidence_score,
+            "experiment_summary": (record.sync_data or {}).get("experiment_summary", {}),
+            "applied_at": datetime.now(timezone.utc).isoformat(),
+        }
+        content["experiment_sync"] = sync_section
+        truth_node.content = content
+        truth_node.version = int(truth_node.version or 1) + 1
diff --git a/backend/app/services/feishu_integration.py b/backend/app/services/feishu_integration.py
new file mode 100644
index 00000000..df2a01ed
--- /dev/null
+++ b/backend/app/services/feishu_integration.py
@@ -0,0 +1,1873 @@
+"""
+Feishu Integration Service - 飞书集成服务
+
+功能：
+1. 接收飞书群聊消息事件
+2. AI 判断消息是否为有价值的信号
+3. 自动创建信号并关联到项目
+4. 推送通知给相关角色
+
+飞书开放平台配置：
+- App ID: 从飞书开放平台获取
+- App Secret: 从飞书开放平台获取
+- 事件接收: 统一使用长连接模式
+"""
+import asyncio
+import json
+import hashlib
+import re
+import time
+from uuid import uuid4
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional
+import httpx
+from sqlalchemy import and_, or_
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+
+from ..models import (
+    User,
+    Project,
+    Signal,
+    ProjectSignalLink,
+    Integration,
+    FeishuReplyOutbox,
+    AIProcessingLog,
+)
+from ..database import AsyncSessionLocal
+from .notification_service import NotificationService
+from .signal_capture import ProactiveSignalCaptureService
+from .llm import LLMService
+from .prompt_profile_service import prompt_profile_service
+
+CN_TZ = timezone(timedelta(hours=8))
+WEEKDAY_CN = ["周一", "周二", "周三", "周四", "周五", "周六", "周日"]
+OUTBOX_MAX_ATTEMPTS = 3
+OUTBOX_STALE_SENDING_SECONDS = 120
+
+
+class FeishuIntegrationService:
+    """飞书集成服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+        # 从数据库获取飞书配置
+        self.app_id = None
+        self.app_secret = None
+        self.encrypt_key = None
+        self.verification_token = None
+        self.event_mode = "long_connection"
+        self.tenant_id = 1
+
+    @staticmethod
+    def _normalize_event_mode(_: Optional[str] = None) -> str:
+        return "long_connection"
+
+    async def load_config(self, tenant_id: int = 1):
+        """从数据库加载飞书配置"""
+        self.tenant_id = tenant_id
+
+        result = await self.db.execute(
+            select(Integration).where(
+                Integration.tenant_id == tenant_id,
+                Integration.type == 'feishu',
+                Integration.is_active.is_(True),
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration:
+            if isinstance(integration.config, str):
+                config = json.loads(integration.config) if integration.config else {}
+            else:
+                config = integration.config or {}
+            self.app_id = config.get('app_id')
+            self.app_secret = config.get('app_secret')
+            self.encrypt_key = None
+            self.verification_token = None
+            self.event_mode = self._normalize_event_mode(config.get('event_mode'))
+        else:
+            self.app_id = None
+            self.app_secret = None
+            self.encrypt_key = None
+            self.verification_token = None
+            self.event_mode = 'long_connection'
+
+    def is_configured(self) -> bool:
+        """检查是否已配置"""
+        return bool(self.app_id and self.app_secret)
+
+    async def save_config(
+        self,
+        app_id: str,
+        app_secret: str,
+        encrypt_key: str,
+        verification_token: str,
+        event_mode: str = "long_connection",
+        tenant_id: int = 1
+    ):
+        """保存飞书配置到数据库"""
+        mode = self._normalize_event_mode(event_mode)
+        config = {
+            'app_id': app_id,
+            'app_secret': app_secret,
+            'encrypt_key': "",
+            'verification_token': "",
+            'event_mode': mode,
+        }
+
+        # 查找或创建集成记录
+        result = await self.db.execute(
+            select(Integration).where(
+                Integration.tenant_id == tenant_id,
+                Integration.type == 'feishu'
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration:
+            integration.config = config
+            integration.is_active = True
+        else:
+            integration = Integration(
+                type='feishu',
+                config=config,
+                is_active=True,
+                tenant_id=tenant_id
+            )
+            self.db.add(integration)
+
+        await self.db.commit()
+
+        # 更新内存配置
+        self.app_id = app_id
+        self.app_secret = app_secret
+        self.encrypt_key = None
+        self.verification_token = None
+        self.event_mode = mode
+        self.tenant_id = tenant_id
+
+    async def handle_message_event(
+        self,
+        event_data: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """
+        处理飞书消息事件
+
+        支持的事件类型：
+        - message.receive_v1: 收到群聊消息
+        - im.message.receive_v1: 收到群聊消息（schema 2.0）
+        """
+        try:
+            event_type = (event_data.get('type') or '').strip()
+
+            # 飞书 schema 2.0 会返回 im.message.receive_v1
+            if event_type in {'message.receive_v1', 'im.message.receive_v1'}:
+                return await self._handle_message_receive(event_data)
+
+            return {
+                'code': 0,
+                'msg': f'Event type {event_type} not handled'
+            }
+
+        except Exception as e:
+            print(f"[Feishu] Handle event error: {e}")
+            return {
+                'code': 1,
+                'msg': str(e)
+            }
+
+    async def _handle_message_receive(
+        self,
+        event: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """处理收到的消息事件"""
+        state_trace_id = f"feishu-sm-{uuid4().hex[:12]}"
+        ingress_source = str(event.get("_ingress_source") or "feishu").strip() or "feishu"
+        message = event.get('message', {}) if isinstance(event.get('message', {}), dict) else {}
+        chat = event.get('chat', {}) if isinstance(event.get('chat', {}), dict) else {}
+        sender = event.get('sender', {}) if isinstance(event.get('sender', {}), dict) else {}
+        message_id = (message.get("message_id") or "").strip() or "unknown"
+        self._emit_state_transition(
+            trace_id=state_trace_id,
+            state="RECEIVED",
+            message_id=message_id,
+            message_type=str(message.get("message_type") or "unknown"),
+        )
+        sender_type = str(sender.get("sender_type") or "").strip().lower()
+        if sender_type in {"app", "bot"}:
+            return {'code': 0, 'msg': 'Ignore bot/app message'}
+
+        mentioned_bot_names = self._extract_mentioned_bot_names(message)
+
+        # 解析消息内容与附件元数据
+        message_payload = self._extract_message_payload(event)
+        message_content = str(message_payload.get("text") or "").strip() or None
+        message_attachments = (
+            message_payload.get("attachments")
+            if isinstance(message_payload.get("attachments"), list)
+            else []
+        )
+        await self._record_ingress_trace(
+            ingress_source=ingress_source,
+            trace_id=state_trace_id,
+            message_id=message_id,
+            event=event,
+            message_content=message_content or "",
+        )
+        if not message_content:
+            print(
+                f"[Feishu] message.receive_v1 ignored: no text content, "
+                f"message_type={message.get('message_type')}, keys={list(message.keys())}"
+            )
+            chat_id = chat.get('chat_id') or message.get('chat_id') or ''
+            if mentioned_bot_names and chat_id:
+                await self._safe_reply_text(
+                    chat_id=chat_id,
+                    mention_open_id=(sender.get('sender_id', {}) or {}).get('open_id'),
+                    mention_name=(sender.get('sender_name') or sender.get('name') or '你'),
+                    text="我已收到你的消息，但目前无法直接解析纯图片内容。请补充 1-2 句文字需求，我会立即给出分派建议。",
+                    trace_id=state_trace_id,
+                )
+                self._emit_state_transition(
+                    trace_id=state_trace_id,
+                    state="REPLIED",
+                    route="empty_text_prompt",
+                )
+            return {'code': 0, 'msg': 'No text content'}
+
+        # 获取发送者信息
+        sender_id = sender.get('sender_id', {}) if isinstance(sender.get('sender_id', {}), dict) else {}
+        sender_open_id = (sender_id.get('open_id') or "").strip() or None
+        sender_name = await self._resolve_sender_display_name(
+            sender=sender,
+            sender_id=sender_id,
+        )
+
+        # 获取群聊信息
+        chat_id = chat.get('chat_id') or message.get('chat_id') or ''
+        chat_type = str(chat.get("chat_type") or message.get("chat_type") or "").strip().lower() or "group"
+        captured_at = self._extract_message_datetime(message)
+        occurred_at_iso = (
+            captured_at.isoformat() if isinstance(captured_at, datetime) else None
+        )
+        message_attachments = self._normalize_ingress_attachment_records(
+            message_attachments,
+            occurred_at=occurred_at_iso,
+            source_message_id=(message.get("message_id") or "").strip() or None,
+            source_chat_id=chat_id or None,
+        )
+
+        # 使用信号捕获服务处理
+        capture_service = ProactiveSignalCaptureService(self.db)
+
+        try:
+            result = await capture_service.capture_from_feishu(
+                group_chat_id=chat_id,
+                message_content=message_content,
+                sender=sender_name,
+                tenant_id=self.tenant_id,
+                proposer_open_id=sender_open_id,
+                captured_at=captured_at,
+                mentioned_bot_names=mentioned_bot_names,
+                feishu_message_id=(message.get("message_id") or "").strip() or None,
+                chat_type=chat_type,
+            )
+            await self._persist_signal_attachments_from_ingress(
+                signal_id=result.get("signal_id"),
+                attachments=message_attachments,
+            )
+        except Exception as exc:
+            self._emit_state_transition(
+                trace_id=state_trace_id,
+                state="ERROR_REPLY",
+                reason="capture_exception",
+                error=str(exc),
+            )
+            if chat_id:
+                await self._safe_reply_text(
+                    chat_id=chat_id,
+                    mention_open_id=sender_open_id,
+                    mention_name=sender_name,
+                    text=self._build_error_reply_text(),
+                    trace_id=state_trace_id,
+                )
+                self._emit_state_transition(
+                    trace_id=state_trace_id,
+                    state="REPLIED",
+                    route="error_template",
+                )
+            return {
+                "code": 0,
+                "msg": "error_handled",
+                "data": {
+                    "status": "error",
+                    "classification": "error",
+                    "ai_trace_id": state_trace_id,
+                },
+            }
+
+        print(
+            "[Feishu] Message processed: "
+            f"status={result.get('status')}, "
+            f"classification={result.get('classification')}, "
+            f"linked_projects={result.get('linked_project_ids', [])}, "
+            f"auto_tasks={len(result.get('auto_tasks', []) if isinstance(result.get('auto_tasks', []), list) else [])}, "
+            f"ai_trace_id={result.get('ai_trace_id')}"
+        )
+        scene = self._classify_message_scene(capture_result=result)
+        self._emit_state_transition(
+            trace_id=state_trace_id,
+            state="CLASSIFIED",
+            scene=scene,
+            status=str(result.get("status") or "unknown"),
+            classification=str(result.get("classification") or "unknown"),
+        )
+        if str(result.get("status") or "").lower() != "error":
+            self._emit_state_transition(
+                trace_id=state_trace_id,
+                state="COMMITTED",
+                ai_trace_id=str(result.get("ai_trace_id") or ""),
+            )
+
+        reply_route = "none"
+        if scene == "execute":
+            if self._is_duplicate_capture_result(result):
+                self._emit_state_transition(
+                    trace_id=state_trace_id,
+                    state="ROUTED",
+                    route="duplicate_suppressed",
+                )
+            else:
+                self._emit_state_transition(
+                    trace_id=state_trace_id,
+                    state="ROUTED",
+                    route="tool_first_execute",
+                )
+                deterministic_reply = str(result.get("direct_reply_text") or "").strip()
+                if not deterministic_reply:
+                    deterministic_reply = self._build_execution_receipt_reply(
+                        message_content=message_content,
+                        capture_result=result,
+                    )
+                deterministic_reply, guard_reason = self._guard_execute_reply_text_with_reason(
+                    deterministic_reply,
+                    trace_id=state_trace_id,
+                )
+                if guard_reason:
+                    asyncio.create_task(
+                        self._record_execute_guard_alert(
+                            trace_id=state_trace_id,
+                            message_id=message_id,
+                            message_content=message_content,
+                            capture_result=result,
+                            guard_reason=guard_reason,
+                        )
+                    )
+                if not deterministic_reply and str(result.get("status") or "").lower() == "error":
+                    deterministic_reply = self._build_error_reply_text()
+                if chat_id and deterministic_reply:
+                    await self._safe_reply_text(
+                        chat_id=chat_id,
+                        mention_open_id=sender_open_id,
+                        mention_name=sender_name,
+                        text=deterministic_reply,
+                        trace_id=state_trace_id,
+                    )
+                    reply_route = "execute_template"
+        elif scene == "qa":
+            self._emit_state_transition(
+                trace_id=state_trace_id,
+                state="ROUTED",
+                route="qa_consult",
+            )
+            if chat_id and bool(result.get("direct_reply_required")):
+                await self._safe_reply_text(
+                    chat_id=chat_id,
+                    mention_open_id=sender_open_id,
+                    mention_name=sender_name,
+                    text=str(result.get("direct_reply_text") or "已收到，正在处理。"),
+                    trace_id=state_trace_id,
+                )
+                reply_route = "qa_direct"
+            elif (
+                chat_id
+                and self._should_reply_to_mention(
+                    mentioned_bot_names=mentioned_bot_names,
+                    message_content=message_content,
+                )
+            ):
+                await self._reply_with_active_role_prompt(
+                    chat_id=chat_id,
+                    sender_open_id=sender_open_id,
+                    sender_name=sender_name,
+                    message_content=message_content,
+                    capture_result=result,
+                    trace_id=state_trace_id,
+                )
+                reply_route = "qa_llm"
+        else:
+            self._emit_state_transition(
+                trace_id=state_trace_id,
+                state="ERROR_REPLY",
+                reason="scene_error",
+                capture_status=str(result.get("status") or "unknown"),
+            )
+            if chat_id:
+                await self._safe_reply_text(
+                    chat_id=chat_id,
+                    mention_open_id=sender_open_id,
+                    mention_name=sender_name,
+                    text=self._build_error_reply_text(),
+                    trace_id=state_trace_id,
+                )
+                reply_route = "error_template"
+
+        if reply_route != "none":
+            self._emit_state_transition(
+                trace_id=state_trace_id,
+                state="REPLIED",
+                route=reply_route,
+            )
+        await self._record_message_trace(
+            trace_id=state_trace_id,
+            message_id=message_id,
+            chat_type=chat_type,
+            message_content=message_content,
+            capture_result=result,
+            reply_source=reply_route,
+        )
+
+        return {
+            'code': 0,
+            'msg': 'success',
+            'data': {
+                'status': result.get('status'),
+                'classification': result.get('classification'),
+                'linked_project_ids': result.get('linked_project_ids', []),
+                'auto_tasks_count': len(result.get('auto_tasks', []) if isinstance(result.get('auto_tasks', []), list) else []),
+                'ai_trace_id': result.get('ai_trace_id'),
+                'state_trace_id': state_trace_id,
+                'scene': scene,
+            },
+        }
+
+    async def _record_ingress_trace(
+        self,
+        *,
+        ingress_source: str,
+        trace_id: str,
+        message_id: str,
+        event: Dict[str, Any],
+        message_content: str,
+    ) -> None:
+        try:
+            message = event.get("message", {}) if isinstance(event.get("message", {}), dict) else {}
+            chat = event.get("chat", {}) if isinstance(event.get("chat", {}), dict) else {}
+            sender = event.get("sender", {}) if isinstance(event.get("sender", {}), dict) else {}
+            sender_id = sender.get("sender_id", {}) if isinstance(sender.get("sender_id", {}), dict) else {}
+            sender_name = (
+                str(sender.get("sender_name") or sender.get("name") or "").strip()
+                or str(sender_id.get("open_id") or "").strip()
+                or "unknown"
+            )
+            chat_id = str(chat.get("chat_id") or message.get("chat_id") or "").strip() or None
+            chat_type = str(chat.get("chat_type") or message.get("chat_type") or "").strip() or None
+
+            async with AsyncSessionLocal() as db:
+                db.add(
+                    AIProcessingLog(
+                        tenant_id=int(self.tenant_id or 1),
+                        feishu_message_id=(message_id or "")[:128] or None,
+                        feishu_item_id=None,
+                        raw_message=(message_content or "")[:8000] or None,
+                        ai_raw_output=json.dumps(
+                            {
+                                "trace_id": trace_id,
+                                "ingress_source": ingress_source,
+                                "event_type": str(event.get("type") or "")[:120],
+                                "chat_id": chat_id,
+                                "chat_type": chat_type,
+                                "sender_name": sender_name[:200],
+                            },
+                            ensure_ascii=False,
+                        )[:120000],
+                        project_score=None,
+                        signal_confidence=None,
+                        level="ingress",
+                        normalized_text=(message_content or "")[:8000] or None,
+                        route=ingress_source[:80],
+                        intent=None,
+                        executor="feishu_ingress",
+                        operation_code=None,
+                        operation_result={
+                            "message_type": str(message.get("message_type") or "")[:80],
+                            "chat_id": chat_id,
+                            "chat_type": chat_type,
+                            "sender_name": sender_name[:200],
+                        },
+                        reply_source=None,
+                        routing_decision="event_received",
+                        routing_reason=f"Feishu ingress via {ingress_source}",
+                        reasoning_bullets=[
+                            f"trace_id={trace_id}",
+                            f"message_id={message_id}",
+                            f"source={ingress_source}",
+                        ],
+                        processing_status="success",
+                        error_message=None,
+                        duration_ms=0,
+                    )
+                )
+                await db.commit()
+        except Exception as exc:
+            print(f"[Feishu] ingress trace write failed: trace_id={trace_id}, error={exc}")
+
+    @staticmethod
+    def _classify_message_scene(*, capture_result: Dict[str, Any]) -> str:
+        if not isinstance(capture_result, dict):
+            return "error"
+        status = str(capture_result.get("status") or "").strip().lower()
+        classification = str(capture_result.get("classification") or "").strip().lower()
+        if status in {"captured", "ignored"}:
+            return "execute"
+        if status == "answered" or classification in {"question", "qa"}:
+            return "qa"
+        if status == "error" or classification == "error":
+            return "error"
+        return "qa"
+
+    @staticmethod
+    def _build_error_reply_text() -> str:
+        return "系统处理中出现异常，本次未执行。请稍后重试，或补充更明确的目标、对象和动作。"
+
+    @staticmethod
+    def _emit_state_transition(trace_id: str, state: str, **payload: Any) -> None:
+        safe_trace = str(trace_id or "unknown")
+        safe_state = str(state or "unknown")
+        details = ", ".join([f"{k}={v}" for k, v in payload.items()]) if payload else ""
+        print(f"[Feishu-SM][{safe_trace}] state={safe_state}{', ' + details if details else ''}")
+
+    async def _reply_with_active_role_prompt(
+        self,
+        *,
+        chat_id: str,
+        sender_open_id: Optional[str],
+        sender_name: str,
+        message_content: str,
+        capture_result: Dict[str, Any],
+        trace_id: Optional[str] = None,
+        ) -> None:
+        """被 @ 时回复一条有人味的执行反馈。"""
+        deterministic_time_reply = self._build_time_truth_reply(message_content)
+        if deterministic_time_reply:
+            await self._safe_reply_text(
+                chat_id=chat_id,
+                mention_open_id=sender_open_id,
+                mention_name=sender_name,
+                text=deterministic_time_reply,
+                trace_id=trace_id,
+            )
+            return
+
+        execution_receipt = self._build_execution_receipt_reply(
+            message_content=message_content,
+            capture_result=capture_result,
+        )
+        execution_receipt = self._guard_execute_reply_text(
+            execution_receipt,
+            trace_id=str(trace_id or f"feishu-sm-{uuid4().hex[:12]}"),
+        )
+        if execution_receipt:
+            await self._safe_reply_text(
+                chat_id=chat_id,
+                mention_open_id=sender_open_id,
+                mention_name=sender_name,
+                text=execution_receipt,
+                trace_id=trace_id,
+            )
+            return
+
+        butler_reply = self._build_butler_decision_reply(capture_result=capture_result)
+        if butler_reply:
+            await self._safe_reply_text(
+                chat_id=chat_id,
+                mention_open_id=sender_open_id,
+                mention_name=sender_name,
+                text=butler_reply,
+                trace_id=trace_id,
+            )
+            return
+
+        try:
+            active_role = await prompt_profile_service.get_active_role(self.db, tenant_id=self.tenant_id)
+            role = active_role.get("role") if isinstance(active_role, dict) else {}
+            role_name = str((role or {}).get("role_name") or "产品经理（任务分派）").strip()
+            system_prompt = str((role or {}).get("system_prompt") or "").strip()
+            if len(system_prompt) < 20:
+                system_prompt = (
+                    "你是 PMFlow 的任务分派产品经理。"
+                    "请用简洁、专业、有温度的中文回复用户，先确认需求，再给可执行下一步。"
+                )
+
+            status = str(capture_result.get("status") or "unknown")
+            classification = str(capture_result.get("classification") or "unknown")
+            reason = str(capture_result.get("reason") or "").strip()
+            ai_trace_id = str(capture_result.get("ai_trace_id") or "").strip()
+            linked_project_ids = capture_result.get("linked_project_ids", [])
+            linked_projects_text = ", ".join([str(x) for x in linked_project_ids[:3]]) if isinstance(linked_project_ids, list) else ""
+            now_cn = datetime.now(CN_TZ)
+            now_cn_text = (
+                f"{now_cn.year}年{now_cn.month}月{now_cn.day}日 "
+                f"{WEEKDAY_CN[now_cn.weekday()]} {now_cn.strftime('%H:%M:%S')}"
+            )
+
+            prompt = f"""
+{system_prompt}
+
+你正在飞书群里以「{role_name}」身份回复用户。请基于以下上下文，生成一条 35-140 字中文回复：
+- 当前系统时间（北京时间）：{now_cn_text}
+- 提及用户：{sender_name}
+- 用户消息：{(message_content or '')[:600]}
+- 系统处理状态：{status}
+- 分类结果：{classification}
+- 处理原因：{reason or "无"}
+- 处理追踪ID：{ai_trace_id or "无"}
+- 关联项目ID：{linked_projects_text or '无'}
+
+回复要求：
+1. 语气自然、专业，避免机器口吻。
+2. 先确认你已处理，再给 1-3 条下一步建议。
+3. 若信息不足，请明确要补充什么。
+4. 如果用户询问“日期/今天/时间/几点/星期”，必须基于“当前系统时间（北京时间）”直接回答，不得编造历史日期。
+5. 不要输出 Markdown，不要输出 JSON。
+只输出最终回复正文。
+""".strip()
+
+            llm = LLMService()
+            raw_reply = await llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=self.tenant_id,
+                function_name="feishu_humanized_reply",
+                temperature=0.4,
+                max_tokens=180,
+            )
+            reply_text = self._clean_reply_text(raw_reply)
+            if (not reply_text) or self._looks_like_provider_error(reply_text):
+                reply_text = self._fallback_reply_text(capture_result)
+        except Exception as exc:
+            print(f"[Feishu] generate mention reply failed: {exc}")
+            reply_text = self._fallback_reply_text(capture_result)
+
+        await self._safe_reply_text(
+            chat_id=chat_id,
+            mention_open_id=sender_open_id,
+            mention_name=sender_name,
+            text=reply_text,
+            trace_id=trace_id,
+        )
+
+    async def _safe_reply_text(
+        self,
+        *,
+        chat_id: str,
+        text: str,
+        mention_open_id: Optional[str] = None,
+        mention_name: Optional[str] = None,
+        trace_id: Optional[str] = None,
+    ) -> None:
+        try:
+            plain_text = self._normalize_outgoing_reply_text(text)
+            mention_prefix = self._build_mention_prefix(
+                open_id=mention_open_id,
+                display_name=mention_name,
+            )
+            final_text = f"{mention_prefix}{plain_text}".strip()
+            outbox_id = await self._enqueue_reply_outbox(
+                chat_id=chat_id,
+                message_text=final_text,
+                trace_id=trace_id,
+            )
+            asyncio.create_task(self._dispatch_outbox_entry(outbox_id=outbox_id))
+        except Exception as exc:
+            print(f"[Feishu] enqueue reply outbox failed: chat_id={chat_id}, error={exc}")
+            try:
+                plain_text = self._normalize_outgoing_reply_text(text)
+                mention_prefix = self._build_mention_prefix(
+                    open_id=mention_open_id,
+                    display_name=mention_name,
+                )
+                await self.send_message_to_group(chat_id=chat_id, message=f"{mention_prefix}{plain_text}".strip())
+            except Exception as send_exc:
+                print(f"[Feishu] send fallback failed: chat_id={chat_id}, error={send_exc}")
+
+    async def _enqueue_reply_outbox(
+        self,
+        *,
+        chat_id: str,
+        message_text: str,
+        trace_id: Optional[str],
+    ) -> int:
+        async with AsyncSessionLocal() as db:
+            row = FeishuReplyOutbox(
+                tenant_id=int(self.tenant_id or 1),
+                trace_id=(trace_id or "")[:64] or None,
+                chat_id=(chat_id or "")[:128],
+                message_text=(message_text or "")[:2000] or "已收到，正在处理。",
+                status="pending",
+                attempts=0,
+                next_retry_at=None,
+            )
+            db.add(row)
+            await db.commit()
+            await db.refresh(row)
+            return int(row.id)
+
+    async def _dispatch_outbox_entry(self, *, outbox_id: int) -> None:
+        await self.__class__._dispatch_outbox_entry_by_id(outbox_id=outbox_id)
+
+    @classmethod
+    async def _dispatch_outbox_entry_by_id(cls, *, outbox_id: int) -> None:
+        for _ in range(OUTBOX_MAX_ATTEMPTS + 1):
+            now = datetime.utcnow()
+            should_wait_seconds = 0.0
+            async with AsyncSessionLocal() as db:
+                row = (
+                    await db.execute(
+                        select(FeishuReplyOutbox).where(
+                            FeishuReplyOutbox.id == int(outbox_id)
+                        )
+                    )
+                ).scalar_one_or_none()
+                if row is None:
+                    return
+                status = str(row.status or "").lower()
+                if status == "sent":
+                    return
+                if status == "failed":
+                    return
+                if status == "sending":
+                    stale_seconds = 0.0
+                    marker = row.updated_at or row.created_at
+                    if isinstance(marker, datetime):
+                        try:
+                            marker_naive = marker.replace(tzinfo=None) if marker.tzinfo else marker
+                            stale_seconds = max(0.0, (now - marker_naive).total_seconds())
+                        except Exception:
+                            stale_seconds = 0.0
+                    if stale_seconds >= OUTBOX_STALE_SENDING_SECONDS:
+                        row.status = "retrying"
+                        row.next_retry_at = now
+                        await db.commit()
+                    else:
+                        return
+                if row.next_retry_at and row.next_retry_at > now:
+                    should_wait_seconds = max(0.2, min(30.0, (row.next_retry_at - now).total_seconds()))
+                else:
+                    row.status = "sending"
+                    row.attempts = int(row.attempts or 0) + 1
+                    row.last_error = None
+                    await db.commit()
+
+            if should_wait_seconds > 0:
+                await asyncio.sleep(should_wait_seconds)
+                continue
+
+            send_error: Optional[str] = None
+            async with AsyncSessionLocal() as db_send:
+                row = (
+                    await db_send.execute(
+                        select(FeishuReplyOutbox).where(
+                            FeishuReplyOutbox.id == int(outbox_id)
+                        )
+                    )
+                ).scalar_one_or_none()
+                if row is None:
+                    return
+                try:
+                    service = FeishuIntegrationService(db_send)
+                    await service.load_config(tenant_id=int(row.tenant_id))
+                    await service.send_message_to_group(
+                        chat_id=str(row.chat_id or ""),
+                        message=str(row.message_text or ""),
+                    )
+                    row.status = "sent"
+                    row.sent_at = datetime.utcnow()
+                    row.next_retry_at = None
+                    row.last_error = None
+                    await db_send.commit()
+                    return
+                except Exception as exc:
+                    send_error = str(exc)
+
+            async with AsyncSessionLocal() as db_retry:
+                row = (
+                    await db_retry.execute(
+                        select(FeishuReplyOutbox).where(
+                            FeishuReplyOutbox.id == int(outbox_id)
+                        )
+                    )
+                ).scalar_one_or_none()
+                if row is None:
+                    return
+                attempts = int(row.attempts or 0)
+                if attempts >= OUTBOX_MAX_ATTEMPTS:
+                    row.status = "failed"
+                    row.next_retry_at = None
+                else:
+                    row.status = "retrying"
+                    delay_seconds = min(30, 2 * (2 ** max(0, attempts - 1)))
+                    row.next_retry_at = datetime.utcnow() + timedelta(seconds=delay_seconds)
+                row.last_error = (send_error or "unknown_send_error")[:1000]
+                await db_retry.commit()
+
+            if attempts >= OUTBOX_MAX_ATTEMPTS:
+                return
+            await asyncio.sleep(min(5, 2 * max(1, attempts)))
+
+    @classmethod
+    async def dispatch_pending_outbox_entries(cls, *, limit: int = 30) -> int:
+        now = datetime.utcnow()
+        fetch_limit = max(1, min(200, int(limit)))
+        async with AsyncSessionLocal() as db:
+            due_rows = (
+                await db.execute(
+                    select(FeishuReplyOutbox.id)
+                    .where(
+                        FeishuReplyOutbox.status.in_(("pending", "retrying")),
+                        or_(
+                            FeishuReplyOutbox.next_retry_at.is_(None),
+                            FeishuReplyOutbox.next_retry_at <= now,
+                        ),
+                    )
+                    .order_by(FeishuReplyOutbox.id.asc())
+                    .limit(fetch_limit)
+                )
+            ).scalars().all()
+
+        if not due_rows:
+            return 0
+
+        await asyncio.gather(
+            *(cls._dispatch_outbox_entry_by_id(outbox_id=int(outbox_id)) for outbox_id in due_rows),
+            return_exceptions=True,
+        )
+        return len(due_rows)
+
+    @staticmethod
+    def _clean_reply_text(value: Optional[str]) -> str:
+        text = str(value or "").strip()
+        if text.startswith("```"):
+            text = text.replace("```", "").strip()
+        if text.lower().startswith("json"):
+            text = text[4:].strip()
+        return text[:500]
+
+    @classmethod
+    def _normalize_outgoing_reply_text(cls, value: Optional[str]) -> str:
+        text = cls._clean_reply_text(value)
+        if (not text) or cls._looks_like_provider_error(text):
+            return cls._build_error_reply_text()
+        return text[:500]
+
+    @staticmethod
+    def _looks_like_provider_error(text: str) -> bool:
+        lowered = str(text or "").strip().lower()
+        if not lowered:
+            return False
+        error_signals = (
+            "[error]",
+            "request was aborted",
+            "request aborted",
+            "connection aborted",
+            "operation aborted",
+            "context canceled",
+            "context cancelled",
+            "cancelled",
+            "canceled",
+            "rate limit",
+            "too many requests",
+            "429",
+            "quota",
+            "insufficient_quota",
+            "insufficient quota",
+            "api key",
+            "authentication failed",
+            "service unavailable",
+            "upstream error",
+            "upstream request",
+            "bad gateway",
+            "gateway timeout",
+            "internal server error",
+            "connection reset",
+            "econnreset",
+            "socket hang up",
+        )
+        if any(flag in lowered for flag in error_signals):
+            return True
+
+        if len(lowered) <= 160:
+            compact = lowered.replace("\n", " ").strip()
+            ascii_chars = sum(1 for ch in compact if ord(ch) < 128)
+            ascii_ratio = ascii_chars / max(1, len(compact))
+            generic_error_tokens = ("error", "failed", "abort", "timeout", "exception")
+            if ascii_ratio >= 0.9 and any(token in compact for token in generic_error_tokens):
+                return True
+
+        return False
+
+    @staticmethod
+    def _fallback_reply_text(capture_result: Dict[str, Any]) -> str:
+        status = str(capture_result.get("status") or "")
+        classification = str(capture_result.get("classification") or "")
+        if status == "captured":
+            if classification in {"project", "new_project"}:
+                return "我已收到需求并完成初步识别，已按项目级信息入队处理。你可以继续补充目标、范围和时间，我会进一步细化分派。"
+            return "我已收到需求并完成初步识别，已进入信号处理流程。若你希望我直接分派任务，请补充目标、优先级和期望时间。"
+        return "我已收到消息并开始处理。若你希望我直接分派任务，请补充目标、优先级和时间要求。"
+
+    @staticmethod
+    def _build_butler_decision_reply(capture_result: Dict[str, Any]) -> Optional[str]:
+        if not isinstance(capture_result, dict):
+            return None
+        status = str(capture_result.get("status") or "").strip().lower()
+        if status != "captured":
+            return None
+
+        results = capture_result.get("results")
+        if not isinstance(results, list) or not results:
+            return None
+
+        auto_results = [row for row in results if str((row or {}).get("route_status") or "") == "auto_processed"]
+        pending_results = [row for row in results if str((row or {}).get("route_status") or "") == "pending"]
+        classification = str(capture_result.get("classification") or "signal").strip().lower()
+        signal_id = str(capture_result.get("signal_id") or "").strip()
+        linked_project_ids = capture_result.get("linked_project_ids") or []
+        project_hint = ""
+        if isinstance(linked_project_ids, list) and linked_project_ids:
+            project_hint = f"，并关联项目 #{linked_project_ids[0]}"
+
+        if pending_results:
+            if classification in {"project", "new_project"}:
+                return (
+                    "我已先完成项目候选入队，当前有不确定项需要你确认："
+                    "1) 归属项目/是否新建 2) 优先级 3) 截止时间。"
+                    "确认后我会自动继续推进。"
+                )
+            return (
+                "我已先完成信号入队，当前判断存在不确定项（多半是项目归属或范围边界）。"
+                "请你补充：归属项目、优先级、期望完成时间，我会立即自动续跑。"
+            )
+
+        if auto_results:
+            if classification in {"project", "new_project"}:
+                return (
+                    f"已自动完成项目流处理{project_hint}。"
+                    "后续我会继续按管家流程拆解与分派，除非你要求人工接管。"
+                )
+            if signal_id:
+                return (
+                    f"已自动完成信号流处理（信号#{signal_id}{project_hint}）。"
+                    "这是明确需求场景，我会继续自动推进后续动作。"
+                )
+            return "已自动完成信号流处理。后续我会继续自动推进，只有不确定项才会向你确认。"
+
+        return None
+
+    @staticmethod
+    def _build_execution_receipt_reply(
+        *,
+        message_content: Optional[str],
+        capture_result: Dict[str, Any],
+    ) -> Optional[str]:
+        """
+        对“已执行”场景优先给简短回执，避免 LLM 输出教程式长文或二次处理建议。
+        """
+        if not isinstance(capture_result, dict):
+            return None
+        status = str(capture_result.get("status") or "").strip().lower()
+        if status not in {"captured", "ignored"}:
+            return None
+
+        if status == "ignored":
+            reason = str(capture_result.get("reason") or "").strip().lower()
+            if reason == "duplicate_filtered":
+                return "这条消息我已处理过，未重复入库。"
+            if reason in {"non_requirement_noise", "empty_or_emoji_only"}:
+                return "这条消息不是可执行需求，我没有入库。请直接说要新增什么信号或任务。"
+            return "这条消息已忽略（未入库）。如需执行，请明确目标、对象和动作。"
+
+        text = str(message_content or "").strip()
+        strict_keywords = ("单独信号源", "单个信号源", "只入库", "仅入库", "不需要二次处理")
+        strict_mode = any(k in text for k in strict_keywords)
+        primary_route_status = ""
+        results = capture_result.get("results")
+        if isinstance(results, list) and results:
+            first = results[0] if isinstance(results[0], dict) else {}
+            primary_route_status = str((first or {}).get("route_status") or "").strip().lower()
+
+        if primary_route_status == "auto_done" or primary_route_status == "auto_processed":
+            return "已自动处理完成并入库，后续动作已继续自动推进。"
+        if primary_route_status == "confirmed":
+            return "已按确认结果执行并入库。"
+        if primary_route_status in {"queued_for_decision", "pending"}:
+            clarification_message = str(capture_result.get("clarification_message") or "").strip()
+            if clarification_message:
+                return clarification_message
+            return "已完成入库，当前进入待确认队列。"
+        if primary_route_status == "dismissed":
+            return "该消息已记录，但当前标记为忽略，不会继续执行。"
+        if primary_route_status == "failed":
+            return "未执行：受控执行器拒绝了这次动作，请补充缺失字段或检查项目上下文。"
+        if primary_route_status == "undone":
+            return "上一次自动结果已撤销，当前不会继续执行。"
+
+        signal_id = str(capture_result.get("signal_id") or "").strip()
+        linked_project_ids = capture_result.get("linked_project_ids")
+        project_id: Optional[str] = None
+        if isinstance(linked_project_ids, list) and linked_project_ids:
+            project_id = str(linked_project_ids[0]).strip() or None
+        if not project_id:
+            candidate = capture_result.get("created_project_id")
+            if candidate is not None:
+                project_id = str(candidate).strip() or None
+
+        if signal_id and project_id:
+            if strict_mode:
+                return f"已按你的要求仅执行入库：信号#{signal_id}，项目#{project_id}。"
+            return f"已执行完成：信号#{signal_id} 已入库并关联项目#{project_id}。"
+        if signal_id:
+            if strict_mode:
+                return f"已按你的要求仅执行入库：信号#{signal_id}。"
+            return f"已执行完成：信号#{signal_id} 已入库。"
+        if project_id:
+            if strict_mode:
+                return f"已按你的要求执行：项目#{project_id} 已关联。"
+            return f"已执行完成：已关联项目#{project_id}。"
+        return "已执行完成。"
+
+    @staticmethod
+    def _detect_forbidden_execute_keyword(text: Optional[str]) -> Optional[str]:
+        lowered = str(text or "").strip().lower()
+        if not lowered:
+            return None
+        forbidden = (
+            "curl",
+            "调用方式",
+            "接口教程",
+            "api 调用",
+            "请求示例",
+            "postman",
+            "sdk 调用",
+        )
+        for token in forbidden:
+            if token in lowered:
+                return token
+        return None
+
+    @classmethod
+    def _contains_forbidden_execute_keywords(cls, text: Optional[str]) -> bool:
+        return cls._detect_forbidden_execute_keyword(text) is not None
+
+    def _guard_execute_reply_text_with_reason(
+        self,
+        text: Optional[str],
+        *,
+        trace_id: str,
+    ) -> tuple[Optional[str], Optional[str]]:
+        if not text:
+            return text, None
+        hit = self._detect_forbidden_execute_keyword(text)
+        if not hit:
+            return text, None
+        self._emit_state_transition(
+            trace_id=trace_id,
+            state="ROUTED",
+            route="execute_template_guard",
+            reason="forbidden_keyword_detected",
+            keyword=hit,
+        )
+        return (
+            "已执行完成并入库。如需我继续动作，请直接补充目标、对象和时限。",
+            hit,
+        )
+
+    def _guard_execute_reply_text(self, text: Optional[str], *, trace_id: str) -> Optional[str]:
+        sanitized, _ = self._guard_execute_reply_text_with_reason(text=text, trace_id=trace_id)
+        return sanitized
+
+    async def _record_execute_guard_alert(
+        self,
+        *,
+        trace_id: str,
+        message_id: str,
+        message_content: str,
+        capture_result: Dict[str, Any],
+        guard_reason: str,
+    ) -> None:
+        try:
+            async with AsyncSessionLocal() as db:
+                db.add(
+                    AIProcessingLog(
+                        tenant_id=int(self.tenant_id or 1),
+                        feishu_message_id=(message_id or "")[:128] or None,
+                        feishu_item_id=None,
+                        raw_message=(message_content or "")[:8000] or None,
+                        ai_raw_output=json.dumps(
+                            {
+                                "trace_id": trace_id,
+                                "guard_reason": guard_reason,
+                                "status": capture_result.get("status"),
+                                "classification": capture_result.get("classification"),
+                            },
+                            ensure_ascii=False,
+                        )[:120000],
+                        project_score=None,
+                        signal_confidence=None,
+                        level="execute_guard",
+                        routing_decision="execute_reply_guard_triggered",
+                        routing_reason=f"执行态回复命中禁词({guard_reason})，已切换模板模式。",
+                        reasoning_bullets=[
+                            "执行态不允许输出调用教程类文本。",
+                            "本次回复已被熔断为固定模板。",
+                            f"trace_id={trace_id}",
+                        ],
+                        processing_status="success",
+                        error_message=None,
+                        duration_ms=0,
+                    )
+                )
+                recipients = (
+                    await db.execute(
+                        select(User.id).where(
+                            and_(
+                                User.tenant_id == int(self.tenant_id or 1),
+                                User.is_active.is_(True),
+                            )
+                        ).limit(5)
+                    )
+                ).scalars().all()
+                await NotificationService(db).create_notifications(
+                    tenant_id=int(self.tenant_id or 1),
+                    user_ids=[int(uid) for uid in recipients],
+                    category="alert",
+                    title="执行态回复熔断已触发",
+                    content=(
+                        f"trace_id={trace_id}\n"
+                        f"message_id={message_id}\n"
+                        f"原因：命中禁词 {guard_reason}\n"
+                        "系统已自动切换为固定模板回复。"
+                    )[:2000],
+                    source="feishu_guard",
+                    entity_type="feishu_message",
+                    entity_id=None,
+                    target_url="/ai-logs",
+                    commit=False,
+                )
+                await db.commit()
+        except Exception as exc:
+            print(f"[Feishu] execute guard alert write failed: trace_id={trace_id}, error={exc}")
+
+    async def _record_message_trace(
+        self,
+        *,
+        trace_id: str,
+        message_id: str,
+        chat_type: str,
+        message_content: str,
+        capture_result: Dict[str, Any],
+        reply_source: str,
+    ) -> None:
+        try:
+            async with AsyncSessionLocal() as db:
+                db.add(
+                    AIProcessingLog(
+                        tenant_id=int(self.tenant_id or 1),
+                        feishu_message_id=(message_id or "")[:128] or None,
+                        feishu_item_id=None,
+                        raw_message=(str(capture_result.get("normalized_text") or message_content) or "")[:8000] or None,
+                        ai_raw_output=json.dumps(
+                            {
+                                "trace_id": trace_id,
+                                "chat_type": chat_type,
+                                "normalized_text": str(capture_result.get("normalized_text") or message_content)[:2000],
+                                "route": capture_result.get("route"),
+                                "db_mutations": capture_result.get("db_mutations") or [],
+                                "reply_source": reply_source or "none",
+                                "reply_text": str(capture_result.get("direct_reply_text") or "")[:500],
+                            },
+                            ensure_ascii=False,
+                        )[:120000],
+                        project_score=None,
+                        signal_confidence=None,
+                        level="reply_trace",
+                        routing_decision=str(capture_result.get("route") or "reply_trace")[:64],
+                        routing_reason="飞书消息 reply trace",
+                        reasoning_bullets=[
+                            f"chat_type={chat_type}",
+                            f"reply_source={reply_source or 'none'}",
+                            f"trace_id={trace_id}",
+                        ],
+                        processing_status="success",
+                        error_message=None,
+                        duration_ms=0,
+                    )
+                )
+                await db.commit()
+        except Exception as exc:
+            print(f"[Feishu] message trace write failed: trace_id={trace_id}, error={exc}")
+
+    @staticmethod
+    def _build_mention_prefix(*, open_id: Optional[str], display_name: Optional[str]) -> str:
+        user_open_id = str(open_id or "").strip()
+        if not user_open_id:
+            return ""
+        safe_name = str(display_name or "你").replace("<", "").replace(">", "").strip()[:24] or "你"
+        return f'<at user_id="{user_open_id}">{safe_name}</at> '
+
+    @staticmethod
+    def _is_duplicate_capture_result(result: Dict[str, Any]) -> bool:
+        if not isinstance(result, dict):
+            return False
+        status = str(result.get("status") or "").strip().lower()
+        if status != "ignored":
+            return False
+        reason = str(result.get("reason") or "").strip().lower()
+        classification = str(result.get("classification") or "").strip().lower()
+        return reason == "duplicate_filtered" or classification == "duplicate"
+
+    @staticmethod
+    def _should_reply_to_mention(*, mentioned_bot_names: List[str], message_content: Optional[str]) -> bool:
+        # 仅在明确 @PMFlow（而不是 @all / @其他成员）时触发自动回复
+        normalized_names = {str(x or "").strip().lower() for x in mentioned_bot_names if str(x or "").strip()}
+        if any(name in {"pmflow", "_user_1"} for name in normalized_names):
+            return True
+
+        # 兼容文本中的 @ 语法（不同消息类型返回格式不一致）
+        text = str(message_content or "").strip().lower()
+        if not text:
+            return False
+        if "@pmflow" in text or "@_user_1" in text:
+            return True
+        return False
+
+    @staticmethod
+    def _build_time_truth_reply(message_content: Optional[str]) -> Optional[str]:
+        text = str(message_content or "").strip().lower()
+        if not text:
+            return None
+        time_keywords = ("今天", "日期", "时间", "几号", "几点", "星期", "周几", "当前日期")
+        if not any(keyword in text for keyword in time_keywords):
+            return None
+
+        now_cn = datetime.now(CN_TZ)
+        today_text = (
+            f"{now_cn.year}年{now_cn.month}月{now_cn.day}日"
+            f"（{WEEKDAY_CN[now_cn.weekday()]}）{now_cn.strftime('%H:%M:%S')}"
+        )
+        if "2023年11月7日" in text or "2023-11-07" in text:
+            return (
+                f"我这边已校准为北京时间 {today_text}。"
+                "你看到“2023年11月7日”通常是客户端缓存或本机时区异常，"
+                "建议刷新会话并检查系统时区是否为 Asia/Shanghai。"
+            )
+        return (
+            f"我这边当前北京时间是 {today_text}。"
+            "如果你的界面显示不一致，优先检查设备时区和飞书客户端缓存。"
+        )
+
+    async def _resolve_sender_display_name(
+        self,
+        *,
+        sender: Dict[str, Any],
+        sender_id: Dict[str, Any],
+    ) -> str:
+        sender_name = (sender.get("sender_name") or "").strip()
+        if sender_name:
+            return sender_name
+        for key in ("name", "en_name"):
+            value = (sender.get(key) or "").strip()
+            if value:
+                return value
+
+        open_id = (sender_id.get("open_id") or "").strip()
+        if open_id:
+            remote_name = await self._fetch_user_name_by_open_id(open_id)
+            if remote_name:
+                return remote_name
+            return f"飞书用户({open_id[:8]})"
+
+        user_id = (sender_id.get("user_id") or "").strip()
+        if user_id:
+            return f"飞书用户({user_id[:8]})"
+        return "未知用户"
+
+    async def _fetch_user_name_by_open_id(self, open_id: str) -> Optional[str]:
+        safe_open_id = (open_id or "").strip()
+        if not safe_open_id:
+            return None
+        try:
+            access_token = await self._get_tenant_access_token()
+        except Exception as exc:
+            print(f"[Feishu] fetch user name skipped: token unavailable: {exc}")
+            return None
+
+        url = f"https://open.feishu.cn/open-apis/contact/v3/users/{safe_open_id}"
+        headers = {"Authorization": f"Bearer {access_token}"}
+        params = {"user_id_type": "open_id"}
+        try:
+            async with httpx.AsyncClient(timeout=10.0, trust_env=False) as client:
+                resp = await client.get(url, headers=headers, params=params)
+                resp.raise_for_status()
+                payload = resp.json() or {}
+            if payload.get("code") != 0:
+                print(
+                    "[Feishu] fetch user name failed: "
+                    f"code={payload.get('code')} msg={payload.get('msg')}"
+                )
+                return None
+            user = payload.get("data", {}).get("user", {}) if isinstance(payload.get("data", {}), dict) else {}
+            name = (user.get("name") or user.get("en_name") or "").strip()
+            return name or None
+        except Exception as exc:
+            print(f"[Feishu] fetch user name failed: {exc}")
+            return None
+
+    @staticmethod
+    def _extract_message_datetime(message: Dict[str, Any]) -> Optional[datetime]:
+        if not isinstance(message, dict):
+            return None
+        raw = message.get("create_time") or message.get("update_time")
+        if raw is None:
+            return None
+        try:
+            ts = int(str(raw).strip())
+            if ts > 10_000_000_000:
+                ts = ts / 1000.0
+            return datetime.fromtimestamp(ts, tz=timezone.utc)
+        except Exception:
+            return None
+
+    @staticmethod
+    def _extract_mentioned_bot_names(message: Dict[str, Any]) -> List[str]:
+        if not isinstance(message, dict):
+            return []
+        mentions = message.get("mentions", [])
+        names: List[str] = []
+        if isinstance(mentions, list):
+            for item in mentions:
+                if not isinstance(item, dict):
+                    continue
+                name = str(item.get("name") or "").strip()
+                if not name:
+                    continue
+                if name not in names:
+                    names.append(name)
+
+        content = message.get("content")
+        text = ""
+        if isinstance(content, dict):
+            text = str(content.get("text") or "")
+        elif isinstance(content, str):
+            try:
+                parsed = json.loads(content)
+                if isinstance(parsed, dict):
+                    text = str(parsed.get("text") or "")
+                else:
+                    text = content
+            except Exception:
+                text = content
+
+        if text:
+            for raw in re.findall(r"<at[^>]*>([^<]+)</at>", text, flags=re.IGNORECASE):
+                name = str(raw or "").strip()
+                if name and name not in names:
+                    names.append(name)
+            for raw in re.findall(r"@([^\s@]+)", text):
+                name = str(raw or "").strip()
+                if name and name not in names:
+                    names.append(name)
+        return names
+
+    @staticmethod
+    def _build_attachment_placeholder_text(attachments: List[Dict[str, Any]]) -> str:
+        if not attachments:
+            return ""
+        names: List[str] = []
+        for item in attachments[:3]:
+            attachment_type = str(item.get("attachment_type") or "附件").strip()
+            file_name = str(item.get("file_name") or "").strip()
+            names.append(file_name or attachment_type)
+        suffix = "、".join(names)
+        return f"[附件消息] {suffix}" if suffix else "[附件消息]"
+
+    @staticmethod
+    def _normalize_ingress_attachment_records(
+        items: List[Dict[str, Any]],
+        *,
+        occurred_at: Optional[str] = None,
+        source_message_id: Optional[str] = None,
+        source_chat_id: Optional[str] = None,
+    ) -> List[Dict[str, Any]]:
+        normalized: List[Dict[str, Any]] = []
+        seen_ids: set[str] = set()
+        for idx, raw in enumerate(items):
+            if not isinstance(raw, dict):
+                continue
+            attachment_type = str(raw.get("attachment_type") or raw.get("type") or "file").strip().lower()
+            if attachment_type in {"img", "picture"}:
+                attachment_type = "image"
+            if attachment_type in {"doc", "docs"}:
+                attachment_type = "document"
+            if attachment_type not in {"image", "file", "media", "document"}:
+                attachment_type = "file"
+
+            file_key = str(raw.get("file_key") or "").strip() or None
+            image_key = str(raw.get("image_key") or "").strip() or None
+            media_id = str(raw.get("media_id") or "").strip() or None
+            download_url = str(raw.get("download_url") or raw.get("url") or "").strip() or None
+            attachment_id = (
+                str(raw.get("attachment_id") or raw.get("id") or file_key or image_key or media_id or "")
+                .strip()
+            )
+            if not attachment_id:
+                attachment_id = f"attachment-{idx + 1}"
+            if not any([file_key, image_key, media_id, download_url]):
+                continue
+            if attachment_id in seen_ids:
+                continue
+            seen_ids.add(attachment_id)
+
+            file_size = raw.get("file_size") if raw.get("file_size") is not None else raw.get("size")
+            try:
+                file_size = int(str(file_size).strip()) if file_size is not None else None
+            except (TypeError, ValueError):
+                file_size = None
+            file_name = (
+                str(raw.get("file_name") or raw.get("name") or raw.get("title") or "").strip()
+                or f"{attachment_type}-{attachment_id}"
+            )
+            mime_type = str(raw.get("mime_type") or raw.get("mime") or "").strip() or None
+            normalized.append(
+                {
+                    "id": attachment_id,
+                    "attachment_id": attachment_id,
+                    "attachment_type": attachment_type,
+                    "file_name": file_name,
+                    "mime_type": mime_type,
+                    "file_size": file_size,
+                    "file_key": file_key,
+                    "image_key": image_key,
+                    "media_id": media_id,
+                    "download_url": download_url,
+                    "source": str(raw.get("source") or "feishu").strip() or "feishu",
+                    "occurred_at": str(raw.get("occurred_at") or occurred_at or "").strip() or None,
+                    "source_message_id": str(raw.get("source_message_id") or source_message_id or "").strip() or None,
+                    "source_chat_id": str(raw.get("source_chat_id") or source_chat_id or "").strip() or None,
+                }
+            )
+        return normalized
+
+    def _extract_message_payload(self, event: Dict[str, Any]) -> Dict[str, Any]:
+        """提取消息文本与附件元数据。"""
+        message = event.get("message", {})
+        if not isinstance(message, dict) and isinstance(event.get("event", {}), dict):
+            message = event.get("event", {}).get("message", {})
+        if not isinstance(message, dict):
+            return {"text": None, "attachments": [], "message_type": ""}
+
+        message_type = str(message.get("message_type") or "").strip().lower()
+        content = message.get("content", "")
+        content_json: Dict[str, Any] = {}
+        if isinstance(content, dict):
+            content_json = content
+        elif isinstance(content, str) and content.strip():
+            try:
+                parsed = json.loads(content)
+                if isinstance(parsed, dict):
+                    content_json = parsed
+            except Exception:
+                if message_type == "text":
+                    return {"text": content.strip() or None, "attachments": [], "message_type": message_type}
+
+        text_parts: List[str] = []
+        attachments: List[Dict[str, Any]] = []
+
+        if message_type == "text":
+            text = str(content_json.get("text") or "").strip()
+            return {"text": text or None, "attachments": [], "message_type": message_type}
+
+        if message_type == "post":
+            post = content_json.get("post", {}) if isinstance(content_json, dict) else {}
+            zh_cn = post.get("zh_cn", {}) if isinstance(post, dict) else {}
+            elements = zh_cn.get("content", []) if isinstance(zh_cn, dict) else []
+            if (not elements) and isinstance(content_json, dict):
+                direct_elements = content_json.get("content", [])
+                if isinstance(direct_elements, list):
+                    elements = direct_elements
+            for row in elements:
+                row_items = row if isinstance(row, list) else [row]
+                for elem in row_items:
+                    if not isinstance(elem, dict):
+                        continue
+                    tag = str(elem.get("tag") or "").strip().lower()
+                    if tag == "text":
+                        text = str(elem.get("text") or "").strip()
+                        if text:
+                            text_parts.append(text)
+                    elif tag in {"img", "image"}:
+                        attachments.append(
+                            {
+                                "attachment_type": "image",
+                                "image_key": elem.get("image_key"),
+                                "file_name": elem.get("title") or "图片附件",
+                            }
+                        )
+                    elif tag in {"file", "media"}:
+                        attachments.append(
+                            {
+                                "attachment_type": "media" if tag == "media" else "file",
+                                "file_key": elem.get("file_key"),
+                                "media_id": elem.get("media_id"),
+                                "file_name": elem.get("file_name") or elem.get("title"),
+                                "file_size": elem.get("file_size"),
+                            }
+                        )
+        elif message_type in {"image", "file", "media"}:
+            attachments.append(
+                {
+                    "attachment_type": message_type,
+                    "image_key": content_json.get("image_key"),
+                    "file_key": content_json.get("file_key"),
+                    "media_id": content_json.get("media_id"),
+                    "file_name": content_json.get("file_name") or content_json.get("title"),
+                    "file_size": content_json.get("file_size"),
+                    "mime_type": content_json.get("file_type") or content_json.get("mime_type"),
+                }
+            )
+            text = str(content_json.get("text") or content_json.get("caption") or "").strip()
+            if text:
+                text_parts.append(text)
+        elif isinstance(content_json, dict):
+            guessed_keys = {
+                "image_key": content_json.get("image_key"),
+                "file_key": content_json.get("file_key"),
+                "media_id": content_json.get("media_id"),
+                "download_url": content_json.get("download_url"),
+            }
+            if any(v for v in guessed_keys.values()):
+                attachments.append(
+                    {
+                        "attachment_type": "file",
+                        "image_key": guessed_keys.get("image_key"),
+                        "file_key": guessed_keys.get("file_key"),
+                        "media_id": guessed_keys.get("media_id"),
+                        "download_url": guessed_keys.get("download_url"),
+                        "file_name": content_json.get("file_name") or content_json.get("title"),
+                        "file_size": content_json.get("file_size"),
+                        "mime_type": content_json.get("mime_type"),
+                    }
+                )
+            text = str(content_json.get("text") or "").strip()
+            if text:
+                text_parts.append(text)
+
+        normalized_attachments = self._normalize_ingress_attachment_records(attachments)
+        merged_text = " ".join([part for part in text_parts if part]).strip()
+        if not merged_text and normalized_attachments:
+            merged_text = self._build_attachment_placeholder_text(normalized_attachments)
+        return {
+            "text": merged_text or None,
+            "attachments": normalized_attachments,
+            "message_type": message_type,
+        }
+
+    def _extract_message_content(self, event: Dict[str, Any]) -> Optional[str]:
+        payload = self._extract_message_payload(event)
+        text = str(payload.get("text") or "").strip()
+        return text or None
+
+    async def _persist_signal_attachments_from_ingress(
+        self,
+        *,
+        signal_id: Any,
+        attachments: List[Dict[str, Any]],
+    ) -> None:
+        if not attachments:
+            return
+        try:
+            signal_int = int(str(signal_id).strip())
+        except (TypeError, ValueError, AttributeError):
+            return
+        signal = (
+            await self.db.execute(
+                select(Signal).where(
+                    and_(
+                        Signal.tenant_id == int(self.tenant_id or 1),
+                        Signal.id == signal_int,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if signal is None:
+            return
+
+        try:
+            payload = signal.strategic_analysis
+            strategic: Dict[str, Any] = {}
+            if isinstance(payload, dict):
+                strategic = dict(payload)
+            elif isinstance(payload, str) and payload.strip():
+                parsed = json.loads(payload)
+                if isinstance(parsed, dict):
+                    strategic = parsed
+
+            source_context = strategic.get("source_context")
+            if not isinstance(source_context, dict):
+                source_context = {}
+            existing = source_context.get("attachments")
+            existing_rows = existing if isinstance(existing, list) else []
+            merged = self._normalize_ingress_attachment_records(
+                [*(row for row in existing_rows if isinstance(row, dict)), *attachments]
+            )
+            if not merged:
+                return
+            source_context["attachments"] = merged
+            strategic["source_context"] = source_context
+            signal.strategic_analysis = json.dumps(strategic, ensure_ascii=False)
+            await self.db.commit()
+        except Exception as exc:
+            print(f"[Feishu] persist signal attachments failed: signal_id={signal_int}, error={exc}")
+            try:
+                await self.db.rollback()
+            except Exception:
+                pass
+
+    async def send_message_to_group(
+        self,
+        chat_id: str,
+        message: str
+    ) -> Dict[str, Any]:
+        """
+        发送消息到飞书群聊
+
+        用于推送通知、简报等
+        """
+        if not self.is_configured():
+            raise ValueError("飞书未配置")
+
+        # 获取 tenant_access_token
+        access_token = await self._get_tenant_access_token()
+
+        # 发送消息 API
+        url = "https://open.feishu.cn/open-apis/im/v1/messages"
+        headers = {
+            'Authorization': f'Bearer {access_token}',
+            'Content-Type': 'application/json'
+        }
+
+        message_data = {
+            'receive_id': chat_id,
+            'msg_type': 'text',
+            'content': json.dumps({'text': message})
+        }
+
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False) as client:
+            resp = await client.post(
+                url,
+                headers=headers,
+                params={'receive_id_type': 'chat_id'},
+                json=message_data,
+            )
+            payload: Dict[str, Any]
+            try:
+                payload = resp.json()
+            except Exception:
+                payload = {"http_status": resp.status_code, "raw": resp.text[:500]}
+            if resp.status_code >= 400:
+                raise ValueError(
+                    f"飞书发送失败(http={resp.status_code}): {payload}"
+                )
+            if int(payload.get("code", -1)) != 0:
+                raise ValueError(
+                    f"飞书发送失败: code={payload.get('code')} msg={payload.get('msg') or payload.get('message')}"
+                )
+            return payload
+
+    async def _get_tenant_access_token(self) -> str:
+        """获取 tenant_access_token"""
+        if not self.app_id or not self.app_secret:
+            raise ValueError("飞书未配置")
+
+        url = "https://open.feishu.cn/open-apis/auth/v3/tenant_access_token/internal"
+        headers = {'Content-Type': 'application/json'}
+        data = {
+            'app_id': self.app_id,
+            'app_secret': self.app_secret
+        }
+
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False) as client:
+            try:
+                resp = await client.post(url, headers=headers, json=data)
+            except httpx.RequestError as exc:
+                raise ValueError(f"访问飞书 auth 接口失败: {exc}") from exc
+
+            try:
+                result = resp.json()
+            except Exception:
+                result = {"http_status": resp.status_code, "raw": resp.text[:500]}
+
+            if resp.status_code >= 400:
+                raise ValueError(
+                    f"获取 tenant_access_token 失败(http={resp.status_code}): {result}"
+                )
+
+            if result.get('code') == 0 and result.get('tenant_access_token'):
+                return result.get('tenant_access_token')
+
+            raise ValueError(
+                "获取 tenant_access_token 失败: "
+                f"code={result.get('code')} msg={result.get('msg') or result.get('message')} payload={result}"
+            )
+
+    async def download_attachment_bytes(self, attachment: Dict[str, Any]) -> Dict[str, Any]:
+        """通过飞书开放平台代理下载附件。"""
+        if not self.is_configured():
+            raise ValueError("飞书未配置")
+        access_token = await self._get_tenant_access_token()
+        attachment_type = str(attachment.get("attachment_type") or "file").strip().lower()
+        file_name = str(attachment.get("file_name") or "").strip() or None
+        media_type = str(attachment.get("mime_type") or "").strip() or None
+
+        image_key = str(attachment.get("image_key") or "").strip()
+        file_key = str(attachment.get("file_key") or "").strip()
+        media_id = str(attachment.get("media_id") or "").strip()
+        download_url = str(attachment.get("download_url") or "").strip()
+
+        url = ""
+        if download_url:
+            url = download_url
+        elif attachment_type == "image" and image_key:
+            url = f"https://open.feishu.cn/open-apis/im/v1/images/{image_key}"
+        elif file_key:
+            url = f"https://open.feishu.cn/open-apis/im/v1/files/{file_key}"
+        elif media_id:
+            url = f"https://open.feishu.cn/open-apis/im/v1/files/{media_id}"
+
+        if not url:
+            raise ValueError("附件缺少可下载 key，无法代理下载")
+
+        headers = {"Authorization": f"Bearer {access_token}"}
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False) as client:
+            try:
+                resp = await client.get(url, headers=headers, follow_redirects=True)
+            except httpx.RequestError as exc:
+                raise ValueError(f"飞书附件下载请求失败: {exc}") from exc
+
+        if resp.status_code >= 400:
+            body_preview = resp.text[:300] if resp.text else ""
+            raise ValueError(
+                f"飞书附件下载失败(http={resp.status_code}): {body_preview}"
+            )
+
+        content = resp.content or b""
+        if not content:
+            raise ValueError("飞书附件下载为空内容")
+
+        response_content_type = str(resp.headers.get("content-type") or "").strip()
+        content_type = media_type or response_content_type or "application/octet-stream"
+        return {
+            "content": content,
+            "content_type": content_type,
+            "file_name": file_name,
+        }
+
+    async def send_rich_card(
+        self,
+        chat_id: str,
+        title: str,
+        content: List[Dict[str, Any]]
+    ) -> Dict[str, Any]:
+        """
+        发送富文本卡片消息
+
+        用于发送格式化的简报、通知等
+        """
+        if not self.is_configured():
+            raise ValueError("飞书未配置")
+
+        access_token = await self._get_tenant_access_token()
+
+        url = "https://open.feishu.cn/open-apis/im/v1/messages"
+        headers = {
+            'Authorization': f'Bearer {access_token}',
+            'Content-Type': 'application/json'
+        }
+
+        # 构造卡片内容
+        card_content = {
+            'config': {'wide_screen_mode': True},
+            'header': {
+                'title': {
+                    'content': title,
+                    'tag': 'plain_text'
+                }
+            },
+            'elements': []
+        }
+
+        for item in content:
+            if item.get('type') == 'text':
+                card_content['elements'].append({
+                    'tag': 'div',
+                    'text': {
+                        'content': item.get('text', ''),
+                        'tag': 'lark_md'
+                    }
+                })
+            elif item.get('type') == 'divider':
+                card_content['elements'].append({'tag': 'hr'})
+
+        message_data = {
+            'receive_id': chat_id,
+            'msg_type': 'interactive',
+            'content': json.dumps(card_content)
+        }
+
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False) as client:
+            resp = await client.post(
+                url,
+                headers=headers,
+                params={'receive_id_type': 'chat_id'},
+                json=message_data,
+            )
+            payload: Dict[str, Any]
+            try:
+                payload = resp.json()
+            except Exception:
+                payload = {"http_status": resp.status_code, "raw": resp.text[:500]}
+            if resp.status_code >= 400:
+                raise ValueError(
+                    f"飞书发送失败(http={resp.status_code}): {payload}"
+                )
+            if int(payload.get("code", -1)) != 0:
+                raise ValueError(
+                    f"飞书发送失败: code={payload.get('code')} msg={payload.get('msg') or payload.get('message')}"
+                )
+            return payload
+
+
+# 全局服务实例
+_feishu_service: Optional[FeishuIntegrationService] = None
+
+
+async def dispatch_pending_feishu_replies(limit: int = 30) -> int:
+    """后台补偿发送：扫描 pending/retrying outbox 并触发派发。"""
+    return await FeishuIntegrationService.dispatch_pending_outbox_entries(limit=limit)
+
+
+def get_feishu_service(db: AsyncSession) -> FeishuIntegrationService:
+    """获取飞书服务实例"""
+    return FeishuIntegrationService(db)
diff --git a/backend/app/services/feishu_long_connection.py b/backend/app/services/feishu_long_connection.py
new file mode 100644
index 00000000..44e50544
--- /dev/null
+++ b/backend/app/services/feishu_long_connection.py
@@ -0,0 +1,546 @@
+"""Feishu 长连接事件接收管理器。"""
+
+from __future__ import annotations
+
+import asyncio
+from datetime import datetime, timezone
+import importlib.util
+import json
+import os
+import threading
+import time
+from typing import Any, Optional
+
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+from sqlalchemy.future import select
+
+from ..core.logging import get_logger
+from ..models import Integration
+from .feishu_integration import FeishuIntegrationService
+
+logger = get_logger(__name__)
+DEFAULT_FEISHU_LONG_CONNECTION_RETRY_MAX_ATTEMPTS = 3
+DEFAULT_FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS = 1.0
+
+
+class FeishuLongConnectionManager:
+    """管理飞书长连接 SDK 生命周期与事件消费。"""
+
+    def __init__(self, session_factory: async_sessionmaker[AsyncSession]) -> None:
+        self._session_factory = session_factory
+        self._loop: Optional[asyncio.AbstractEventLoop] = None
+        self._queue: asyncio.Queue[dict[str, Any]] = asyncio.Queue()
+        self._worker_task: Optional[asyncio.Task[None]] = None
+        self._supervisor_task: Optional[asyncio.Task[None]] = None
+        self._thread: Optional[threading.Thread] = None
+        self._client: Any = None
+        self._boot_lock = asyncio.Lock()
+
+        self._running = False
+        self._sdk_available = self._check_sdk_available()
+        self._configured = False
+        self._event_mode = "long_connection"
+        self._tenant_id = 1
+        self._last_error: Optional[str] = None
+        self._last_error_code: Optional[str] = None
+        self._last_error_retryable = False
+        self._last_event_at: Optional[str] = None
+        self._last_auth_at: Optional[str] = None
+        self._last_connect_at: Optional[str] = None
+        self._last_disconnect_at: Optional[str] = None
+        self._restart_attempts = 0
+        self._next_retry_at = 0.0
+        self._stopping = False
+        self._max_retry_attempts = self._read_retry_max_attempts()
+        self._retry_base_seconds = self._read_retry_base_seconds()
+
+    async def start(self) -> None:
+        """启动长连接（若未配置或未安装 SDK 则仅记录状态）。"""
+        if self._running:
+            return
+
+        self._stopping = False
+        self._running = True
+        self._loop = asyncio.get_running_loop()
+        self._worker_task = asyncio.create_task(
+            self._event_worker(),
+            name="feishu-long-connection-worker",
+        )
+        self._supervisor_task = asyncio.create_task(
+            self._connection_supervisor(),
+            name="feishu-long-connection-supervisor",
+        )
+
+        started = await self._ensure_connection(reason="startup")
+        if started:
+            logger.info("feishu_long_connection_started", tenant_id=self._tenant_id, event_mode=self._event_mode)
+
+    async def stop(self) -> None:
+        """停止长连接。"""
+        self._stopping = True
+        self._running = False
+
+        current_client = self._client
+        if current_client is not None:
+            try:
+                await asyncio.to_thread(self._stop_client_blocking, current_client)
+            except Exception as exc:  # pragma: no cover
+                logger.warning(f"feishu_long_connection_stop_failed: {exc}")
+
+        self._client = None
+
+        if self._supervisor_task:
+            self._supervisor_task.cancel()
+            try:
+                await self._supervisor_task
+            except asyncio.CancelledError:
+                pass
+            self._supervisor_task = None
+
+        if self._worker_task:
+            self._worker_task.cancel()
+            try:
+                await self._worker_task
+            except asyncio.CancelledError:
+                pass
+            self._worker_task = None
+
+        if self._thread and self._thread.is_alive():
+            await asyncio.to_thread(self._thread.join, 3)
+        self._thread = None
+
+        while not self._queue.empty():
+            try:
+                self._queue.get_nowait()
+                self._queue.task_done()
+            except asyncio.QueueEmpty:
+                break
+
+        logger.info("feishu_long_connection_stopped")
+        self._stopping = False
+
+    async def restart(self) -> None:
+        """重启长连接（用于配置变更后重连）。"""
+        await self.stop()
+        await self.start()
+
+    def status(self) -> dict[str, Any]:
+        """返回当前运行状态，供 API 展示。"""
+        return {
+            "running": self._running,
+            "tenant_id": self._tenant_id,
+            "event_mode": self._event_mode,
+            "sdk_available": self._sdk_available,
+            "configured": self._configured,
+            "worker_alive": bool(self._worker_task and not self._worker_task.done()),
+            "supervisor_alive": bool(self._supervisor_task and not self._supervisor_task.done()),
+            "thread_alive": bool(self._thread and self._thread.is_alive()),
+            "last_error": self._last_error,
+            "last_error_code": self._last_error_code,
+            "last_error_retryable": self._last_error_retryable,
+            "last_event_at": self._last_event_at,
+            "last_auth_at": self._last_auth_at,
+            "last_connect_at": self._last_connect_at,
+            "last_disconnect_at": self._last_disconnect_at,
+            "restart_attempts": self._restart_attempts,
+            "next_retry_at": (
+                datetime.fromtimestamp(self._next_retry_at, tz=timezone.utc).isoformat()
+                if self._next_retry_at > 0
+                else None
+            ),
+            "retry_policy": {
+                "max_attempts": self._max_retry_attempts,
+                "base_seconds": self._retry_base_seconds,
+            },
+        }
+
+    def _check_sdk_available(self) -> bool:
+        # 只做可用性探测，避免在主线程提前 import lark_oapi，
+        # 否则 SDK 会把全局 event loop 绑定到 uvicorn loop，导致线程内 run_until_complete 报错。
+        return importlib.util.find_spec("lark_oapi") is not None
+
+    @staticmethod
+    def _read_retry_max_attempts() -> int:
+        raw = os.getenv(
+            "FEISHU_LONG_CONNECTION_RETRY_MAX_ATTEMPTS",
+            str(DEFAULT_FEISHU_LONG_CONNECTION_RETRY_MAX_ATTEMPTS),
+        )
+        try:
+            parsed = int(str(raw).strip())
+        except Exception:
+            parsed = DEFAULT_FEISHU_LONG_CONNECTION_RETRY_MAX_ATTEMPTS
+        # 强制最多 3 次，避免无限重连雪崩。
+        return min(3, max(1, parsed))
+
+    @staticmethod
+    def _read_retry_base_seconds() -> float:
+        raw = os.getenv(
+            "FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS",
+            str(DEFAULT_FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS),
+        )
+        try:
+            parsed = float(str(raw).strip())
+        except Exception:
+            parsed = DEFAULT_FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS
+        if parsed <= 0:
+            parsed = DEFAULT_FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS
+        return parsed
+
+    @staticmethod
+    def _stop_client_blocking(client: Any) -> None:
+        stop_fn = getattr(client, "stop", None)
+        if callable(stop_fn):
+            stop_fn()
+
+    async def _load_active_config(self) -> Optional[dict[str, Any]]:
+        async with self._session_factory() as db:
+            result = await db.execute(
+                select(Integration).where(
+                    Integration.type == "feishu",
+                    Integration.is_active.is_(True),
+                ).order_by(Integration.updated_at.desc(), Integration.id.desc())
+            )
+            integration = result.scalars().first()
+            if not integration:
+                return None
+
+            config = integration.config
+            parsed: dict[str, Any]
+            if isinstance(config, str):
+                try:
+                    parsed = json.loads(config) if config else {}
+                except Exception:
+                    parsed = {}
+            else:
+                parsed = config or {}
+            parsed["tenant_id"] = int(getattr(integration, "tenant_id", 1) or 1)
+            return parsed
+
+    async def _ensure_connection(self, *, reason: str) -> bool:
+        async with self._boot_lock:
+            if not self._running:
+                return False
+            if self._thread and self._thread.is_alive():
+                return True
+            if self._client is not None:
+                try:
+                    await asyncio.to_thread(self._stop_client_blocking, self._client)
+                except Exception as exc:  # pragma: no cover
+                    logger.warning(f"feishu_long_connection_stale_client_stop_failed: {exc}")
+                finally:
+                    self._client = None
+
+            config = await self._load_active_config()
+            self._event_mode = "long_connection"
+            self._tenant_id = int((config or {}).get("tenant_id") or 1)
+            app_id = str((config or {}).get("app_id") or "").strip()
+            app_secret = str((config or {}).get("app_secret") or "").strip()
+            self._configured = bool(app_id and app_secret)
+
+            if not self._configured:
+                self._clear_error()
+                logger.info("feishu_long_connection_not_configured")
+                return False
+
+            if not self._sdk_available:
+                self._set_error("lark_oapi not installed", code="sdk_missing", retryable=False)
+                logger.warning("feishu_long_connection_sdk_missing: pip install lark-oapi")
+                return False
+
+            try:
+                await self._preflight_auth(tenant_id=self._tenant_id)
+                self._last_auth_at = datetime.now(timezone.utc).isoformat()
+            except Exception as exc:
+                code, retryable = self._classify_error(exc)
+                self._set_error(str(exc), code=code, retryable=retryable)
+                if retryable:
+                    self._schedule_retry()
+                logger.warning(
+                    "feishu_long_connection_preflight_failed",
+                    tenant_id=self._tenant_id,
+                    error_code=code,
+                    retryable=retryable,
+                    reason=reason,
+                    error=str(exc),
+                )
+                return False
+
+            self._clear_error()
+            self._last_connect_at = datetime.now(timezone.utc).isoformat()
+            self._last_disconnect_at = None
+            self._restart_attempts = 0
+            self._next_retry_at = 0.0
+            self._start_ws_thread(app_id, app_secret)
+            return True
+
+    async def _preflight_auth(self, *, tenant_id: int) -> None:
+        async with self._session_factory() as db:
+            service = FeishuIntegrationService(db)
+            await service.load_config(tenant_id=tenant_id)
+            await service._get_tenant_access_token()
+
+    async def _connection_supervisor(self) -> None:
+        while self._running:
+            try:
+                await asyncio.sleep(3)
+            except asyncio.CancelledError:
+                break
+
+            if not self._running or self._stopping:
+                continue
+            if not self._configured or self._event_mode != "long_connection":
+                continue
+            if not self._sdk_available:
+                continue
+            if self._thread and self._thread.is_alive():
+                continue
+            if self._last_error_code and not self._last_error_retryable:
+                continue
+
+            now = time.time()
+            if self._next_retry_at > now:
+                continue
+
+            await self._ensure_connection(reason="supervisor")
+
+    @staticmethod
+    def _classify_error(exc: Exception) -> tuple[str, bool]:
+        text = str(exc or "").strip().lower()
+        if "401" in text or "403" in text or "unauthorized" in text or "invalid app" in text:
+            return "auth_failed", False
+        if any(marker in text for marker in ("timed out", "timeout", "econnreset", "connection reset", "network", "dns", "temporary failure")):
+            return "network_error", True
+        if "certificate" in text or "ssl" in text or "tls" in text:
+            return "network_error", True
+        return "client_error", True
+
+    def _set_error(self, message: str, *, code: str, retryable: bool) -> None:
+        self._last_error = str(message or "").strip() or None
+        self._last_error_code = code
+        self._last_error_retryable = retryable
+
+    def _clear_error(self) -> None:
+        self._last_error = None
+        self._last_error_code = None
+        self._last_error_retryable = False
+
+    def _schedule_retry(self) -> None:
+        if self._restart_attempts >= self._max_retry_attempts:
+            self._last_error_retryable = False
+            self._next_retry_at = 0.0
+            logger.error(
+                "feishu_long_connection_retry_exhausted",
+                tenant_id=self._tenant_id,
+                attempts=self._restart_attempts,
+                max_attempts=self._max_retry_attempts,
+                last_error_code=self._last_error_code,
+                last_error=self._last_error,
+            )
+            return
+        self._restart_attempts += 1
+        backoff = self._retry_base_seconds * (2 ** max(0, self._restart_attempts - 1))
+        self._next_retry_at = time.time() + backoff
+
+    def _start_ws_thread(self, app_id: str, app_secret: str) -> None:
+        self._thread = threading.Thread(
+            target=self._run_ws_client,
+            args=(app_id, app_secret),
+            name="feishu-long-connection",
+            daemon=True,
+        )
+        self._thread.start()
+
+    def _run_ws_client(self, app_id: str, app_secret: str) -> None:
+        """在独立线程中运行 SDK 客户端（SDK start 是阻塞调用）。"""
+        try:
+            import lark_oapi as lark  # type: ignore
+
+            builder = lark.EventDispatcherHandler.builder("", "")
+            register_receive = getattr(builder, "register_p2_im_message_receive_v1", None)
+            if not callable(register_receive):
+                raise RuntimeError("当前 lark_oapi 版本不支持 register_p2_im_message_receive_v1")
+            builder = register_receive(self._on_message_event)
+
+            # 明确注册高频非业务事件，避免 SDK 抛 processor not found 干扰可观测性。
+            register_read = getattr(builder, "register_p2_im_message_message_read_v1", None)
+            if callable(register_read):
+                builder = register_read(self._on_ignored_event)
+            register_reaction_create = getattr(builder, "register_p2_im_message_reaction_created_v1", None)
+            if callable(register_reaction_create):
+                builder = register_reaction_create(self._on_ignored_event)
+            register_reaction_delete = getattr(builder, "register_p2_im_message_reaction_deleted_v1", None)
+            if callable(register_reaction_delete):
+                builder = register_reaction_delete(self._on_ignored_event)
+            register_recalled = getattr(builder, "register_p2_im_message_recalled_v1", None)
+            if callable(register_recalled):
+                builder = register_recalled(self._on_ignored_event)
+
+            event_handler = builder.build()
+            self._client = lark.ws.Client(
+                app_id,
+                app_secret,
+                event_handler=event_handler,
+                log_level=lark.LogLevel.INFO,
+            )
+            self._client.start()
+        except Exception as exc:  # pragma: no cover
+            code, retryable = self._classify_error(exc)
+            self._set_error(str(exc), code=code, retryable=retryable)
+            if retryable and self._running and not self._stopping:
+                self._schedule_retry()
+            logger.error(
+                "feishu_long_connection_client_failed",
+                tenant_id=self._tenant_id,
+                error_code=code,
+                retryable=retryable,
+                error=str(exc),
+            )
+        finally:
+            if self._client is not None:
+                try:
+                    self._stop_client_blocking(self._client)
+                except Exception as exc:  # pragma: no cover
+                    logger.warning(f"feishu_long_connection_client_cleanup_failed: {exc}")
+            self._client = None
+            if self._running and not self._stopping:
+                self._last_disconnect_at = datetime.now(timezone.utc).isoformat()
+                if not self._last_error_code:
+                    self._set_error("feishu long connection thread exited", code="thread_exited", retryable=True)
+                    self._schedule_retry()
+                    logger.warning("feishu_long_connection_thread_exited", tenant_id=self._tenant_id)
+            self._thread = None
+
+    def _on_message_event(self, data: Any) -> None:
+        """SDK 回调：将事件安全投递到主事件循环。"""
+        if not self._running or not self._loop:
+            return
+
+        try:
+            payload = self._to_dict(data)
+            message_id = ""
+            event_type = ""
+            try:
+                event = payload.get("event", {}) if isinstance(payload, dict) else {}
+                header = payload.get("header", {}) if isinstance(payload, dict) else {}
+                message = event.get("message", {}) if isinstance(event, dict) else {}
+                message_id = str(message.get("message_id") or "").strip()
+                event_type = str(
+                    (header.get("event_type") if isinstance(header, dict) else "")
+                    or payload.get("event_type")
+                    or payload.get("type")
+                    or ""
+                ).strip()
+            except Exception:
+                pass
+            logger.info(
+                "feishu_long_connection_event_received: "
+                f"event_type={event_type or 'unknown'}, message_id={message_id or 'unknown'}"
+            )
+            self._loop.call_soon_threadsafe(self._queue.put_nowait, payload)
+        except Exception as exc:  # pragma: no cover
+            self._set_error(str(exc), code="enqueue_error", retryable=True)
+            logger.warning(f"feishu_long_connection_enqueue_failed: {exc}")
+
+    def _on_ignored_event(self, data: Any) -> None:
+        """显式忽略非业务事件，避免 SDK 报 processor not found。"""
+        if not self._running:
+            return
+        payload = self._to_dict(data)
+        event_type = ""
+        try:
+            header = payload.get("header", {}) if isinstance(payload, dict) else {}
+            event_type = str(
+                (header.get("event_type") if isinstance(header, dict) else "")
+                or payload.get("event_type")
+                or payload.get("type")
+                or ""
+            ).strip()
+        except Exception:
+            pass
+        logger.info(f"feishu_long_connection_event_ignored: event_type={event_type or 'unknown'}")
+
+    async def _event_worker(self) -> None:
+        while self._running:
+            try:
+                payload = await self._queue.get()
+            except asyncio.CancelledError:
+                break
+
+            try:
+                await self._process_payload(payload)
+                self._last_event_at = datetime.now(timezone.utc).isoformat()
+            except Exception as exc:
+                self._set_error(str(exc), code="event_process_failed", retryable=True)
+                logger.warning(f"feishu_long_connection_process_failed: {exc}")
+            finally:
+                self._queue.task_done()
+
+    async def _process_payload(self, payload: dict[str, Any]) -> None:
+        event_data = self._extract_event_data(payload)
+        if not event_data:
+            logger.info("feishu_long_connection_event_ignored: empty_event")
+            return
+
+        event_type = str(event_data.get("type") or "").strip()
+        if event_type and event_type not in {"im.message.receive_v1", "message.receive_v1"}:
+            logger.info(f"feishu_long_connection_event_skipped: event_type={event_type}")
+            return
+
+        async with self._session_factory() as db:
+            service = FeishuIntegrationService(db)
+            await service.load_config(tenant_id=self._tenant_id)
+            await service.handle_message_event(event_data)
+
+    def _extract_event_data(self, payload: dict[str, Any]) -> dict[str, Any]:
+        event_data = payload.get("event")
+        if not isinstance(event_data, dict):
+            event_data = {}
+
+        header = payload.get("header")
+        event_type: Optional[str] = None
+        if isinstance(header, dict):
+            event_type = header.get("event_type")
+        if not event_type:
+            event_type = payload.get("event_type") or payload.get("type")
+
+        if event_type:
+            event_data = {**event_data, "type": event_type}
+
+        if event_data:
+            event_data = {**event_data, "_ingress_source": "long_connection"}
+
+        return event_data
+
+    def _to_dict(self, value: Any) -> Any:
+        if value is None or isinstance(value, (str, int, float, bool)):
+            return value
+        if isinstance(value, list):
+            return [self._to_dict(item) for item in value]
+        if isinstance(value, dict):
+            return {k: self._to_dict(v) for k, v in value.items()}
+        to_dict = getattr(value, "to_dict", None)
+        if callable(to_dict):
+            return self._to_dict(to_dict())
+        if hasattr(value, "__dict__"):
+            return {
+                k: self._to_dict(v)
+                for k, v in value.__dict__.items()
+                if not k.startswith("_")
+            }
+        return str(value)
+
+
+_manager: Optional[FeishuLongConnectionManager] = None
+
+
+def init_feishu_long_connection_manager(
+    session_factory: async_sessionmaker[AsyncSession],
+) -> FeishuLongConnectionManager:
+    global _manager
+    if _manager is None:
+        _manager = FeishuLongConnectionManager(session_factory)
+    return _manager
+
+
+def get_feishu_long_connection_manager() -> Optional[FeishuLongConnectionManager]:
+    return _manager
diff --git a/backend/app/services/feishu_signal_pipeline.py b/backend/app/services/feishu_signal_pipeline.py
new file mode 100644
index 00000000..407608eb
--- /dev/null
+++ b/backend/app/services/feishu_signal_pipeline.py
@@ -0,0 +1,5986 @@
+"""飞书消息分流与决策队列服务。"""
+from __future__ import annotations
+
+import json
+import hashlib
+import re
+import time
+from datetime import datetime, timedelta, timezone
+from typing import Any, Literal, Optional
+
+from pydantic import BaseModel, Field, ValidationError
+from sqlalchemy import and_, delete, func, or_, select, text
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import (
+    PROMPT_FEISHU_LEVEL_CLASSIFICATION,
+    PROMPT_FEISHU_PROJECT_PREFILL,
+    PROMPT_FEISHU_SIGNAL_EXTRACTION,
+)
+from ..core.logging import get_logger
+from ..models import (
+    FeishuItem,
+    AIProcessingLog,
+    FeishuReplyOutbox,
+    Integration,
+    Project,
+    ProjectSignalLink,
+    ProjectSuggestion,
+    Signal,
+    SignalEdit,
+    User,
+)
+from .llm import LLMService
+from .adaptive_rag_service import AdaptiveRAGService
+from .notification_service import NotificationService
+from .openclaw_commander_service import OpenClawCommanderService, OperationExecutorService
+from ..utils.project_members import ensure_project_default_members
+from ..utils.sequence import ensure_project_id_sequence
+
+logger = get_logger(__name__)
+
+PROJECT_PRIORITY_ORDER = {
+    "P0": 4,
+    "P1": 3,
+    "P2": 2,
+    "P3": 1,
+    "high": 3,
+    "medium": 2,
+    "low": 1,
+}
+
+SIGNAL_TYPE_OPTIONS = {"竞品动态", "用户反馈", "项目风险", "新需求"}
+PROJECT_INTENT_RE = re.compile(r"(立项|新项目|我们要做|需要开发|启动|规划)", re.IGNORECASE)
+SIGNAL_FACT_RE = re.compile(r"(竞品|上线|发布|反馈|风险|bug|故障|告警|投诉)", re.IGNORECASE)
+TIMELINE_RE = re.compile(r"(\d+\s*(天|周|月)|本周|本月|月底|下周|下月|Q[1-4]|\d{1,2}月\d{1,2}日)")
+OWNER_RE = re.compile(r"(由|负责人|owner|owner:|@)(\S+)")
+INDEPENDENT_REQUIREMENT_RE = re.compile(
+    r"(独立需求|单独需求|单开|另起|新增需求|新建项目|新增项目|单独立项|开个新|做个新)",
+    re.IGNORECASE,
+)
+SIMPLE_SIGNAL_REQUEST_RE = re.compile(
+    r"(新增|增加|加一个|加个|补充|支持|实现|接入|优化|修复|改进).{0,18}(功能|能力|接口|页面|按钮|流程|识别|提醒|校验|报表|字段|规则|模块)",
+    re.IGNORECASE,
+)
+PROJECT_COMPLEXITY_RE = re.compile(
+    r"(立项|新项目|项目章程|里程碑|预算|排期|roadmap|跨团队|跨部门|季度目标|年度目标|okr)",
+    re.IGNORECASE,
+)
+QUESTION_REQUIREMENT_VERB_RE = re.compile(
+    r"(新增|增加|加一个|加个|支持|接入|实现|优化|修复|改进|调整|补充|完善|同步|对齐|落库|入库)",
+    re.IGNORECASE,
+)
+QUESTION_REQUIREMENT_OBJECT_RE = re.compile(
+    r"(功能|能力|接口|页面|按钮|流程|校验|报表|字段|规则|模块|任务|项目|需求|通知|提醒|配置|归属|队列|信号)",
+    re.IGNORECASE,
+)
+PROJECT_LIST_QUERY_RE = re.compile(
+    r"(现在|当前|目前|系统里|系统中|项目里|项目中)?.{0,4}(有哪些项目|项目列表|项目清单|哪(些|几个)项目)",
+    re.IGNORECASE,
+)
+PROJECT_ID_QUERY_RE = re.compile(
+    r"项目\s*id\s*[:：]?\s*(\d+).{0,12}(有吗|存在吗|还在吗|是否存在|是不是)",
+    re.IGNORECASE,
+)
+OPENCLAW_PROJECT_CREATE_RE = re.compile(
+    r"(创建|新建|建立|开个|开一个|建个|建一个|做个|做一个).{0,6}项目",
+    re.IGNORECASE,
+)
+OPENCLAW_COMMAND_VERB_RE = re.compile(
+    r"(创建|新建|建立|开个|开一个|建个|建一个|做个|做一个|修改|更新|改成|改为|设为|设成|分配|指派|归档|重开|重新打开|恢复|标记完成|完成|提醒|通知|调整工时|调整工作量)",
+    re.IGNORECASE,
+)
+OPENCLAW_COMMAND_OBJECT_RE = re.compile(
+    r"(项目|任务|负责人|优先级|状态|通知|提醒|工时|工作量)",
+    re.IGNORECASE,
+)
+FOLLOWUP_CONTINUE_RE = re.compile(r"(继续处理|继续推进|继续执行|接着处理|继续跑)", re.IGNORECASE)
+FOLLOWUP_CREATE_CONTINUE_RE = re.compile(r"(继续创建|接着创建|继续建|接着建)", re.IGNORECASE)
+FOLLOWUP_STATUS_RE = re.compile(
+    r"(进度|状态|处理到哪|处理了吗|现在怎么样|怎么样了|结果呢|是否已经创建|已经创建了吗|创建了吗|创建成功了吗|建好了吗|建成了吗)",
+    re.IGNORECASE,
+)
+FOLLOWUP_REFERENCE_RE = re.compile(
+    r"(刚才|刚刚|前面|之前|上次|上一条|我刚才要求|我刚才说|我说的|那个)",
+    re.IGNORECASE,
+)
+OBVIOUS_QUESTION_RE = re.compile(
+    r"([?？]$|吗$|呢$|呀$|吧$|怎么|如何|为什么|多少|几号|几点|周几|有没有|是否|是不是|是什么|在哪|哪一个|哪几个)",
+    re.IGNORECASE,
+)
+FEISHU_BOT_MENTION_RE = re.compile(r"(?:<at[^>]*>[^<]*</at>|@pmflow\b|@_user_1\b)", re.IGNORECASE)
+FEISHU_QUOTE_PREFIX_RE = re.compile(r"(?m)^\s*(?:回复|引用|转发)[：:]\s*")
+FEISHU_QUOTE_LINE_RE = re.compile(r"(?m)^\s*(?:>+|【引用】|引用：|回复：).*$")
+FEISHU_NOISE_TOKEN_RE = re.compile(r"\[(?:附带图片|图片|表情|文件|语音|消息卡片)\]", re.IGNORECASE)
+EXTERNAL_INTEGRATION_RE = re.compile(
+    r"(github|gitlab|jira|slack|webhook|飞书|企业微信|企微|lark|api|接入|授权|同步)",
+    re.IGNORECASE,
+)
+CROSS_SYSTEM_RE = re.compile(r"(跨系统|跨平台|跨团队|外部系统|第三方系统)", re.IGNORECASE)
+TASK_CREATE_RE = re.compile(
+    r"(新增|新建|创建|加一个|加个|补一个|补个).{0,8}任务",
+    re.IGNORECASE,
+)
+PROJECT_CREATE_GATE_RE = re.compile(
+    r"(新增|新建|创建|建立|立项|放到|放进|加入|录入|进).{0,18}(pmflow|项目列表|项目库|项目台账|项目主数据|一个项目|项目)",
+    re.IGNORECASE,
+)
+HEALTHCHECK_RE = re.compile(
+    r"^(?:ping|pong|health(?:check)?|在线吗|活着吗|存活吗|还在吗|服务正常吗)\??$",
+    re.IGNORECASE,
+)
+SHORT_CONFIRM_RE = re.compile(
+    r"^(?:收到|好的|好|ok|okay|明白|了解|知道了|收到啦|嗯嗯|行|辛苦了|thanks?|thx)[!！.。]?$",
+    re.IGNORECASE,
+)
+CREATE_ACTION_RE = re.compile(r"(新增|新建|创建|建立|接入|配置|录入|加入|放到|放进|立项)", re.IGNORECASE)
+CORRECTION_CUE_RE = re.compile(
+    r"(不对|错了|理解错|不是这个|不是这个意思|我说的是|我要的是|不是啊|不是新增 md|不是写 markdown)",
+    re.IGNORECASE,
+)
+
+ROUTING_AUTO_CREATE_PROJECT = "auto_create_project"
+ROUTING_AUTO_ENTER_SIGNAL = "auto_enter_signal"
+ROUTING_PENDING_LOW_CONFIDENCE = "pending_low_confidence"
+ROUTING_PENDING_COLD_START = "pending_cold_start"
+ROUTING_PENDING_UNCERTAIN_PROJECT = "pending_uncertain_project"
+ROUTING_DUPLICATE_FILTERED = "duplicate_filtered"
+ROUTING_NOISE_FILTERED = "noise_filtered"
+ROUTING_FAILED = "failed"
+ROUTING_DIRECT_QA = "direct_qa"
+ROUTING_OPENCLAW_EXECUTED = "openclaw_executed"
+ROUTING_OPENCLAW_BLOCKED = "openclaw_blocked"
+ROUTING_QUEUE_FOLLOWUP_QUERY = "queue_followup_query"
+ROUTING_QUEUE_FOLLOWUP_CONFIRM = "queue_followup_confirm"
+ROUTING_QUEUE_FOLLOWUP_BLOCKED = "queue_followup_blocked"
+ROUTING_INTENT_GATE_QUERY = "intent_gate_query"
+ROUTING_INTENT_GATE_HEALTHCHECK = "intent_gate_healthcheck"
+ROUTING_INTENT_GATE_CLARIFICATION = "intent_gate_clarification"
+ROUTING_INTENT_GATE_EXTERNAL_CONNECT = "intent_gate_external_connect"
+ROUTING_OPERATION_COMMITTED = "operation_committed"
+ROUTING_OPERATION_FAILED = "operation_failed"
+
+INTENT_QUERY = "query"
+INTENT_SMALLTALK = "smalltalk"
+INTENT_HEALTHCHECK = "healthcheck"
+INTENT_CREATE_PROJECT = "create_project"
+INTENT_CREATE_TASK = "create_task"
+INTENT_SIGNAL_INTAKE = "signal_intake"
+INTENT_EXTERNAL_CONNECT = "external_connect"
+INTENT_CLARIFICATION_REQUIRED = "clarification_required"
+
+ROUTE_STATUS_CLASSIFIED = "classified"
+ROUTE_STATUS_QUEUED = "queued_for_decision"
+ROUTE_STATUS_AUTO_RUNNING = "auto_running"
+ROUTE_STATUS_AUTO_DONE = "auto_done"
+ROUTE_STATUS_FAILED = "failed"
+ROUTE_STATUS_DISMISSED = "dismissed"
+ROUTE_STATUS_UNDONE = "undone"
+
+ACTIVE_ITEM_ROUTE_STATUSES = {
+    ROUTE_STATUS_CLASSIFIED,
+    ROUTE_STATUS_QUEUED,
+    ROUTE_STATUS_AUTO_RUNNING,
+    "pending",
+}
+AUTO_DONE_ITEM_ROUTE_STATUSES = {
+    ROUTE_STATUS_AUTO_DONE,
+    "auto_processed",
+    "confirmed",
+}
+
+DEFAULT_SIGNAL_AUTO_ENTER_THRESHOLD = 0.80
+DEFAULT_PROJECT_AUTO_CREATE_THRESHOLD = 0.85
+DEFAULT_PROJECT_UNCERTAIN_THRESHOLD = 0.60
+
+
+class _LevelSplitItemSchema(BaseModel):
+    level: Literal["project", "signal"]
+    content: str = Field(min_length=1, max_length=4000)
+
+
+class _LevelClassificationSchema(BaseModel):
+    build_intent_score: float = Field(ge=0.0, le=1.0)
+    scope_score: float = Field(ge=0.0, le=1.0)
+    timeline_score: float = Field(ge=0.0, le=1.0)
+    owner_score: float = Field(ge=0.0, le=1.0)
+    project_score: float = Field(ge=0.0, le=1.0)
+    level: Literal["project", "signal"]
+    level_confidence: float = Field(ge=0.0, le=1.0)
+    level_reasoning: str = Field(min_length=1, max_length=600)
+    should_split: bool = False
+    split_items: list[_LevelSplitItemSchema] = Field(default_factory=list, max_length=2)
+
+
+class _SignalExtractionSchema(BaseModel):
+    title: str = Field(min_length=1, max_length=200)
+    type: Literal["竞品动态", "用户反馈", "项目风险", "新需求"]
+    priority: Literal["P0", "P1", "P2", "P3"]
+    matched_project_id: Optional[int] = None
+    summary: str = Field(min_length=1, max_length=3000)
+    reasoning: list[str] = Field(min_length=1, max_length=3)
+    confidence: float = Field(ge=0.0, le=1.0)
+    recommended_action: Literal["enter_only", "create_charter", "trigger_alert"]
+
+
+class _ProjectPrefillSchema(BaseModel):
+    suggested_name: str = Field(min_length=1, max_length=255)
+    suggested_description: str = Field(min_length=1, max_length=4000)
+    suggested_owner: str = Field(min_length=1, max_length=120)
+    suggested_timeline: str = Field(min_length=1, max_length=120)
+    confidence: float = Field(ge=0.0, le=1.0)
+    reasoning: list[str] = Field(min_length=1, max_length=3)
+
+
+def _normalize_reasoning(value: Any, max_items: int = 3) -> list[str]:
+    if isinstance(value, list):
+        rows = [str(x).strip() for x in value if str(x).strip()]
+        return rows[:max_items]
+    if isinstance(value, str):
+        parts = [p.strip(" -•\t") for p in value.splitlines() if p.strip()]
+        return parts[:max_items]
+    return []
+
+
+def _safe_json_object(value: str) -> dict[str, Any]:
+    raw = (value or "").strip()
+    if raw.startswith("```"):
+        chunks = raw.split("```")
+        if len(chunks) >= 2:
+            raw = chunks[1].strip()
+        if raw.startswith("json"):
+            raw = raw[4:].strip()
+    try:
+        parsed = json.loads(raw)
+        return parsed if isinstance(parsed, dict) else {}
+    except Exception:
+        return {}
+
+
+def _validate_schema(
+    schema: type[BaseModel],
+    payload: dict[str, Any],
+    *,
+    context: str,
+) -> dict[str, Any]:
+    if not isinstance(payload, dict):
+        raise ValueError(f"{context}_schema_invalid: payload_not_object")
+    try:
+        if hasattr(schema, "model_validate"):
+            model = schema.model_validate(payload)  # type: ignore[attr-defined]
+        else:
+            model = schema.parse_obj(payload)  # type: ignore[attr-defined]
+    except ValidationError as exc:
+        raise ValueError(f"{context}_schema_invalid: {str(exc)[:400]}") from exc
+    if hasattr(model, "model_dump"):
+        return model.model_dump()  # type: ignore[no-any-return]
+    return model.dict()  # type: ignore[no-any-return]
+
+
+def _safe_float(value: Any, default: float = 0.0) -> float:
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return float(default)
+
+
+def _safe_bool(value: Any, default: bool = False) -> bool:
+    if isinstance(value, bool):
+        return value
+    if value is None:
+        return default
+    text = str(value).strip().lower()
+    if text in {"1", "true", "yes", "y", "on", "是", "真"}:
+        return True
+    if text in {"0", "false", "no", "n", "off", "否", "假"}:
+        return False
+    return default
+
+
+def _safe_int(value: Any) -> Optional[int]:
+    if isinstance(value, bool):
+        return None
+    try:
+        return int(str(value).strip())
+    except (TypeError, ValueError):
+        return None
+
+
+def _normalize_level_value(value: Any, default: str = "signal") -> str:
+    text = str(value or "").strip().lower()
+    if not text:
+        return default
+    if text in {"project", "project-level", "project_level", "new_project"}:
+        return "project"
+    if text in {"signal", "signal-level", "signal_level"}:
+        return "signal"
+    if text in {"项目", "项目级", "立项", "新项目"}:
+        return "project"
+    if text in {"信号", "信号级", "需求", "需求级", "事件"}:
+        return "signal"
+    if "project" in text or "项目" in text or "立项" in text:
+        return "project"
+    if "signal" in text or "信号" in text or "需求" in text:
+        return "signal"
+    return default
+
+
+def _first_non_empty(data: dict[str, Any], keys: tuple[str, ...]) -> Any:
+    for key in keys:
+        if key not in data:
+            continue
+        value = data.get(key)
+        if value is None:
+            continue
+        if isinstance(value, str) and not value.strip():
+            continue
+        return value
+    return None
+
+
+def _utc_now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def _compact_text(value: Any, max_len: int = 160) -> str:
+    text = str(value or "").strip()
+    if len(text) <= max_len:
+        return text
+    return f"{text[: max_len - 3]}..."
+
+
+def _safe_dict(value: Any) -> dict[str, Any]:
+    if isinstance(value, dict):
+        return value
+    if isinstance(value, str):
+        return _safe_json_object(value)
+    return {}
+
+
+def _build_undo_deadline(value: Optional[datetime], *, minutes: int = 10) -> Optional[datetime]:
+    if value is None:
+        return None
+    if value.tzinfo is not None:
+        value = value.astimezone(timezone.utc).replace(tzinfo=None)
+    return value + timedelta(minutes=minutes)
+
+
+class FeishuSignalPipelineService:
+    """飞书消息的两段式分流与路由。"""
+
+    def __init__(self, db: AsyncSession, *, tenant_id: int):
+        self.db = db
+        self.tenant_id = int(tenant_id)
+        self.llm = LLMService()
+
+    @staticmethod
+    def _normalize_route_status(value: Any) -> str:
+        text = str(value or "").strip().lower()
+        if text in {ROUTE_STATUS_CLASSIFIED, ROUTE_STATUS_QUEUED, ROUTE_STATUS_AUTO_RUNNING, ROUTE_STATUS_AUTO_DONE, ROUTE_STATUS_FAILED, ROUTE_STATUS_DISMISSED, ROUTE_STATUS_UNDONE}:
+            return text
+        if text == "pending":
+            return ROUTE_STATUS_QUEUED
+        if text in {"auto_processed", "confirmed"}:
+            return ROUTE_STATUS_AUTO_DONE
+        if text == "dismissed":
+            return ROUTE_STATUS_DISMISSED
+        return text or ROUTE_STATUS_FAILED
+
+    @classmethod
+    def _is_active_route_status(cls, value: Any) -> bool:
+        return cls._normalize_route_status(value) in {
+            ROUTE_STATUS_CLASSIFIED,
+            ROUTE_STATUS_QUEUED,
+            ROUTE_STATUS_AUTO_RUNNING,
+        }
+
+    @staticmethod
+    def _is_healthcheck_message(message: str) -> bool:
+        return bool(HEALTHCHECK_RE.search(str(message or "").strip()))
+
+    @staticmethod
+    def _is_short_confirmation_message(message: str) -> bool:
+        return bool(SHORT_CONFIRM_RE.search(str(message or "").strip()))
+
+    @staticmethod
+    def _is_correction_message(message: str) -> bool:
+        return bool(CORRECTION_CUE_RE.search(str(message or "").strip()))
+
+    @staticmethod
+    def _looks_like_continue_followup(message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        if FOLLOWUP_CREATE_CONTINUE_RE.search(text):
+            return bool(FOLLOWUP_REFERENCE_RE.search(text))
+        return bool(FOLLOWUP_CONTINUE_RE.search(text))
+
+    @staticmethod
+    def _looks_like_task_create_request(message: str) -> bool:
+        return bool(TASK_CREATE_RE.search(str(message or "").strip()))
+
+    @staticmethod
+    def _looks_like_project_create_request(message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        if PROJECT_CREATE_GATE_RE.search(text):
+            return True
+        return bool(OPENCLAW_PROJECT_CREATE_RE.search(text))
+
+    @staticmethod
+    def _looks_like_external_connect_request(message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        return bool(CREATE_ACTION_RE.search(text) and EXTERNAL_INTEGRATION_RE.search(text))
+
+    @staticmethod
+    def _can_execute_high_value_action(user: Optional[User]) -> bool:
+        if user is None:
+            return False
+        return str(getattr(user, "role", "") or "").upper() in {"BOSS", "ADMIN", "PM"}
+
+    @staticmethod
+    def _stable_positive_int(*parts: Any) -> int:
+        raw = "::".join(str(part or "").strip() for part in parts if str(part or "").strip())
+        digest = hashlib.sha1(raw.encode("utf-8")).hexdigest()
+        return int(digest[:12], 16)
+
+    @staticmethod
+    def _normalize_key_fragment(value: str, *, max_len: int = 80) -> str:
+        normalized = re.sub(r"[^0-9a-z\u4e00-\u9fff]+", "", str(value or "").strip().lower())
+        return normalized[:max_len]
+
+    @classmethod
+    def _build_intent_fragment(cls, message: str) -> str:
+        text = str(message or "").strip()
+        if not text:
+            return "generic"
+        if EXTERNAL_INTEGRATION_RE.search(text):
+            return "external_integration"
+        if CROSS_SYSTEM_RE.search(text):
+            return "cross_system"
+        if PROJECT_COMPLEXITY_RE.search(text):
+            return "project_setup"
+        if SIMPLE_SIGNAL_REQUEST_RE.search(text):
+            return "feature_change"
+        fragment = cls._normalize_key_fragment(text, max_len=64)
+        return fragment or "generic"
+
+    def _build_queue_dedupe_key(
+        self,
+        *,
+        level: str,
+        intent: Optional[str],
+        message: str,
+        source_chat_id: str,
+        matched_project_id: Optional[int],
+        source_object: Optional[str] = None,
+    ) -> str:
+        project_key = str(int(matched_project_id)) if matched_project_id is not None else "global"
+        source_key = self._normalize_key_fragment(source_chat_id or "direct", max_len=48) or "direct"
+        intent_key = self._normalize_key_fragment(str(intent or "").strip().lower(), max_len=48) or self._build_intent_fragment(message)
+        object_key = self._normalize_key_fragment(source_object or message, max_len=64) or "generic"
+        return f"{level}:{project_key}:{source_key}:{intent_key}:{object_key}"[:255]
+
+    async def _find_active_dedupe_item(self, dedupe_key: Optional[str]) -> Optional[FeishuItem]:
+        normalized_key = (dedupe_key or "").strip()
+        if not normalized_key:
+            return None
+        return (
+            await self.db.execute(
+                select(FeishuItem)
+                .where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.dedupe_key == normalized_key,
+                        FeishuItem.route_status.in_(tuple(ACTIVE_ITEM_ROUTE_STATUSES)),
+                    )
+                )
+                .order_by(FeishuItem.updated_at.desc(), FeishuItem.id.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+
+    async def _build_existing_item_result(self, item: FeishuItem) -> dict[str, Any]:
+        level = str(item.level or "").lower()
+        result: dict[str, Any] = {
+            "item_id": int(item.id),
+            "level": level,
+            "project_id": int(item.matched_project_id) if item.matched_project_id is not None else None,
+            "signal_id": None,
+            "route_status": self._normalize_route_status(item.route_status),
+            "clarification_message": item.blocking_reason or None,
+        }
+        if level == "project":
+            suggestion = await self._get_project_suggestion(item.id)
+            if suggestion and suggestion.created_project_id:
+                result["project_id"] = int(suggestion.created_project_id)
+            return result
+
+        if level == "command":
+            payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+            project_id = _safe_int(payload.get("project_id") or item.matched_project_id)
+            if project_id is not None:
+                result["project_id"] = int(project_id)
+            return result
+
+        signal = await self._get_signal_by_item(item.id)
+        if signal:
+            result["signal_id"] = int(signal.id)
+            if signal.matched_project_id is not None:
+                result["project_id"] = int(signal.matched_project_id)
+        return result
+
+    async def _resolve_intent_gate(
+        self,
+        *,
+        message: str,
+        inferred_project_id: Optional[int],
+        source_chat_id: str = "",
+        source_sender_open_id: Optional[str] = None,
+        sender: str = "",
+    ) -> dict[str, Any]:
+        text = str(message or "").strip()
+        if not text:
+            return {
+                "intent": INTENT_SMALLTALK,
+                "canonical_action": None,
+                "clarification_template": "",
+                "risk_level": "low",
+                "fields": {},
+                "correction_match": None,
+            }
+
+        correction_match: Optional[dict[str, Any]] = None
+        if self.db is not None and len(text) >= 4:
+            try:
+                recall_rows = await AdaptiveRAGService(self.db, tenant_id=self.tenant_id).retrieve_intent_correction_samples(
+                    utterance=text,
+                    top_k=1,
+                    project_id=inferred_project_id,
+                )
+                if recall_rows:
+                    correction_match = recall_rows[0]
+            except Exception as exc:
+                logger.warning(f"intent_correction_recall_failed: tenant_id={self.tenant_id}, error={exc}")
+
+        if self._is_healthcheck_message(text):
+            return {
+                "intent": INTENT_HEALTHCHECK,
+                "canonical_action": None,
+                "clarification_template": "",
+                "risk_level": "low",
+                "fields": {},
+                "correction_match": correction_match,
+            }
+        if self._is_short_confirmation_message(text):
+            return {
+                "intent": INTENT_SMALLTALK,
+                "canonical_action": None,
+                "clarification_template": "",
+                "risk_level": "low",
+                "fields": {},
+                "correction_match": correction_match,
+            }
+
+        if correction_match and float(correction_match.get("score") or 0.0) >= 0.84:
+            gate = {
+                "intent": str(correction_match.get("correct_intent") or INTENT_SIGNAL_INTAKE).strip().lower(),
+                "canonical_action": str(correction_match.get("canonical_action") or "").strip() or None,
+                "clarification_template": str(correction_match.get("clarification_template") or "").strip(),
+                "risk_level": str(correction_match.get("risk_level") or "medium").strip().lower() or "medium",
+                "fields": {},
+                "correction_match": correction_match,
+            }
+            if gate["intent"] not in {
+                INTENT_QUERY,
+                INTENT_SMALLTALK,
+                INTENT_HEALTHCHECK,
+                INTENT_CREATE_PROJECT,
+                INTENT_CREATE_TASK,
+                INTENT_SIGNAL_INTAKE,
+                INTENT_EXTERNAL_CONNECT,
+                INTENT_CLARIFICATION_REQUIRED,
+            }:
+                gate["intent"] = INTENT_SIGNAL_INTAKE
+            return gate
+
+        if self._looks_like_project_create_request(text):
+            fields = self._extract_project_intent_fields(text)
+            fields = await self._maybe_prefill_project_fields_from_embedded_content(
+                message=text,
+                fields=fields,
+            )
+            fields = await self._maybe_prefill_project_fields_from_context(
+                message=text,
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                sender=sender,
+                fields=fields,
+            )
+            if fields["ready"]:
+                return {
+                    "intent": INTENT_CREATE_PROJECT,
+                    "canonical_action": "project.create",
+                    "clarification_template": "",
+                    "risk_level": "high",
+                    "fields": fields,
+                    "correction_match": correction_match,
+                }
+            return {
+                "intent": INTENT_CLARIFICATION_REQUIRED,
+                "canonical_action": "project.create",
+                "clarification_template": self._build_project_clarification_message(fields["missing_fields"]),
+                "risk_level": "high",
+                "fields": fields,
+                "correction_match": correction_match,
+            }
+
+        if self._looks_like_task_create_request(text):
+            fields = self._extract_task_intent_fields(text=text, inferred_project_id=inferred_project_id)
+            if fields["ready"]:
+                return {
+                    "intent": INTENT_CREATE_TASK,
+                    "canonical_action": "task.create",
+                    "clarification_template": "",
+                    "risk_level": "high",
+                    "fields": fields,
+                    "correction_match": correction_match,
+                }
+            return {
+                "intent": INTENT_CLARIFICATION_REQUIRED,
+                "canonical_action": "task.create",
+                "clarification_template": self._build_task_clarification_message(fields["missing_fields"]),
+                "risk_level": "high",
+                "fields": fields,
+                "correction_match": correction_match,
+            }
+
+        if self._looks_like_external_connect_request(text):
+            return {
+                "intent": INTENT_EXTERNAL_CONNECT,
+                "canonical_action": "source.connect",
+                "clarification_template": "外部系统接入默认进入待决策队列。请补充接入目标、授权边界、回流项目和幂等要求。",
+                "risk_level": "high",
+                "fields": {"project_id": inferred_project_id},
+                "correction_match": correction_match,
+            }
+
+        if self._is_obvious_question(text):
+            return {
+                "intent": INTENT_QUERY,
+                "canonical_action": None,
+                "clarification_template": "",
+                "risk_level": "low",
+                "fields": {},
+                "correction_match": correction_match,
+            }
+
+        return {
+            "intent": INTENT_SIGNAL_INTAKE,
+            "canonical_action": "signal.enter",
+            "clarification_template": "",
+            "risk_level": "medium",
+            "fields": {"project_id": inferred_project_id},
+            "correction_match": correction_match,
+        }
+
+    @staticmethod
+    def _extract_project_scope(text: str, *, description: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip()
+        for pattern in (
+            r"(?:范围|scope|面向|目标|用于|负责|服务于)[：:\s]*(.+)",
+            r"(?:描述|背景|场景|做什么)[：:\s]*(.+)",
+        ):
+            match = re.search(pattern, cleaned, re.IGNORECASE)
+            if not match:
+                continue
+            candidate = re.sub(r"[。；;!！?？].*$", "", match.group(1)).strip(" ，,:：")
+            if candidate:
+                return candidate[:400]
+        if len(description) >= 12:
+            return description[:400]
+        return ""
+
+    def _extract_project_intent_fields(self, message: str) -> dict[str, Any]:
+        title = OpenClawCommanderService._extract_project_create_title(message)
+        if not title:
+            cleaned = re.sub(r"\s+", " ", str(message or "")).strip()
+            for pattern in (
+                r"(?:名称|名字|项目名)(?:叫|是|为)?[：:\s]*(.+)",
+                r"项目(?:叫|名为|名称叫|名称是|名称为)[：:\s]*(.+)",
+            ):
+                match = re.search(pattern, cleaned, re.IGNORECASE)
+                if not match:
+                    continue
+                candidate = re.sub(r"[，。,.;；!?？].*$", "", match.group(1)).strip(" ，,:：\"'“”")
+                if candidate:
+                    title = candidate[:24]
+                    break
+        description = OpenClawCommanderService._extract_project_create_description(
+            instruction=message,
+            title=title,
+        )
+        project_scope = self._extract_project_scope(message, description=description)
+        missing_fields: list[str] = []
+        if not title:
+            missing_fields.append("title")
+        if not description:
+            missing_fields.append("description")
+        if not project_scope:
+            missing_fields.append("project_scope")
+        return {
+            "title": title,
+            "description": description,
+            "project_scope": project_scope,
+            "missing_fields": missing_fields,
+            "ready": not missing_fields,
+        }
+
+    @staticmethod
+    def _looks_like_contextual_project_request(message: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(message or "")).lower()
+        if not normalized:
+            return False
+        return any(
+            token in normalized
+            for token in (
+                "根据这个",
+                "根据上面",
+                "基于这个",
+                "基于上面",
+                "按这个",
+                "按上面",
+                "汇总以上",
+                "汇总上面",
+                "以上内容",
+                "上面内容",
+                "这个汇总",
+                "这个总结",
+                "刚才内容",
+            )
+        )
+
+    @staticmethod
+    def _strip_feishu_mention_prefix(text: str) -> str:
+        cleaned = re.sub(r"^<at[^>]*>[^<]*</at>\s*", "", str(text or ""), flags=re.IGNORECASE).strip()
+        return re.sub(r"\s+", " ", cleaned).strip()
+
+    @staticmethod
+    def _looks_like_placeholder_project_text(text: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(text or "")).lower()
+        if not normalized:
+            return True
+        if len(normalized) > 48:
+            return False
+        return any(
+            token in normalized
+            for token in (
+                "创建个项目",
+                "创建一个项目",
+                "新建一个项目",
+                "新增一个项目",
+                "给我创建个项目",
+                "项目在项目列表",
+                "项目在pmflow",
+                "放进pmflow",
+                "放到pmflow",
+                "放进项目列表",
+                "放到项目列表",
+                "汇总以上内容",
+                "根据这个",
+                "根据上面",
+            )
+        )
+
+    @staticmethod
+    def _looks_like_project_request_artifact(text: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(text or "")).lower()
+        if not normalized:
+            return False
+        return any(
+            token in normalized
+            for token in (
+                "给我创建个项目",
+                "给我创建一个项目",
+                "创建个项目",
+                "创建一个项目",
+                "新建一个项目",
+                "新增一个项目",
+                "项目在项目列表",
+                "项目列表@pmflow",
+                "进项目列表",
+                "放进项目列表",
+                "放到项目列表",
+                "放进pmflow",
+                "放到pmflow",
+                "根据这个汇总",
+                "根据上面汇总",
+                "汇总以上内容",
+            )
+        )
+
+    @staticmethod
+    def _looks_like_unusable_project_context(text: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(text or "")).lower()
+        if not normalized:
+            return True
+        return any(
+            token in normalized
+            for token in (
+                "系统处理中出现异常",
+                "本次未执行",
+                "请稍后重试",
+                "requestwasaborted",
+                "connectionaborted",
+                "429",
+                "模型服务异常",
+                "无法生成可靠回答",
+                "无法直接给结论",
+                "没有输出，可能后端有问题",
+                "让我看看main.py",
+            )
+        )
+
+    @staticmethod
+    def _guess_project_title_from_text(text: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip()
+        if not cleaned:
+            return ""
+        normalized = re.sub(r"\s+", "", cleaned).lower()
+        if "十五五" in cleaned and any(token in cleaned for token in ("医药", "医疗", "器械", "中医药")):
+            return "十五五医药方向研究"
+        if "医药" in cleaned and any(
+            token in normalized
+            for token in ("方向", "值得关注", "创新药", "中医药", "老龄化", "供应链安全", "出海")
+        ):
+            return "医药方向研究"
+        if "医疗器械" in cleaned and "国产替代" in cleaned:
+            return "医疗器械国产替代研究"
+        return ""
+
+    @staticmethod
+    def _extract_embedded_project_context(message: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(message or "")).strip()
+        if len(cleaned) < 80:
+            return ""
+        markers = []
+        for pattern in (
+            r"(?:根据这个|根据上面|基于这个|基于上面|按这个|按上面|汇总以上内容|汇总上面内容|汇总一下|这个汇总|这个总结).{0,40}(?:创建|新建|建立|建个|建一个).{0,8}项目",
+            r"(?:给我|帮我|请|麻烦)?(?:创建|新建|建立|建个|建一个|做个|做一个)(?:一个|个)?项目(?:在|进|到)?(?:pmflow|项目列表)",
+        ):
+            match = re.search(pattern, cleaned, re.IGNORECASE)
+            if match:
+                markers.append(match.start())
+        if not markers:
+            return ""
+        context = cleaned[: min(markers)].strip(" ，,:：")
+        if len(context) < 80:
+            return ""
+        return context[:4000]
+
+    @staticmethod
+    def _derive_project_title_from_context(*, question: str, answer: str) -> str:
+        guessed = FeishuSignalPipelineService._guess_project_title_from_text(f"{question}\n{answer}".strip())
+        if guessed:
+            return guessed
+
+        def _normalize(raw: str) -> str:
+            candidate = re.sub(r"\s+", " ", str(raw or "")).strip(" ，,。.!！？?：:;；")
+            candidate = re.sub(r"^(?:请问|麻烦|帮我|我想知道|我想了解|想了解一下|目前|现在|关于|对于|针对)", "", candidate)
+            candidate = re.sub(
+                r"(?:还有哪些|有哪些|值得关注什么|值得关注哪些|值得关注|应该投什么|该投什么|该怎么做|怎么做|是什么|是啥|可以吗|吗|呢)$",
+                "",
+                candidate,
+            )
+            candidate = re.sub(r"[，,].*$", "", candidate).strip(" ，,。.!！？?：:;；")
+            candidate = re.sub(r"(?:里面|里)$", "", candidate).strip(" ，,。.!！？?：:;；")
+            candidate = re.sub(r"[。；;!！?？].*$", "", candidate).strip(" ，,。.!！？?：:;；")
+            normalized = OpenClawCommanderService._normalize_project_title_candidate(candidate)
+            if normalized.endswith(("方向", "规划", "创新")):
+                normalized = OpenClawCommanderService._normalize_project_title_candidate(f"{normalized}研究")
+            return normalized
+
+        title = _normalize(question)
+        if title:
+            return title
+
+        cleaned_answer = re.sub(r"\s+", " ", str(answer or "")).strip()
+        for pattern in (
+            r"((?:十五五|医药|医疗|器械|公共卫生|老龄化)[^。；;!！?？]{0,16}(?:方向|规划|创新|体系))",
+            r"((?:[A-Za-z0-9\u4e00-\u9fff]{2,18})(?:方向|规划|创新|中台|系统|平台|治理|流程|方案|体系))",
+        ):
+            match = re.search(pattern, cleaned_answer)
+            if not match:
+                continue
+            title = _normalize(match.group(1))
+            if title:
+                return title
+        return ""
+
+    async def _load_recent_chat_turns(
+        self,
+        *,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        sender: str,
+        limit: int = 3,
+    ) -> list[dict[str, str]]:
+        if self.db is None or not source_chat_id:
+            return []
+        conversation_key = self._build_rag_conversation_key(
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            sender=sender,
+        )
+        external_session_id = AdaptiveRAGService._external_session_id_from_key(conversation_key)
+        if external_session_id is None:
+            return []
+        rag_service = AdaptiveRAGService(self.db, tenant_id=self.tenant_id)
+        messages = await rag_service.context_cache.get_session_messages(
+            tenant_id=self.tenant_id,
+            session_id=int(external_session_id),
+        )
+        if not messages:
+            return []
+        turns = rag_service.context_cache.messages_to_turns(messages)
+        return [row for row in turns if isinstance(row, dict)][-limit:]
+
+    async def _load_recent_chat_reply_fallback_text(self, *, source_chat_id: str, limit: int = 5) -> str:
+        if self.db is None or not source_chat_id:
+            return ""
+        rows = (
+            await self.db.execute(
+                select(FeishuReplyOutbox)
+                .where(
+                    and_(
+                        FeishuReplyOutbox.chat_id == (source_chat_id or "")[:128],
+                        FeishuReplyOutbox.status == "sent",
+                    )
+                )
+                .order_by(FeishuReplyOutbox.id.desc())
+                .limit(limit)
+            )
+        ).scalars().all()
+        for row in rows:
+            text = self._strip_feishu_mention_prefix(str(row.message_text or ""))
+            if not text:
+                continue
+            if self._looks_like_unusable_project_context(text):
+                continue
+            return text[:4000]
+        return ""
+
+    async def _maybe_prefill_project_fields_from_embedded_content(
+        self,
+        *,
+        message: str,
+        fields: dict[str, Any],
+    ) -> dict[str, Any]:
+        if fields.get("ready"):
+            return fields
+
+        embedded_context = self._extract_embedded_project_context(message)
+        if not embedded_context or self._looks_like_unusable_project_context(embedded_context):
+            return fields
+
+        current_description = str(fields.get("description") or "").strip()
+        if (
+            self._looks_like_placeholder_project_text(current_description)
+            or self._looks_like_project_request_artifact(current_description)
+        ):
+            current_description = ""
+        current_scope = str(fields.get("project_scope") or "").strip()
+        if (
+            self._looks_like_placeholder_project_text(current_scope)
+            or self._looks_like_project_request_artifact(current_scope)
+        ):
+            current_scope = ""
+
+        title = str(fields.get("title") or "").strip() or self._guess_project_title_from_text(embedded_context)
+        description = current_description or embedded_context[:4000]
+        project_scope = current_scope or self._extract_project_scope(embedded_context, description=description)
+        if title and description and project_scope:
+            return {
+                "title": title,
+                "description": description,
+                "project_scope": project_scope,
+                "missing_fields": [],
+                "ready": True,
+            }
+
+        project_info = await self._extract_project_prefill(
+            f"当前消息里的项目材料：{embedded_context}\n\n用户动作：{message}\n\n请仅基于当前消息里的材料补全项目名称、项目描述、项目范围。"
+        )
+        suggested_name = OpenClawCommanderService._normalize_project_title_candidate(
+            str(project_info.get("suggested_name") or "").strip()
+        )
+        if self._looks_like_placeholder_project_text(suggested_name):
+            suggested_name = ""
+        title = title or suggested_name
+        description = description or str(project_info.get("suggested_description") or "").strip()[:4000]
+        if self._looks_like_unusable_project_context(description):
+            description = ""
+        project_scope = project_scope or self._extract_project_scope(embedded_context, description=description)
+
+        missing_fields: list[str] = []
+        if not title:
+            missing_fields.append("title")
+        if not description:
+            missing_fields.append("description")
+        if not project_scope:
+            missing_fields.append("project_scope")
+        return {
+            "title": title,
+            "description": description,
+            "project_scope": project_scope,
+            "missing_fields": missing_fields,
+            "ready": not missing_fields,
+        }
+
+    async def _maybe_prefill_project_fields_from_context(
+        self,
+        *,
+        message: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        sender: str,
+        fields: dict[str, Any],
+    ) -> dict[str, Any]:
+        if fields.get("ready"):
+            return fields
+        if not source_chat_id or not self._looks_like_contextual_project_request(message):
+            return fields
+
+        turns = await self._load_recent_chat_turns(
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            sender=sender,
+            limit=3,
+        )
+        context_blocks: list[str] = []
+        question_hint = ""
+        answer_hint = ""
+        if turns:
+            anchor = turns[-1]
+            recent_question = str(anchor.get("question") or "").strip()
+            recent_answer = str(anchor.get("answer") or "").strip()
+            if recent_question:
+                context_blocks.append(f"最近用户问题：{recent_question}")
+                question_hint = recent_question
+            if recent_answer and not self._looks_like_unusable_project_context(recent_answer):
+                context_blocks.append(f"最近系统回答：{recent_answer}")
+                answer_hint = recent_answer
+
+        if not context_blocks:
+            reply_fallback = await self._load_recent_chat_reply_fallback_text(source_chat_id=source_chat_id)
+            if reply_fallback:
+                context_blocks.append(f"最近系统回答：{reply_fallback}")
+                answer_hint = reply_fallback
+
+        if not context_blocks:
+            return fields
+
+        base_description = str(fields.get("description") or "").strip()
+        if (
+            self._looks_like_placeholder_project_text(base_description)
+            or self._looks_like_project_request_artifact(base_description)
+        ):
+            base_description = ""
+        if not answer_hint and not base_description:
+            return fields
+
+        heuristic_title = self._derive_project_title_from_context(question=question_hint, answer=answer_hint)
+        heuristic_description = base_description or answer_hint[:4000]
+        heuristic_scope = str(fields.get("project_scope") or "").strip()
+        if (
+            self._looks_like_placeholder_project_text(heuristic_scope)
+            or self._looks_like_project_request_artifact(heuristic_scope)
+        ):
+            heuristic_scope = ""
+        if not heuristic_scope and heuristic_description:
+            heuristic_scope = self._extract_project_scope(
+                f"{question_hint}\n\n{answer_hint}".strip(),
+                description=heuristic_description,
+            )
+        if heuristic_title and heuristic_description and heuristic_scope:
+            return {
+                "title": str(fields.get("title") or "").strip() or heuristic_title,
+                "description": heuristic_description,
+                "project_scope": heuristic_scope,
+                "missing_fields": [],
+                "ready": True,
+            }
+
+        prefill_source = (
+            f"当前建项目请求：{message}\n\n"
+            + "\n\n".join(context_blocks)
+            + "\n\n请基于最近上下文补全项目名称、项目描述、项目范围。"
+        )
+        project_info = await self._extract_project_prefill(prefill_source)
+        current_description = str(fields.get("description") or "").strip()
+        if (
+            self._looks_like_placeholder_project_text(current_description)
+            or self._looks_like_project_request_artifact(current_description)
+        ):
+            current_description = ""
+        current_scope = str(fields.get("project_scope") or "").strip()
+        if (
+            self._looks_like_placeholder_project_text(current_scope)
+            or self._looks_like_project_request_artifact(current_scope)
+        ):
+            current_scope = ""
+        title = (
+            str(fields.get("title") or "").strip()
+            or heuristic_title
+            or OpenClawCommanderService._normalize_project_title_candidate(str(project_info.get("suggested_name") or ""))
+        )
+        description = current_description or str(project_info.get("suggested_description") or "").strip()
+        if not description and answer_hint:
+            description = answer_hint[:4000]
+        if self._looks_like_unusable_project_context(description):
+            description = ""
+        project_scope = current_scope
+        if not project_scope:
+            scope_source = f"{description}\n\n{answer_hint}".strip()
+            project_scope = self._extract_project_scope(scope_source, description=description)
+
+        missing_fields: list[str] = []
+        if not title:
+            missing_fields.append("title")
+        if not description:
+            missing_fields.append("description")
+        if not project_scope:
+            missing_fields.append("project_scope")
+        return {
+            "title": title,
+            "description": description,
+            "project_scope": project_scope,
+            "missing_fields": missing_fields,
+            "ready": not missing_fields,
+        }
+
+    def _resolve_project_candidate_fields(
+        self,
+        *,
+        item: FeishuItem,
+        suggestion: Optional[ProjectSuggestion] = None,
+    ) -> dict[str, Any]:
+        payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+        operation_payload = payload.get("operation_payload") if isinstance(payload.get("operation_payload"), dict) else {}
+        title = str(
+            (suggestion.suggested_name if suggestion is not None else None)
+            or operation_payload.get("title")
+            or ""
+        ).strip()
+        description = str(
+            (suggestion.suggested_description if suggestion is not None else None)
+            or operation_payload.get("description")
+            or ""
+        ).strip()
+        project_scope = str(payload.get("project_scope") or "").strip()
+        if not title or not description:
+            extracted = self._extract_project_intent_fields(item.raw_message or "")
+            if not title:
+                title = str(extracted.get("title") or "").strip()
+            if not description:
+                description = str(extracted.get("description") or "").strip()
+            if not project_scope:
+                project_scope = str(extracted.get("project_scope") or "").strip()
+        if not project_scope:
+            project_scope = self._extract_project_scope(item.raw_message or "", description=description)
+
+        missing_fields: list[str] = []
+        if not title:
+            missing_fields.append("title")
+        if not description:
+            missing_fields.append("description")
+        if not project_scope:
+            missing_fields.append("project_scope")
+        return {
+            "title": title,
+            "description": description,
+            "project_scope": project_scope,
+            "missing_fields": missing_fields,
+            "ready": not missing_fields,
+        }
+
+    @staticmethod
+    def _build_project_clarification_message(missing_fields: list[str]) -> str:
+        labels = {
+            "title": "项目名称",
+            "description": "项目描述",
+            "project_scope": "项目范围",
+        }
+        ordered = [labels[field] for field in missing_fields if field in labels]
+        suffix = "、".join(ordered) if ordered else "项目名称、项目描述、项目范围"
+        return f"未执行：当前建项目请求还缺少 {suffix}。请直接补充这三项后我再创建。"
+
+    @classmethod
+    def _build_project_preview_title(
+        cls,
+        *,
+        title: str,
+        description: str,
+        project_scope: str,
+    ) -> str:
+        normalized_title = OpenClawCommanderService._normalize_project_title_candidate(title)
+        if normalized_title and not cls._looks_like_placeholder_project_text(normalized_title):
+            return normalized_title
+
+        for candidate in (project_scope, description):
+            cleaned = str(candidate or "").strip()
+            if not cleaned or cls._looks_like_placeholder_project_text(cleaned):
+                continue
+            normalized = OpenClawCommanderService._normalize_project_title_candidate(cleaned[:48])
+            if normalized:
+                return normalized
+        return "项目创建请求"
+
+    @classmethod
+    def _build_project_preview_summary(
+        cls,
+        *,
+        description: str,
+        project_scope: str,
+        blocking_reason: Optional[str],
+        fallback_text: str,
+    ) -> str:
+        normalized_description = str(description or "").strip()
+        if cls._looks_like_placeholder_project_text(normalized_description):
+            normalized_description = ""
+        normalized_scope = str(project_scope or "").strip()
+        if cls._looks_like_placeholder_project_text(normalized_scope):
+            normalized_scope = ""
+
+        fragments: list[str] = []
+        if normalized_description:
+            fragments.append(_compact_text(normalized_description, max_len=140))
+        if normalized_scope and normalized_scope not in normalized_description:
+            fragments.append(f"范围：{_compact_text(normalized_scope, max_len=90)}")
+        if fragments:
+            return "；".join(fragments[:2])
+        if str(blocking_reason or "").strip():
+            return str(blocking_reason).strip()
+        return _compact_text(fallback_text, max_len=140)
+
+    def _extract_task_intent_fields(self, *, text: str, inferred_project_id: Optional[int]) -> dict[str, Any]:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip()
+        title = ""
+        quoted_match = re.search(r"[\"'“](.+?)[\"'”]", cleaned)
+        if quoted_match:
+            title = quoted_match.group(1).strip()
+        if not title:
+            for pattern in (
+                r"(?:新增|新建|创建|加一个|加个).{0,8}任务(?:叫|名为|标题是|是|：|:)?\s*(.+)",
+                r"任务(?:叫|名为|标题是|是|：|:)\s*(.+)",
+            ):
+                match = re.search(pattern, cleaned, re.IGNORECASE)
+                if not match:
+                    continue
+                candidate = re.sub(r"[。；;!！?？].*$", "", match.group(1)).strip(" ，,:：")
+                candidate = re.sub(r"^(?:到|进|放到)\s*项目.*$", "", candidate, flags=re.IGNORECASE).strip(" ，,:：")
+                if candidate:
+                    title = candidate[:120]
+                    break
+
+        description = ""
+        if title:
+            description = cleaned.replace(title, "", 1).strip(" ，,:：")
+            description = re.sub(r"^(?:新增|新建|创建|加一个|加个).{0,8}任务", "", description, flags=re.IGNORECASE).strip(" ，,:：")
+
+        missing_fields: list[str] = []
+        if inferred_project_id is None:
+            missing_fields.append("project_id")
+        if not title:
+            missing_fields.append("title")
+        return {
+            "project_id": int(inferred_project_id) if inferred_project_id is not None else None,
+            "title": title,
+            "description": description[:5000] or None,
+            "missing_fields": missing_fields,
+            "ready": not missing_fields,
+        }
+
+    @staticmethod
+    def _build_task_clarification_message(missing_fields: list[str]) -> str:
+        labels = {
+            "project_id": "归属项目",
+            "title": "任务标题",
+        }
+        ordered = [labels[field] for field in missing_fields if field in labels]
+        suffix = "、".join(ordered) if ordered else "归属项目和任务标题"
+        return f"未执行：当前建任务请求还缺少 {suffix}。请补充后我再创建。"
+
+    @staticmethod
+    def _build_pending_operation_summary(*, operation_code: str, payload: dict[str, Any]) -> str:
+        if operation_code == "project.create":
+            return f"待创建项目：{str(payload.get('title') or '未命名项目').strip()}"
+        if operation_code == "task.create":
+            return f"待创建任务：{str(payload.get('title') or '未命名任务').strip()}"
+        return f"待执行动作：{operation_code}"
+
+    @staticmethod
+    def _build_pending_operation_next_step(*, operation_code: str) -> str:
+        if operation_code == "project.create":
+            return "补齐信息或确认后，系统会创建项目并绑定真实项目 ID。"
+        if operation_code == "task.create":
+            return "确认后，系统会在目标项目中创建任务。"
+        return "补齐上下文后，系统会继续执行该动作。"
+
+    @staticmethod
+    def _build_gate_source_object(*, intent: str, fields: dict[str, Any], message: str) -> str:
+        if intent == INTENT_CREATE_PROJECT:
+            return str(fields.get("title") or fields.get("project_scope") or message)
+        if intent == INTENT_CREATE_TASK:
+            return str(fields.get("title") or fields.get("project_id") or message)
+        return str(fields.get("project_id") or intent or message)
+
+    async def _queue_project_intent_item(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        occurred_at: datetime,
+        intent_gate: dict[str, Any],
+        blocking_reason: str,
+        decision_reason: str,
+    ) -> dict[str, Any]:
+        fields = intent_gate.get("fields") if isinstance(intent_gate.get("fields"), dict) else {}
+        source_object = self._build_gate_source_object(
+            intent=str(intent_gate.get("intent") or INTENT_CREATE_PROJECT),
+            fields=fields,
+            message=message,
+        )
+        dedupe_key = self._build_queue_dedupe_key(
+            level="project",
+            intent=str(intent_gate.get("intent") or INTENT_CREATE_PROJECT),
+            message=message,
+            source_chat_id=source_chat_id,
+            matched_project_id=None,
+            source_object=source_object,
+        )
+        existing_item = await self._find_active_dedupe_item(dedupe_key)
+        if existing_item is not None:
+            existing_item.occurrence_count = max(1, int(existing_item.occurrence_count or 1)) + 1
+            existing_item.blocking_reason = blocking_reason[:2000]
+            existing_item.decision_reason = decision_reason[:1000]
+            existing_item.raw_message = message[:4000]
+            await self.db.flush()
+            result = await self._build_existing_item_result(existing_item)
+            result["clarification_message"] = blocking_reason
+            return result
+
+        item = FeishuItem(
+            tenant_id=self.tenant_id,
+            feishu_message_id=feishu_message_id,
+            project_score=1.0,
+            signal_confidence=0.0,
+            raw_message=message[:4000],
+            split_from_item_id=None,
+            level="project",
+            level_confidence=1.0,
+            level_reasoning="intent_gate:create_project",
+            card_type="new_project_candidate",
+            route_status=ROUTE_STATUS_QUEUED,
+            matched_project_id=None,
+            source_chat_id=(source_chat_id or "")[:128] or None,
+            source_sender=(sender or "")[:120] or None,
+            source_sender_open_id=(source_sender_open_id or "")[:128] or None,
+            source_occurred_at=occurred_at.replace(tzinfo=None),
+            ai_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False)[:8000],
+            decision_reason=decision_reason[:1000] or None,
+            blocking_reason=blocking_reason[:2000] or None,
+            safe_to_auto_execute=False,
+            auto_execute_state="blocked",
+            dedupe_key=dedupe_key,
+            message_class="action_required",
+            occurrence_count=1,
+            next_step_text=self._build_pending_operation_next_step(operation_code="project.create"),
+            queue_visible=True,
+            result_payload={
+                "intent": INTENT_CREATE_PROJECT,
+                "executor": "operation_executor",
+                "operation_code": "project.create",
+                "operation_payload": {
+                    "title": str(fields.get("title") or "").strip(),
+                    "description": str(fields.get("description") or "").strip(),
+                },
+                "project_scope": str(fields.get("project_scope") or "").strip() or None,
+                "result_summary": self._build_pending_operation_summary(
+                    operation_code="project.create",
+                    payload={
+                        "title": str(fields.get("title") or "").strip(),
+                    },
+                ),
+            },
+        )
+        self.db.add(item)
+        await self.db.flush()
+
+        suggestion = ProjectSuggestion(
+            tenant_id=self.tenant_id,
+            feishu_item_id=int(item.id),
+            suggested_name=(str(fields.get("title") or "").strip() or message[:20] or "新项目建议")[:255],
+            suggested_description=(str(fields.get("description") or "").strip() or message[:400])[:4000],
+            suggested_owner=(sender or "待指定")[:120],
+            suggested_timeline="待确认",
+            confidence=1.0,
+            reasoning=json.dumps(["命中 intent gate:create_project，等待确认或补齐字段。"], ensure_ascii=False),
+            status="pending",
+        )
+        self.db.add(suggestion)
+        await self.db.flush()
+        return {
+            "item_id": int(item.id),
+            "level": "project",
+            "project_id": None,
+            "signal_id": None,
+            "route_status": ROUTE_STATUS_QUEUED,
+            "clarification_message": blocking_reason,
+        }
+
+    async def _queue_command_intent_item(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        occurred_at: datetime,
+        intent_gate: dict[str, Any],
+        operation_code: str,
+        operation_payload: dict[str, Any],
+        blocking_reason: str,
+        decision_reason: str,
+        matched_project_id: Optional[int],
+    ) -> dict[str, Any]:
+        fields = intent_gate.get("fields") if isinstance(intent_gate.get("fields"), dict) else {}
+        source_object = self._build_gate_source_object(
+            intent=str(intent_gate.get("intent") or ""),
+            fields=fields,
+            message=message,
+        )
+        dedupe_key = self._build_queue_dedupe_key(
+            level="command",
+            intent=str(intent_gate.get("intent") or ""),
+            message=message,
+            source_chat_id=source_chat_id,
+            matched_project_id=matched_project_id,
+            source_object=source_object,
+        )
+        existing_item = await self._find_active_dedupe_item(dedupe_key)
+        if existing_item is not None:
+            existing_item.occurrence_count = max(1, int(existing_item.occurrence_count or 1)) + 1
+            existing_item.blocking_reason = blocking_reason[:2000]
+            existing_item.decision_reason = decision_reason[:1000]
+            existing_item.raw_message = message[:4000]
+            existing_item.result_payload = {
+                **(existing_item.result_payload if isinstance(existing_item.result_payload, dict) else {}),
+                "intent": str(intent_gate.get("intent") or ""),
+                "executor": "operation_executor",
+                "operation_code": operation_code,
+                "operation_payload": operation_payload,
+                "result_summary": self._build_pending_operation_summary(
+                    operation_code=operation_code,
+                    payload=operation_payload,
+                ),
+            }
+            await self.db.flush()
+            result = await self._build_existing_item_result(existing_item)
+            result["clarification_message"] = blocking_reason
+            return result
+
+        item = FeishuItem(
+            tenant_id=self.tenant_id,
+            feishu_message_id=feishu_message_id,
+            project_score=1.0 if matched_project_id is not None else None,
+            signal_confidence=0.0 if matched_project_id is not None else None,
+            raw_message=message[:4000],
+            split_from_item_id=None,
+            level="command",
+            level_confidence=1.0,
+            level_reasoning=f"intent_gate:{str(intent_gate.get('intent') or '')}",
+            card_type=operation_code.replace(".", "_")[:30],
+            route_status=ROUTE_STATUS_QUEUED,
+            matched_project_id=int(matched_project_id) if matched_project_id is not None else None,
+            source_chat_id=(source_chat_id or "")[:128] or None,
+            source_sender=(sender or "")[:120] or None,
+            source_sender_open_id=(source_sender_open_id or "")[:128] or None,
+            source_occurred_at=occurred_at.replace(tzinfo=None),
+            ai_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False)[:8000],
+            decision_reason=decision_reason[:1000] or None,
+            blocking_reason=blocking_reason[:2000] or None,
+            safe_to_auto_execute=False,
+            auto_execute_state="blocked",
+            dedupe_key=dedupe_key,
+            message_class="action_required",
+            occurrence_count=1,
+            next_step_text=self._build_pending_operation_next_step(operation_code=operation_code),
+            queue_visible=True,
+            result_payload={
+                "intent": str(intent_gate.get("intent") or ""),
+                "executor": "operation_executor",
+                "operation_code": operation_code,
+                "operation_payload": operation_payload,
+                "result_summary": self._build_pending_operation_summary(
+                    operation_code=operation_code,
+                    payload=operation_payload,
+                ),
+                "project_id": int(matched_project_id) if matched_project_id is not None else None,
+            },
+        )
+        self.db.add(item)
+        await self.db.flush()
+        return {
+            "item_id": int(item.id),
+            "level": "command",
+            "project_id": int(matched_project_id) if matched_project_id is not None else None,
+            "signal_id": None,
+            "route_status": ROUTE_STATUS_QUEUED,
+            "clarification_message": blocking_reason,
+        }
+
+    async def _record_committed_operation_item(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        captured_at: Optional[datetime],
+        intent: str,
+        operation_code: str,
+        execution_result: dict[str, Any],
+        matched_project_id: Optional[int],
+    ) -> Optional[int]:
+        occurred_at = captured_at or _utc_now()
+        if occurred_at.tzinfo is not None:
+            occurred_at = occurred_at.astimezone(timezone.utc).replace(tzinfo=None)
+        item = FeishuItem(
+            tenant_id=self.tenant_id,
+            feishu_message_id=feishu_message_id,
+            project_score=1.0 if matched_project_id is not None else None,
+            signal_confidence=0.0 if matched_project_id is not None else None,
+            raw_message=message[:4000],
+            split_from_item_id=None,
+            level="command",
+            level_confidence=1.0,
+            level_reasoning=f"intent_gate:{intent}",
+            card_type=operation_code.replace(".", "_")[:30],
+            route_status=ROUTE_STATUS_AUTO_DONE,
+            matched_project_id=int(matched_project_id) if matched_project_id is not None else None,
+            source_chat_id=(source_chat_id or "")[:128] or None,
+            source_sender=(sender or "")[:120] or None,
+            source_sender_open_id=(source_sender_open_id or "")[:128] or None,
+            source_occurred_at=occurred_at,
+            ai_output=json.dumps({"execution_result": execution_result}, ensure_ascii=False)[:8000],
+            decision_reason="deterministic_operation_committed",
+            blocking_reason=None,
+            safe_to_auto_execute=True,
+            auto_execute_state="done",
+            message_class="auto_done",
+            occurrence_count=1,
+            next_step_text="真实提交已完成，结果已绑定 DB commit。",
+            queue_visible=False,
+            processed_at=_utc_now().replace(tzinfo=None),
+            result_payload={
+                "intent": intent,
+                "executor": "operation_executor",
+                "operation_code": operation_code,
+                "result_summary": str(execution_result.get("message") or "").strip() or operation_code,
+                "operation_result": execution_result.get("result") if isinstance(execution_result.get("result"), dict) else {},
+                "project_id": int(matched_project_id) if matched_project_id is not None else None,
+                "target_url": f"/projects/{int(matched_project_id)}/board" if matched_project_id is not None else None,
+            },
+        )
+        self.db.add(item)
+        await self.db.flush()
+        return int(item.id)
+
+    async def _execute_controlled_operation(
+        self,
+        *,
+        operator_user: User,
+        operation_code: str,
+        payload: dict[str, Any],
+        trace_id: str,
+    ) -> dict[str, Any]:
+        executor = OperationExecutorService(self.db)
+        project_context = None
+        project_id = _safe_int(payload.get("project_id"))
+        if operation_code != "project.create" and project_id is not None:
+            project_context = (
+                await self.db.execute(
+                    select(Project).where(
+                        and_(
+                            Project.tenant_id == self.tenant_id,
+                            Project.id == int(project_id),
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if project_context is None:
+                raise ValueError("project_context_missing")
+
+        logger.info(
+            "operation_started",
+            tenant_id=self.tenant_id,
+            trace_id=trace_id,
+            operation_code=operation_code,
+            payload=payload,
+            executor="operation_executor",
+        )
+        execution = await executor.execute_operations(
+            tenant_id=self.tenant_id,
+            operator_user=operator_user,
+            project=project_context,
+            operations=[
+                {
+                    "operation_code": operation_code,
+                    "payload": payload,
+                    "reason": "intent_gate",
+                }
+            ],
+            dry_run=False,
+            trace_id=trace_id,
+        )
+        first = execution["results"][0] if execution.get("results") else {}
+        event_name = "operation_committed" if str(first.get("status") or "") in {"applied", "skipped"} else "operation_failed"
+        logger.info(
+            event_name,
+            tenant_id=self.tenant_id,
+            trace_id=trace_id,
+            operation_code=operation_code,
+            result=first,
+            executor="operation_executor",
+        )
+        return execution
+
+    @staticmethod
+    def _derive_wrong_interpretation_from_recent_item(item: Optional[FeishuItem]) -> str:
+        if item is None:
+            return "unknown"
+        level = str(item.level or "").strip().lower()
+        if level == "project":
+            return "project_candidate"
+        if level == "signal":
+            return "signal_intake"
+        if level == "command":
+            return str(item.card_type or "command")
+        return level or "unknown"
+
+    async def _maybe_write_intent_correction_sample(
+        self,
+        *,
+        message: str,
+        source_chat_id: str,
+        intent_gate: dict[str, Any],
+    ) -> None:
+        if self.db is None or not self._is_correction_message(message):
+            return
+        recent_items = await self._load_recent_chat_items(source_chat_id=source_chat_id, limit=3) if source_chat_id else []
+        wrong_interpretation = self._derive_wrong_interpretation_from_recent_item(recent_items[0] if recent_items else None)
+        fields = intent_gate.get("fields") if isinstance(intent_gate.get("fields"), dict) else {}
+        clarification_template = str(intent_gate.get("clarification_template") or "").strip()
+        if not clarification_template:
+            if str(intent_gate.get("intent") or "") == INTENT_CREATE_PROJECT:
+                clarification_template = self._build_project_clarification_message(
+                    list(fields.get("missing_fields") or []),
+                )
+            elif str(intent_gate.get("intent") or "") == INTENT_CREATE_TASK:
+                clarification_template = self._build_task_clarification_message(
+                    list(fields.get("missing_fields") or []),
+                )
+        try:
+            await AdaptiveRAGService(self.db, tenant_id=self.tenant_id).upsert_intent_correction_sample(
+                user_utterance=message,
+                wrong_interpretation=wrong_interpretation,
+                correct_intent=str(intent_gate.get("intent") or INTENT_CLARIFICATION_REQUIRED),
+                canonical_action=str(intent_gate.get("canonical_action") or "").strip() or "clarify",
+                clarification_template=clarification_template,
+                risk_level=str(intent_gate.get("risk_level") or "medium"),
+                project_id=_safe_int(fields.get("project_id")),
+            )
+        except Exception as exc:
+            logger.warning(f"intent_correction_write_failed: tenant_id={self.tenant_id}, error={exc}")
+
+    async def _build_healthcheck_capture(
+        self,
+        *,
+        message: str,
+        feishu_message_id: Optional[str],
+        process_started_at: float,
+    ) -> dict[str, Any]:
+        duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=None,
+            raw_message=message,
+            ai_raw_output=json.dumps({"intent": INTENT_HEALTHCHECK}, ensure_ascii=False),
+            project_score=None,
+            signal_confidence=None,
+            level="healthcheck",
+            routing_decision=ROUTING_INTENT_GATE_HEALTHCHECK,
+            routing_reason="命中 healthcheck，直接返回执行链可用状态，不进入 Inbox。",
+            reasoning_bullets=[
+                "intent gate 已识别为 healthcheck。",
+                "不涉及写库动作。",
+                "回执固定由后端生成。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=duration_ms,
+            normalized_text=message,
+            route=ROUTING_INTENT_GATE_HEALTHCHECK,
+            intent=INTENT_HEALTHCHECK,
+            executor="router",
+            reply_source="deterministic_reply",
+        )
+        return {
+            "status": "answered",
+            "classification": "healthcheck",
+            "project_scope": "healthcheck",
+            "linked_project_ids": [],
+            "signal_id": None,
+            "created_project_id": None,
+            "auto_tasks": [],
+            "ai_trace_id": f"feishu-healthcheck-{(feishu_message_id or 'noid')[:32]}",
+            "results": [],
+            "intent": INTENT_HEALTHCHECK,
+            "direct_reply_required": True,
+            "direct_reply_text": "healthcheck=ok；执行链在线；仅真实 commit 后才会回“已执行”。",
+        }
+
+    async def _handle_create_project_gate(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        captured_at: Optional[datetime],
+        intent_gate: dict[str, Any],
+        process_started_at: float,
+    ) -> dict[str, Any]:
+        fields = intent_gate.get("fields") if isinstance(intent_gate.get("fields"), dict) else {}
+        clarification_message = str(intent_gate.get("clarification_template") or "").strip() or self._build_project_clarification_message(
+            list(fields.get("missing_fields") or []),
+        )
+        occurred_at = captured_at or _utc_now()
+        if occurred_at.tzinfo is None:
+            occurred_at = occurred_at.replace(tzinfo=timezone.utc)
+
+        operator_user = await self._resolve_command_operator_user(sender=sender)
+        if not fields.get("ready"):
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False),
+                project_score=1.0,
+                signal_confidence=0.0,
+                level="project",
+                routing_decision=ROUTING_INTENT_GATE_CLARIFICATION,
+                routing_reason="建项目缺少必填字段，已转待澄清，禁止猜测执行。",
+                reasoning_bullets=[
+                    "intent gate 命中 create_project。",
+                    "当前缺少 project.create 必填字段。",
+                    "仅保留结构化待澄清项，不触发执行器。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                normalized_text=message,
+                route=ROUTING_INTENT_GATE_CLARIFICATION,
+                intent=INTENT_CLARIFICATION_REQUIRED,
+                executor="operation_executor",
+                operation_code="project.create",
+                reply_source="deterministic_reply",
+            )
+            return {
+                "status": "captured",
+                "classification": "project",
+                "project_scope": "project",
+                "linked_project_ids": [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-project-clarify-{(feishu_message_id or 'noid')[:32]}",
+                "results": [],
+                "intent": INTENT_CLARIFICATION_REQUIRED,
+                "clarification_message": clarification_message,
+                "direct_reply_required": True,
+                "direct_reply_text": clarification_message,
+            }
+
+        if not self._can_execute_high_value_action(operator_user):
+            queued = await self._queue_project_intent_item(
+                message=message,
+                sender=sender,
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                feishu_message_id=feishu_message_id,
+                occurred_at=occurred_at,
+                intent_gate=intent_gate,
+                blocking_reason=clarification_message if not fields.get("ready") else "建项目属于高风险动作，当前需要你确认后再执行。",
+                decision_reason="high_value_action_guard",
+            )
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=_safe_int(queued.get("item_id")),
+                raw_message=message,
+                ai_raw_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False),
+                project_score=1.0,
+                signal_confidence=0.0,
+                level="project",
+                routing_decision=ROUTING_INTENT_GATE_CLARIFICATION,
+                routing_reason="建项目命中高风险保护，已转待决策/待澄清。",
+                reasoning_bullets=[
+                    "intent gate 命中 create_project。",
+                    "当前未满足自动执行条件。",
+                    "仅保留结构化待决策项，不做猜测执行。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                normalized_text=message,
+                route=ROUTING_INTENT_GATE_CLARIFICATION,
+                intent=INTENT_CLARIFICATION_REQUIRED if not fields.get("ready") else INTENT_CREATE_PROJECT,
+                executor="operation_executor",
+                operation_code="project.create",
+                reply_source="deterministic_reply",
+            )
+            return {
+                "status": "captured",
+                "classification": "project",
+                "project_scope": "project",
+                "linked_project_ids": [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-project-pending-{queued.get('item_id')}",
+                "results": [queued],
+                "intent": INTENT_CLARIFICATION_REQUIRED if not fields.get("ready") else INTENT_CREATE_PROJECT,
+                "clarification_message": queued.get("clarification_message"),
+                "direct_reply_required": True,
+                "direct_reply_text": str(queued.get("clarification_message") or "已转待决策。"),
+            }
+
+        payload_description = str(fields.get("description") or "").strip()
+        project_scope = str(fields.get("project_scope") or "").strip()
+        if project_scope and project_scope not in payload_description:
+            payload_description = f"{payload_description}\n\n项目范围：{project_scope}".strip()
+        payload = {
+            "title": str(fields.get("title") or "").strip(),
+            "description": payload_description,
+        }
+        trace_id = f"intent-project-{self._stable_positive_int(message, sender, feishu_message_id)}"
+        execution = await self._execute_controlled_operation(
+            operator_user=operator_user,
+            operation_code="project.create",
+            payload=payload,
+            trace_id=trace_id,
+        )
+        first = execution["results"][0] if execution.get("results") else {}
+        status = str(first.get("status") or "").strip().lower()
+        if status not in {"applied", "skipped"}:
+            raise ValueError(str(first.get("message") or "project.create failed"))
+
+        project_result = first.get("result") if isinstance(first.get("result"), dict) else {}
+        project_id = _safe_int(project_result.get("project_id") or first.get("entity_id"))
+        item_id = await self._record_committed_operation_item(
+            message=message,
+            sender=sender,
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            feishu_message_id=feishu_message_id,
+            captured_at=captured_at,
+            intent=INTENT_CREATE_PROJECT,
+            operation_code="project.create",
+            execution_result=first,
+            matched_project_id=project_id,
+        )
+        reply_text = self._build_openclaw_command_reply({"execution": {"results": [first], "applied_count": 1 if status == "applied" else 0, "skipped_count": 1 if status == "skipped" else 0, "failed_count": 0}})
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=item_id,
+            raw_message=message,
+            ai_raw_output=json.dumps({"intent_gate": intent_gate, "execution": execution}, ensure_ascii=False)[:120000],
+            project_score=1.0,
+            signal_confidence=0.0,
+            level="project",
+            routing_decision=ROUTING_OPERATION_COMMITTED,
+            routing_reason="建项目已通过 deterministic operation executor 提交，并以真实 DB 结果回执。",
+            reasoning_bullets=[
+                "intent gate 命中 create_project。",
+                "执行器仅走 project.create。",
+                "回复绑定真实 commit 结果。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+            normalized_text=message,
+            route=ROUTING_OPERATION_COMMITTED,
+            intent=INTENT_CREATE_PROJECT,
+            executor="operation_executor",
+            operation_code="project.create",
+            operation_result=first if isinstance(first, dict) else None,
+            reply_source="db_commit",
+        )
+        return {
+            "status": "captured",
+            "classification": "project",
+            "project_scope": "project",
+            "linked_project_ids": [int(project_id)] if project_id is not None else [],
+            "signal_id": None,
+            "created_project_id": int(project_id) if project_id is not None else None,
+            "auto_tasks": [],
+            "ai_trace_id": trace_id,
+            "results": [
+                {
+                    "item_id": item_id,
+                    "level": "command",
+                    "project_id": int(project_id) if project_id is not None else None,
+                    "signal_id": None,
+                    "route_status": ROUTE_STATUS_AUTO_DONE,
+                }
+            ],
+            "intent": INTENT_CREATE_PROJECT,
+            "direct_reply_required": True,
+            "direct_reply_text": reply_text,
+            "command_result": {"execution": execution},
+        }
+
+    async def _handle_create_task_gate(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        captured_at: Optional[datetime],
+        intent_gate: dict[str, Any],
+        process_started_at: float,
+    ) -> dict[str, Any]:
+        fields = intent_gate.get("fields") if isinstance(intent_gate.get("fields"), dict) else {}
+        operation_payload = {
+            "project_id": _safe_int(fields.get("project_id")),
+            "title": str(fields.get("title") or "").strip(),
+            "description": str(fields.get("description") or "").strip() or None,
+        }
+        clarification_message = str(intent_gate.get("clarification_template") or "").strip() or self._build_task_clarification_message(
+            list(fields.get("missing_fields") or []),
+        )
+        occurred_at = captured_at or _utc_now()
+        if occurred_at.tzinfo is None:
+            occurred_at = occurred_at.replace(tzinfo=timezone.utc)
+
+        operator_user = await self._resolve_command_operator_user(sender=sender)
+        if not fields.get("ready"):
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False),
+                project_score=1.0 if fields.get("project_id") is not None else None,
+                signal_confidence=0.0 if fields.get("project_id") is not None else None,
+                level="command",
+                routing_decision=ROUTING_INTENT_GATE_CLARIFICATION,
+                routing_reason="建任务缺少必填字段，已转待澄清，禁止猜测执行。",
+                reasoning_bullets=[
+                    "intent gate 命中 create_task。",
+                    "当前缺少 task.create 必填字段。",
+                    "仅保留结构化待澄清项，不触发执行器。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                normalized_text=message,
+                route=ROUTING_INTENT_GATE_CLARIFICATION,
+                intent=INTENT_CLARIFICATION_REQUIRED,
+                executor="operation_executor",
+                operation_code="task.create",
+                reply_source="deterministic_reply",
+            )
+            return {
+                "status": "captured",
+                "classification": "command",
+                "project_scope": "command",
+                "linked_project_ids": [int(fields["project_id"])] if fields.get("project_id") is not None else [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-task-clarify-{(feishu_message_id or 'noid')[:32]}",
+                "results": [],
+                "intent": INTENT_CLARIFICATION_REQUIRED,
+                "clarification_message": clarification_message,
+                "direct_reply_required": True,
+                "direct_reply_text": clarification_message,
+            }
+
+        if not self._can_execute_high_value_action(operator_user):
+            queued = await self._queue_command_intent_item(
+                message=message,
+                sender=sender,
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                feishu_message_id=feishu_message_id,
+                occurred_at=occurred_at,
+                intent_gate=intent_gate,
+                operation_code="task.create",
+                operation_payload={k: v for k, v in operation_payload.items() if v is not None},
+                blocking_reason=clarification_message if not fields.get("ready") else "建任务属于高风险动作，当前需要你确认后再执行。",
+                decision_reason="high_value_action_guard",
+                matched_project_id=_safe_int(fields.get("project_id")),
+            )
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=_safe_int(queued.get("item_id")),
+                raw_message=message,
+                ai_raw_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False),
+                project_score=1.0 if fields.get("project_id") is not None else None,
+                signal_confidence=0.0 if fields.get("project_id") is not None else None,
+                level="command",
+                routing_decision=ROUTING_INTENT_GATE_CLARIFICATION,
+                routing_reason="建任务命中高风险保护，已转待决策/待澄清。",
+                reasoning_bullets=[
+                    "intent gate 命中 create_task。",
+                    "当前未满足自动执行条件。",
+                    "仅保留结构化待决策项，不做猜测执行。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                normalized_text=message,
+                route=ROUTING_INTENT_GATE_CLARIFICATION,
+                intent=INTENT_CLARIFICATION_REQUIRED if not fields.get("ready") else INTENT_CREATE_TASK,
+                executor="operation_executor",
+                operation_code="task.create",
+                reply_source="deterministic_reply",
+            )
+            return {
+                "status": "captured",
+                "classification": "command",
+                "project_scope": "command",
+                "linked_project_ids": [int(fields["project_id"])] if fields.get("project_id") is not None else [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-task-pending-{queued.get('item_id')}",
+                "results": [queued],
+                "intent": INTENT_CLARIFICATION_REQUIRED if not fields.get("ready") else INTENT_CREATE_TASK,
+                "clarification_message": queued.get("clarification_message"),
+                "direct_reply_required": True,
+                "direct_reply_text": str(queued.get("clarification_message") or "已转待决策。"),
+            }
+
+        trace_id = f"intent-task-{self._stable_positive_int(message, sender, feishu_message_id)}"
+        execution = await self._execute_controlled_operation(
+            operator_user=operator_user,
+            operation_code="task.create",
+            payload={k: v for k, v in operation_payload.items() if v is not None},
+            trace_id=trace_id,
+        )
+        first = execution["results"][0] if execution.get("results") else {}
+        status = str(first.get("status") or "").strip().lower()
+        if status not in {"applied", "skipped"}:
+            raise ValueError(str(first.get("message") or "task.create failed"))
+        project_id = _safe_int(fields.get("project_id"))
+        item_id = await self._record_committed_operation_item(
+            message=message,
+            sender=sender,
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            feishu_message_id=feishu_message_id,
+            captured_at=captured_at,
+            intent=INTENT_CREATE_TASK,
+            operation_code="task.create",
+            execution_result=first,
+            matched_project_id=project_id,
+        )
+        reply_text = self._build_openclaw_command_reply({"execution": {"results": [first], "applied_count": 1 if status == "applied" else 0, "skipped_count": 1 if status == "skipped" else 0, "failed_count": 0}})
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=item_id,
+            raw_message=message,
+            ai_raw_output=json.dumps({"intent_gate": intent_gate, "execution": execution}, ensure_ascii=False)[:120000],
+            project_score=1.0 if project_id is not None else None,
+            signal_confidence=0.0 if project_id is not None else None,
+            level="command",
+            routing_decision=ROUTING_OPERATION_COMMITTED,
+            routing_reason="建任务已通过 deterministic operation executor 提交，并以真实 DB 结果回执。",
+            reasoning_bullets=[
+                "intent gate 命中 create_task。",
+                "执行器仅走 task.create。",
+                "回复绑定真实 commit 结果。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+            normalized_text=message,
+            route=ROUTING_OPERATION_COMMITTED,
+            intent=INTENT_CREATE_TASK,
+            executor="operation_executor",
+            operation_code="task.create",
+            operation_result=first if isinstance(first, dict) else None,
+            reply_source="db_commit",
+        )
+        return {
+            "status": "captured",
+            "classification": "command",
+            "project_scope": "command",
+            "linked_project_ids": [int(project_id)] if project_id is not None else [],
+            "signal_id": None,
+            "created_project_id": None,
+            "auto_tasks": [],
+            "ai_trace_id": trace_id,
+            "results": [
+                {
+                    "item_id": item_id,
+                    "level": "command",
+                    "project_id": int(project_id) if project_id is not None else None,
+                    "signal_id": None,
+                    "route_status": ROUTE_STATUS_AUTO_DONE,
+                }
+            ],
+            "intent": INTENT_CREATE_TASK,
+            "direct_reply_required": True,
+            "direct_reply_text": reply_text,
+            "command_result": {"execution": execution},
+        }
+
+    async def _handle_external_connect_gate(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        captured_at: Optional[datetime],
+        intent_gate: dict[str, Any],
+        process_started_at: float,
+    ) -> dict[str, Any]:
+        occurred_at = captured_at or _utc_now()
+        if occurred_at.tzinfo is None:
+            occurred_at = occurred_at.replace(tzinfo=timezone.utc)
+        fields = intent_gate.get("fields") if isinstance(intent_gate.get("fields"), dict) else {}
+        blocking_reason = str(intent_gate.get("clarification_template") or "").strip() or "外部系统接入默认进入待决策队列。"
+        queued = await self._queue_command_intent_item(
+            message=message,
+            sender=sender,
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            feishu_message_id=feishu_message_id,
+            occurred_at=occurred_at,
+            intent_gate=intent_gate,
+            operation_code="source.connect",
+            operation_payload={"project_id": _safe_int(fields.get("project_id"))},
+            blocking_reason=blocking_reason,
+            decision_reason="external_connect_requires_manual_decision",
+            matched_project_id=_safe_int(fields.get("project_id")),
+        )
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=_safe_int(queued.get("item_id")),
+            raw_message=message,
+            ai_raw_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False),
+            project_score=1.0 if fields.get("project_id") is not None else None,
+            signal_confidence=0.0 if fields.get("project_id") is not None else None,
+            level="command",
+            routing_decision=ROUTING_INTENT_GATE_EXTERNAL_CONNECT,
+            routing_reason="外部系统接入被收口到待决策队列，不允许自由执行。",
+            reasoning_bullets=[
+                "intent gate 命中 external_connect。",
+                "当前仅允许受控工具链执行。",
+                "系统已转待决策，不做任意 HTTP/shell。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+            normalized_text=message,
+            route=ROUTING_INTENT_GATE_EXTERNAL_CONNECT,
+            intent=INTENT_EXTERNAL_CONNECT,
+            executor="router",
+            operation_code="source.connect",
+            reply_source="deterministic_reply",
+        )
+        return {
+            "status": "captured",
+            "classification": "command",
+            "project_scope": "command",
+            "linked_project_ids": [int(fields["project_id"])] if fields.get("project_id") is not None else [],
+            "signal_id": None,
+            "created_project_id": None,
+            "auto_tasks": [],
+            "ai_trace_id": f"feishu-external-{queued.get('item_id')}",
+            "results": [queued],
+            "intent": INTENT_EXTERNAL_CONNECT,
+            "clarification_message": queued.get("clarification_message"),
+            "direct_reply_required": True,
+            "direct_reply_text": str(queued.get("clarification_message") or blocking_reason),
+        }
+
+    async def process_message(
+        self,
+        *,
+        message_content: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str] = None,
+        feishu_message_id: Optional[str] = None,
+        captured_at: Optional[datetime] = None,
+        chat_type: Optional[str] = None,
+    ) -> dict[str, Any]:
+        process_started_at = time.perf_counter()
+        normalized_chat_type = self._normalize_chat_type(chat_type=chat_type, source_chat_id=source_chat_id)
+        message = self._normalize_incoming_text(message_content)
+        if not message:
+            return self._decorate_capture_result(
+                {"status": "ignored", "reason": "empty_message", "results": []},
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.empty_message",
+                db_mutations=[],
+            )
+
+        if self._is_water_talk_message(message):
+            duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=None,
+                project_score=None,
+                signal_confidence=None,
+                level=None,
+                routing_decision=ROUTING_NOISE_FILTERED,
+                routing_reason="识别为闲聊/问答水话，已过滤且不进入待处理队列。",
+                reasoning_bullets=[
+                    "消息缺少可执行交付信息。",
+                    "不满足需求信号的最小结构（目标/对象/动作）。",
+                    "已按噪音消息忽略，不触发入库与排队。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=duration_ms,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result({
+                "status": "ignored",
+                "reason": ROUTING_NOISE_FILTERED,
+                "classification": "non_requirement",
+                "project_scope": "non_requirement",
+                "linked_project_ids": [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-noise-{(feishu_message_id or 'noid')[:32]}",
+                "results": [],
+            }, chat_type=normalized_chat_type, normalized_text=message, route=f"{normalized_chat_type}.noise_filtered", db_mutations=[])
+
+        if feishu_message_id:
+            await self._acquire_message_lock(feishu_message_id)
+
+        if feishu_message_id and await self._has_processed_message(feishu_message_id):
+            duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=None,
+                project_score=None,
+                signal_confidence=None,
+                level=None,
+                routing_decision=ROUTING_DUPLICATE_FILTERED,
+                routing_reason="检测到重复飞书消息 ID，已过滤且不再重复执行 AI 处理。",
+                reasoning_bullets=[
+                    "消息唯一标识已存在于处理日志中。",
+                    "本次请求被标记为重复消息。",
+                    "系统已跳过后续分类和入库动作。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=duration_ms,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result({
+                "status": "ignored",
+                "reason": ROUTING_DUPLICATE_FILTERED,
+                "classification": "duplicate",
+                "project_scope": "duplicate",
+                "linked_project_ids": [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-duplicate-{feishu_message_id}",
+                "results": [],
+            }, chat_type=normalized_chat_type, normalized_text=message, route=f"{normalized_chat_type}.duplicate_filtered", db_mutations=[])
+
+        projects = await self._load_active_projects()
+        chat_context_resolution = {
+            "project_id": None,
+            "ambiguous": False,
+            "candidate_ids": [],
+            "top_ratio": 0.0,
+        }
+        if source_chat_id:
+            try:
+                chat_context_resolution = await self._resolve_chat_context_project(source_chat_id=source_chat_id)
+            except Exception:
+                chat_context_resolution = {
+                    "project_id": None,
+                    "ambiguous": False,
+                    "candidate_ids": [],
+                    "top_ratio": 0.0,
+                }
+        inferred_project_id = _safe_int(chat_context_resolution.get("project_id"))
+
+        followup_capture = await self._maybe_handle_queue_followup(
+            message=message,
+            sender=sender,
+            source_chat_id=source_chat_id,
+            feishu_message_id=feishu_message_id,
+            process_started_at=process_started_at,
+        )
+        if followup_capture is not None:
+            await self.db.commit()
+            return self._decorate_capture_result(
+                followup_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.followup",
+                db_mutations=self._summarize_db_mutations(followup_capture),
+            )
+
+        lookup_capture = await self._maybe_answer_project_lookup_question(
+            message=message,
+            feishu_message_id=feishu_message_id,
+            process_started_at=process_started_at,
+            projects=projects,
+        )
+        if lookup_capture is not None:
+            await self.db.commit()
+            return self._decorate_capture_result(
+                lookup_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.project_lookup",
+                db_mutations=[],
+            )
+
+        intent_gate = await self._resolve_intent_gate(
+            message=message,
+            inferred_project_id=inferred_project_id,
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            sender=sender,
+        )
+        logger.info(
+            "intent_resolved",
+            tenant_id=self.tenant_id,
+            feishu_message_id=feishu_message_id,
+            intent=str(intent_gate.get("intent") or ""),
+            canonical_action=str(intent_gate.get("canonical_action") or ""),
+            normalized_text=message[:400],
+        )
+        await self._maybe_write_intent_correction_sample(
+            message=message,
+            source_chat_id=source_chat_id,
+            intent_gate=intent_gate,
+        )
+
+        gate_intent = str(intent_gate.get("intent") or INTENT_SIGNAL_INTAKE)
+        if gate_intent == INTENT_SMALLTALK:
+            duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=json.dumps({"intent_gate": intent_gate}, ensure_ascii=False),
+                project_score=None,
+                signal_confidence=None,
+                level="smalltalk",
+                routing_decision=ROUTING_NOISE_FILTERED,
+                routing_reason="intent gate 识别为 smalltalk/短确认，不进入 Inbox。",
+                reasoning_bullets=[
+                    "消息不包含可执行主数据动作。",
+                    "命中 short confirmation / smalltalk。",
+                    "系统直接忽略，不写通知。",
+                ],
+                processing_status="success",
+                error_message=None,
+                duration_ms=duration_ms,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result(
+                {
+                    "status": "ignored",
+                    "reason": ROUTING_NOISE_FILTERED,
+                    "classification": "smalltalk",
+                    "project_scope": "smalltalk",
+                    "linked_project_ids": [],
+                    "signal_id": None,
+                    "created_project_id": None,
+                    "auto_tasks": [],
+                    "ai_trace_id": f"feishu-smalltalk-{(feishu_message_id or 'noid')[:32]}",
+                    "results": [],
+                },
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.intent_gate.smalltalk",
+                db_mutations=[],
+            )
+
+        if gate_intent == INTENT_HEALTHCHECK:
+            healthcheck_capture = await self._build_healthcheck_capture(
+                message=message,
+                feishu_message_id=feishu_message_id,
+                process_started_at=process_started_at,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result(
+                healthcheck_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.intent_gate.healthcheck",
+                db_mutations=[],
+            )
+
+        if gate_intent == INTENT_QUERY:
+            rag_service = AdaptiveRAGService(self.db, tenant_id=self.tenant_id)
+            conversation_key = self._build_rag_conversation_key(
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                sender=sender,
+            )
+            rag_result = await rag_service.answer_question(
+                question=message,
+                asked_by=None,
+                asked_via="feishu",
+                project_id=inferred_project_id,
+                source_chat_id=source_chat_id,
+                conversation_key=conversation_key,
+            )
+            qa_capture = await self._build_direct_qa_capture(
+                message=message,
+                rag_result=rag_result,
+                inferred_project_id=inferred_project_id,
+                feishu_message_id=feishu_message_id,
+                process_started_at=process_started_at,
+                route_reason="intent gate 判定为 query，直接走 RAG 直答通道，不进入决策队列。",
+                intent_label=INTENT_QUERY,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result(
+                qa_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.intent_gate.query",
+                db_mutations=[],
+            )
+
+        if gate_intent == INTENT_CREATE_PROJECT or (
+            gate_intent == INTENT_CLARIFICATION_REQUIRED and str(intent_gate.get("canonical_action") or "") == "project.create"
+        ):
+            project_capture = await self._handle_create_project_gate(
+                message=message,
+                sender=sender,
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                feishu_message_id=feishu_message_id,
+                captured_at=captured_at,
+                intent_gate=intent_gate,
+                process_started_at=process_started_at,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result(
+                project_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.intent_gate.project_create",
+                db_mutations=self._summarize_db_mutations(project_capture),
+            )
+
+        if gate_intent == INTENT_CREATE_TASK or (
+            gate_intent == INTENT_CLARIFICATION_REQUIRED and str(intent_gate.get("canonical_action") or "") == "task.create"
+        ):
+            task_capture = await self._handle_create_task_gate(
+                message=message,
+                sender=sender,
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                feishu_message_id=feishu_message_id,
+                captured_at=captured_at,
+                intent_gate=intent_gate,
+                process_started_at=process_started_at,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result(
+                task_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.intent_gate.task_create",
+                db_mutations=self._summarize_db_mutations(task_capture),
+            )
+
+        if gate_intent == INTENT_EXTERNAL_CONNECT:
+            external_capture = await self._handle_external_connect_gate(
+                message=message,
+                sender=sender,
+                source_chat_id=source_chat_id,
+                source_sender_open_id=source_sender_open_id,
+                feishu_message_id=feishu_message_id,
+                captured_at=captured_at,
+                intent_gate=intent_gate,
+                process_started_at=process_started_at,
+            )
+            await self.db.commit()
+            return self._decorate_capture_result(
+                external_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.intent_gate.external_connect",
+                db_mutations=self._summarize_db_mutations(external_capture),
+            )
+
+        command_capture = await self._maybe_execute_openclaw_command(
+            message=message,
+            sender=sender,
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            feishu_message_id=feishu_message_id,
+            captured_at=captured_at,
+            inferred_project_id=inferred_project_id,
+            projects=projects,
+            process_started_at=process_started_at,
+        )
+        if command_capture is not None:
+            await self.db.commit()
+            return self._decorate_capture_result(
+                command_capture,
+                chat_type=normalized_chat_type,
+                normalized_text=message,
+                route=f"{normalized_chat_type}.command",
+                db_mutations=self._summarize_db_mutations(command_capture),
+            )
+
+        rag_service = AdaptiveRAGService(self.db, tenant_id=self.tenant_id)
+        conversation_key = self._build_rag_conversation_key(
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            sender=sender,
+        )
+
+        sidecar_rag_result: Optional[dict[str, Any]] = None
+        intent_for_response = "statement"
+        force_signal_fastpath = False
+        rag_question_part = message
+
+        if self._is_obvious_question(message):
+            rag_result = await rag_service.answer_question(
+                question=rag_question_part,
+                asked_by=None,
+                asked_via="feishu",
+                project_id=inferred_project_id,
+                source_chat_id=source_chat_id,
+                conversation_key=conversation_key,
+            )
+            if self._is_requirement_like_question(message):
+                sidecar_rag_result = rag_result
+                intent_for_response = "question_requirement"
+                message = self._extract_requirement_statement_from_question(message)
+                if not message:
+                    message = self._normalize_incoming_text(message_content)
+                force_signal_fastpath = True
+            else:
+                qa_capture = await self._build_direct_qa_capture(
+                    message=message,
+                    rag_result=rag_result,
+                    inferred_project_id=inferred_project_id,
+                    feishu_message_id=feishu_message_id,
+                    process_started_at=process_started_at,
+                    route_reason="明显问句直接走 RAG 直答通道，不进入决策队列。",
+                    intent_label="question_obvious",
+                )
+                await self.db.commit()
+                return self._decorate_capture_result(
+                    qa_capture,
+                    chat_type=normalized_chat_type,
+                    normalized_text=message,
+                    route=f"{normalized_chat_type}.qa_direct",
+                    db_mutations=[],
+                )
+        else:
+            intent_result = await rag_service.classify_intent(message)
+            intent = str(intent_result.get("intent") or "statement").lower()
+            intent_for_response = intent
+            rag_question_part = str(intent_result.get("question_part") or "").strip() or message
+            rag_statement_part = str(intent_result.get("statement_part") or "").strip()
+
+            if intent == "question":
+                rag_result = await rag_service.answer_question(
+                    question=rag_question_part,
+                    asked_by=None,
+                    asked_via="feishu",
+                    project_id=inferred_project_id,
+                    source_chat_id=source_chat_id,
+                    conversation_key=conversation_key,
+                )
+                if self._is_requirement_like_question(message):
+                    sidecar_rag_result = rag_result
+                    intent_for_response = "question_requirement"
+                    message = self._extract_requirement_statement_from_question(message)
+                    if not message:
+                        message = self._normalize_incoming_text(message_content)
+                    force_signal_fastpath = True
+                else:
+                    qa_capture = await self._build_direct_qa_capture(
+                        message=message,
+                        rag_result=rag_result,
+                        inferred_project_id=inferred_project_id,
+                        feishu_message_id=feishu_message_id,
+                        process_started_at=process_started_at,
+                        route_reason="识别为问答消息，已走 RAG 直答通道，不进入决策队列。",
+                        intent_label=intent,
+                    )
+                    await self.db.commit()
+                    return self._decorate_capture_result(
+                        qa_capture,
+                        chat_type=normalized_chat_type,
+                        normalized_text=message,
+                        route=f"{normalized_chat_type}.qa_direct",
+                        db_mutations=[],
+                    )
+
+            if intent == "mixed":
+                mixed_rag_result = await rag_service.answer_question(
+                    question=rag_question_part,
+                    asked_by=None,
+                    asked_via="feishu",
+                    project_id=inferred_project_id,
+                    source_chat_id=source_chat_id,
+                    conversation_key=conversation_key,
+                )
+                if rag_statement_part:
+                    message = rag_statement_part
+                sidecar_rag_result = mixed_rag_result
+
+        try:
+            if force_signal_fastpath:
+                level_result = self._heuristic_level_fallback(message)
+            else:
+                level_result = await self._classify_level(message=message, projects=projects)
+            project_score = _safe_float(level_result.get("project_score"), 0.0)
+            level = _normalize_level_value(
+                level_result.get("level"),
+                default="project" if project_score >= 0.6 else "signal",
+            )
+            level_confidence = _safe_float(level_result.get("level_confidence"), project_score)
+            level_reasoning = str(level_result.get("level_reasoning") or "").strip()
+
+            split_items = []
+            seen_split_items: set[tuple[str, str]] = set()
+            raw_split_items = level_result.get("split_items")
+            should_split = bool(level_result.get("should_split"))
+            if should_split and isinstance(raw_split_items, list):
+                for item in raw_split_items[:2]:
+                    if not isinstance(item, dict):
+                        continue
+                    item_level = _normalize_level_value(item.get("level"), default=level)
+                    item_content = str(item.get("content") or "").strip()
+                    if item_level in {"project", "signal"} and item_content:
+                        dedupe_key = (item_level, item_content)
+                        if dedupe_key in seen_split_items:
+                            continue
+                        seen_split_items.add(dedupe_key)
+                        split_items.append({"level": item_level, "content": item_content})
+            if not split_items:
+                split_items = [{"level": level if level in {"project", "signal"} else "signal", "content": message}]
+
+            routing_policy = await self._load_routing_policy()
+            occurred_at = captured_at or _utc_now()
+            if occurred_at.tzinfo is None:
+                occurred_at = occurred_at.replace(tzinfo=timezone.utc)
+
+            results: list[dict[str, Any]] = []
+            parent_item_id: Optional[int] = None
+            for idx, split in enumerate(split_items):
+                item_result = await self._process_one_item(
+                    content=split["content"],
+                    level=split["level"],
+                    sender=sender,
+                    source_chat_id=source_chat_id,
+                    source_sender_open_id=source_sender_open_id,
+                    feishu_message_id=feishu_message_id,
+                    occurred_at=occurred_at,
+                    level_result=level_result,
+                    level_confidence=level_confidence,
+                    level_reasoning=level_reasoning,
+                    project_score=project_score,
+                    cold_start_enabled=bool(routing_policy["cold_start_enabled"]),
+                    signal_auto_enter_threshold=float(routing_policy["signal_auto_enter_threshold"]),
+                    project_auto_create_threshold=float(routing_policy["project_auto_create_threshold"]),
+                    project_uncertain_threshold=float(routing_policy["project_uncertain_threshold"]),
+                    chat_context_auto_link_enabled=bool(routing_policy["chat_context_auto_link_enabled"]),
+                    projects=projects,
+                    split_from_item_id=parent_item_id if idx > 0 else None,
+                    chat_type=normalized_chat_type,
+                    chat_context_resolution=chat_context_resolution,
+                )
+                if idx == 0 and item_result.get("item_id"):
+                    parent_item_id = int(item_result["item_id"])
+                results.append(item_result)
+
+            await self.db.commit()
+
+            signal_ids = [int(x["signal_id"]) for x in results if x.get("signal_id")]
+            linked_projects = sorted(
+                {
+                    int(x["project_id"])
+                    for x in results
+                    if x.get("project_id") is not None
+                }
+            )
+            classification = "split" if len(results) > 1 else (results[0].get("level") or "signal")
+            clarification_message = next(
+                (
+                    str(row.get("clarification_message") or "").strip()
+                    for row in results
+                    if isinstance(row, dict) and str(row.get("clarification_message") or "").strip()
+                ),
+                "",
+            )
+            direct_reply_text: Optional[str] = None
+            if sidecar_rag_result:
+                sidecar_answer = str(sidecar_rag_result.get("answer") or "").strip()
+                if sidecar_answer:
+                    direct_reply_text = (
+                        self._build_rag_reply_text(
+                            answer=sidecar_answer,
+                            sources=sidecar_rag_result.get("sources") or [],
+                        )
+                        + "\n\n我已回答你的问题，并把需求部分记录为信号进入处理队列。"
+                    )
+            return self._decorate_capture_result({
+                "status": "captured",
+                "classification": classification,
+                "project_scope": classification,
+                "linked_project_ids": linked_projects,
+                "signal_id": str(signal_ids[0]) if signal_ids else None,
+                "created_project_id": linked_projects[0] if linked_projects else None,
+                "auto_tasks": [],
+                "ai_trace_id": f"feishu-item-{results[0].get('item_id')}" if results else None,
+                "results": results,
+                "intent": intent_for_response,
+                "rag_result": sidecar_rag_result,
+                "direct_reply_required": bool(direct_reply_text),
+                "direct_reply_text": direct_reply_text,
+                "clarification_message": clarification_message or None,
+            }, chat_type=normalized_chat_type, normalized_text=message, route=f"{normalized_chat_type}.{classification}", db_mutations=self._summarize_db_mutations({"results": results, "signal_id": str(signal_ids[0]) if signal_ids else None, "created_project_id": linked_projects[0] if linked_projects else None}))
+        except IntegrityError as exc:
+            await self.db.rollback()
+            if feishu_message_id and self._is_duplicate_integrity_error(exc):
+                duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+                await self._write_processing_log(
+                    feishu_message_id=feishu_message_id,
+                    feishu_item_id=None,
+                    raw_message=message,
+                    ai_raw_output=None,
+                    project_score=None,
+                    signal_confidence=None,
+                    level=None,
+                    routing_decision=ROUTING_DUPLICATE_FILTERED,
+                    routing_reason="唯一约束命中，判定为重复飞书消息，已跳过重复处理。",
+                    reasoning_bullets=[
+                        "数据库唯一约束阻止了重复入库。",
+                        "当前请求被收敛到 duplicate 模板回执。",
+                        "后续不再触发 AI 分类与入库动作。",
+                    ],
+                    processing_status="success",
+                    error_message=None,
+                    duration_ms=duration_ms,
+                )
+                await self.db.commit()
+                return self._decorate_capture_result({
+                    "status": "ignored",
+                    "reason": ROUTING_DUPLICATE_FILTERED,
+                    "classification": "duplicate",
+                    "project_scope": "duplicate",
+                    "linked_project_ids": [],
+                    "signal_id": None,
+                    "created_project_id": None,
+                    "auto_tasks": [],
+                    "ai_trace_id": f"feishu-duplicate-{feishu_message_id}",
+                    "results": [],
+                }, chat_type=normalized_chat_type, normalized_text=message, route=f"{normalized_chat_type}.duplicate_integrity", db_mutations=[])
+
+            duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=None,
+                project_score=None,
+                signal_confidence=None,
+                level=None,
+                routing_decision=ROUTING_FAILED,
+                routing_reason="数据库完整性校验失败，已停止当前消息处理。",
+                reasoning_bullets=[
+                    f"异常类型：{exc.__class__.__name__}",
+                    "可能原因：唯一约束/外键约束冲突。",
+                    "请结合 SQL 日志排查数据一致性问题。",
+                ],
+                processing_status="failed",
+                error_message=str(exc)[:2000] or "integrity_error",
+                duration_ms=duration_ms,
+            )
+            await self.db.commit()
+            raise
+        except Exception as exc:
+            await self.db.rollback()
+            duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=None,
+                raw_message=message,
+                ai_raw_output=None,
+                project_score=None,
+                signal_confidence=None,
+                level=None,
+                routing_decision=ROUTING_FAILED,
+                routing_reason="AI 处理链路执行失败，已停止当前消息处理并写入错误日志。",
+                reasoning_bullets=[
+                    f"异常类型：{exc.__class__.__name__}",
+                    f"异常信息：{str(exc)[:120] or '未知错误'}",
+                    "请在日志详情中查看完整错误并重试。",
+                ],
+                processing_status="failed",
+                error_message=str(exc)[:2000] or "unknown_error",
+                duration_ms=duration_ms,
+            )
+            await self.db.commit()
+            raise
+
+    @staticmethod
+    def _is_duplicate_integrity_error(exc: IntegrityError) -> bool:
+        message = str(exc or "").lower()
+        if "feishu_items" not in message:
+            return False
+        duplicate_hints = (
+            "uq_feishu_items_tenant_message_root",
+            "uq_feishu_items_tenant_message_split",
+            "unique constraint",
+            "duplicate key value",
+            "is not unique",
+        )
+        return any(hint in message for hint in duplicate_hints)
+
+    @staticmethod
+    def _build_rag_reply_text(*, answer: str, sources: list[dict[str, Any]]) -> str:
+        clean_answer = str(answer or "").strip() or "我未找到足够证据，建议补充更多上下文。"
+        labels: list[str] = []
+        deep_link = ""
+        for source in sources[:4]:
+            if not isinstance(source, dict):
+                continue
+            label = str(source.get("label") or source.get("source") or "").strip()
+            if label and label not in labels:
+                labels.append(label)
+            if not deep_link:
+                meta = source.get("metadata") if isinstance(source.get("metadata"), dict) else {}
+                target_url = str(meta.get("target_url") or "").strip()
+                if not target_url:
+                    project_id = meta.get("project_id")
+                    if project_id is not None:
+                        try:
+                            target_url = f"/projects/{int(project_id)}/board"
+                        except Exception:
+                            target_url = ""
+                if target_url.startswith("/"):
+                    deep_link = target_url
+        source_text = " / ".join(labels) if labels else "未命中可引用来源"
+        if deep_link:
+            return f"{clean_answer}\n\n来源: {source_text}\nView details → {deep_link}"
+        return f"{clean_answer}\n\n来源: {source_text}"
+
+    @staticmethod
+    def _normalize_chat_type(*, chat_type: Optional[str], source_chat_id: str) -> str:
+        text = str(chat_type or "").strip().lower()
+        if text in {"p2p", "single", "single_chat", "dm", "direct"}:
+            return "p2p"
+        if text in {"group", "chat", "group_chat"}:
+            return "group"
+        return "group" if source_chat_id else "p2p"
+
+    @staticmethod
+    def _normalize_incoming_text(message_content: Optional[str]) -> str:
+        text = str(message_content or "").replace("\r\n", "\n").replace("\r", "\n")
+        if not text:
+            return ""
+        text = FEISHU_BOT_MENTION_RE.sub(" ", text)
+        text = FEISHU_QUOTE_LINE_RE.sub(" ", text)
+        text = FEISHU_QUOTE_PREFIX_RE.sub("", text)
+        text = FEISHU_NOISE_TOKEN_RE.sub(" ", text)
+        text = re.sub(r"@所有人|@everyone", " ", text, flags=re.IGNORECASE)
+        text = re.sub(r"[ \t]+", " ", text)
+        text = re.sub(r"\n{2,}", "\n", text)
+        text = re.sub(r"\s+", " ", text).strip()
+        return text[:2000]
+
+    @classmethod
+    def _is_obvious_question(cls, message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        if cls._is_openclaw_business_command(text):
+            return False
+        if cls._looks_like_project_lookup_question(text):
+            return True
+        return bool(OBVIOUS_QUESTION_RE.search(text))
+
+    @staticmethod
+    def _decorate_capture_result(
+        result: dict[str, Any],
+        *,
+        chat_type: str,
+        normalized_text: str,
+        route: str,
+        db_mutations: list[dict[str, Any]],
+    ) -> dict[str, Any]:
+        payload = dict(result or {})
+        direct_reply_text = str(payload.get("direct_reply_text") or "").strip()
+        if direct_reply_text:
+            payload["direct_reply_text"] = FeishuSignalPipelineService._normalize_direct_reply_text(direct_reply_text)
+        clarification_message = str(payload.get("clarification_message") or "").strip()
+        if clarification_message:
+            payload["clarification_message"] = FeishuSignalPipelineService._normalize_direct_reply_text(clarification_message)
+        followup_reply = str(payload.get("followup_reply") or "").strip()
+        if followup_reply:
+            payload["followup_reply"] = FeishuSignalPipelineService._normalize_direct_reply_text(followup_reply)
+        payload["chat_type"] = chat_type
+        payload["normalized_text"] = normalized_text
+        payload["route"] = route
+        payload["db_mutations"] = db_mutations
+        payload.setdefault("results", [])
+        return payload
+
+    @staticmethod
+    def _normalize_direct_reply_text(text: str) -> str:
+        cleaned = str(text or "").strip()
+        if cleaned.startswith("```"):
+            cleaned = cleaned.replace("```", "").strip()
+        if cleaned.lower().startswith("json"):
+            cleaned = cleaned[4:].strip()
+        if not cleaned:
+            return "系统处理中出现异常，本次未执行。请稍后重试，或补充更明确的目标、对象和动作。"
+        lowered = cleaned.lower()
+        provider_error_signals = (
+            "request was aborted",
+            "request aborted",
+            "connection aborted",
+            "operation aborted",
+            "context canceled",
+            "context cancelled",
+            "cancelled",
+            "canceled",
+            "rate limit",
+            "too many requests",
+            "quota",
+            "insufficient_quota",
+            "insufficient quota",
+            "authentication failed",
+            "service unavailable",
+            "upstream error",
+            "bad gateway",
+            "gateway timeout",
+            "internal server error",
+            "connection reset",
+            "econnreset",
+            "socket hang up",
+        )
+        if any(flag in lowered for flag in provider_error_signals):
+            return "系统处理中出现异常，本次未执行。请稍后重试，或补充更明确的目标、对象和动作。"
+        return cleaned[:500]
+
+    @staticmethod
+    def _summarize_db_mutations(result: dict[str, Any]) -> list[dict[str, Any]]:
+        if not isinstance(result, dict):
+            return []
+        mutations: list[dict[str, Any]] = []
+        command_result = result.get("command_result") if isinstance(result.get("command_result"), dict) else None
+        execution = command_result.get("execution") if isinstance(command_result, dict) else None
+        execution_results = execution.get("results") if isinstance(execution, dict) else None
+        if isinstance(execution_results, list):
+            for row in execution_results[:6]:
+                if not isinstance(row, dict):
+                    continue
+                mutations.append(
+                    {
+                        "operation_code": str(row.get("operation_code") or ""),
+                        "status": str(row.get("status") or ""),
+                        "entity_type": str(row.get("entity_type") or ""),
+                        "entity_id": str(row.get("entity_id") or ""),
+                    }
+                )
+            if mutations:
+                return mutations
+
+        rows = result.get("results")
+        if isinstance(rows, list):
+            for row in rows[:6]:
+                if not isinstance(row, dict):
+                    continue
+                mutations.append(
+                    {
+                        "item_id": row.get("item_id"),
+                        "level": row.get("level"),
+                        "route_status": row.get("route_status"),
+                        "project_id": row.get("project_id"),
+                        "signal_id": row.get("signal_id"),
+                    }
+                )
+        return mutations
+
+    async def _build_direct_qa_capture(
+        self,
+        *,
+        message: str,
+        rag_result: dict[str, Any],
+        inferred_project_id: Optional[int],
+        feishu_message_id: Optional[str],
+        process_started_at: float,
+        route_reason: str,
+        intent_label: str,
+    ) -> dict[str, Any]:
+        duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=None,
+            raw_message=message,
+            ai_raw_output=json.dumps({"intent": intent_label, "rag": rag_result}, ensure_ascii=False)[:4000],
+            project_score=None,
+            signal_confidence=None,
+            level="question",
+            routing_decision=ROUTING_DIRECT_QA,
+            routing_reason=route_reason,
+            reasoning_bullets=[
+                f"intent={intent_label}",
+                "命中 question 路由，跳过 signal/project 处理。",
+                "回答已记录 QA 训练数据。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=duration_ms,
+        )
+        return {
+            "status": "answered",
+            "classification": "question",
+            "project_scope": "question",
+            "linked_project_ids": [int(inferred_project_id)] if inferred_project_id is not None else [],
+            "signal_id": None,
+            "created_project_id": int(inferred_project_id) if inferred_project_id is not None else None,
+            "auto_tasks": [],
+            "ai_trace_id": f"feishu-qa-{(feishu_message_id or 'noid')[:32]}",
+            "results": [],
+            "rag_result": rag_result,
+            "direct_reply_required": True,
+            "direct_reply_text": self._build_rag_reply_text(
+                answer=str(rag_result.get("answer") or ""),
+                sources=rag_result.get("sources") or [],
+            ),
+        }
+
+    async def _maybe_handle_queue_followup(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        feishu_message_id: Optional[str],
+        process_started_at: float,
+    ) -> Optional[dict[str, Any]]:
+        if not source_chat_id:
+            return None
+        wants_continue = self._looks_like_continue_followup(message)
+        wants_status = bool(FOLLOWUP_STATUS_RE.search(message))
+        if not wants_continue and not wants_status:
+            return None
+
+        recent_items = await self._load_recent_chat_items(source_chat_id=source_chat_id)
+        if not recent_items:
+            reply_text = "当前群里还没有可追踪的飞书处理项。"
+            decision = ROUTING_QUEUE_FOLLOWUP_BLOCKED
+            reason = "群上下文中不存在可追踪的最近处理项。"
+            classification = "question"
+            status = "answered"
+            mutations: list[dict[str, Any]] = []
+            results: list[dict[str, Any]] = []
+        elif wants_continue:
+            pending_items = [item for item in recent_items if self._normalize_route_status(item.route_status) == ROUTE_STATUS_QUEUED]
+            if len(pending_items) != 1:
+                reply_text = self._build_followup_pending_clarification(pending_items or recent_items[:3])
+                decision = ROUTING_QUEUE_FOLLOWUP_BLOCKED
+                reason = "群上下文中存在多个待处理项，无法自动确定继续处理目标。"
+                classification = "command"
+                status = "captured"
+                mutations = []
+                results = []
+            else:
+                operator_user = await self._resolve_command_operator_user(sender=sender)
+                if operator_user is None:
+                    reply_text = "未执行：当前租户下没有可用操作者，无法继续处理该项。"
+                    decision = ROUTING_QUEUE_FOLLOWUP_BLOCKED
+                    reason = "继续处理被阻断：当前租户下没有可用操作者。"
+                    classification = "command"
+                    status = "captured"
+                    mutations = []
+                    results = []
+                else:
+                    target_item = pending_items[0]
+                    if (target_item.level or "").lower() == "project":
+                        suggestion = await self._get_project_suggestion(int(target_item.id))
+                        if suggestion is None:
+                            reply_text = "未执行：最近待处理项目候选缺少建议记录，无法继续处理。"
+                            decision = ROUTING_QUEUE_FOLLOWUP_BLOCKED
+                            reason = "继续处理被阻断：最近待处理项目候选缺少建议记录。"
+                            classification = "command"
+                            status = "captured"
+                            mutations = []
+                            results = []
+                        else:
+                            candidate_fields = self._resolve_project_candidate_fields(item=target_item, suggestion=suggestion)
+                            if not candidate_fields["ready"]:
+                                reply_text = (
+                                    str(target_item.blocking_reason or "").strip()
+                                    or self._build_project_clarification_message(candidate_fields["missing_fields"])
+                                )
+                                decision = ROUTING_QUEUE_FOLLOWUP_BLOCKED
+                                reason = "继续处理被阻断：最近待处理建项目项仍缺少必填字段。"
+                                classification = "command"
+                                status = "captured"
+                                mutations = []
+                                results = []
+                            else:
+                                confirm_result = await self.confirm_item(item_id=int(target_item.id), user_id=int(operator_user.id))
+                                reply_text = self._build_followup_confirmation_reply(target_item=target_item, confirm_result=confirm_result)
+                                decision = ROUTING_QUEUE_FOLLOWUP_CONFIRM
+                                reason = "命中继续处理指令，已按最近唯一待处理项继续执行。"
+                                classification = "command"
+                                status = "captured"
+                                mutations = self._summarize_db_mutations(
+                                    {
+                                        "results": [
+                                            {
+                                                "item_id": int(target_item.id),
+                                                "level": target_item.level,
+                                                "route_status": ROUTE_STATUS_AUTO_DONE,
+                                                "project_id": confirm_result.get("project_id"),
+                                                "signal_id": confirm_result.get("signal_id"),
+                                            }
+                                        ]
+                                    }
+                                )
+                                results = [
+                                    {
+                                        "item_id": int(target_item.id),
+                                        "level": target_item.level,
+                                        "route_status": ROUTE_STATUS_AUTO_DONE,
+                                        "project_id": _safe_int(confirm_result.get("project_id")),
+                                        "signal_id": _safe_int(confirm_result.get("signal_id")),
+                                    }
+                                ]
+                    else:
+                        signal = await self._get_signal_by_item(int(target_item.id))
+                        if (target_item.level or "").lower() == "signal" and not (signal and (signal.matched_project_id or target_item.matched_project_id)):
+                            reply_text = "未执行：最近待处理信号仍缺少明确项目归属，请先补充项目名或项目 ID。"
+                            decision = ROUTING_QUEUE_FOLLOWUP_BLOCKED
+                            reason = "继续处理被阻断：最近待处理信号缺少项目归属。"
+                            classification = "command"
+                            status = "captured"
+                            mutations = []
+                            results = []
+                        else:
+                            confirm_result = await self.confirm_item(item_id=int(target_item.id), user_id=int(operator_user.id))
+                            reply_text = self._build_followup_confirmation_reply(target_item=target_item, confirm_result=confirm_result)
+                            decision = ROUTING_QUEUE_FOLLOWUP_CONFIRM
+                            reason = "命中继续处理指令，已按最近唯一待处理项继续执行。"
+                            classification = "command"
+                            status = "captured"
+                            mutations = self._summarize_db_mutations(
+                                {
+                                    "results": [
+                                        {
+                                            "item_id": int(target_item.id),
+                                            "level": target_item.level,
+                                            "route_status": ROUTE_STATUS_AUTO_DONE,
+                                            "project_id": confirm_result.get("project_id"),
+                                            "signal_id": confirm_result.get("signal_id"),
+                                        }
+                                    ]
+                                }
+                            )
+                            results = [
+                                {
+                                    "item_id": int(target_item.id),
+                                    "level": target_item.level,
+                                    "route_status": ROUTE_STATUS_AUTO_DONE,
+                                    "project_id": _safe_int(confirm_result.get("project_id")),
+                                    "signal_id": _safe_int(confirm_result.get("signal_id")),
+                                }
+                            ]
+        else:
+            target_item = self._select_followup_status_target(message=message, recent_items=recent_items)
+            reply_text = await self._build_followup_status_reply(target_item)
+            decision = ROUTING_QUEUE_FOLLOWUP_QUERY
+            reason = "命中进度/状态追问，已直接基于真实 DB 返回最近处理项状态。"
+            classification = "question"
+            status = "answered"
+            mutations = []
+            results = []
+
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=None,
+            raw_message=message,
+            ai_raw_output=json.dumps(
+                {
+                    "followup_reply": reply_text,
+                    "recent_item_ids": [int(item.id) for item in recent_items[:3]],
+                    "wants_continue": wants_continue,
+                    "wants_status": wants_status,
+                },
+                ensure_ascii=False,
+            )[:4000],
+            project_score=None,
+            signal_confidence=None,
+            level="followup",
+            routing_decision=decision,
+            routing_reason=reason,
+            reasoning_bullets=[
+                "follow-up 查询/指令命中最近飞书上下文。",
+                "回复内容只基于真实 DB 状态，不使用自由生成。",
+                f"recent_items={len(recent_items)}",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+        )
+
+        latest = recent_items[0] if recent_items else None
+        linked_project_ids: list[int] = []
+        if results and results[0].get("project_id") is not None:
+            linked_project_ids.append(int(results[0]["project_id"]))
+        if latest and latest.matched_project_id is not None:
+            linked_project_ids.append(int(latest.matched_project_id))
+        linked_project_ids = sorted(set(linked_project_ids))
+        return {
+            "status": status,
+            "classification": classification,
+            "project_scope": classification,
+            "linked_project_ids": linked_project_ids,
+            "signal_id": str(results[0].get("signal_id")) if results else None,
+            "created_project_id": str(results[0].get("project_id")) if results else None,
+            "auto_tasks": [],
+            "ai_trace_id": f"feishu-followup-{(feishu_message_id or 'noid')[:32]}",
+            "results": results,
+            "direct_reply_required": True,
+            "direct_reply_text": reply_text,
+            "db_mutations": mutations,
+        }
+
+    async def _load_recent_chat_items(self, *, source_chat_id: str, limit: int = 5) -> list[FeishuItem]:
+        rows = (
+            await self.db.execute(
+                select(FeishuItem)
+                .where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.source_chat_id == (source_chat_id or "")[:128],
+                    )
+                )
+                .order_by(FeishuItem.id.desc())
+                .limit(limit)
+            )
+        ).scalars().all()
+        return list(rows)
+
+    @staticmethod
+    def _extract_item_operation_code(item: FeishuItem) -> str:
+        payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+        return str(payload.get("operation_code") or "").strip().lower()
+
+    @classmethod
+    def _select_followup_status_target(cls, *, message: str, recent_items: list[FeishuItem]) -> FeishuItem:
+        if not recent_items:
+            raise ValueError("recent_items_empty")
+        normalized = re.sub(r"\s+", "", str(message or "")).lower()
+        wants_project_create = any(token in normalized for token in ("创建", "建好", "建成", "项目"))
+        wants_source_status = any(token in normalized for token in ("信号源", "web搜索", "webhook", "source"))
+
+        if wants_source_status:
+            for item in recent_items:
+                operation_code = cls._extract_item_operation_code(item)
+                if operation_code == "source.connect" or "信号源" in str(item.raw_message or ""):
+                    return item
+
+        if wants_project_create:
+            for item in recent_items:
+                operation_code = cls._extract_item_operation_code(item)
+                level = str(item.level or "").strip().lower()
+                card_type = str(item.card_type or "").strip().lower()
+                if (
+                    level == "project"
+                    or operation_code == "project.create"
+                    or card_type in {"project_create", "new_project", "new_project_candidate"}
+                ):
+                    return item
+
+        return recent_items[0]
+
+    @staticmethod
+    def _build_followup_pending_clarification(items: list[FeishuItem]) -> str:
+        if not items:
+            return "当前群没有待处理项可继续。"
+        fragments = []
+        for item in items[:3]:
+            level = str(item.level or "").lower()
+            label = "项目候选" if level == "project" else ("命令" if level == "command" else "信号")
+            fragments.append(f"{label}#{int(item.id)}")
+        return f"当前群最近有多个待处理项：{'、'.join(fragments)}。请先明确要继续哪一条。"
+
+    def _build_followup_confirmation_reply(self, *, target_item: FeishuItem, confirm_result: dict[str, Any]) -> str:
+        level = str(target_item.level or "").lower()
+        if level == "project":
+            project_id = confirm_result.get("project_id")
+            return f"已继续处理：项目候选#{int(target_item.id)} 已确认并创建项目#{project_id}。"
+        if level == "command":
+            project_id = confirm_result.get("project_id")
+            task_id = confirm_result.get("task_id")
+            if task_id and project_id:
+                return f"已继续处理：任务#{task_id} 已在项目#{project_id} 中创建。"
+            if project_id:
+                return f"已继续处理：命令队列项#{int(target_item.id)} 已落到项目#{project_id}。"
+            return f"已继续处理：命令队列项#{int(target_item.id)} 已确认。"
+        signal_id = confirm_result.get("signal_id")
+        project_id = confirm_result.get("project_id")
+        if signal_id and project_id:
+            return f"已继续处理：信号#{signal_id} 已确认并关联项目#{project_id}。"
+        return f"已继续处理：队列项#{int(target_item.id)} 已确认。"
+
+    async def _build_followup_status_reply(self, item: FeishuItem) -> str:
+        route_status = self._normalize_route_status(item.route_status)
+        level = str(item.level or "").strip().lower()
+        if level == "project":
+            suggestion = await self._get_project_suggestion(int(item.id))
+            project_id = suggestion.created_project_id if suggestion and suggestion.created_project_id else item.matched_project_id
+            if route_status == ROUTE_STATUS_AUTO_DONE and project_id:
+                return f"最近一条项目候选已确认，项目#{int(project_id)} 已落库。"
+            if route_status == ROUTE_STATUS_UNDONE:
+                return "最近一条项目候选已撤销，系统没有保留该次自动结果。"
+            return "最近一条项目候选仍在待确认，系统还没有落正式项目。"
+
+        if level == "command":
+            payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+            operation_code = str(payload.get("operation_code") or "").strip()
+            project_id = _safe_int(payload.get("project_id") or item.matched_project_id)
+            summary = str(payload.get("result_summary") or "").strip()
+            if route_status == ROUTE_STATUS_AUTO_DONE:
+                return summary or (
+                    f"最近一条命令已执行完成，归属项目#{int(project_id)}。"
+                    if project_id is not None
+                    else "最近一条命令已执行完成。"
+                )
+            if route_status == ROUTE_STATUS_UNDONE:
+                return "最近一条命令结果已撤销。"
+            if route_status == ROUTE_STATUS_QUEUED:
+                return (
+                    f"最近一条命令 {operation_code or '待处理动作'} 已入队待确认。"
+                    if operation_code
+                    else "最近一条命令已入队待确认。"
+                )
+            return "最近一条命令仍在处理中。"
+
+        signal = await self._get_signal_by_item(int(item.id))
+        signal_id = int(signal.id) if signal is not None else None
+        project_id = int(signal.matched_project_id or item.matched_project_id) if (signal and (signal.matched_project_id or item.matched_project_id)) else None
+        if route_status == ROUTE_STATUS_AUTO_DONE and signal_id and project_id:
+            return f"最近一条信号#{signal_id} 已确认并关联项目#{project_id}。"
+        if route_status == ROUTE_STATUS_QUEUED and signal_id and project_id:
+            return f"最近一条信号#{signal_id} 已入队待确认，候选项目#{project_id}。"
+        if route_status == ROUTE_STATUS_QUEUED and signal_id:
+            return f"最近一条信号#{signal_id} 已入队待确认，但项目归属仍不明确。"
+        if route_status == ROUTE_STATUS_UNDONE:
+            return "最近一条信号结果已撤销。"
+        return "最近一条飞书处理项仍在处理中，尚未得到可确认的 DB 状态。"
+
+    @staticmethod
+    def _looks_like_project_create_command(message: str) -> bool:
+        return bool(OPENCLAW_PROJECT_CREATE_RE.search(str(message or "")))
+
+    @classmethod
+    def _looks_like_project_lookup_question(cls, message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        if cls._is_openclaw_business_command(text):
+            return False
+        normalized = re.sub(r"\s+", "", text.lower())
+        if PROJECT_LIST_QUERY_RE.search(text):
+            return True
+        if PROJECT_ID_QUERY_RE.search(text):
+            return True
+        if "项目" not in normalized:
+            return False
+        return any(token in normalized for token in ("有吗", "有没有", "有无", "是否有", "是否存在", "存在吗", "是什么", "哪些"))
+
+    @classmethod
+    def _is_openclaw_business_command(cls, message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        if cls._looks_like_project_create_command(text):
+            return True
+        return bool(OPENCLAW_COMMAND_VERB_RE.search(text) and OPENCLAW_COMMAND_OBJECT_RE.search(text))
+
+    async def _maybe_execute_openclaw_command(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        captured_at: Optional[datetime],
+        inferred_project_id: Optional[int],
+        projects: list[dict[str, Any]],
+        process_started_at: float,
+    ) -> Optional[dict[str, Any]]:
+        if not self._is_openclaw_business_command(message):
+            return None
+
+        operator_user = await self._resolve_command_operator_user(sender=sender)
+        if operator_user is None:
+            return await self._build_openclaw_blocked_capture(
+                message=message,
+                feishu_message_id=feishu_message_id,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                reason="当前租户下没有可用操作者，系统未执行写操作。",
+            )
+
+        target_project_id = inferred_project_id or self._guess_matched_project_id_from_text(message, projects)
+        commander = OpenClawCommanderService(self.db)
+        if self._looks_like_project_create_command(message):
+            command_result = await commander.dispatch_global_command(
+                tenant_id=self.tenant_id,
+                operator_user=operator_user,
+                instruction=message,
+                dry_run=False,
+                max_operations=4,
+                response_mode="fast",
+                persist_run=False,
+            )
+        elif target_project_id is not None:
+            command_result = await commander.dispatch_command(
+                tenant_id=self.tenant_id,
+                operator_user=operator_user,
+                project_id=int(target_project_id),
+                instruction=message,
+                dry_run=False,
+                max_operations=6,
+                response_mode="fast",
+                persist_run=False,
+            )
+        else:
+            return await self._build_openclaw_blocked_capture(
+                message=message,
+                feishu_message_id=feishu_message_id,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                reason="缺少明确项目上下文。请指明项目名，或在对应项目群内发送该指令。",
+            )
+
+        execution = command_result.get("execution") if isinstance(command_result, dict) else {}
+        requested_count = int((execution or {}).get("requested_count") or 0)
+        if requested_count <= 0 and int(command_result.get("planned_operations") or 0) <= 0:
+            return await self._build_openclaw_blocked_capture(
+                message=message,
+                feishu_message_id=feishu_message_id,
+                duration_ms=int((time.perf_counter() - process_started_at) * 1000),
+                reason="未识别到可安全执行的业务操作，系统未写库。",
+            )
+
+        linked_project_ids = self._collect_command_project_ids(command_result)
+        created_project_id = self._extract_created_project_id(command_result)
+        primary_project_id = created_project_id or (linked_project_ids[0] if linked_project_ids else None)
+        route_status = self._resolve_command_route_status(command_result)
+        item_id = await self._record_command_context_item(
+            message=message,
+            sender=sender,
+            source_chat_id=source_chat_id,
+            source_sender_open_id=source_sender_open_id,
+            feishu_message_id=feishu_message_id,
+            captured_at=captured_at,
+            matched_project_id=primary_project_id,
+            route_status=route_status,
+            command_result=command_result,
+        )
+        reply_text = self._build_openclaw_command_reply(command_result)
+        duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=item_id,
+            raw_message=message,
+            ai_raw_output=json.dumps({"command_result": command_result}, ensure_ascii=False)[:120000],
+            project_score=1.0 if primary_project_id is not None else None,
+            signal_confidence=0.0,
+            level="command",
+            routing_decision=ROUTING_OPENCLAW_EXECUTED,
+            routing_reason="识别为明确业务命令，已直接进入 OpenClaw command 层并按真实执行结果回执。",
+            reasoning_bullets=[
+                "命中 command-first 路由，跳过 RAG 问答与信号分流。",
+                f"planner_status={str(command_result.get('planner_status') or '') or 'unknown'}",
+                f"openclaw_status={str(command_result.get('openclaw_status') or '') or 'unknown'}",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=duration_ms,
+        )
+        return {
+            "status": "captured",
+            "classification": "command",
+            "project_scope": "command",
+            "linked_project_ids": linked_project_ids,
+            "signal_id": None,
+            "created_project_id": created_project_id or primary_project_id,
+            "auto_tasks": [],
+            "ai_trace_id": str(command_result.get("trace_id") or ""),
+            "results": [
+                {
+                    "item_id": item_id,
+                    "level": "command",
+                    "project_id": primary_project_id,
+                    "signal_id": None,
+                    "route_status": route_status,
+                }
+            ],
+            "intent": "command",
+            "direct_reply_required": True,
+            "direct_reply_text": reply_text,
+            "command_result": command_result,
+        }
+
+    async def _maybe_answer_project_lookup_question(
+        self,
+        *,
+        message: str,
+        feishu_message_id: Optional[str],
+        process_started_at: float,
+        projects: list[dict[str, Any]],
+    ) -> Optional[dict[str, Any]]:
+        if not self._looks_like_project_lookup_question(message):
+            return None
+
+        linked_project_ids: list[int] = []
+        reply_text = ""
+        id_match = PROJECT_ID_QUERY_RE.search(message)
+        if id_match:
+            target_id = int(id_match.group(1))
+            matched = next((item for item in projects if _safe_int(item.get("id")) == target_id), None)
+            if matched is None:
+                reply_text = f"项目#{target_id} 不存在，或当前租户下不可见。"
+            else:
+                linked_project_ids = [target_id]
+                reply_text = self._format_project_lookup_hit_rows([matched], prefix="查到了")
+        elif PROJECT_LIST_QUERY_RE.search(message):
+            if not projects:
+                reply_text = "当前没有可见项目。"
+            else:
+                reply_text = self._format_project_lookup_hit_rows(projects[:8], prefix="当前项目")
+                linked_project_ids = [int(item["id"]) for item in projects[:8] if _safe_int(item.get("id"))]
+        else:
+            keyword = self._extract_project_lookup_keyword(message)
+            matched_rows = self._match_projects_by_keyword(projects=projects, keyword=keyword)
+            if matched_rows:
+                linked_project_ids = [int(item["id"]) for item in matched_rows if _safe_int(item.get("id"))]
+                reply_text = self._format_project_lookup_hit_rows(matched_rows[:5], prefix="查到了")
+            elif keyword:
+                reply_text = f"没有查到包含“{keyword}”的项目。"
+            else:
+                reply_text = "这条问题缺少可检索的项目关键词。"
+
+        duration_ms = int((time.perf_counter() - process_started_at) * 1000)
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=None,
+            raw_message=message,
+            ai_raw_output=json.dumps({"deterministic_lookup": True, "linked_project_ids": linked_project_ids}, ensure_ascii=False),
+            project_score=None,
+            signal_confidence=None,
+            level="question",
+            routing_decision=ROUTING_DIRECT_QA,
+            routing_reason="命中项目查询问句，已直接基于真实 DB 返回结果，跳过 LLM 分流。",
+            reasoning_bullets=[
+                "命中 deterministic project lookup 规则。",
+                "直接查询 projects 表，不走 signal/project 分流。",
+                "回复内容只基于真实 DB 结果。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=duration_ms,
+        )
+        return {
+            "status": "answered",
+            "classification": "question",
+            "project_scope": "question",
+            "linked_project_ids": linked_project_ids,
+            "signal_id": None,
+            "created_project_id": None,
+            "auto_tasks": [],
+            "ai_trace_id": f"feishu-project-lookup-{(feishu_message_id or 'noid')[:32]}",
+            "results": [],
+            "intent": "question",
+            "direct_reply_required": True,
+            "direct_reply_text": reply_text,
+        }
+
+    async def _build_openclaw_blocked_capture(
+        self,
+        *,
+        message: str,
+        feishu_message_id: Optional[str],
+        duration_ms: int,
+        reason: str,
+    ) -> dict[str, Any]:
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=None,
+            raw_message=message,
+            ai_raw_output=None,
+            project_score=None,
+            signal_confidence=None,
+            level="command",
+            routing_decision=ROUTING_OPENCLAW_BLOCKED,
+            routing_reason=reason,
+            reasoning_bullets=[
+                "命中 command-first 路由，但当前缺少安全执行所需的上下文。",
+                "系统未执行任何写库动作。",
+                "回执只基于确定性拦截结果，不使用 LLM 概括。",
+            ],
+            processing_status="success",
+            error_message=None,
+            duration_ms=duration_ms,
+        )
+        return {
+            "status": "captured",
+            "classification": "command",
+            "project_scope": "command",
+            "linked_project_ids": [],
+            "signal_id": None,
+            "created_project_id": None,
+            "auto_tasks": [],
+            "ai_trace_id": None,
+            "results": [],
+            "intent": "command",
+            "direct_reply_required": True,
+            "direct_reply_text": f"未执行：{reason}",
+            "command_result": None,
+        }
+
+    async def _resolve_command_operator_user(self, *, sender: str) -> Optional[User]:
+        sender_id = await self._resolve_user_id_by_name(sender)
+        if sender_id is not None:
+            sender_user = await self._load_active_user_by_id(user_id=int(sender_id))
+            if sender_user is not None:
+                return sender_user
+
+        manager = (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.is_active.is_(True),
+                        User.status == "active",
+                        User.role.in_(["BOSS", "ADMIN", "PM"]),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+            )
+        ).scalar_one_or_none()
+        if manager is not None:
+            return manager
+        return await self._load_active_user_by_id(user_id=await self._fallback_user_id())
+
+    async def _load_active_user_by_id(self, user_id: Optional[int]) -> Optional[User]:
+        if user_id is None:
+            return None
+        return (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.id == int(user_id),
+                        User.is_active.is_(True),
+                        User.status == "active",
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+
+    async def _record_command_context_item(
+        self,
+        *,
+        message: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        captured_at: Optional[datetime],
+        matched_project_id: Optional[int],
+        route_status: str,
+        command_result: dict[str, Any],
+    ) -> Optional[int]:
+        if matched_project_id is None and not source_chat_id:
+            return None
+        occurred_at = captured_at or _utc_now()
+        if occurred_at.tzinfo is not None:
+            occurred_at = occurred_at.astimezone(timezone.utc).replace(tzinfo=None)
+        item = FeishuItem(
+            tenant_id=self.tenant_id,
+            feishu_message_id=feishu_message_id,
+            project_score=1.0 if matched_project_id is not None else None,
+            signal_confidence=0.0 if matched_project_id is not None else None,
+            raw_message=message[:4000],
+            split_from_item_id=None,
+            level="command",
+            level_confidence=1.0,
+            level_reasoning="OpenClaw command-first execution",
+            card_type="command",
+            route_status=self._normalize_route_status(route_status or ROUTE_STATUS_AUTO_DONE)[:30],
+            matched_project_id=int(matched_project_id) if matched_project_id is not None else None,
+            source_chat_id=(source_chat_id or "")[:128] or None,
+            source_sender=(sender or "")[:120] or None,
+            source_sender_open_id=(source_sender_open_id or "")[:128] or None,
+            source_occurred_at=occurred_at,
+            ai_output=json.dumps({"command_result": command_result}, ensure_ascii=False)[:8000],
+            message_class="auto_done",
+            safe_to_auto_execute=True,
+            auto_execute_state="done",
+            result_payload={
+                "intent": "command",
+                "executor": "openclaw_commander",
+                "operation_code": None,
+                "result_summary": self._build_openclaw_command_reply(command_result),
+                "command_result": command_result,
+                "project_id": int(matched_project_id) if matched_project_id is not None else None,
+                "target_url": f"/projects/{int(matched_project_id)}/board" if matched_project_id is not None else None,
+            },
+            queue_visible=False,
+            processed_at=_utc_now().replace(tzinfo=None),
+        )
+        self.db.add(item)
+        await self.db.flush()
+        return int(item.id)
+
+    @staticmethod
+    def _resolve_command_route_status(command_result: dict[str, Any]) -> str:
+        status = str(command_result.get("openclaw_status") or "").strip().lower()
+        if status in {"completed", "partial", "planned"}:
+            return ROUTE_STATUS_AUTO_DONE
+        return ROUTE_STATUS_FAILED
+
+    @staticmethod
+    def _collect_command_project_ids(command_result: dict[str, Any]) -> list[int]:
+        project_ids: set[int] = set()
+        project_info = command_result.get("project") if isinstance(command_result, dict) else None
+        if isinstance(project_info, dict):
+            project_id = project_info.get("id")
+            try:
+                if project_id is not None and int(project_id) > 0:
+                    project_ids.add(int(project_id))
+            except (TypeError, ValueError):
+                pass
+        execution = command_result.get("execution") if isinstance(command_result, dict) else None
+        results = execution.get("results") if isinstance(execution, dict) else None
+        if isinstance(results, list):
+            for row in results:
+                if not isinstance(row, dict):
+                    continue
+                if str(row.get("entity_type") or "").strip() != "project":
+                    continue
+                payload = row.get("result") if isinstance(row.get("result"), dict) else {}
+                candidate = payload.get("project_id") or row.get("entity_id")
+                try:
+                    if candidate is not None and int(candidate) > 0:
+                        project_ids.add(int(candidate))
+                except (TypeError, ValueError):
+                    continue
+        return sorted(project_ids)
+
+    @staticmethod
+    def _extract_created_project_id(command_result: dict[str, Any]) -> Optional[int]:
+        execution = command_result.get("execution") if isinstance(command_result, dict) else None
+        results = execution.get("results") if isinstance(execution, dict) else None
+        if not isinstance(results, list):
+            return None
+        for row in results:
+            if not isinstance(row, dict):
+                continue
+            if str(row.get("operation_code") or "").strip() != "project.create":
+                continue
+            payload = row.get("result") if isinstance(row.get("result"), dict) else {}
+            candidate = payload.get("project_id") or row.get("entity_id")
+            try:
+                if candidate is not None and int(candidate) > 0:
+                    return int(candidate)
+            except (TypeError, ValueError):
+                continue
+        return None
+
+    @classmethod
+    def _build_openclaw_command_reply(cls, command_result: dict[str, Any]) -> str:
+        execution = command_result.get("execution") if isinstance(command_result, dict) else {}
+        results = execution.get("results") if isinstance(execution, dict) else []
+        if not isinstance(results, list) or not results:
+            return "未执行：未识别到可安全落库的业务操作。"
+
+        fragments = [cls._format_openclaw_result_fragment(row) for row in results[:3] if isinstance(row, dict)]
+        fragments = [item for item in fragments if item]
+        applied_count = int((execution or {}).get("applied_count") or 0)
+        skipped_count = int((execution or {}).get("skipped_count") or 0)
+        failed_count = int((execution or {}).get("failed_count") or 0)
+
+        if applied_count <= 0 and skipped_count <= 0 and failed_count > 0:
+            first_failed = next((row for row in results if isinstance(row, dict) and str(row.get("status") or "") == "failed"), {})
+            message = str(first_failed.get("message") or "执行失败").strip()
+            return f"未执行：{message}"
+        if not fragments:
+            return "已接收指令，但未生成可展示的执行回执。"
+
+        prefix = "未重复执行：" if applied_count <= 0 and skipped_count > 0 else "已执行完成："
+        reply = f"{prefix}{'；'.join(fragments[:2])}"
+        if failed_count > 0 and applied_count > 0:
+            reply = f"{reply}；另有 {failed_count} 项失败"
+        return reply + "。"
+
+    @staticmethod
+    def _format_openclaw_result_fragment(row: dict[str, Any]) -> str:
+        operation_code = str(row.get("operation_code") or "").strip()
+        status = str(row.get("status") or "").strip().lower()
+        message = str(row.get("message") or "").strip()
+        payload = row.get("result") if isinstance(row.get("result"), dict) else {}
+        if status == "failed":
+            return ""
+
+        if operation_code.startswith("project."):
+            project_id = payload.get("project_id") or row.get("entity_id")
+            title = str(payload.get("title") or payload.get("project_title") or "").strip()
+            status_text = str(payload.get("status") or "").strip()
+            label = f"项目#{project_id}"
+            if title:
+                label = f"{label}「{title}」"
+            if operation_code == "project.create" and status == "applied":
+                return f"{label}已创建，状态 {status_text or 'active'}"
+            if operation_code == "project.create" and status == "skipped":
+                return f"{label}已存在，状态 {status_text or 'active'}"
+            if status_text:
+                return f"{label}状态已更新为 {status_text}"
+            return f"{label}{message or '已处理'}"
+
+        if operation_code == "task.create":
+            return f"任务#{row.get('entity_id') or ''}已创建".strip()
+        if operation_code == "task.assign":
+            return f"任务#{row.get('entity_id') or ''}已分配".strip()
+        if operation_code == "task.prioritize":
+            return f"任务#{row.get('entity_id') or ''}优先级已更新".strip()
+        if operation_code == "task.status_transition":
+            return f"任务#{row.get('entity_id') or ''}状态已更新".strip()
+        if operation_code == "personnel.workload_update":
+            return f"成员#{row.get('entity_id') or ''}工时已更新".strip()
+        if operation_code == "notification.create":
+            return f"通知#{row.get('entity_id') or ''}已创建".strip()
+        return message or operation_code
+
+    @staticmethod
+    def _extract_project_lookup_keyword(message: str) -> str:
+        text = str(message or "").strip()
+        if not text:
+            return ""
+        cleaned = re.sub(
+            r"(请问|麻烦问下|问下|现在|当前|目前|系统里|系统中|项目里|项目中|有没有|有无|有吗|是否有|是否存在|存在吗|这个|那个|一个|个|项目|列表|清单|有|吗|呢|呀|啊|么)",
+            " ",
+            text,
+            flags=re.IGNORECASE,
+        )
+        cleaned = re.sub(r"[，。,.;；:：!?？（）()\\[\\]【】]+", " ", cleaned)
+        cleaned = re.sub(r"\s+", " ", cleaned).strip()
+        return cleaned[:40]
+
+    @staticmethod
+    def _match_projects_by_keyword(*, projects: list[dict[str, Any]], keyword: str) -> list[dict[str, Any]]:
+        normalized_keyword = re.sub(r"\s+", "", str(keyword or "")).lower()
+        if len(normalized_keyword) < 2:
+            return []
+        matched: list[dict[str, Any]] = []
+        for item in projects:
+            title = re.sub(r"\s+", "", str(item.get("title") or item.get("name") or "")).lower()
+            if normalized_keyword and normalized_keyword in title:
+                matched.append(item)
+        return matched
+
+    @staticmethod
+    def _format_project_lookup_hit_rows(rows: list[dict[str, Any]], *, prefix: str) -> str:
+        if not rows:
+            return "当前没有可见项目。"
+        fragments: list[str] = []
+        for item in rows[:8]:
+            project_id = item.get("id")
+            title = str(item.get("title") or item.get("name") or "").strip() or f"项目#{project_id}"
+            status = str(item.get("status") or "").strip() or "unknown"
+            fragments.append(f"#{project_id} {title}({status})")
+        return f"{prefix}：{'、'.join(fragments)}。"
+
+    @classmethod
+    def _should_hide_pending_item_from_inbox(cls, item: FeishuItem) -> bool:
+        if str(item.decision_reason or "").strip() == "missing_required_fields":
+            return True
+        raw_message = str(item.raw_message or "").strip()
+        if not raw_message:
+            return False
+        normalized = re.sub(r"\s+", "", raw_message).lower()
+        if cls._is_water_talk_message(raw_message):
+            return True
+        if cls._is_healthcheck_message(raw_message):
+            return True
+        if cls._is_short_confirmation_message(raw_message):
+            return True
+        if cls._looks_like_project_lookup_question(raw_message):
+            return True
+        if "长连接" in normalized and any(token in normalized for token in ("已通", "都已通", "正常", "在线", "已恢复")):
+            return True
+        return False
+
+    async def list_queue(self, *, limit: int = 30) -> dict[str, Any]:
+        routing_policy = await self._load_routing_policy()
+        pending_rows = (
+            await self.db.execute(
+                select(FeishuItem)
+                .where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.queue_visible.is_(True),
+                        FeishuItem.route_status.in_((ROUTE_STATUS_QUEUED, "pending")),
+                    )
+                )
+                .order_by(FeishuItem.created_at.desc())
+                .limit(limit)
+            )
+        ).scalars().all()
+        pending_rows = self._dedupe_queue_items(pending_rows)
+        pending_rows = [row for row in pending_rows if not self._should_hide_pending_item_from_inbox(row)]
+
+        # DB 列是 timestamp without time zone，比较值必须使用 naive UTC，避免 asyncpg aware/naive 报错。
+        today_start = _utc_now().astimezone(timezone.utc).replace(
+            hour=0,
+            minute=0,
+            second=0,
+            microsecond=0,
+            tzinfo=None,
+        )
+        auto_rows = (
+            await self.db.execute(
+                select(FeishuItem)
+                .where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.route_status.in_((ROUTE_STATUS_AUTO_DONE, ROUTE_STATUS_UNDONE, "auto_processed", "confirmed")),
+                        FeishuItem.processed_at.is_not(None),
+                        FeishuItem.processed_at >= today_start,
+                    )
+                )
+                .order_by(FeishuItem.processed_at.desc())
+                .limit(limit)
+            )
+        ).scalars().all()
+        auto_rows = self._dedupe_queue_items(auto_rows)
+        auto_rows = [row for row in auto_rows if not self._is_water_talk_message(row.raw_message or "")]
+
+        pending_items = [await self._serialize_item(row) for row in pending_rows]
+        pending_items.sort(
+            key=lambda x: (
+                -int(x.get("priority_rank") or 0),
+                str(x.get("created_at") or ""),
+            )
+        )
+        auto_items = [await self._serialize_item(row) for row in auto_rows]
+
+        signal_auto = len([x for x in auto_items if x.get("level") == "signal"])
+        project_auto = len(
+            [
+                x for x in auto_items
+                if x.get("level") == "project"
+                or (
+                    x.get("level") == "command"
+                    and isinstance(x.get("command"), dict)
+                    and str(x["command"].get("operation_code") or "") == "project.create"
+                )
+            ]
+        )
+        return {
+            "pending_count": len(pending_items),
+            "pending_items": pending_items,
+            "auto_processed_items": auto_items,
+            "auto_summary": {
+                "signals_auto_entered_today": signal_auto,
+                "projects_auto_created_today": project_auto,
+            },
+            "cold_start": bool(routing_policy["cold_start_enabled"]),
+        }
+
+    async def get_item_detail(self, item_id: int) -> dict[str, Any]:
+        item = await self._get_item(item_id)
+        if not item:
+            raise ValueError("item_not_found")
+        return await self._serialize_item(item, include_raw=True, include_history=True)
+
+    async def update_item(
+        self,
+        *,
+        item_id: int,
+        payload: dict[str, Any],
+        user_id: int,
+    ) -> dict[str, Any]:
+        item = await self._get_item(item_id)
+        if not item:
+            raise ValueError("item_not_found")
+
+        changed = False
+        if (item.level or "").lower() == "project":
+            suggestion = await self._get_project_suggestion(item.id)
+            if not suggestion:
+                raise ValueError("project_suggestion_not_found")
+            for field in ("suggested_name", "suggested_description", "suggested_owner", "suggested_timeline"):
+                if field in payload:
+                    new_value = str(payload.get(field) or "").strip() or None
+                    if getattr(suggestion, field) != new_value:
+                        setattr(suggestion, field, new_value)
+                        changed = True
+            if changed:
+                suggestion.status = "pending"
+        else:
+            signal = await self._get_signal_by_item(item.id)
+            if not signal:
+                raise ValueError("signal_not_found")
+            editable_fields = (
+                "title",
+                "signal_type",
+                "priority",
+                "summary",
+                "recommended_action",
+                "matched_project_id",
+            )
+            for field in editable_fields:
+                if field not in payload:
+                    continue
+                old_value = getattr(signal, field)
+                raw_new = payload.get(field)
+                if field == "matched_project_id":
+                    new_value = int(raw_new) if raw_new is not None else None
+                else:
+                    new_value = str(raw_new).strip() if raw_new is not None else None
+                    if new_value == "":
+                        new_value = None
+                if old_value == new_value:
+                    continue
+                setattr(signal, field, new_value)
+                self.db.add(
+                    SignalEdit(
+                        tenant_id=self.tenant_id,
+                        signal_id=int(signal.id),
+                        field=field,
+                        old_value=None if old_value is None else str(old_value),
+                        new_value=None if new_value is None else str(new_value),
+                        edited_by=int(user_id),
+                        edited_at=_utc_now().replace(tzinfo=None),
+                    )
+                )
+                changed = True
+
+        if changed:
+            await self.db.commit()
+        return await self.get_item_detail(item_id)
+
+    async def dismiss_item(self, *, item_id: int, reason: str) -> dict[str, Any]:
+        item = await self._get_item(item_id)
+        if not item:
+            raise ValueError("item_not_found")
+        item.route_status = ROUTE_STATUS_DISMISSED
+        item.queue_visible = False
+        item.dismissed_reason = (reason or "").strip()[:300] or "用户手动忽略"
+        item.processed_at = _utc_now().replace(tzinfo=None)
+        item.message_class = "ignored"
+        item.auto_execute_state = "dismissed"
+
+        if (item.level or "").lower() == "project":
+            suggestion = await self._get_project_suggestion(item.id)
+            if suggestion:
+                suggestion.status = "dismissed"
+        else:
+            signal = await self._get_signal_by_item(item.id)
+            if signal:
+                signal.processed = True
+                signal.status = "已关闭"
+
+        await self.db.commit()
+        return {"id": str(item.id), "status": item.route_status}
+
+    async def confirm_item(self, *, item_id: int, user_id: int, action: Optional[str] = None) -> dict[str, Any]:
+        item = await self._get_item(item_id)
+        if not item:
+            raise ValueError("item_not_found")
+
+        level = (item.level or "").lower()
+        result: dict[str, Any] = {"id": str(item.id), "level": level}
+        if level == "project":
+            suggestion = await self._get_project_suggestion(item.id)
+            if not suggestion:
+                raise ValueError("project_suggestion_not_found")
+            candidate_fields = self._resolve_project_candidate_fields(item=item, suggestion=suggestion)
+            if not candidate_fields["ready"]:
+                raise ValueError(
+                    str(item.blocking_reason or "").strip()
+                    or self._build_project_clarification_message(candidate_fields["missing_fields"])
+                )
+            operator_user = await self._load_active_user_by_id(user_id=int(user_id))
+            if operator_user is None:
+                raise ValueError("operator_not_found")
+            execution = await self._execute_controlled_operation(
+                operator_user=operator_user,
+                operation_code="project.create",
+                payload={
+                    "title": candidate_fields["title"],
+                    "description": candidate_fields["description"],
+                },
+                trace_id=f"confirm-project-{int(item.id)}-{int(user_id)}",
+            )
+            first = execution["results"][0] if execution.get("results") else {}
+            status = str(first.get("status") or "").strip().lower()
+            if status not in {"applied", "skipped"}:
+                raise ValueError(str(first.get("message") or "project_create_failed"))
+            project_result = first.get("result") if isinstance(first.get("result"), dict) else {}
+            project_id = _safe_int(project_result.get("project_id") or first.get("entity_id"))
+            suggestion.status = "confirmed"
+            suggestion.created_project_id = int(project_id) if project_id is not None else None
+            item.matched_project_id = int(project_id) if project_id is not None else item.matched_project_id
+            item.result_payload = {
+                "intent": INTENT_CREATE_PROJECT,
+                "executor": "operation_executor",
+                "operation_code": "project.create",
+                "result_summary": f"已创建项目 #{project_id}" if project_id else "已创建项目",
+                "operation_result": project_result,
+                "target_url": f"/projects/{project_id}/board" if project_id else None,
+                "project_id": project_id,
+            }
+            result["project_id"] = str(project_id) if project_id is not None else None
+        elif level == "command":
+            operator_user = await self._load_active_user_by_id(user_id=int(user_id))
+            if operator_user is None:
+                raise ValueError("operator_not_found")
+            payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+            operation_code = str(payload.get("operation_code") or "").strip()
+            operation_payload = payload.get("operation_payload") if isinstance(payload.get("operation_payload"), dict) else {}
+            if operation_code not in {"task.create"}:
+                raise ValueError("command_operation_not_supported")
+            execution = await self._execute_controlled_operation(
+                operator_user=operator_user,
+                operation_code=operation_code,
+                payload=operation_payload,
+                trace_id=f"confirm-command-{int(item.id)}-{int(user_id)}",
+            )
+            first = execution["results"][0] if execution.get("results") else {}
+            status = str(first.get("status") or "").strip().lower()
+            if status not in {"applied", "skipped"}:
+                raise ValueError(str(first.get("message") or "command_execution_failed"))
+            project_id = _safe_int(operation_payload.get("project_id") or payload.get("project_id") or item.matched_project_id)
+            task_id = _safe_int(first.get("entity_id"))
+            item.matched_project_id = int(project_id) if project_id is not None else item.matched_project_id
+            item.result_payload = {
+                **payload,
+                "executor": "operation_executor",
+                "result_summary": f"已创建任务 #{task_id}" if task_id is not None else "已创建任务",
+                "operation_result": first.get("result") if isinstance(first.get("result"), dict) else {},
+                "target_url": f"/projects/{project_id}/board" if project_id else None,
+                "project_id": project_id,
+            }
+            result["project_id"] = str(project_id) if project_id is not None else None
+            result["task_id"] = str(task_id) if task_id is not None else None
+        else:
+            signal = await self._get_signal_by_item(item.id)
+            if not signal:
+                raise ValueError("signal_not_found")
+            matched_project_id = signal.matched_project_id or item.matched_project_id
+            if not matched_project_id:
+                raise ValueError("signal_missing_project")
+            await self._ensure_signal_link(signal_id=int(signal.id), project_id=int(matched_project_id))
+            signal.processed = False
+            signal.status = "待处理"
+            if action in {"create_charter", "trigger_alert"}:
+                signal.recommended_action = action
+            result["signal_id"] = str(signal.id)
+            result["project_id"] = str(matched_project_id)
+
+        item.route_status = ROUTE_STATUS_AUTO_DONE
+        item.queue_visible = False
+        item.processed_at = _utc_now().replace(tzinfo=None)
+        item.message_class = "auto_done"
+        item.safe_to_auto_execute = False
+        item.auto_execute_state = "done"
+        item.blocking_reason = None
+        item.next_step_text = "你已授权继续执行，结果已落地，可进入结果页查看。"
+        item.undoable_until = _build_undo_deadline(item.processed_at) if level in {"project", "signal"} else None
+        if level == "project":
+            created_project_id = _safe_int(result.get("project_id"))
+            item.result_payload = {
+                **(item.result_payload if isinstance(item.result_payload, dict) else {}),
+                "result_count": 1,
+                "impact_objects": [f"项目#{created_project_id}"] if created_project_id else [],
+            }
+        elif level == "signal":
+            project_id = _safe_int(result.get("project_id"))
+            signal_id = _safe_int(result.get("signal_id"))
+            item.result_payload = {
+                "result_summary": f"已确认信号 #{signal_id} 并关联项目 #{project_id}" if project_id and signal_id else "已确认信号并完成入库",
+                "result_count": 1,
+                "impact_objects": [f"信号#{signal_id}"] if signal_id else [],
+                "target_url": f"/projects/{project_id}/board" if project_id else None,
+                "project_id": project_id,
+            }
+        else:
+            payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+            item.result_payload = {
+                **payload,
+                "result_count": 1,
+            }
+        await self.db.commit()
+        return result
+
+    async def undo_auto_processed(self, *, item_id: int) -> dict[str, Any]:
+        item = await self._get_item(item_id)
+        if not item:
+            raise ValueError("item_not_found")
+        if self._normalize_route_status(item.route_status) != ROUTE_STATUS_AUTO_DONE:
+            raise ValueError("item_not_auto_processed")
+
+        level = (item.level or "").lower()
+        if level == "project":
+            suggestion = await self._get_project_suggestion(item.id)
+            if suggestion and suggestion.created_project_id:
+                project = (
+                    await self.db.execute(
+                        select(Project).where(
+                            and_(
+                                Project.tenant_id == self.tenant_id,
+                                Project.id == int(suggestion.created_project_id),
+                            )
+                        )
+                    )
+                ).scalar_one_or_none()
+                if project:
+                    project.status = "inactive"
+            if suggestion:
+                suggestion.status = "pending"
+        elif level == "signal":
+            signal = await self._get_signal_by_item(item.id)
+            if signal:
+                await self.db.execute(
+                    delete(ProjectSignalLink).where(
+                        and_(
+                            ProjectSignalLink.tenant_id == self.tenant_id,
+                            ProjectSignalLink.signal_id == int(signal.id),
+                        )
+                    )
+                )
+                signal.status = "待处理"
+                signal.processed = False
+
+        item.route_status = ROUTE_STATUS_UNDONE
+        item.queue_visible = False
+        item.processed_at = _utc_now().replace(tzinfo=None)
+        item.message_class = "auto_done"
+        item.auto_execute_state = "undone"
+        item.undoable_until = None
+        existing_payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+        item.result_payload = {
+            **existing_payload,
+            "result_summary": "该结果已撤销",
+            "undone": True,
+        }
+        await self.db.commit()
+        return {"id": str(item.id), "status": item.route_status}
+
+    async def _process_one_item(
+        self,
+        *,
+        content: str,
+        level: str,
+        sender: str,
+        source_chat_id: str,
+        source_sender_open_id: Optional[str],
+        feishu_message_id: Optional[str],
+        occurred_at: datetime,
+        level_result: dict[str, Any],
+        level_confidence: float,
+        level_reasoning: str,
+        project_score: float,
+        cold_start_enabled: bool,
+        signal_auto_enter_threshold: float,
+        project_auto_create_threshold: float,
+        project_uncertain_threshold: float,
+        chat_context_auto_link_enabled: bool,
+        projects: list[dict[str, Any]],
+        split_from_item_id: Optional[int],
+        chat_type: str,
+        chat_context_resolution: dict[str, Any],
+    ) -> dict[str, Any]:
+        item_started_at = time.perf_counter()
+        level_val = "project" if level == "project" else "signal"
+        matched_project_hint = _safe_int(chat_context_resolution.get("project_id"))
+        dedupe_key = self._build_queue_dedupe_key(
+            level=level_val,
+            intent=INTENT_SIGNAL_INTAKE if level_val == "signal" else INTENT_CREATE_PROJECT,
+            message=content,
+            source_chat_id=source_chat_id,
+            matched_project_id=matched_project_hint,
+            source_object=content,
+        )
+        existing_item = await self._find_active_dedupe_item(dedupe_key)
+        if existing_item is not None:
+            existing_item.occurrence_count = max(1, int(existing_item.occurrence_count or 1)) + 1
+            existing_item.raw_message = content[:4000]
+            existing_item.source_sender = (sender or "")[:120] or None
+            existing_item.source_sender_open_id = (source_sender_open_id or "")[:128] or None
+            existing_item.source_chat_id = (source_chat_id or "")[:128] or None
+            existing_item.source_occurred_at = occurred_at.replace(tzinfo=None)
+            existing_item.level_reasoning = level_reasoning[:1000]
+            existing_item.ai_output = json.dumps(
+                {
+                    "chat_type": chat_type,
+                    "normalized_text": content,
+                    "level_result": level_result,
+                },
+                ensure_ascii=False,
+            )[:8000]
+            await self.db.flush()
+            return await self._build_existing_item_result(existing_item)
+
+        item = FeishuItem(
+            tenant_id=self.tenant_id,
+            feishu_message_id=feishu_message_id,
+            project_score=project_score,
+            signal_confidence=max(0.0, 1 - project_score),
+            raw_message=content[:4000],
+            split_from_item_id=split_from_item_id,
+            level=level_val,
+            level_confidence=level_confidence,
+            level_reasoning=level_reasoning[:1000],
+            card_type="signal" if level_val == "signal" else "new_project",
+            route_status=ROUTE_STATUS_CLASSIFIED,
+            source_chat_id=(source_chat_id or "")[:128] or None,
+            source_sender=(sender or "")[:120] or None,
+            source_sender_open_id=(source_sender_open_id or "")[:128] or None,
+            source_occurred_at=occurred_at.replace(tzinfo=None),
+            ai_output=json.dumps(
+                {
+                    "chat_type": chat_type,
+                    "normalized_text": content,
+                    "level_result": level_result,
+                },
+                ensure_ascii=False,
+            )[:8000],
+            decision_reason=level_reasoning[:1000] or None,
+            safe_to_auto_execute=False,
+            auto_execute_state="blocked",
+            dedupe_key=dedupe_key,
+            message_class="action_required",
+            occurrence_count=1,
+            next_step_text="AI 正在校验上下文，确认后会继续推进后续动作。",
+            queue_visible=True,
+        )
+        self.db.add(item)
+        await self.db.flush()
+
+        if level_val == "project":
+            project_info = await self._extract_project_prefill(content)
+            suggestion = ProjectSuggestion(
+                tenant_id=self.tenant_id,
+                feishu_item_id=int(item.id),
+                suggested_name=(project_info.get("suggested_name") or content[:20] or "新项目建议")[:255],
+                suggested_description=(project_info.get("suggested_description") or content[:400])[:4000],
+                suggested_owner=(project_info.get("suggested_owner") or sender or "待分配")[:120],
+                suggested_timeline=(project_info.get("suggested_timeline") or "待确认")[:120],
+                confidence=_safe_float(project_info.get("confidence"), level_confidence),
+                reasoning=json.dumps(_normalize_reasoning(project_info.get("reasoning")), ensure_ascii=False),
+                status="pending",
+            )
+            self.db.add(suggestion)
+            await self.db.flush()
+            project_id = await self._route_project_item(
+                item=item,
+                suggestion=suggestion,
+                sender=sender,
+                project_score=project_score,
+                cold_start_enabled=cold_start_enabled,
+                project_auto_create_threshold=project_auto_create_threshold,
+            )
+            routing_decision, routing_reason, reasoning_bullets = self._build_project_routing_log(
+                project_score=project_score,
+                cold_start_enabled=cold_start_enabled,
+                project_auto_create_threshold=project_auto_create_threshold,
+                project_uncertain_threshold=project_uncertain_threshold,
+            )
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=int(item.id),
+                raw_message=content,
+                ai_raw_output=json.dumps(
+                    {
+                        "level_classification": level_result,
+                        "project_prefill": project_info,
+                    },
+                    ensure_ascii=False,
+                ),
+                project_score=project_score,
+                signal_confidence=max(0.0, 1 - project_score),
+                level="project",
+                routing_decision=routing_decision,
+                routing_reason=routing_reason,
+                reasoning_bullets=reasoning_bullets,
+                processing_status="success",
+                error_message=None,
+                duration_ms=int((time.perf_counter() - item_started_at) * 1000),
+            )
+            return {
+                "item_id": int(item.id),
+                "level": "project",
+                "project_id": project_id,
+                "signal_id": None,
+                "route_status": item.route_status,
+            }
+
+        simple_signal_request = self._is_simple_signal_source_request(content)
+        signal_info = await self._extract_signal(content, projects=projects)
+        matched_project_id = signal_info.get("matched_project_id")
+        if matched_project_id is not None:
+            try:
+                matched_project_id = int(matched_project_id)
+            except (TypeError, ValueError):
+                matched_project_id = None
+
+        confidence = _safe_float(signal_info.get("confidence"), max(0.55, 1 - project_score))
+        if simple_signal_request:
+            confidence = max(confidence, min(0.95, signal_auto_enter_threshold + 0.05))
+        inferred_project_id: Optional[int] = None
+        clarification_message: Optional[str] = None
+        if (chat_context_auto_link_enabled or simple_signal_request) and matched_project_id is None and source_chat_id:
+            inferred_project_id = _safe_int(chat_context_resolution.get("project_id"))
+            if inferred_project_id is not None:
+                matched_project_id = int(inferred_project_id)
+            elif bool(chat_context_resolution.get("ambiguous")):
+                candidate_ids = [int(x) for x in (chat_context_resolution.get("candidate_ids") or [])[:3] if _safe_int(x)]
+                if candidate_ids:
+                    clarification_message = (
+                        f"当前群最近同时涉及项目#{'、#'.join([str(pid) for pid in candidate_ids])}，"
+                        "请明确这条消息归属哪个项目。"
+                    )
+
+        if matched_project_id is None and self._should_promote_signal_to_project_candidate(
+            content=content,
+            signal_info=signal_info,
+            confidence=confidence,
+            signal_auto_enter_threshold=signal_auto_enter_threshold,
+            simple_signal_request=simple_signal_request,
+        ):
+            promoted_project_score = max(float(project_score or 0.0), float(project_uncertain_threshold))
+            item.level = "project"
+            item.project_score = promoted_project_score
+            item.signal_confidence = max(0.0, 1 - promoted_project_score)
+            item.card_type = "new_project_candidate"
+            item.matched_project_id = None
+
+            project_info = await self._extract_project_prefill(content)
+            suggestion = ProjectSuggestion(
+                tenant_id=self.tenant_id,
+                feishu_item_id=int(item.id),
+                suggested_name=(project_info.get("suggested_name") or content[:20] or "新项目建议")[:255],
+                suggested_description=(project_info.get("suggested_description") or content[:400])[:4000],
+                suggested_owner=(project_info.get("suggested_owner") or sender or "待分配")[:120],
+                suggested_timeline=(project_info.get("suggested_timeline") or "待确认")[:120],
+                confidence=_safe_float(project_info.get("confidence"), confidence),
+                reasoning=json.dumps(_normalize_reasoning(project_info.get("reasoning")), ensure_ascii=False),
+                status="pending",
+            )
+            self.db.add(suggestion)
+            await self.db.flush()
+
+            project_id = await self._route_project_item(
+                item=item,
+                suggestion=suggestion,
+                sender=sender,
+                project_score=promoted_project_score,
+                cold_start_enabled=cold_start_enabled,
+                project_auto_create_threshold=project_auto_create_threshold,
+            )
+            routing_decision, routing_reason, reasoning_bullets = self._build_project_routing_log(
+                project_score=promoted_project_score,
+                cold_start_enabled=cold_start_enabled,
+                project_auto_create_threshold=project_auto_create_threshold,
+                project_uncertain_threshold=project_uncertain_threshold,
+            )
+            await self._write_processing_log(
+                feishu_message_id=feishu_message_id,
+                feishu_item_id=int(item.id),
+                raw_message=content,
+                ai_raw_output=json.dumps(
+                    {
+                        "level_classification": level_result,
+                        "signal_extraction": signal_info,
+                        "promoted_to_project_candidate": True,
+                        "simple_signal_request": simple_signal_request,
+                        "project_prefill": project_info,
+                    },
+                    ensure_ascii=False,
+                ),
+                project_score=promoted_project_score,
+                signal_confidence=confidence,
+                level="project",
+                routing_decision=routing_decision,
+                routing_reason=routing_reason,
+                reasoning_bullets=reasoning_bullets,
+                processing_status="success",
+                error_message=None,
+                duration_ms=int((time.perf_counter() - item_started_at) * 1000),
+            )
+            return {
+                "item_id": int(item.id),
+                "level": "project",
+                "project_id": project_id,
+                "signal_id": None,
+                "route_status": item.route_status,
+            }
+
+        signal = Signal(
+            tenant_id=self.tenant_id,
+            source="Feishu",
+            content=f"[{sender or '未知用户'}]: {content[:800]}",
+            impact_score=max(0.1, min(1.0, confidence)),
+            priority=self._to_signal_priority(str(signal_info.get("priority") or "P2")),
+            processed=False,
+            strategic_analysis=json.dumps(
+                {
+                    "source_context": {
+                        "group_id": source_chat_id,
+                        "sender": sender,
+                        "sender_open_id": source_sender_open_id,
+                    },
+                    "ai_analysis": {
+                        "classification": "signal",
+                        "reason": signal_info.get("summary"),
+                    },
+                },
+                ensure_ascii=False,
+            ),
+            tasks_generated="[]",
+            feishu_item_id=int(item.id),
+            matched_project_id=matched_project_id,
+            title=(signal_info.get("title") or content[:40] or "未命名信号")[:200],
+            signal_type=self._normalize_signal_type(signal_info.get("type")),
+            summary=(signal_info.get("summary") or "")[:3000] or None,
+            reasoning=json.dumps(_normalize_reasoning(signal_info.get("reasoning")), ensure_ascii=False),
+            confidence=confidence,
+            recommended_action=self._normalize_recommended_action(signal_info.get("recommended_action")),
+            status="待处理",
+        )
+        self.db.add(signal)
+        await self.db.flush()
+
+        project_id = await self._route_signal_item(
+            item=item,
+            signal=signal,
+            matched_project_id=matched_project_id,
+            confidence=confidence,
+            cold_start_enabled=cold_start_enabled,
+            signal_auto_enter_threshold=signal_auto_enter_threshold,
+        )
+        routing_decision, routing_reason, reasoning_bullets = self._build_signal_routing_log(
+            confidence=confidence,
+            cold_start_enabled=cold_start_enabled,
+            matched_project_id=matched_project_id,
+            signal_auto_enter_threshold=signal_auto_enter_threshold,
+        )
+        await self._write_processing_log(
+            feishu_message_id=feishu_message_id,
+            feishu_item_id=int(item.id),
+            raw_message=content,
+            ai_raw_output=json.dumps(
+                {
+                        "level_classification": level_result,
+                        "signal_extraction": signal_info,
+                        "simple_signal_request": simple_signal_request,
+                        "chat_context_inferred_project_id": inferred_project_id,
+                        "chat_context_resolution": chat_context_resolution,
+                        "clarification_message": clarification_message,
+                    },
+                    ensure_ascii=False,
+                ),
+            project_score=project_score,
+            signal_confidence=confidence,
+            level="signal",
+            routing_decision=routing_decision,
+            routing_reason=routing_reason,
+            reasoning_bullets=reasoning_bullets,
+            processing_status="success",
+            error_message=None,
+            duration_ms=int((time.perf_counter() - item_started_at) * 1000),
+        )
+        return {
+            "item_id": int(item.id),
+            "level": "signal",
+            "project_id": project_id,
+            "signal_id": int(signal.id),
+            "route_status": item.route_status,
+            "clarification_message": clarification_message,
+        }
+
+    async def _route_project_item(
+        self,
+        *,
+        item: FeishuItem,
+        suggestion: ProjectSuggestion,
+        sender: str,
+        project_score: float,
+        cold_start_enabled: bool,
+        project_auto_create_threshold: float,
+    ) -> Optional[int]:
+        _ = project_auto_create_threshold
+        high_score = project_score >= max(project_auto_create_threshold, 0.7)
+        item.card_type = "new_project" if high_score else "new_project_candidate"
+        item.route_status = ROUTE_STATUS_QUEUED
+        item.queue_visible = True
+        item.message_class = "action_required"
+        item.safe_to_auto_execute = False
+        item.auto_execute_state = "blocked"
+        item.blocking_reason = (
+            "当前处于冷启动保护期，建项目默认需要你确认后再执行。"
+            if cold_start_enabled
+            else "建项目属于高风险动作，默认需要你授权后再执行。"
+        )
+        item.decision_reason = (
+            "intent gate/自由分类均不再直接 auto create，避免误把自然语言写成项目主数据。"
+        )
+        item.next_step_text = "确认后，系统会走受控 project.create，并绑定真实项目 ID。"
+        return None
+
+    async def _route_signal_item(
+        self,
+        *,
+        item: FeishuItem,
+        signal: Signal,
+        matched_project_id: Optional[int],
+        confidence: float,
+        cold_start_enabled: bool,
+        signal_auto_enter_threshold: float,
+    ) -> Optional[int]:
+        item.matched_project_id = int(matched_project_id) if matched_project_id else None
+        if cold_start_enabled or confidence < signal_auto_enter_threshold or not matched_project_id:
+            item.card_type = "signal"
+            item.route_status = ROUTE_STATUS_QUEUED
+            item.queue_visible = True
+            item.message_class = "action_required"
+            item.safe_to_auto_execute = False
+            item.auto_execute_state = "blocked"
+            if not matched_project_id:
+                item.blocking_reason = "当前缺少明确项目归属，继续自动执行会产生错误副作用。"
+                item.next_step_text = "补齐项目归属后，AI 将自动把信号入库并继续后续链路。"
+                item.decision_reason = "上下文尚不完整，系统暂停在需你决策队列。"
+            elif cold_start_enabled:
+                item.blocking_reason = "当前处于冷启动保护期，信号入库暂不自动执行。"
+                item.next_step_text = "确认后，AI 将继续完成信号入库并推进后续动作。"
+                item.decision_reason = "保守自动化已拦截自动入库，避免早期误关联。"
+            else:
+                item.blocking_reason = "当前置信度不足以自动入库，需要你拍板后再继续。"
+                item.next_step_text = "确认后，AI 将把信号入库并进入后续生成流程。"
+                item.decision_reason = "信号虽然可执行，但归属判断仍存在不确定性。"
+            return matched_project_id
+
+        await self._ensure_signal_link(signal_id=int(signal.id), project_id=int(matched_project_id))
+        item.card_type = "signal_auto"
+        item.route_status = ROUTE_STATUS_AUTO_DONE
+        item.queue_visible = False
+        item.processed_at = _utc_now().replace(tzinfo=None)
+        item.message_class = "auto_done"
+        item.safe_to_auto_execute = True
+        item.auto_execute_state = "done"
+        item.blocking_reason = None
+        item.decision_reason = "低风险且上下文完整，AI 已自动完成信号入库。"
+        item.next_step_text = "下一步可基于该信号继续生成 charter 或分发任务。"
+        item.undoable_until = _build_undo_deadline(item.processed_at)
+        item.result_payload = {
+            "result_summary": f"已自动入库信号 {signal.title or '未命名信号'}",
+            "result_count": 1,
+            "impact_objects": [signal.title or f"信号#{signal.id}", f"项目#{int(matched_project_id)}"],
+            "target_url": f"/projects/{int(matched_project_id)}/board",
+            "project_id": int(matched_project_id),
+        }
+        await self._notify_signal_auto_entered(
+            project_id=int(matched_project_id),
+            signal=signal,
+        )
+        return int(matched_project_id)
+
+    async def _extract_project_prefill(self, message: str) -> dict[str, Any]:
+        prompt = PROMPT_FEISHU_PROJECT_PREFILL.format(message=message[:1500])
+        try:
+            raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=self.tenant_id,
+                function_name="feishu_project_prefill",
+            )
+            parsed = _safe_json_object(raw)
+            if not parsed:
+                raise ValueError("project_prefill_empty_json")
+            parsed["reasoning"] = _normalize_reasoning(parsed.get("reasoning"))
+            return _validate_schema(_ProjectPrefillSchema, parsed, context="project_prefill")
+        except Exception as exc:
+            logger.warning(f"feishu_project_prefill_fallback: error={exc}")
+            fallback = self._heuristic_project_prefill_fallback(message)
+            return _validate_schema(_ProjectPrefillSchema, fallback, context="project_prefill_fallback")
+
+    async def _extract_signal(self, message: str, *, projects: list[dict[str, Any]]) -> dict[str, Any]:
+        if self._is_simple_signal_source_request(message):
+            fallback = self._heuristic_signal_fallback(message=message, projects=projects)
+            return _validate_schema(_SignalExtractionSchema, fallback, context="signal_extraction_fastpath")
+
+        prompt = PROMPT_FEISHU_SIGNAL_EXTRACTION.format(
+            message=message[:1500],
+            projects_json=json.dumps(projects, ensure_ascii=False),
+        )
+        try:
+            raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=self.tenant_id,
+                function_name="feishu_signal_extraction",
+            )
+            parsed = _safe_json_object(raw)
+            if not parsed:
+                raise ValueError("signal_extraction_empty_json")
+            parsed["reasoning"] = _normalize_reasoning(parsed.get("reasoning"))
+            return _validate_schema(_SignalExtractionSchema, parsed, context="signal_extraction")
+        except Exception as exc:
+            logger.warning(f"feishu_signal_extraction_fallback: error={exc}")
+            fallback = self._heuristic_signal_fallback(message=message, projects=projects)
+            return _validate_schema(_SignalExtractionSchema, fallback, context="signal_extraction_fallback")
+
+    async def _classify_level(self, *, message: str, projects: list[dict[str, Any]]) -> dict[str, Any]:
+        if self._is_simple_signal_source_request(message):
+            fallback = self._heuristic_level_fallback(message)
+            return _validate_schema(_LevelClassificationSchema, fallback, context="level_classification_fastpath")
+
+        prompt = PROMPT_FEISHU_LEVEL_CLASSIFICATION.format(
+            message=message[:1500],
+            projects_json=json.dumps(projects, ensure_ascii=False),
+        )
+        try:
+            raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=self.tenant_id,
+                function_name="feishu_level_classification",
+            )
+            parsed = _safe_json_object(raw)
+            if not parsed:
+                raise ValueError("level_classification_empty_json")
+
+            # 兼容不同模型输出别名，避免只返回总分/中文 level 时被误降级到 signal。
+            bi_raw = _first_non_empty(parsed, ("build_intent_score", "buildIntentScore", "intent_score", "build_intent"))
+            sc_raw = _first_non_empty(parsed, ("scope_score", "scopeScore", "scope"))
+            tl_raw = _first_non_empty(parsed, ("timeline_score", "timelineScore", "timeline"))
+            ow_raw = _first_non_empty(parsed, ("owner_score", "ownerScore", "owner"))
+            provided_project_score_raw = _first_non_empty(parsed, ("project_score", "projectScore", "score"))
+
+            bi = _safe_float(bi_raw, 0.0)
+            sc = _safe_float(sc_raw, 0.0)
+            tl = _safe_float(tl_raw, 0.0)
+            ow = _safe_float(ow_raw, 0.0)
+            weighted = round((bi * 0.4) + (sc * 0.25) + (tl * 0.2) + (ow * 0.15), 4)
+
+            component_count = sum(value is not None for value in (bi_raw, sc_raw, tl_raw, ow_raw))
+            has_project_score = provided_project_score_raw is not None
+            provided_project_score = _safe_float(provided_project_score_raw, 0.0)
+
+            final_project_score = weighted
+            if component_count < 4 and has_project_score:
+                final_project_score = provided_project_score
+            elif component_count > 0 and has_project_score:
+                # 当模型给出了总分与维度分，优先采用更保守的较高值，降低误判为 signal 的风险。
+                final_project_score = max(weighted, provided_project_score)
+            final_project_score = max(0.0, min(1.0, final_project_score))
+
+            default_level = "project" if final_project_score >= 0.6 else "signal"
+            normalized_split_items: list[dict[str, str]] = []
+            raw_split_items = parsed.get("split_items")
+            if isinstance(raw_split_items, list):
+                for item in raw_split_items[:2]:
+                    if not isinstance(item, dict):
+                        continue
+                    item_level = _normalize_level_value(item.get("level"), default=default_level)
+                    item_content = str(item.get("content") or "").strip()
+                    if not item_content:
+                        continue
+                    normalized_split_items.append({"level": item_level, "content": item_content})
+
+            level_reasoning = str(parsed.get("level_reasoning") or "").strip() or "模型返回缺少理由，已按结构化兜底。"
+            payload: dict[str, Any] = {
+                "build_intent_score": max(0.0, min(1.0, bi)),
+                "scope_score": max(0.0, min(1.0, sc)),
+                "timeline_score": max(0.0, min(1.0, tl)),
+                "owner_score": max(0.0, min(1.0, ow)),
+                "project_score": final_project_score,
+                "level": _normalize_level_value(parsed.get("level"), default=default_level),
+                "level_confidence": max(
+                    0.0,
+                    min(1.0, _safe_float(parsed.get("level_confidence"), max(final_project_score, 1 - final_project_score))),
+                ),
+                "level_reasoning": level_reasoning[:600],
+                "should_split": _safe_bool(parsed.get("should_split"), False),
+                "split_items": normalized_split_items,
+            }
+            if self._is_simple_signal_source_request(message):
+                payload["project_score"] = min(_safe_float(payload.get("project_score"), final_project_score), 0.35)
+                payload["level"] = "signal"
+                payload["should_split"] = False
+                payload["split_items"] = []
+                payload["level_reasoning"] = (
+                    f"{payload['level_reasoning']}；命中简单需求模式，按信号处理避免误升级为项目。"
+                )[:600]
+            return _validate_schema(_LevelClassificationSchema, payload, context="level_classification")
+        except Exception as exc:
+            logger.warning(f"feishu_level_classification_fallback: error={exc}")
+            fallback = self._heuristic_level_fallback(message)
+            return _validate_schema(_LevelClassificationSchema, fallback, context="level_classification_fallback")
+
+    def _heuristic_level_fallback(self, message: str) -> dict[str, Any]:
+        text = message or ""
+        build_intent_score = 1.0 if PROJECT_INTENT_RE.search(text) else 0.2
+        scope_score = 0.75 if ("系统" in text or "平台" in text or "功能" in text) else 0.35
+        timeline_score = 0.8 if TIMELINE_RE.search(text) else 0.1
+        owner_score = 0.7 if OWNER_RE.search(text) else 0.15
+        project_score = round(
+            (build_intent_score * 0.4)
+            + (scope_score * 0.25)
+            + (timeline_score * 0.2)
+            + (owner_score * 0.15),
+            4,
+        )
+        has_project_intent = bool(PROJECT_INTENT_RE.search(text))
+        has_signal_fact = bool(SIGNAL_FACT_RE.search(text))
+        should_split = bool(has_project_intent and has_signal_fact)
+        simple_signal_request = self._is_simple_signal_source_request(text)
+        if simple_signal_request:
+            project_score = min(project_score, 0.35)
+            should_split = False
+            level = "signal"
+            level_reasoning = "命中简单需求规则，按信号处理，避免误升级项目。"
+        else:
+            level = "project" if project_score >= 0.6 else "signal"
+            level_reasoning = "基于关键词与上下文规则的兜底判定。"
+        return {
+            "build_intent_score": build_intent_score,
+            "scope_score": scope_score,
+            "timeline_score": timeline_score,
+            "owner_score": owner_score,
+            "project_score": project_score,
+            "level": level,
+            "level_confidence": max(project_score, 1 - project_score),
+            "level_reasoning": level_reasoning,
+            "should_split": should_split,
+            "split_items": (
+                [
+                    {"level": "project", "content": text},
+                    {"level": "signal", "content": text},
+                ]
+                if should_split
+                else []
+            ),
+        }
+
+    @staticmethod
+    def _guess_matched_project_id_from_text(message: str, projects: list[dict[str, Any]]) -> Optional[int]:
+        normalized_message = str(message or "").strip().lower()
+        if not normalized_message:
+            return None
+        for row in projects:
+            if not isinstance(row, dict):
+                continue
+            pid_raw = row.get("id")
+            title = str(row.get("name") or row.get("title") or "").strip()
+            if not title or len(title) < 2:
+                continue
+            try:
+                pid = int(pid_raw) if pid_raw is not None else None
+            except Exception:
+                pid = None
+            if pid is None:
+                continue
+            if title.lower() in normalized_message:
+                return pid
+        return None
+
+    def _heuristic_signal_fallback(self, *, message: str, projects: list[dict[str, Any]]) -> dict[str, Any]:
+        text = str(message or "").strip()
+        normalized = text.lower()
+        cleaned = re.sub(r"\s+", " ", text).strip()
+        summary = cleaned[:280] if cleaned else "飞书消息（空内容）"
+
+        signal_type = "新需求"
+        priority = "P2"
+        recommended_action = "enter_only"
+        if any(k in normalized for k in ("故障", "告警", "宕机", "报错", "bug", "风险")):
+            signal_type = "项目风险"
+            priority = "P1"
+            recommended_action = "trigger_alert"
+        elif any(k in normalized for k in ("反馈", "吐槽", "体验", "投诉")):
+            signal_type = "用户反馈"
+            priority = "P2"
+        elif any(k in normalized for k in ("竞品", "对标", "发布", "上线")):
+            signal_type = "竞品动态"
+            priority = "P2"
+
+        title_seed = cleaned or "飞书需求信号"
+        title = title_seed[:48]
+        if not title:
+            title = "飞书需求信号"
+        confidence = 0.78 if signal_type in {"项目风险", "用户反馈"} else 0.72
+
+        return {
+            "title": title,
+            "type": signal_type,
+            "priority": priority,
+            "matched_project_id": self._guess_matched_project_id_from_text(text, projects),
+            "summary": summary,
+            "reasoning": ["LLM 返回异常，已按关键词规则生成兜底信号结构。"],
+            "confidence": confidence,
+            "recommended_action": recommended_action,
+        }
+
+    @staticmethod
+    def _heuristic_project_prefill_fallback(message: str) -> dict[str, Any]:
+        text = str(message or "").strip()
+        cleaned = re.sub(r"\s+", " ", text).strip()
+        suggested_name = cleaned[:30] if cleaned else "飞书新需求项目"
+        if len(suggested_name) < 4:
+            suggested_name = "飞书新需求项目"
+        suggested_description = (
+            f"由飞书需求自动生成的项目草案：{cleaned[:200]}"
+            if cleaned
+            else "由飞书需求自动生成的项目草案，待补充目标、范围与里程碑。"
+        )
+        return {
+            "suggested_name": suggested_name,
+            "suggested_description": suggested_description,
+            "suggested_owner": "待指定",
+            "suggested_timeline": "待评估",
+            "confidence": 0.62,
+            "reasoning": ["LLM 返回异常，已按规则生成项目预填草案。"],
+        }
+
+    async def _load_routing_policy(self) -> dict[str, float | bool]:
+        integration = (
+            await self.db.execute(
+                select(Integration).where(
+                    and_(
+                        Integration.tenant_id == self.tenant_id,
+                        Integration.type == "feishu",
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        config: dict[str, Any] = {}
+        if integration is not None:
+            if isinstance(integration.config, dict):
+                config = integration.config
+            elif isinstance(integration.config, str):
+                config = _safe_json_object(integration.config)
+
+        cold_start_enabled = bool(config.get("cold_start_enabled"))
+        chat_context_auto_link_enabled = bool(config.get("chat_context_auto_link_enabled", False))
+
+        signal_auto_enter_threshold = min(
+            1.0,
+            max(
+                0.0,
+                _safe_float(
+                    config.get("signal_auto_enter_threshold"),
+                    DEFAULT_SIGNAL_AUTO_ENTER_THRESHOLD,
+                ),
+            ),
+        )
+        project_auto_create_threshold = min(
+            1.0,
+            max(
+                0.0,
+                _safe_float(
+                    config.get("project_auto_create_threshold"),
+                    DEFAULT_PROJECT_AUTO_CREATE_THRESHOLD,
+                ),
+            ),
+        )
+        project_uncertain_threshold = min(
+            project_auto_create_threshold,
+            max(
+                0.0,
+                _safe_float(
+                    config.get("project_uncertain_threshold"),
+                    DEFAULT_PROJECT_UNCERTAIN_THRESHOLD,
+                ),
+            ),
+        )
+        return {
+            "cold_start_enabled": cold_start_enabled,
+            "signal_auto_enter_threshold": signal_auto_enter_threshold,
+            "project_auto_create_threshold": project_auto_create_threshold,
+            "project_uncertain_threshold": project_uncertain_threshold,
+            "chat_context_auto_link_enabled": chat_context_auto_link_enabled,
+        }
+
+    async def _load_active_projects(self) -> list[dict[str, Any]]:
+        rows = (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.status != "inactive",
+                    )
+                )
+            )
+        ).scalars().all()
+        return [
+            {
+                "id": int(p.id),
+                "name": p.name,
+                "title": p.title or p.name,
+                "status": p.status,
+                "description": p.description or "",
+            }
+            for p in rows
+        ]
+
+    async def _create_project_from_suggestion(
+        self,
+        suggestion: ProjectSuggestion,
+        *,
+        actor_user_id: Optional[int],
+    ) -> Project:
+        owner_id = await self._resolve_user_id_by_name(suggestion.suggested_owner) or actor_user_id
+        project_name = (suggestion.suggested_name or "新项目").strip()[:255] or "新项目"
+        project_desc = (suggestion.suggested_description or "").strip()[:4000] or None
+        await ensure_project_id_sequence(self.db)
+        project = Project(
+            tenant_id=self.tenant_id,
+            name=project_name,
+            title=project_name,
+            description=project_desc,
+            owner_id=owner_id,
+            status="active",
+            domain_id=None,
+            assigned_role="ai_engineer",
+            signal_count=0,
+        )
+        self.db.add(project)
+        await self.db.flush()
+        await ensure_project_default_members(
+            self.db,
+            tenant_id=self.tenant_id,
+            project_id=int(project.id),
+            owner_id=int(project.owner_id) if project.owner_id is not None else None,
+            assigned_role=project.assigned_role,
+        )
+        return project
+
+    async def _ensure_signal_link(self, *, signal_id: int, project_id: int) -> None:
+        existed = (
+            await self.db.execute(
+                select(ProjectSignalLink.id).where(
+                    and_(
+                        ProjectSignalLink.tenant_id == self.tenant_id,
+                        ProjectSignalLink.project_id == project_id,
+                        ProjectSignalLink.signal_id == signal_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if existed is not None:
+            return
+        row = ProjectSignalLink(
+            tenant_id=self.tenant_id,
+            project_id=int(project_id),
+            signal_id=int(signal_id),
+            link_type="feishu_signal",
+            relevance_score=0.8,
+        )
+        self.db.add(row)
+
+    async def _notify_project_created(self, *, project: Project) -> None:
+        if not project.owner_id:
+            return
+        project_name = project.title or project.name or f"项目#{project.id}"
+        await NotificationService(self.db).create_notification(
+            tenant_id=self.tenant_id,
+            user_id=int(project.owner_id),
+            category="requirement",
+            title=f"创建项目：{project_name}",
+            content=f"AI 已根据飞书消息自动创建项目 [{project_name}]，可直接进入 SSTC 看板。",
+            source="feishu",
+            entity_type="project",
+            entity_id=int(project.id),
+            target_url=f"/projects/{project.id}/board",
+            kind="auto_done",
+            priority="high",
+            status="auto_done",
+            recommended_action="view_result",
+            dedupe_key=f"feishu-project-auto:{int(project.id)}",
+            payload={
+                "message_class": "auto_done",
+                "summary": f"AI 已自动创建项目 {project_name}",
+                "ai_plan": "下一步可进入项目看板补齐 charter 与执行任务。",
+                "project_id": int(project.id),
+                "project_name": project_name,
+                "result_count": 1,
+                "impact_objects": [project_name],
+                "can_undo": True,
+                "undo_deadline": _build_undo_deadline(_utc_now().replace(tzinfo=None)).isoformat(),
+            },
+            commit=False,
+        )
+
+    async def _notify_signal_auto_entered(self, *, project_id: int, signal: Signal) -> None:
+        project = (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.id == project_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not project or not project.owner_id:
+            return
+        project_name = project.title or project.name or f"项目#{project.id}"
+        await NotificationService(self.db).create_notification(
+            tenant_id=self.tenant_id,
+            user_id=int(project.owner_id),
+            category="requirement",
+            title=f"信号已自动入库：{signal.title or '未命名信号'}",
+            content="高置信度信号已自动进入项目信号源，可直接继续生成宪章。",
+            source="feishu",
+            entity_type="signal",
+            entity_id=int(signal.id),
+            target_url=f"/projects/{project_id}/board",
+            kind="auto_done",
+            priority=str(signal.priority or "medium"),
+            status="auto_done",
+            recommended_action="view_result",
+            dedupe_key=f"feishu-signal-auto:{int(signal.id)}:{int(project_id)}",
+            payload={
+                "message_class": "auto_done",
+                "summary": f"AI 已把信号 {signal.title or f'#{signal.id}'} 自动入库到 {project_name}",
+                "ai_plan": "下一步可继续生成 charter、任务或通知负责人。",
+                "project_id": int(project_id),
+                "project_name": project_name,
+                "result_count": 1,
+                "impact_objects": [signal.title or f"信号#{signal.id}", project_name],
+                "can_undo": True,
+                "undo_deadline": _build_undo_deadline(_utc_now().replace(tzinfo=None)).isoformat(),
+            },
+            commit=False,
+        )
+
+    async def _load_project_name(self, project_id: Optional[int]) -> Optional[str]:
+        if project_id is None:
+            return None
+        project = (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.id == int(project_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if project is None:
+            return None
+        return str(project.title or project.name or "").strip() or f"项目#{int(project_id)}"
+
+    @staticmethod
+    def _resolve_message_class(item: FeishuItem) -> str:
+        value = str(item.message_class or "").strip().lower()
+        if value in {"action_required", "auto_running", "auto_done", "info", "chat", "query", "ignored"}:
+            return value
+        route_status = FeishuSignalPipelineService._normalize_route_status(item.route_status)
+        if route_status == ROUTE_STATUS_QUEUED:
+            return "action_required"
+        if route_status == ROUTE_STATUS_AUTO_RUNNING:
+            return "auto_running"
+        if route_status in {ROUTE_STATUS_AUTO_DONE, ROUTE_STATUS_UNDONE}:
+            return "auto_done"
+        return "info"
+
+    @staticmethod
+    def _resolve_auto_execute_state(item: FeishuItem) -> str:
+        value = str(item.auto_execute_state or "").strip().lower()
+        if value:
+            return value
+        route_status = FeishuSignalPipelineService._normalize_route_status(item.route_status)
+        if route_status == ROUTE_STATUS_AUTO_DONE:
+            return "done"
+        if route_status == ROUTE_STATUS_QUEUED:
+            return "blocked"
+        if route_status == ROUTE_STATUS_UNDONE:
+            return "undone"
+        if route_status == ROUTE_STATUS_FAILED:
+            return "failed"
+        return "idle"
+
+    async def _serialize_item(
+        self,
+        item: FeishuItem,
+        *,
+        include_raw: bool = False,
+        include_history: bool = False,
+    ) -> dict[str, Any]:
+        result_payload = item.result_payload if isinstance(item.result_payload, dict) else {}
+        ai_output = _safe_dict(item.ai_output)
+        project_name = await self._load_project_name(_safe_int(item.matched_project_id))
+        row: dict[str, Any] = {
+            "id": str(item.id),
+            "feishu_message_id": item.feishu_message_id,
+            "level": item.level,
+            "card_type": item.card_type,
+            "route_status": self._normalize_route_status(item.route_status),
+            "project_score": item.project_score,
+            "signal_confidence": item.signal_confidence,
+            "level_confidence": item.level_confidence,
+            "level_reasoning": item.level_reasoning,
+            "matched_project_id": item.matched_project_id,
+            "project_name": project_name,
+            "sender": item.source_sender,
+            "created_at": item.created_at.isoformat() if item.created_at else None,
+            "processed_at": item.processed_at.isoformat() if item.processed_at else None,
+            "updated_at": item.updated_at.isoformat() if item.updated_at else None,
+            "decision_reason": item.decision_reason,
+            "blocking_reason": item.blocking_reason,
+            "safe_to_auto_execute": bool(item.safe_to_auto_execute),
+            "auto_execute_state": self._resolve_auto_execute_state(item),
+            "dedupe_key": item.dedupe_key,
+            "message_class": self._resolve_message_class(item),
+            "occurrence_count": int(item.occurrence_count or 1),
+            "next_step_text": item.next_step_text,
+            "undoable_until": item.undoable_until.isoformat() if item.undoable_until else None,
+            "result_payload": result_payload or None,
+            "display_kind": self._resolve_message_class(item),
+            "target_url": result_payload.get("target_url") if isinstance(result_payload, dict) else None,
+            "priority_rank": 0,
+        }
+        if not row["target_url"] and item.matched_project_id:
+            row["target_url"] = f"/projects/{int(item.matched_project_id)}/board"
+
+        if include_raw:
+            row["raw_message"] = item.raw_message
+
+        if (item.level or "").lower() == "project":
+            suggestion = await self._get_project_suggestion(item.id)
+            operation_payload = result_payload.get("operation_payload") if isinstance(result_payload.get("operation_payload"), dict) else {}
+            project_fields = {
+                "title": str(
+                    (suggestion.suggested_name if suggestion is not None else None)
+                    or operation_payload.get("title")
+                    or ""
+                ).strip(),
+                "description": str(
+                    (suggestion.suggested_description if suggestion is not None else None)
+                    or operation_payload.get("description")
+                    or ""
+                ).strip(),
+                "project_scope": str(result_payload.get("project_scope") or "").strip(),
+                "missing_fields": [],
+                "ready": False,
+            }
+            if (
+                self._looks_like_placeholder_project_text(project_fields["title"])
+                or self._looks_like_placeholder_project_text(project_fields["description"])
+                or self._looks_like_placeholder_project_text(project_fields["project_scope"])
+            ):
+                project_fields = await self._maybe_prefill_project_fields_from_context(
+                    message=str(item.raw_message or ""),
+                    source_chat_id=str(item.source_chat_id or ""),
+                    source_sender_open_id=str(item.source_sender_open_id or "") or None,
+                    sender=str(item.source_sender or ""),
+                    fields=project_fields,
+                )
+            if suggestion:
+                created_project_id = _safe_int(suggestion.created_project_id)
+                created_project_name = await self._load_project_name(created_project_id)
+                row["project_suggestion"] = {
+                    "id": str(suggestion.id),
+                    "suggested_name": suggestion.suggested_name,
+                    "suggested_description": suggestion.suggested_description,
+                    "suggested_owner": suggestion.suggested_owner,
+                    "suggested_timeline": suggestion.suggested_timeline,
+                    "confidence": suggestion.confidence,
+                    "reasoning": _normalize_reasoning(suggestion.reasoning),
+                    "status": suggestion.status,
+                    "created_project_id": str(created_project_id) if created_project_id else None,
+                }
+                row["project_name"] = created_project_name or row["project_name"]
+                row["matched_project_id"] = created_project_id or row["matched_project_id"]
+                row["priority"] = "high" if (item.project_score or 0) >= 0.85 else "medium"
+                row["priority_rank"] = 4 if (item.project_score or 0) >= 0.85 else 3
+                if created_project_id and not row["target_url"]:
+                    row["target_url"] = f"/projects/{created_project_id}/board"
+            row["title"] = self._build_project_preview_title(
+                title=str(project_fields.get("title") or ""),
+                description=str(project_fields.get("description") or ""),
+                project_scope=str(project_fields.get("project_scope") or ""),
+            )
+            row["summary"] = self._build_project_preview_summary(
+                description=str(project_fields.get("description") or ""),
+                project_scope=str(project_fields.get("project_scope") or ""),
+                blocking_reason=item.blocking_reason,
+                fallback_text=str(ai_output.get("normalized_text") or item.raw_message or ""),
+            )
+            row["detail_summary"] = row["summary"]
+            if not row.get("priority"):
+                row["priority"] = "high" if (item.project_score or 0) >= 0.85 else "medium"
+                row["priority_rank"] = 4 if (item.project_score or 0) >= 0.85 else 3
+            return row
+
+        if (item.level or "").lower() == "command":
+            operation_code = str(result_payload.get("operation_code") or "").strip()
+            row["title"] = (
+                str(result_payload.get("result_summary") or "").strip()
+                or operation_code
+                or "受控动作"
+            )
+            row["summary"] = str(item.blocking_reason or row["title"] or "").strip()
+            row["priority"] = "high" if row["display_kind"] == "action_required" else "medium"
+            row["priority_rank"] = 4 if row["display_kind"] == "action_required" else 2
+            if not row.get("next_step_text"):
+                row["next_step_text"] = self._build_pending_operation_next_step(operation_code=operation_code)
+            if result_payload:
+                row["command"] = {
+                    "operation_code": operation_code or None,
+                    "executor": str(result_payload.get("executor") or "").strip() or None,
+                    "operation_payload": result_payload.get("operation_payload") if isinstance(result_payload.get("operation_payload"), dict) else None,
+                    "operation_result": result_payload.get("operation_result") if isinstance(result_payload.get("operation_result"), dict) else None,
+                }
+            return row
+
+        signal = await self._get_signal_by_item(item.id)
+        if signal:
+            matched_project_id = _safe_int(signal.matched_project_id) or _safe_int(item.matched_project_id)
+            matched_project_name = await self._load_project_name(matched_project_id)
+            row["signal"] = {
+                "id": str(signal.id),
+                "title": signal.title,
+                "signal_type": signal.signal_type,
+                "priority": signal.priority,
+                "summary": signal.summary,
+                "reasoning": _normalize_reasoning(signal.reasoning),
+                "confidence": signal.confidence,
+                "recommended_action": signal.recommended_action,
+                "matched_project_id": signal.matched_project_id,
+                "status": signal.status,
+            }
+            row["title"] = signal.title or "未命名信号"
+            row["summary"] = signal.summary
+            row["priority"] = signal.priority
+            row["priority_rank"] = PROJECT_PRIORITY_ORDER.get(str(signal.priority or "").upper(), 1)
+            row["matched_project_id"] = matched_project_id
+            row["project_name"] = matched_project_name or row["project_name"]
+            if matched_project_id and not row["target_url"]:
+                row["target_url"] = f"/projects/{matched_project_id}/board"
+            if not row.get("next_step_text"):
+                if row["display_kind"] == "auto_done":
+                    row["next_step_text"] = "下一步可继续生成 charter、任务或通知负责人。"
+                elif not matched_project_id:
+                    row["next_step_text"] = "补齐项目归属后，AI 将自动继续后续链路。"
+            if not row.get("summary"):
+                row["summary"] = _compact_text(signal.content or item.raw_message, max_len=140)
+            if include_history:
+                history_rows = (
+                    await self.db.execute(
+                        select(SignalEdit)
+                        .where(
+                            and_(
+                                SignalEdit.tenant_id == self.tenant_id,
+                                SignalEdit.signal_id == int(signal.id),
+                            )
+                        )
+                        .order_by(SignalEdit.edited_at.desc())
+                        .limit(20)
+                    )
+                ).scalars().all()
+                row["edit_history"] = [
+                    {
+                        "field": h.field,
+                        "old_value": h.old_value,
+                        "new_value": h.new_value,
+                        "edited_by": str(h.edited_by) if h.edited_by else None,
+                        "edited_at": h.edited_at.isoformat() if h.edited_at else None,
+                    }
+                    for h in history_rows
+                    ]
+        return row
+
+    async def _resolve_user_id_by_name(self, name: Optional[str]) -> Optional[int]:
+        key = (name or "").strip()
+        if not key:
+            return None
+        user = (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.is_active.is_(True),
+                        func.lower(User.name) == key.lower(),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if user:
+            return int(user.id)
+        fuzzy = (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.is_active.is_(True),
+                        User.name.ilike(f"%{key}%"),
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+        return int(fuzzy.id) if fuzzy else None
+
+    async def _fallback_user_id(self) -> Optional[int]:
+        user = (
+            await self.db.execute(
+                select(User.id).where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.is_active.is_(True),
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+        return int(user) if user else None
+
+    async def _resolve_chat_context_project(self, *, source_chat_id: str) -> dict[str, Any]:
+        chat_id = (source_chat_id or "").strip()
+        if not chat_id:
+            return {"project_id": None, "ambiguous": False, "candidate_ids": [], "top_ratio": 0.0}
+
+        recent_rows = (
+            await self.db.execute(
+                select(FeishuItem.matched_project_id)
+                .where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.source_chat_id == chat_id,
+                        FeishuItem.matched_project_id.is_not(None),
+                        FeishuItem.route_status.in_((ROUTE_STATUS_AUTO_DONE, "auto_processed", "confirmed")),
+                    )
+                )
+                .order_by(FeishuItem.id.desc())
+                .limit(20)
+            )
+        ).all()
+        project_ids = [int(row[0]) for row in recent_rows if row and row[0] is not None]
+        if not project_ids:
+            return {"project_id": None, "ambiguous": False, "candidate_ids": [], "top_ratio": 0.0}
+
+        unique_ids = set(project_ids)
+        candidate_ids = sorted(unique_ids, key=lambda pid: (-project_ids.count(pid), pid))
+        if len(unique_ids) == 1:
+            return {"project_id": int(project_ids[0]), "ambiguous": False, "candidate_ids": candidate_ids, "top_ratio": 1.0}
+
+        top_pid = max(unique_ids, key=lambda pid: project_ids.count(pid))
+        top_count = project_ids.count(top_pid)
+        ratio = top_count / len(project_ids)
+        if top_count >= 2 and ratio >= 0.6:
+            return {"project_id": int(top_pid), "ambiguous": False, "candidate_ids": candidate_ids, "top_ratio": ratio}
+        return {"project_id": None, "ambiguous": True, "candidate_ids": candidate_ids, "top_ratio": ratio}
+
+    async def _infer_project_id_from_chat_context(self, *, source_chat_id: str) -> Optional[int]:
+        resolution = await self._resolve_chat_context_project(source_chat_id=source_chat_id)
+        return _safe_int(resolution.get("project_id"))
+
+    async def _get_item(self, item_id: int) -> Optional[FeishuItem]:
+        return (
+            await self.db.execute(
+                select(FeishuItem).where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.id == int(item_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _get_project_suggestion(self, item_id: int) -> Optional[ProjectSuggestion]:
+        return (
+            await self.db.execute(
+                select(ProjectSuggestion).where(
+                    and_(
+                        ProjectSuggestion.tenant_id == self.tenant_id,
+                        ProjectSuggestion.feishu_item_id == int(item_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _get_signal_by_item(self, item_id: int) -> Optional[Signal]:
+        return (
+            await self.db.execute(
+                select(Signal).where(
+                    and_(
+                        Signal.tenant_id == self.tenant_id,
+                        Signal.feishu_item_id == int(item_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    def _dedupe_queue_items(self, rows: list[FeishuItem]) -> list[FeishuItem]:
+        deduped: list[FeishuItem] = []
+        seen_signal_keys: set[str] = set()
+        aggregated: dict[str, FeishuItem] = {}
+        for row in rows:
+            explicit_key = str(row.dedupe_key or "").strip()
+            if explicit_key:
+                existing = aggregated.get(explicit_key)
+                if existing is None:
+                    aggregated[explicit_key] = row
+                    deduped.append(row)
+                else:
+                    existing.occurrence_count = max(1, int(existing.occurrence_count or 1)) + max(
+                        1, int(row.occurrence_count or 1)
+                    )
+                continue
+            level = str(row.level or "").lower()
+            if level != "signal":
+                deduped.append(row)
+                continue
+            message_id = str(row.feishu_message_id or "").strip()
+            key = message_id or f"{str(row.raw_message or '')[:120]}::{str(row.source_sender or '')}"
+            if key in seen_signal_keys:
+                continue
+            seen_signal_keys.add(key)
+            deduped.append(row)
+        return deduped
+
+    @staticmethod
+    def _is_water_talk_message(message: str) -> bool:
+        content = str(message or "").strip().lower()
+        if not content:
+            return True
+        normalized = re.sub(r"\s+", "", content)
+        question_hints = (
+            "?",
+            "？",
+            "吗",
+            "怎么",
+            "如何",
+            "为什么",
+            "谁",
+            "多少",
+            "是否",
+            "有没有",
+            "几号",
+            "几点",
+            "周几",
+        )
+        water_keywords = (
+            "你是谁",
+            "你是誰",
+            "什么身份",
+            "身份确认",
+            "今天什么日期",
+            "今天几号",
+            "当前日期",
+            "日期疑问",
+            "询问日期",
+            "现在几点",
+            "几点了",
+            "今天周几",
+            "今天星期几",
+            "说话",
+            "在吗",
+            "在不在",
+        )
+        requirement_hints = (
+            "需求",
+            "优化",
+            "调整",
+            "修复",
+            "bug",
+            "故障",
+            "报错",
+            "项目",
+            "页面",
+            "接口",
+            "功能",
+            "排版",
+            "样式",
+            "任务",
+            "立项",
+            "上线",
+            "配置",
+        )
+        if any(k in normalized for k in water_keywords) and not any(h in normalized for h in requirement_hints):
+            return True
+        if any(k in normalized for k in question_hints):
+            return False
+        return False
+
+    @staticmethod
+    def _build_rag_conversation_key(
+        *,
+        source_chat_id: Optional[str],
+        source_sender_open_id: Optional[str],
+        sender: Optional[str],
+    ) -> str:
+        chat = str(source_chat_id or "").strip()
+        if not chat:
+            return ""
+        actor = str(source_sender_open_id or "").strip()
+        if not actor:
+            actor = str(sender or "").strip().lower()
+            actor = re.sub(r"\s+", "", actor)
+        actor = actor[:96] if actor else "unknown"
+        return f"feishu:{chat}:{actor}"
+
+    @staticmethod
+    def _is_requirement_like_question(message: str) -> bool:
+        text = str(message or "").strip()
+        if not text:
+            return False
+        normalized = re.sub(r"\s+", "", text.lower())
+        has_question = (
+            "?" in normalized
+            or "？" in normalized
+            or any(
+                cue in normalized
+                for cue in ("吗", "是否", "是不是", "要不要", "能不能", "可不可以", "需不需要", "有没有必要")
+            )
+        )
+        if not has_question:
+            return False
+        if INDEPENDENT_REQUIREMENT_RE.search(normalized):
+            return True
+        if SIMPLE_SIGNAL_REQUEST_RE.search(normalized):
+            return True
+        has_verb = bool(QUESTION_REQUIREMENT_VERB_RE.search(normalized))
+        has_object = bool(QUESTION_REQUIREMENT_OBJECT_RE.search(normalized))
+        return has_verb and has_object
+
+    @staticmethod
+    def _extract_requirement_statement_from_question(message: str) -> str:
+        text = str(message or "").strip()
+        if not text:
+            return ""
+        candidate = re.sub(r"^(请问|问下|想问下|我想问下|麻烦问下|请教下)\s*", "", text, flags=re.IGNORECASE).strip()
+        candidate = re.sub(r"\s*[？?！!。\.]+\s*$", "", candidate).strip()
+        candidate = re.sub(r"(吗|呢|呀|吧)\s*$", "", candidate).strip()
+        replacements = (
+            (r"能不能", "需要"),
+            (r"可不可以", "需要"),
+            (r"要不要", "需要"),
+            (r"需不需要", "需要"),
+            (r"有没有必要", "需要"),
+            (r"是不是", ""),
+            (r"是否", ""),
+        )
+        for pattern, repl in replacements:
+            candidate = re.sub(pattern, repl, candidate, flags=re.IGNORECASE)
+        candidate = re.sub(r"\s+", " ", candidate).strip(" ，,:：")
+        return candidate or text
+
+    async def _has_processed_message(self, feishu_message_id: str) -> bool:
+        key = (feishu_message_id or "").strip()
+        if not key:
+            return False
+        existed_in_items = (
+            await self.db.execute(
+                select(FeishuItem.id).where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.feishu_message_id == key,
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+        if existed_in_items is not None:
+            return True
+        existed = (
+            await self.db.execute(
+                select(AIProcessingLog.id).where(
+                    and_(
+                        AIProcessingLog.tenant_id == self.tenant_id,
+                        AIProcessingLog.feishu_message_id == key,
+                        or_(
+                            AIProcessingLog.level.is_(None),
+                            AIProcessingLog.level != "ingress",
+                        ),
+                        or_(
+                            AIProcessingLog.routing_decision.is_(None),
+                            AIProcessingLog.routing_decision != "event_received",
+                        ),
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+        return existed is not None
+
+    async def _acquire_message_lock(self, feishu_message_id: str) -> None:
+        key = (feishu_message_id or "").strip()
+        if not key:
+            return
+        try:
+            bind = self.db.get_bind()
+            dialect_name = str(getattr(getattr(bind, "dialect", None), "name", "") or "").lower()
+        except Exception:
+            dialect_name = ""
+        if dialect_name not in {"postgresql", "postgres"}:
+            return
+        try:
+            await self.db.execute(
+                text("SELECT pg_advisory_xact_lock(hashtext(:key))"),
+                {"key": key},
+            )
+        except Exception as exc:
+            logger.warning(f"feishu_pipeline_message_lock_failed: key={key[:32]}, error={exc}")
+
+    async def _write_processing_log(
+        self,
+        *,
+        feishu_message_id: Optional[str],
+        feishu_item_id: Optional[int],
+        raw_message: Optional[str],
+        ai_raw_output: Optional[str],
+        project_score: Optional[float],
+        signal_confidence: Optional[float],
+        level: Optional[str],
+        routing_decision: str,
+        routing_reason: Optional[str],
+        reasoning_bullets: Optional[list[str]],
+        processing_status: str,
+        error_message: Optional[str],
+        duration_ms: Optional[int],
+        normalized_text: Optional[str] = None,
+        route: Optional[str] = None,
+        intent: Optional[str] = None,
+        executor: Optional[str] = None,
+        operation_code: Optional[str] = None,
+        operation_result: Optional[dict[str, Any]] = None,
+        reply_source: Optional[str] = None,
+    ) -> None:
+        normalized_route = (route or routing_decision or "")[:80] or None
+        normalized_intent = (intent or level or "")[:40] or None
+        normalized_executor = (executor or ("operation_executor" if operation_code else "router"))[:40] or None
+        normalized_reply_source = (
+            reply_source
+            or ("db_commit" if operation_code else "deterministic_reply")
+        )
+        self.db.add(
+            AIProcessingLog(
+                tenant_id=self.tenant_id,
+                feishu_message_id=(feishu_message_id or "")[:128] or None,
+                feishu_item_id=int(feishu_item_id) if feishu_item_id is not None else None,
+                raw_message=(raw_message or "")[:8000] or None,
+                ai_raw_output=(ai_raw_output or "")[:120000] or None,
+                project_score=project_score,
+                signal_confidence=signal_confidence,
+                level=(level or "")[:20] or None,
+                normalized_text=(normalized_text or raw_message or "")[:8000] or None,
+                route=normalized_route,
+                intent=normalized_intent,
+                executor=normalized_executor,
+                operation_code=(operation_code or "")[:80] or None,
+                operation_result=operation_result if isinstance(operation_result, dict) else None,
+                reply_source=(normalized_reply_source or "")[:40] or None,
+                routing_decision=(routing_decision or ROUTING_FAILED)[:64],
+                routing_reason=(routing_reason or "")[:1000] or None,
+                reasoning_bullets=[str(x).strip()[:120] for x in (reasoning_bullets or [])[:3]],
+                processing_status=(processing_status or "success")[:20],
+                error_message=(error_message or "")[:4000] or None,
+                duration_ms=max(0, int(duration_ms)) if duration_ms is not None else None,
+            )
+        )
+
+    def _build_project_routing_log(
+        self,
+        *,
+        project_score: float,
+        cold_start_enabled: bool,
+        project_auto_create_threshold: float,
+        project_uncertain_threshold: float,
+    ) -> tuple[str, str, list[str]]:
+        _ = project_auto_create_threshold
+        if cold_start_enabled:
+            return (
+                ROUTING_PENDING_COLD_START,
+                "当前处于冷启动期，项目候选会统一进入待处理队列。",
+                [
+                    "冷启动期开启，自动创建策略暂缓。",
+                    f"project_score={project_score:.2f}，仅用于排序参考。",
+                    "等待人工确认后再执行创建动作。",
+                ],
+            )
+        if project_score >= project_uncertain_threshold:
+            return (
+                ROUTING_PENDING_UNCERTAIN_PROJECT,
+                "项目分数较高，但建项已收口到待决策链，不再自动创建。",
+                [
+                    f"project_score={project_score:.2f}，高于 {project_uncertain_threshold:.2f}。",
+                    "即使意图强，也必须等待受控 project.create。",
+                    "已进入待处理队列等待确认。",
+                ],
+            )
+        return (
+            ROUTING_PENDING_LOW_CONFIDENCE,
+            "项目分数偏低，暂不自动创建，进入待处理队列。",
+            [
+                f"project_score={project_score:.2f}，低于 {project_uncertain_threshold:.2f}。",
+                "消息更接近执行信号而非完整立项需求。",
+                "需要人工补充信息后再决策。",
+            ],
+        )
+
+    def _build_signal_routing_log(
+        self,
+        *,
+        confidence: float,
+        cold_start_enabled: bool,
+        matched_project_id: Optional[int],
+        signal_auto_enter_threshold: float,
+    ) -> tuple[str, str, list[str]]:
+        if cold_start_enabled:
+            return (
+                ROUTING_PENDING_COLD_START,
+                "当前处于冷启动期，信号统一进入待处理队列。",
+                [
+                    "冷启动期开启，自动入库策略暂缓。",
+                    f"signal_confidence={confidence:.2f}，仅用于排序参考。",
+                    "等待人工审核后再入库。",
+                ],
+            )
+        if confidence < signal_auto_enter_threshold:
+            return (
+                ROUTING_PENDING_LOW_CONFIDENCE,
+                "信号置信度低于自动入库阈值，已进入待处理队列。",
+                [
+                    f"signal_confidence={confidence:.2f}，低于 {signal_auto_enter_threshold:.2f} 阈值。",
+                    "当前信息不足以支持自动决策。",
+                    "需要人工确认后再执行后续动作。",
+                ],
+            )
+        if not matched_project_id:
+            return (
+                ROUTING_PENDING_UNCERTAIN_PROJECT,
+                "置信度达标但项目归属不确定，转人工确认。",
+                [
+                    f"signal_confidence={confidence:.2f}，达到自动判定线。",
+                    "未识别到明确项目归属。",
+                    "已进入待处理队列等待项目确认。",
+                ],
+            )
+        return (
+            ROUTING_AUTO_ENTER_SIGNAL,
+            "信号置信度达标且项目归属明确，系统已自动入库。",
+            [
+                f"signal_confidence={confidence:.2f}，高于 {signal_auto_enter_threshold:.2f} 阈值。",
+                f"已匹配项目 #{int(matched_project_id)}。",
+                "已自动建立信号关联并发送通知。",
+            ],
+        )
+
+    @staticmethod
+    def _normalize_recommended_action(value: Any) -> str:
+        action = str(value or "enter_only").strip().lower()
+        if action not in {"enter_only", "create_charter", "trigger_alert"}:
+            return "enter_only"
+        return action
+
+    @staticmethod
+    def _normalize_signal_type(value: Any) -> str:
+        signal_type = str(value or "新需求").strip()
+        if signal_type in SIGNAL_TYPE_OPTIONS:
+            return signal_type
+        return "新需求"
+
+    @staticmethod
+    def _to_signal_priority(value: str) -> str:
+        key = str(value or "").strip().upper()
+        if key in {"P0", "HIGH"}:
+            return "high"
+        if key in {"P1", "MEDIUM"}:
+            return "medium"
+        if key in {"P2", "LOW"}:
+            return "low"
+        if key == "P3":
+            return "low"
+        return "medium"
+
+    def _should_promote_signal_to_project_candidate(
+        self,
+        *,
+        content: str,
+        signal_info: dict[str, Any],
+        confidence: float,
+        signal_auto_enter_threshold: float,
+        simple_signal_request: bool,
+    ) -> bool:
+        _ = (content, signal_info, confidence, signal_auto_enter_threshold, simple_signal_request)
+        # 项目创建只能通过 intent gate:create_project + 受控 operation executor 落库；
+        # 不再把 signal 自由升级为 project candidate。
+        return False
+
+    @staticmethod
+    def _is_simple_signal_source_request(message: str) -> bool:
+        text = re.sub(r"\s+", "", str(message or ""))
+        if len(text) < 6:
+            return False
+        if PROJECT_COMPLEXITY_RE.search(text):
+            return False
+        if SIMPLE_SIGNAL_REQUEST_RE.search(text):
+            return True
+        action_tokens = ("新增", "增加", "实现", "支持", "修复", "优化", "接入", "改造", "调整")
+        object_tokens = ("功能", "能力", "页面", "接口", "流程", "按钮", "提醒", "识别", "字段", "规则", "逻辑")
+        return (
+            len(text) <= 180
+            and any(token in text for token in action_tokens)
+            and any(token in text for token in object_tokens)
+        )
diff --git a/backend/app/services/fetcher.py b/backend/app/services/fetcher.py
new file mode 100644
index 00000000..3dbf97cc
--- /dev/null
+++ b/backend/app/services/fetcher.py
@@ -0,0 +1,9 @@
+"""外部信号抓取兼容实现。"""
+
+from __future__ import annotations
+
+
+async def fetch_hn_stories(limit: int = 30) -> list[dict]:
+    """返回空列表，保留接口契约。"""
+    _ = limit
+    return []
diff --git a/backend/app/services/integration.py b/backend/app/services/integration.py
new file mode 100644
index 00000000..12c2b7e9
--- /dev/null
+++ b/backend/app/services/integration.py
@@ -0,0 +1,540 @@
+"""
+Neural Sync 2.0 - Integration Service
+
+集成服务 - 管理 Plane 和飞书集成
+- 配置管理
+- 状态监控
+- 健康检查
+- 故障降级
+"""
+
+import asyncio
+import logging
+from typing import Optional, Dict, List, Any
+from pydantic import BaseModel, Field
+from datetime import datetime, timedelta
+from enum import Enum
+import httpx
+
+logger = logging.getLogger(__name__)
+
+
+class IntegrationType(str, Enum):
+    """集成类型"""
+    PLANE = "plane"
+    LARK = "lark"
+    GITHUB = "github"
+    GITLAB = "gitlab"
+
+
+class IntegrationStatus(str, Enum):
+    """集成状态"""
+    ACTIVE = "active"
+    INACTIVE = "inactive"
+    ERROR = "error"
+    DEGRADED = "degraded"
+    PENDING = "pending"
+
+
+class PlaneConfig(BaseModel):
+    """Plane 配置"""
+    api_url: str = "https://api.plane.so/api/v1"
+    api_key: str = ""
+    workspace_slug: str = ""
+    project_id: str = ""
+    enabled: bool = True
+
+
+class LarkConfig(BaseModel):
+    """飞书配置"""
+    app_id: str = ""
+    app_secret: str = ""
+    verification_token: str = ""
+    encrypt_key: str = ""
+    default_chat_id: str = ""
+    enabled: bool = True
+
+
+class IntegrationHealth(BaseModel):
+    """集成健康状态"""
+    type: IntegrationType
+    status: IntegrationStatus
+    last_check: datetime
+    last_success: Optional[datetime] = None
+    last_error: Optional[str] = None
+    error_count: int = 0
+    consecutive_failures: int = 0
+    latency_ms: Optional[float] = None
+    message: str = ""
+
+
+class IntegrationConfig(BaseModel):
+    """集成配置"""
+    plane: PlaneConfig = Field(default_factory=PlaneConfig)
+    lark: LarkConfig = Field(default_factory=LarkConfig)
+
+
+class TaskSyncResult(BaseModel):
+    """任务同步结果"""
+    success: bool
+    integration_type: IntegrationType
+    tasks_created: int = 0
+    tasks_updated: int = 0
+    errors: List[str] = Field(default_factory=list)
+    details: Dict[str, Any] = Field(default_factory=dict)
+
+
+class NotificationResult(BaseModel):
+    """通知发送结果"""
+    success: bool
+    integration_type: IntegrationType
+    recipients: int = 0
+    error: Optional[str] = None
+
+
+class IntegrationService:
+    """
+    集成服务 - 管理外部集成
+
+    特性：
+    - 配置管理（动态加载/保存）
+    - 健康检查和监控
+    - 自动故障降级
+    - 重试机制
+    - 事件通知
+    """
+
+    # 健康检查配置
+    HEALTH_CHECK_INTERVAL = 60  # 秒
+    MAX_CONSECUTIVE_FAILURES = 3
+    DEGRADE_THRESHOLD = 5  # 错误次数后降级
+
+    def __init__(self, config: Optional[IntegrationConfig] = None, event_bus=None):
+        self.config = config or IntegrationConfig()
+        self.event_bus = event_bus
+
+        # 健康状态
+        self._health_status: Dict[IntegrationType, IntegrationHealth] = {
+            IntegrationType.PLANE: IntegrationHealth(
+                type=IntegrationType.PLANE,
+                status=IntegrationStatus.PENDING,
+                last_check=datetime.utcnow(),
+            ),
+            IntegrationType.LARK: IntegrationHealth(
+                type=IntegrationType.LARK,
+                status=IntegrationStatus.PENDING,
+                last_check=datetime.utcnow(),
+            ),
+        }
+
+        # 统计信息
+        self._stats = {
+            IntegrationType.PLANE: {"synced": 0, "failed": 0},
+            IntegrationType.LARK: {"sent": 0, "failed": 0},
+        }
+
+        # 服务实例（延迟加载）
+        self._plane_service = None
+        self._lark_service = None
+
+        # 健康检查任务
+        self._health_check_task: Optional[asyncio.Task] = None
+
+    async def start(self):
+        """启动集成服务"""
+        self._health_check_task = asyncio.create_task(self._health_check_loop())
+        logger.info("[Integration] Service started")
+
+    async def stop(self):
+        """停止集成服务"""
+        if self._health_check_task:
+            self._health_check_task.cancel()
+            try:
+                await self._health_check_task
+            except asyncio.CancelledError:
+                pass
+        logger.info("[Integration] Service stopped")
+
+    async def sync_tasks_to_plane(
+        self,
+        tasks: List[Dict[str, Any]],
+        is_urgent: bool = False
+    ) -> TaskSyncResult:
+        """
+        同步任务到 Plane
+
+        增强功能：
+        - 批量创建
+        - 标签管理
+        - 优先级映射
+        - 错误重试
+        """
+        integration_type = IntegrationType.PLANE
+        health = self._health_status[integration_type]
+
+        # 检查服务是否可用
+        if not self._is_integration_available(integration_type):
+            return TaskSyncResult(
+                success=False,
+                integration_type=integration_type,
+                errors=[f"Plane integration is {health.status.value}"],
+            )
+
+        start_time = datetime.utcnow()
+
+        try:
+            # 获取 Plane 服务
+            plane = await self._get_plane_service()
+
+            # 批量同步
+            result = await plane.sync_tasks(tasks, is_urgent)
+
+            # 更新统计
+            if result.get("status") == "success":
+                self._stats[integration_type]["synced"] += result.get("created_count", 0)
+                await self._record_success(integration_type)
+            else:
+                self._stats[integration_type]["failed"] += len(result.get("errors", []))
+                await self._record_failure(integration_type, ", ".join(result.get("errors", [])))
+
+            # 计算延迟
+            latency = (datetime.utcnow() - start_time).total_seconds() * 1000
+            health.latency_ms = latency
+
+            return TaskSyncResult(
+                success=result.get("status") == "success",
+                integration_type=integration_type,
+                tasks_created=result.get("created_count", 0),
+                errors=result.get("errors", []),
+                details=result,
+            )
+
+        except Exception as e:
+            logger.error(f"[Integration] Plane sync failed: {e}")
+            await self._record_failure(integration_type, str(e))
+            return TaskSyncResult(
+                success=False,
+                integration_type=integration_type,
+                errors=[str(e)],
+            )
+
+    async def send_strategic_alert(
+        self,
+        chat_id: str,
+        signal_content: str,
+        decision: Dict[str, Any],
+        tasks: List[Dict[str, Any]]
+    ) -> NotificationResult:
+        """
+        发送战略预警到飞书
+
+        增强功能：
+        - 富文本卡片
+        - 可操作按钮
+        - @提及支持
+        - 消息去重
+        """
+        integration_type = IntegrationType.LARK
+        health = self._health_status[integration_type]
+
+        # 检查服务是否可用
+        if not self._is_integration_available(integration_type):
+            return NotificationResult(
+                success=False,
+                integration_type=integration_type,
+                error=f"Lark integration is {health.status.value}",
+            )
+
+        start_time = datetime.utcnow()
+
+        try:
+            # 获取飞书服务
+            lark = await self._get_lark_service()
+
+            # 发送卡片消息
+            await lark.send_strategic_alert(chat_id, signal_content, decision, tasks)
+
+            # 更新统计
+            self._stats[integration_type]["sent"] += 1
+            await self._record_success(integration_type)
+
+            # 计算延迟
+            latency = (datetime.utcnow() - start_time).total_seconds() * 1000
+            health.latency_ms = latency
+
+            # 发送事件
+            if self.event_bus:
+                from ..events import EventType, Event
+                await self.event_bus.publish(
+                    Event(
+                        type=EventType.LARK_NOTIFY_SUCCESS,
+                        data={
+                            "type": "strategic_alert",
+                            "chat_id": chat_id,
+                        },
+                    )
+                )
+
+            return NotificationResult(
+                success=True,
+                integration_type=integration_type,
+                recipients=1,
+            )
+
+        except Exception as e:
+            logger.error(f"[Integration] Lark notification failed: {e}")
+            await self._record_failure(integration_type, str(e))
+
+            # 发送失败事件
+            if self.event_bus:
+                from ..events import EventType, Event
+                await self.event_bus.publish(
+                    Event(
+                        type=EventType.LARK_NOTIFY_FAILED,
+                        data={
+                            "type": "strategic_alert",
+                            "error": str(e),
+                        },
+                    )
+                )
+
+            return NotificationResult(
+                success=False,
+                integration_type=integration_type,
+                error=str(e),
+            )
+
+    async def send_card_message(
+        self,
+        chat_id: str,
+        content: str,
+        tasks: List[Dict[str, Any]]
+    ) -> NotificationResult:
+        """
+        发送卡片消息到飞书
+        """
+        integration_type = IntegrationType.LARK
+        health = self._health_status[integration_type]
+
+        if not self._is_integration_available(integration_type):
+            return NotificationResult(
+                success=False,
+                integration_type=integration_type,
+                error=f"Lark integration is {health.status.value}",
+            )
+
+        try:
+            lark = await self._get_lark_service()
+            await lark.send_card_message(chat_id, content, tasks)
+
+            self._stats[integration_type]["sent"] += 1
+            await self._record_success(integration_type)
+
+            return NotificationResult(
+                success=True,
+                integration_type=integration_type,
+                recipients=1,
+            )
+
+        except Exception as e:
+            logger.error(f"[Integration] Lark card message failed: {e}")
+            await self._record_failure(integration_type, str(e))
+            return NotificationResult(
+                success=False,
+                integration_type=integration_type,
+                error=str(e),
+            )
+
+    async def update_plane_issue_status(
+        self,
+        issue_id: str,
+        status: str = "Done"
+    ) -> bool:
+        """更新 Plane 工单状态"""
+        integration_type = IntegrationType.PLANE
+
+        if not self._is_integration_available(integration_type):
+            return False
+
+        try:
+            plane = await self._get_plane_service()
+            result = await plane.update_issue_status(issue_id, status)
+
+            if "error" not in result:
+                await self._record_success(integration_type)
+                return True
+
+            await self._record_failure(integration_type, result.get("error", "Unknown error"))
+            return False
+
+        except Exception as e:
+            logger.error(f"[Integration] Plane status update failed: {e}")
+            await self._record_failure(integration_type, str(e))
+            return False
+
+    def get_health_status(self) -> Dict[str, IntegrationHealth]:
+        """获取所有集成健康状态"""
+        return {k.value: v for k, v in self._health_status.items()}
+
+    def get_stats(self) -> Dict[str, Dict[str, int]]:
+        """获取集成统计信息"""
+        return {k.value: v for k, v in self._stats.items()}
+
+    async def check_plane_health(self) -> IntegrationHealth:
+        """检查 Plane 集成健康"""
+        integration_type = IntegrationType.PLANE
+        health = self._health_status[integration_type]
+        start_time = datetime.utcnow()
+
+        if not self.config.plane.enabled or not self.config.plane.api_key:
+            health.status = IntegrationStatus.INACTIVE
+            health.message = "Plane integration is disabled"
+            health.last_check = datetime.utcnow()
+            return health
+
+        try:
+            # 简单的健康检查：尝试获取项目标签
+            plane = await self._get_plane_service()
+            label_id = await plane.get_label_id("__health_check__")
+
+            health.status = IntegrationStatus.ACTIVE
+            health.message = "Connected"
+            health.last_check = datetime.utcnow()
+            health.latency_ms = (datetime.utcnow() - start_time).total_seconds() * 1000
+
+        except Exception as e:
+            logger.warning(f"[Integration] Plane health check failed: {e}")
+            health.status = IntegrationStatus.ERROR
+            health.message = f"Health check failed: {str(e)}"
+            health.last_check = datetime.utcnow()
+
+        return health
+
+    async def check_lark_health(self) -> IntegrationHealth:
+        """检查飞书集成健康"""
+        integration_type = IntegrationType.LARK
+        health = self._health_status[integration_type]
+        start_time = datetime.utcnow()
+
+        if not self.config.lark.enabled or not self.config.lark.app_id:
+            health.status = IntegrationStatus.INACTIVE
+            health.message = "Lark integration is disabled"
+            health.last_check = datetime.utcnow()
+            return health
+
+        try:
+            # 简单的健康检查：尝试获取访问令牌
+            lark = await self._get_lark_service()
+            token = await lark.get_tenant_access_token()
+
+            if token and token != "mock_token":
+                health.status = IntegrationStatus.ACTIVE
+                health.message = "Connected"
+            else:
+                health.status = IntegrationStatus.DEGRADED
+                health.message = "Running in simulation mode"
+
+            health.last_check = datetime.utcnow()
+            health.latency_ms = (datetime.utcnow() - start_time).total_seconds() * 1000
+
+        except Exception as e:
+            logger.warning(f"[Integration] Lark health check failed: {e}")
+            health.status = IntegrationStatus.ERROR
+            health.message = f"Health check failed: {str(e)}"
+            health.last_check = datetime.utcnow()
+
+        return health
+
+    async def update_config(self, config: IntegrationConfig) -> None:
+        """更新集成配置"""
+        self.config = config
+
+        # 重新评估健康状态
+        await self.check_plane_health()
+        await self.check_lark_health()
+
+        logger.info("[Integration] Configuration updated")
+
+    async def _get_plane_service(self):
+        """获取 Plane 服务实例"""
+        if self._plane_service is None:
+            from .plane import PlaneService
+            self._plane_service = PlaneService()
+            # 应用配置
+            if self.config.plane.api_url:
+                self._plane_service.api_url = self.config.plane.api_url
+            if self.config.plane.api_key:
+                self._plane_service.api_key = self.config.plane.api_key
+            if self.config.plane.workspace_slug:
+                self._plane_service.workspace_slug = self.config.plane.workspace_slug
+            if self.config.plane.project_id:
+                self._plane_service.project_id = self.config.plane.project_id
+            self._plane_service.is_simulation = not self.config.plane.api_key
+        return self._plane_service
+
+    async def _get_lark_service(self):
+        """获取飞书服务实例"""
+        if self._lark_service is None:
+            from .lark import LarkService
+            self._lark_service = LarkService()
+            # 应用配置
+            if self.config.lark.app_id:
+                self._lark_service.app_id = self.config.lark.app_id
+            if self.config.lark.app_secret:
+                self._lark_service.app_secret = self.config.lark.app_secret
+            if self.config.lark.verification_token:
+                self._lark_service.verification_token = self.config.lark.verification_token
+        return self._lark_service
+
+    def _is_integration_available(self, integration_type: IntegrationType) -> bool:
+        """检查集成是否可用"""
+        health = self._health_status[integration_type]
+        return health.status in [IntegrationStatus.ACTIVE, IntegrationStatus.DEGRADED]
+
+    async def _record_success(self, integration_type: IntegrationType) -> None:
+        """记录成功操作"""
+        health = self._health_status[integration_type]
+        health.last_success = datetime.utcnow()
+        health.consecutive_failures = 0
+        if health.status == IntegrationStatus.ERROR:
+            health.status = IntegrationStatus.ACTIVE
+
+    async def _record_failure(self, integration_type: IntegrationType, error: str) -> None:
+        """记录失败操作"""
+        health = self._health_status[integration_type]
+        health.last_error = error
+        health.error_count += 1
+        health.consecutive_failures += 1
+        health.last_check = datetime.utcnow()
+
+        # 达到降级阈值
+        if health.consecutive_failures >= self.MAX_CONSECUTIVE_FAILURES:
+            health.status = IntegrationStatus.ERROR
+            health.message = f"Service degraded after {health.consecutive_failures} consecutive failures"
+            logger.warning(f"[Integration] {integration_type.value} degraded after {health.consecutive_failures} failures")
+
+    async def _health_check_loop(self) -> None:
+        """健康检查循环"""
+        while True:
+            try:
+                await asyncio.sleep(self.HEALTH_CHECK_INTERVAL)
+
+                await self.check_plane_health()
+                await self.check_lark_health()
+
+            except asyncio.CancelledError:
+                break
+            except Exception as e:
+                logger.error(f"[Integration] Health check loop error: {e}")
+
+
+# 全局服务实例
+_integration_service = None
+
+
+def get_integration_service(config: Optional[IntegrationConfig] = None, event_bus=None) -> IntegrationService:
+    """获取集成服务实例"""
+    global _integration_service
+    if _integration_service is None:
+        _integration_service = IntegrationService(config, event_bus)
+    return _integration_service
diff --git a/backend/app/services/lark.py b/backend/app/services/lark.py
new file mode 100644
index 00000000..ac93a697
--- /dev/null
+++ b/backend/app/services/lark.py
@@ -0,0 +1,36 @@
+"""飞书服务兼容层。"""
+
+from __future__ import annotations
+
+
+class LarkService:
+    def __init__(self) -> None:
+        self.app_id: str | None = None
+        self.app_secret: str | None = None
+        self.verification_token: str | None = None
+
+    async def send_strategic_alert(
+        self,
+        chat_id: str,
+        signal_content: str,
+        decision_proposal: dict,
+        tasks: list[dict],
+    ) -> dict:
+        return {
+            "chat_id": chat_id,
+            "status": "sent",
+            "tasks": len(tasks or []),
+            "summary": (signal_content or "")[:80],
+            "decision": decision_proposal,
+        }
+
+    async def send_card_message(self, chat_id: str, content: str, tasks: list[dict]) -> dict:
+        return {
+            "chat_id": chat_id,
+            "status": "sent",
+            "tasks": len(tasks or []),
+            "content_preview": (content or "")[:80],
+        }
+
+
+lark_service = LarkService()
diff --git a/backend/app/services/llm.py b/backend/app/services/llm.py
new file mode 100644
index 00000000..d6cef769
--- /dev/null
+++ b/backend/app/services/llm.py
@@ -0,0 +1,2560 @@
+"""统一 LLM 服务（DeepSeek / GLM / OpenAI / Ollama）。
+
+能力：
+1. 多引擎配置持久化（Integration.type=llm_engine）
+2. 动态选择激活引擎
+3. OpenAI-Compatible chat/completions 调用
+4. 基础 AI 调用日志写入（ai_call_logs）
+"""
+
+from __future__ import annotations
+
+import asyncio
+from dataclasses import dataclass
+import hashlib
+import json
+import os
+import re
+import time
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from typing import Any, Optional
+from uuid import uuid4
+
+import httpx
+from sqlalchemy import and_, select
+from sqlalchemy.exc import OperationalError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import build_default_role_prompt_bundle
+from ..database import AsyncSessionLocal
+from ..models import AICallLog, Integration
+
+try:
+    import redis.asyncio as redis_async
+
+    REDIS_AVAILABLE = True
+except Exception:  # pragma: no cover
+    REDIS_AVAILABLE = False
+    redis_async = None
+
+SUPPORTED_PROVIDERS = {"deepseek", "glm", "openai", "ollama"}
+OPENCLAW_PRIMARY_GATEWAY_URL = "http://openclaw:18789/v1"
+OPENCLAW_DEFAULT_LOW_MODEL = "zai/glm-4-flash"
+OPENCLAW_DEFAULT_HIGH_MODEL = "minimax/MiniMax-M2.5"
+OPENCLAW_DEFAULT_SAFETY_MODEL = "deepseek/deepseek-chat"
+PROVIDER_DEFAULTS: dict[str, dict[str, str]] = {
+    "deepseek": {
+        "name": "DeepSeek",
+        "base_url": "https://api.deepseek.com/v1",
+        "model": "deepseek-chat",
+        "api_env": "DEEPSEEK_API_KEY",
+    },
+    "glm": {
+        "name": "GLM",
+        "base_url": "https://open.bigmodel.cn/api/paas/v4",
+        "model": "glm-4-flash",
+        "api_env": "GLM_API_KEY",
+    },
+    "openai": {
+        "name": "OpenAI",
+        "base_url": OPENCLAW_PRIMARY_GATEWAY_URL,
+        "model": OPENCLAW_DEFAULT_LOW_MODEL,
+        "api_env": "OPENAI_API_KEY",
+    },
+    "ollama": {
+        "name": "Ollama",
+        "base_url": "http://127.0.0.1:11434",
+        "model": "llama3:latest",
+        "api_env": "",
+    },
+}
+PROVIDER_ALIASES = {
+    "zhipu": "glm",
+    "chatglm": "glm",
+    "deep-seek": "deepseek",
+    "local": "ollama",
+}
+PROMPT_PROFILE_CACHE_TTL_SECONDS = 60.0
+BOOL_TRUE_VALUES = {"1", "true", "yes", "y", "on", "是"}
+OPENCLAW_ROUTING_CACHE_TTL_SECONDS = 30.0
+OPENCLAW_BACKEND_RULES_PATH = Path(__file__).resolve().parents[2] / ".openclaw" / "pmflow-routing-rules.json"
+OPENCLAW_PROJECT_RULES_PATH = Path(__file__).resolve().parents[3] / ".openclaw" / "pmflow-routing-rules.json"
+OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS = 120
+OPENCLAW_AGENTS_TIMEOUT_SECONDS_MAX = 120
+DEFAULT_TIER_MODEL_OVERRIDES = {
+    "low": OPENCLAW_DEFAULT_LOW_MODEL,
+    "high": OPENCLAW_DEFAULT_HIGH_MODEL,
+    "safety": OPENCLAW_DEFAULT_SAFETY_MODEL,
+}
+DEFAULT_HIGH_TIER_HINTS = (
+    "project_from_text",
+    "feishu_project_prefill",
+    "feishu_task_orchestration",
+    "feishu_project_signal_classification",
+    "adaptive_rag_answer_deep",
+    "rag_db_project_semantic_retrieval",
+    "analyze_food_calories",
+    "sstc_ai_breakdown",
+    "generate_task_breakdown",
+    "recommend_assignee",
+    "alignment",
+    "charter",
+    "cognitive_conflict",
+    "blocker",
+    "model_requirement",
+)
+DEFAULT_LOW_TIER_HINTS = (
+    "feishu_level_classification",
+    "feishu_signal_extraction",
+    "feishu_humanized_reply",
+    "signal_source_onboarding_plan",
+    "openclaw_command_dispatch",
+    "signal_source_ai_worker_execution",
+    "extract_context_insights",
+    "adaptive_rag_answer_normal",
+    "rag_intent_classification",
+    "classify_intent",
+)
+STRUCTURED_OUTPUT_FUNCTION_HINTS = (
+    "feishu_level_classification",
+    "feishu_signal_extraction",
+    "signal_source_onboarding_plan",
+    "signal_source_ai_worker_execution",
+    "rag_intent_classification",
+    "classify_intent",
+    "feishu_project_prefill",
+    "feishu_task_orchestration",
+    "openclaw_command_dispatch",
+    "generate_task_breakdown",
+    "generate_charter_from_signal",
+    "decompose_tasks_from_charter",
+    "recommend_assignees_for_task",
+    "evaluate_model_requirement",
+    "calculate_alignment_score",
+    "detect_cognitive_conflicts",
+    "rag_db_project_semantic_retrieval",
+    "workload_optimize_prompt",
+    "analyze_cost_roi",
+    "analyze_food_calories",
+    "sstc_ai_breakdown",
+    "sstc_charter_localize_zh",
+    "sstc_charter_prefill",
+)
+LOW_TO_HIGH_ESCALATION_FALLBACK_MARKERS = (
+    "不知道",
+    "不清楚",
+    "无法回答",
+    "无法确定",
+    "无法判断",
+    "信息不足",
+    "缺少信息",
+    "没有足够信息",
+    "请提供更多信息",
+    "抱歉",
+    "不能提供",
+    "无法访问",
+    "as an ai",
+    "i'm sorry",
+    "cannot answer",
+    "don't know",
+)
+FAST_LOCAL_FALLBACK_FUNCTION_HINTS = (
+    "signal_source_onboarding_plan",
+    "openclaw_command_dispatch",
+    "signal_source_ai_worker_execution",
+)
+FAST_LOCAL_PLANNER_FUNCTION_HINTS = (
+    "signal_source_onboarding_plan",
+    "openclaw_command_dispatch",
+)
+FAST_LOCAL_WORKER_FUNCTION_HINTS = ("signal_source_ai_worker_execution",)
+FAST_LOCAL_FALLBACK_FAST_TIMEOUT_SECONDS = 3.0
+FAST_LOCAL_FALLBACK_COMPLETE_TIMEOUT_SECONDS = 6.0
+STRUCTURED_CACHE_TTL_SECONDS = 300
+FAIL_FAST_ERROR_MARKERS = (
+    "primary_engine_empty_response",
+    "high_tier_empty_response",
+    "llm 返回空内容",
+    "返回空内容",
+    "empty_response",
+    "provider_circuit_open",
+)
+CIRCUIT_BREAKER_FAILURE_THRESHOLD = 2
+CIRCUIT_BREAKER_COOLDOWN_SECONDS = 300.0
+STRUCTURED_OBJECT_SCHEMAS: dict[str, dict[str, Any]] = {
+    "signal_source_onboarding_plan": {
+        "required_keys": ("should_onboard", "summary", "worker_roles", "openclaw_instruction"),
+        "type_checks": {
+            "should_onboard": "bool",
+            "summary": "str",
+            "worker_roles": "list",
+            "openclaw_instruction": "str",
+        },
+        "optional_type_checks": {
+            "source": ("dict_or_none",),
+        },
+    },
+    "openclaw_command_dispatch": {
+        "required_keys": ("summary", "confidence", "operations"),
+        "type_checks": {
+            "summary": "str",
+            "confidence": "number",
+            "operations": "list",
+        },
+    },
+    "signal_source_ai_worker_execution": {
+        "required_keys": ("summary", "deliverables", "risks", "next_actions"),
+        "type_checks": {
+            "summary": "str",
+            "deliverables": "list",
+            "risks": "list",
+            "next_actions": "list",
+        },
+    },
+}
+
+
+def _normalize_provider(raw: Optional[str]) -> str:
+    provider = str(raw or "").strip().lower()
+    provider = PROVIDER_ALIASES.get(provider, provider)
+    if provider not in SUPPORTED_PROVIDERS:
+        raise ValueError("provider 仅支持 deepseek/glm/openai/ollama")
+    return provider
+
+
+def _mask_secret(secret: str) -> str:
+    value = str(secret or "").strip()
+    if not value:
+        return ""
+    if len(value) <= 8:
+        return "*" * len(value)
+    return f"{value[:4]}...{value[-4:]}"
+
+
+def _safe_json(value: Any) -> dict[str, Any]:
+    if isinstance(value, dict):
+        try:
+            # 返回深拷贝，避免原地修改导致 SQLAlchemy 不感知 JSON 字段变更
+            return json.loads(json.dumps(value))
+        except Exception:
+            return dict(value)
+    if isinstance(value, str):
+        try:
+            parsed = json.loads(value)
+            return parsed if isinstance(parsed, dict) else {}
+        except Exception:
+            return {}
+    return {}
+
+
+def _env_bool(name: str, default: bool = False) -> bool:
+    raw = os.getenv(name)
+    if raw is None:
+        return default
+    return str(raw).strip().lower() in BOOL_TRUE_VALUES
+
+
+def _parse_csv_items(raw: str) -> tuple[str, ...]:
+    rows = [item.strip().lower() for item in str(raw or "").split(",") if item.strip()]
+    return tuple(rows)
+
+
+def _cn_now_text() -> str:
+    cn_tz = timezone(timedelta(hours=8))
+    return datetime.now(cn_tz).strftime("%Y-%m-%d %H:%M:%S")
+
+
+def _runtime_log(scope: str, message: str) -> None:
+    print(f"[{_cn_now_text()}][{scope}] {message}")
+
+
+@dataclass
+class ChatCompletionResult:
+    text: str
+    raw_response: Any = None
+
+
+@dataclass
+class StructuredExtractionResult:
+    payload: Optional[dict[str, Any]]
+    source_field: str
+    provider: str
+    model: str
+    error_code: Optional[str] = None
+    raw_text: str = ""
+    validation: str = "invalid"
+    from_cache: bool = False
+
+
+class LLMService:
+    _provider_circuit_state: dict[str, dict[str, Any]] = {}
+
+    def __init__(self) -> None:
+        self.request_timeout = float(os.getenv("LLM_REQUEST_TIMEOUT", "45"))
+        self.redis_url = os.getenv("REDIS_URL")
+        self._redis: Optional[Any] = None
+        self._prompt_profile_cache: dict[int, tuple[float, dict[str, str]]] = {}
+        self._openclaw_routing_cache: tuple[float, dict[str, Any]] | None = None
+
+    async def _get_redis(self):
+        if not REDIS_AVAILABLE or not self.redis_url:
+            return None
+        if self._redis is None:
+            self._redis = redis_async.from_url(self.redis_url, decode_responses=True)
+        return self._redis
+
+    @staticmethod
+    def _normalize_prompt_for_hash(prompt: str) -> str:
+        return re.sub(r"\s+", " ", str(prompt or "")).strip()
+
+    @classmethod
+    def _structured_cache_key(cls, *, function_name: str, model: str, prompt: str) -> str:
+        normalized_prompt = cls._normalize_prompt_for_hash(prompt)
+        digest = hashlib.sha256(normalized_prompt.encode("utf-8")).hexdigest()
+        return f"pmflow:structured:{function_name}:{model}:{digest}"
+
+    async def _structured_cache_get(self, *, key: str) -> Optional[dict[str, Any]]:
+        redis = await self._get_redis()
+        if redis is None:
+            return None
+        try:
+            raw = await redis.get(key)
+        except Exception:
+            return None
+        if not raw:
+            return None
+        try:
+            parsed = json.loads(raw)
+        except Exception:
+            return None
+        return parsed if isinstance(parsed, dict) else None
+
+    async def _structured_cache_set(self, *, key: str, value: dict[str, Any], ttl: int = STRUCTURED_CACHE_TTL_SECONDS) -> None:
+        redis = await self._get_redis()
+        if redis is None:
+            return
+        try:
+            await redis.setex(key, ttl, json.dumps(value, ensure_ascii=False))
+        except Exception:
+            return
+
+    @staticmethod
+    def _provider_circuit_key(*, provider: str, base_url: str, model: str) -> str:
+        return f"{str(provider or '').strip().lower()}|{str(base_url or '').strip().rstrip('/')}|{str(model or '').strip()}"
+
+    @staticmethod
+    def _http_status_from_exception(exc: Exception) -> Optional[int]:
+        if isinstance(exc, httpx.HTTPStatusError) and exc.response is not None:
+            try:
+                return int(exc.response.status_code)
+            except Exception:
+                return None
+        return None
+
+    @classmethod
+    def _is_fail_fast_exception(cls, exc: Exception) -> bool:
+        status_code = cls._http_status_from_exception(exc)
+        if status_code in {401, 403}:
+            return True
+        text = str(exc or "").strip().lower()
+        return any(marker in text for marker in FAIL_FAST_ERROR_MARKERS)
+
+    @classmethod
+    def _is_circuit_breaker_worthy_exception(cls, exc: Exception) -> bool:
+        if cls._is_fail_fast_exception(exc):
+            return True
+        if isinstance(exc, httpx.RequestError):
+            return True
+        status_code = cls._http_status_from_exception(exc)
+        return status_code in {408, 409, 429, 500, 502, 503, 504}
+
+    @classmethod
+    def _is_provider_circuit_open(cls, circuit_key: str) -> bool:
+        state = cls._provider_circuit_state.get(str(circuit_key))
+        if not state:
+            return False
+        try:
+            open_until = float(state.get("open_until") or 0.0)
+        except Exception:
+            open_until = 0.0
+        now = time.monotonic()
+        if open_until <= now:
+            if open_until:
+                cls._provider_circuit_state.pop(str(circuit_key), None)
+            return False
+        return True
+
+    @classmethod
+    def _record_provider_success(cls, circuit_key: str) -> None:
+        cls._provider_circuit_state.pop(str(circuit_key), None)
+
+    @classmethod
+    def _record_provider_failure(cls, circuit_key: str, exc: Exception) -> None:
+        if not cls._is_circuit_breaker_worthy_exception(exc):
+            return
+        now = time.monotonic()
+        state = dict(cls._provider_circuit_state.get(str(circuit_key)) or {})
+        consecutive = int(state.get("consecutive_failures") or 0) + 1
+        open_until = float(state.get("open_until") or 0.0)
+        if consecutive >= CIRCUIT_BREAKER_FAILURE_THRESHOLD:
+            open_until = now + CIRCUIT_BREAKER_COOLDOWN_SECONDS
+        cls._provider_circuit_state[str(circuit_key)] = {
+            "consecutive_failures": consecutive,
+            "open_until": open_until,
+            "last_error": str(exc)[:240],
+            "updated_at_monotonic": now,
+        }
+
+    # ---------------------------------------------------------------------
+    # Engine config management
+    # ---------------------------------------------------------------------
+    async def list_engines(self, db: AsyncSession, tenant_id: int = 1) -> dict[str, Any]:
+        _, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        engines = bundle.get("engines", [])
+        return {
+            "active_engine_id": bundle.get("active_engine_id"),
+            "engines": [self._sanitize_engine_for_response(item) for item in engines],
+        }
+
+    async def create_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        name: str,
+        provider: str,
+        api_key: Optional[str],
+        model: Optional[str] = None,
+        base_url: Optional[str] = None,
+        set_active: bool = True,
+    ) -> dict[str, Any]:
+        provider = _normalize_provider(provider)
+        api_key = str(api_key or "").strip()
+        if provider != "ollama" and not api_key:
+            raise ValueError("api_key 不能为空")
+
+        integration, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        now = datetime.now(timezone.utc).isoformat()
+        engine_id = str(uuid4())
+        defaults = PROVIDER_DEFAULTS[provider]
+        engine = {
+            "id": engine_id,
+            "name": (name or defaults["name"]).strip()[:80] or defaults["name"],
+            "provider": provider,
+            "api_key": api_key,
+            "model": (model or defaults["model"]).strip() or defaults["model"],
+            "base_url": (base_url or defaults["base_url"]).strip().rstrip("/") or defaults["base_url"],
+            "created_at": now,
+            "updated_at": now,
+        }
+        engines = bundle.get("engines", [])
+        engines.append(engine)
+        bundle["engines"] = engines
+        if set_active or not bundle.get("active_engine_id"):
+            bundle["active_engine_id"] = engine_id
+        integration.config = json.loads(json.dumps(bundle))
+        integration.is_active = True
+        await db.commit()
+        return self._sanitize_engine_for_response(engine)
+
+    async def update_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        engine_id: str,
+        name: Optional[str] = None,
+        provider: Optional[str] = None,
+        api_key: Optional[str] = None,
+        model: Optional[str] = None,
+        base_url: Optional[str] = None,
+    ) -> dict[str, Any]:
+        integration, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        engines = bundle.get("engines", [])
+        target: Optional[dict[str, Any]] = None
+        for item in engines:
+            if str(item.get("id")) == str(engine_id):
+                target = item
+                break
+        if not target:
+            raise ValueError("engine 不存在")
+
+        if provider is not None:
+            target["provider"] = _normalize_provider(provider)
+        active_provider = _normalize_provider(target.get("provider"))
+        defaults = PROVIDER_DEFAULTS[active_provider]
+
+        if name is not None:
+            target["name"] = (name or "").strip()[:80] or target.get("name") or defaults["name"]
+        if api_key is not None:
+            api_key_val = str(api_key).strip()
+            if api_key_val:
+                target["api_key"] = api_key_val
+        if model is not None:
+            target["model"] = (model or "").strip() or defaults["model"]
+        if base_url is not None:
+            target["base_url"] = (base_url or "").strip().rstrip("/") or defaults["base_url"]
+
+        target["updated_at"] = datetime.now(timezone.utc).isoformat()
+        integration.config = json.loads(json.dumps(bundle))
+        await db.commit()
+        return self._sanitize_engine_for_response(target)
+
+    async def delete_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        engine_id: str,
+    ) -> dict[str, Any]:
+        integration, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        engines = bundle.get("engines", [])
+        kept = [item for item in engines if str(item.get("id")) != str(engine_id)]
+        if len(kept) == len(engines):
+            raise ValueError("engine 不存在")
+        bundle["engines"] = kept
+        active_engine_id = str(bundle.get("active_engine_id") or "")
+        if active_engine_id == str(engine_id):
+            bundle["active_engine_id"] = kept[0]["id"] if kept else None
+        integration.config = json.loads(json.dumps(bundle))
+        await db.commit()
+        return {
+            "deleted_engine_id": str(engine_id),
+            "active_engine_id": bundle.get("active_engine_id"),
+            "remaining_count": len(kept),
+        }
+
+    async def activate_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        engine_id: str,
+    ) -> dict[str, Any]:
+        integration, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        engines = bundle.get("engines", [])
+        exists = any(str(item.get("id")) == str(engine_id) for item in engines)
+        if not exists:
+            raise ValueError("engine 不存在")
+        bundle["active_engine_id"] = str(engine_id)
+        integration.config = json.loads(json.dumps(bundle))
+        await db.commit()
+        return {"active_engine_id": str(engine_id)}
+
+    async def test_engine_connection(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        engine_id: Optional[str] = None,
+        prompt: Optional[str] = None,
+        max_tokens: int = 64,
+    ) -> dict[str, Any]:
+        test_prompt = (prompt or "你是连通性测试助手。请只返回: OK").strip()
+        if not test_prompt:
+            raise ValueError("测试 prompt 不能为空")
+
+        engine = await self._resolve_engine_for_test(
+            db,
+            tenant_id=tenant_id,
+            engine_id=engine_id,
+        )
+        response_text = ""
+        status = "success"
+        started = time.perf_counter()
+        try:
+            response_text = await self._chat_completion(
+                provider=str(engine.get("provider") or ""),
+                prompt=test_prompt,
+                system_prompt="",
+                api_key=str(engine.get("api_key") or ""),
+                base_url=str(engine.get("base_url") or ""),
+                model=str(engine.get("model") or ""),
+                temperature=0.0,
+                max_tokens=max_tokens,
+            )
+            if not response_text.strip():
+                raise ValueError("测试响应为空")
+            latency_ms = int((time.perf_counter() - started) * 1000)
+            return {
+                "ok": True,
+                "engine_id": str(engine.get("id") or ""),
+                "name": str(engine.get("name") or ""),
+                "provider": str(engine.get("provider") or ""),
+                "model": str(engine.get("model") or ""),
+                "base_url": str(engine.get("base_url") or ""),
+                "api_key_masked": _mask_secret(str(engine.get("api_key") or "")),
+                "latency_ms": latency_ms,
+                "response_preview": response_text[:300],
+            }
+        except Exception as exc:
+            status = "failed"
+            response_text = f"[error] {exc}"
+            raise ValueError(f"引擎连通测试失败: {exc}") from exc
+        finally:
+            try:
+                await self._log_ai_call(
+                    tenant_id=tenant_id,
+                    function_name="test_engine_connection",
+                    prompt=test_prompt,
+                    response=response_text,
+                    status=status,
+                    provider=str(engine.get("provider") or ""),
+                    model=str(engine.get("model") or ""),
+                )
+            except Exception as log_exc:
+                _runtime_log("LLM", f"测试调用日志写入失败：{log_exc}")
+
+    # ---------------------------------------------------------------------
+    # LLM calls
+    # ---------------------------------------------------------------------
+    async def generate(self, prompt: str, **kwargs: Any) -> str:
+        call_kwargs = dict(kwargs)
+        db = call_kwargs.pop("db", None) or call_kwargs.pop("db_session", None)
+        if db is None:
+            async with AsyncSessionLocal() as session:
+                return await self._generate_with_db(session, prompt, **call_kwargs)
+        return await self._generate_with_db(db, prompt, **call_kwargs)
+
+    async def generate_structured_object(self, prompt: str, **kwargs: Any) -> StructuredExtractionResult:
+        call_kwargs = dict(kwargs)
+        db = call_kwargs.pop("db", None) or call_kwargs.pop("db_session", None)
+        if db is None:
+            async with AsyncSessionLocal() as session:
+                return await self._generate_structured_with_db(session, prompt, **call_kwargs)
+        return await self._generate_structured_with_db(db, prompt, **call_kwargs)
+
+    async def _generate_with_db(self, db: AsyncSession, prompt: str, **kwargs: Any) -> str:
+        tenant_id = int(kwargs.get("tenant_id") or 1)
+        function_name = str(kwargs.get("function_name") or "llm_generate")
+        temperature = float(kwargs.get("temperature") or 0.2)
+        max_tokens = int(kwargs.get("max_tokens") or 1024)
+        use_prompt_profile = bool(kwargs.get("use_prompt_profile", True))
+        response_mode = str(kwargs.get("response_mode") or "complete").strip().lower() or "complete"
+        response_text = ""
+        status = "success"
+        routing_decision = "primary_engine"
+        engine_meta: dict[str, Any] = {}
+        prompt_profile = {"role_name": "", "system_prompt": ""}
+        merged_prompt = self._merge_prompt_with_profile(prompt=prompt, prompt_profile=prompt_profile)
+        safety_attempted = False
+        try:
+            if use_prompt_profile:
+                prompt_profile = await self._resolve_active_prompt_profile(db, tenant_id=tenant_id)
+            merged_prompt = self._merge_prompt_with_profile(prompt=prompt, prompt_profile=prompt_profile)
+            engine = await self._resolve_engine_for_call(
+                db,
+                tenant_id=tenant_id,
+                function_name=function_name,
+            )
+            engine_meta = engine
+            provider = str(engine.get("provider") or "")
+            model = str(engine.get("model") or "")
+            response_text = await self._chat_completion(
+                provider=provider,
+                prompt=prompt,
+                system_prompt=prompt_profile.get("system_prompt") or "",
+                api_key=str(engine.get("api_key") or ""),
+                base_url=str(engine.get("base_url") or ""),
+                model=model,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                function_name=function_name,
+                response_mode=response_mode,
+            )
+            if not response_text.strip():
+                raise ValueError("primary_engine_empty_response")
+
+            invalid_structured_output = self._is_invalid_structured_output(
+                function_name=function_name,
+                response_text=response_text,
+            )
+            should_upgrade_to_high = False
+            if not self._prefer_fast_local_fallback(function_name):
+                should_upgrade_to_high = self._should_escalate_low_to_high(
+                    function_name=function_name,
+                    response_text=response_text,
+                ) or (invalid_structured_output and self._is_low_tier_function(function_name))
+
+            if should_upgrade_to_high:
+                high_engine = await self._resolve_high_tier_engine_for_escalation(
+                    db,
+                    tenant_id=tenant_id,
+                    primary_engine=engine,
+                )
+                if high_engine is None:
+                    raise ValueError("high_tier_engine_unavailable")
+                high_provider = str(high_engine.get("provider") or "")
+                high_model = str(high_engine.get("model") or "")
+                _runtime_log(
+                    "LLM升级",
+                    (
+                        "低阶引擎触发升级："
+                        f"function={function_name}, from={provider}/{model}, to={high_provider}/{high_model}"
+                    ),
+                )
+                try:
+                    escalated_text = await self._chat_completion(
+                        provider=high_provider,
+                        prompt=prompt,
+                        system_prompt=prompt_profile.get("system_prompt") or "",
+                        api_key=str(high_engine.get("api_key") or ""),
+                        base_url=str(high_engine.get("base_url") or ""),
+                        model=high_model,
+                        temperature=temperature,
+                        max_tokens=max_tokens,
+                        function_name=function_name,
+                        response_mode=response_mode,
+                    )
+                    if not escalated_text.strip():
+                        raise ValueError("high_tier_empty_response")
+                    if self._is_invalid_structured_output(function_name=function_name, response_text=escalated_text):
+                        raise ValueError("high_tier_invalid_structured_output")
+                    response_text = escalated_text
+                    engine_meta = high_engine
+                except Exception as escalation_exc:
+                    safety_attempted = True
+                    safety_text, safety_engine = await self._try_safety_engine(
+                        db,
+                        tenant_id=tenant_id,
+                        function_name=function_name,
+                        prompt=prompt,
+                        system_prompt=prompt_profile.get("system_prompt") or "",
+                        temperature=temperature,
+                        max_tokens=max_tokens,
+                        primary_engine=high_engine,
+                        reason=f"high_tier_retry_failed:{escalation_exc}",
+                    )
+                    if safety_engine is not None and safety_text.strip():
+                        routing_decision = "fallback_safety_engine"
+                        status = "fallback_safety_engine"
+                        engine_meta = safety_engine
+                        response_text = safety_text
+                        return response_text
+                    raise ValueError(f"high_tier_retry_failed:{escalation_exc}") from escalation_exc
+
+            if self._is_invalid_structured_output(function_name=function_name, response_text=response_text):
+                raise ValueError("invalid_structured_output")
+            return response_text
+        except Exception as exc:
+            upgraded_exc: Exception = exc
+            if self._is_fail_fast_exception(exc):
+                status = "failed_fast"
+                response_text = self._stub_response(prompt)
+                _runtime_log("LLM", f"生成快速失败，已跳过升级重试：{exc}")
+                return response_text
+            if self._is_low_tier_function(function_name):
+                if self._prefer_fast_local_fallback(function_name):
+                    raise upgraded_exc
+                high_engine = await self._resolve_high_tier_engine_for_escalation(
+                    db,
+                    tenant_id=tenant_id,
+                    primary_engine=engine_meta,
+                )
+                if high_engine is not None:
+                    high_provider = str(high_engine.get("provider") or "")
+                    high_model = str(high_engine.get("model") or "")
+                    _runtime_log(
+                        "LLM升级",
+                        (
+                            "低阶引擎异常，转高阶兜底："
+                            f"function={function_name}, from={engine_meta.get('provider')}/{engine_meta.get('model')}, "
+                            f"to={high_provider}/{high_model}, error={exc}"
+                        ),
+                    )
+                    try:
+                        escalated_text = await self._chat_completion(
+                            provider=high_provider,
+                            prompt=prompt,
+                            system_prompt=prompt_profile.get("system_prompt") or "",
+                            api_key=str(high_engine.get("api_key") or ""),
+                            base_url=str(high_engine.get("base_url") or ""),
+                            model=high_model,
+                            temperature=temperature,
+                            max_tokens=max_tokens,
+                            function_name=function_name,
+                            response_mode=response_mode,
+                        )
+                        if not escalated_text.strip():
+                            raise ValueError("high_tier_empty_response")
+                        if self._is_invalid_structured_output(function_name=function_name, response_text=escalated_text):
+                            raise ValueError("high_tier_invalid_structured_output")
+                        routing_decision = "fallback_high_tier_engine"
+                        status = "fallback_high_tier_engine"
+                        engine_meta = high_engine
+                        response_text = escalated_text
+                        return response_text
+                    except Exception as high_exc:
+                        upgraded_exc = ValueError(f"high_tier_retry_failed:{high_exc}")
+                        engine_meta = high_engine
+
+            if not safety_attempted:
+                safety_attempted = True
+                safety_text, safety_engine = await self._try_safety_engine(
+                    db,
+                    tenant_id=tenant_id,
+                    function_name=function_name,
+                    prompt=prompt,
+                    system_prompt=prompt_profile.get("system_prompt") or "",
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    primary_engine=engine_meta,
+                    reason=str(upgraded_exc),
+                )
+                if safety_engine is not None and safety_text.strip():
+                    routing_decision = "fallback_safety_engine"
+                    status = "fallback_safety_engine"
+                    engine_meta = safety_engine
+                    response_text = safety_text
+                    return response_text
+
+            retry_text = ""
+            if self._direct_emergency_enabled():
+                fallback_engines = self._build_fallback_engines(engine_meta)
+                for fallback_engine in fallback_engines:
+                    try:
+                        fb_provider = str(fallback_engine.get("provider") or "")
+                        fb_model = str(fallback_engine.get("model") or "")
+                        _runtime_log(
+                            "LLM",
+                            f"应急直连引擎重试：provider={fb_provider}, model={fb_model}, error={exc}",
+                        )
+                        retry_text = await self._chat_completion(
+                            provider=fb_provider,
+                            prompt=prompt,
+                            system_prompt=prompt_profile.get("system_prompt") or "",
+                            api_key=str(fallback_engine.get("api_key") or ""),
+                            base_url=str(fallback_engine.get("base_url") or ""),
+                            model=fb_model,
+                            temperature=temperature,
+                            max_tokens=max_tokens,
+                            function_name=function_name,
+                            response_mode=response_mode,
+                        )
+                        if retry_text.strip():
+                            routing_decision = "fallback_emergency_direct_engine"
+                            status = "fallback_emergency_direct_engine"
+                            engine_meta = fallback_engine
+                            response_text = retry_text
+                            return response_text
+                    except Exception as retry_exc:
+                        _runtime_log(
+                            "LLM",
+                            (
+                                "应急直连重试失败："
+                                f"provider={fallback_engine.get('provider')}, "
+                                f"model={fallback_engine.get('model')}, error={retry_exc}"
+                            ),
+                        )
+
+            status = "failed"
+            response_text = self._stub_response(prompt)
+            _runtime_log("LLM", f"生成失败，已返回兜底 stub：{exc}")
+            return response_text
+        finally:
+            try:
+                merged_prompt = self._merge_prompt_with_profile(prompt=prompt, prompt_profile=prompt_profile)
+                await self._log_ai_call(
+                    tenant_id=tenant_id,
+                    function_name=function_name,
+                    prompt=merged_prompt,
+                    response=response_text,
+                    status=status,
+                    provider=str(engine_meta.get("provider") or ""),
+                    model=str(engine_meta.get("model") or ""),
+                )
+                if routing_decision != "primary_engine":
+                    _runtime_log(
+                        "LLM路由",
+                        (
+                            f"routing_decision={routing_decision}, "
+                            f"function={function_name}, provider={engine_meta.get('provider')}, model={engine_meta.get('model')}"
+                        ),
+                    )
+            except Exception as log_exc:
+                _runtime_log("LLM", f"调用日志写入失败：{log_exc}")
+
+    async def _generate_structured_with_db(
+        self,
+        db: AsyncSession,
+        prompt: str,
+        **kwargs: Any,
+    ) -> StructuredExtractionResult:
+        tenant_id = int(kwargs.get("tenant_id") or 1)
+        function_name = str(kwargs.get("function_name") or "llm_generate").strip().lower() or "llm_generate"
+        temperature = float(kwargs.get("temperature") or 0.0)
+        max_tokens = int(kwargs.get("max_tokens") or 1024)
+        use_prompt_profile = bool(kwargs.get("use_prompt_profile", True))
+        response_mode = str(kwargs.get("response_mode") or "complete").strip().lower() or "complete"
+        prompt_profile = {"role_name": "", "system_prompt": ""}
+        merged_prompt = prompt
+        provider = ""
+        model = ""
+        status = "success"
+        response_text = ""
+        try:
+            if use_prompt_profile:
+                prompt_profile = await self._resolve_active_prompt_profile(db, tenant_id=tenant_id)
+            merged_prompt = self._merge_prompt_with_profile(prompt=prompt, prompt_profile=prompt_profile)
+            engine = await self._resolve_engine_for_call(
+                db,
+                tenant_id=tenant_id,
+                function_name=function_name,
+            )
+            provider = str(engine.get("provider") or "")
+            model = str(engine.get("model") or "")
+            cache_key = ""
+            if self._structured_cache_enabled(function_name):
+                cache_key = self._structured_cache_key(
+                    function_name=function_name,
+                    model=model or "unknown",
+                    prompt=merged_prompt,
+                )
+                cached = await self._structured_cache_get(key=cache_key)
+                if cached is not None:
+                    status = "success_cache_hit"
+                    result = StructuredExtractionResult(
+                        payload=cached,
+                        source_field="cache",
+                        provider=provider,
+                        model=model,
+                        raw_text=json.dumps(cached, ensure_ascii=False),
+                        validation="ok",
+                        from_cache=True,
+                    )
+                    _runtime_log(
+                        "LLM结构化",
+                        f"function={function_name}, structured_source=cache, validation=ok, fallback=false",
+                    )
+                    return result
+
+            completion = await self._chat_completion_raw(
+                provider=provider,
+                prompt=prompt,
+                system_prompt=prompt_profile.get("system_prompt") or "",
+                api_key=str(engine.get("api_key") or ""),
+                base_url=str(engine.get("base_url") or ""),
+                model=model,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                function_name=function_name,
+                response_mode=response_mode,
+            )
+            response_text = completion.text
+            extraction = self._extract_structured_object_from_response(
+                function_name=function_name,
+                response=completion.raw_response,
+                fallback_text=completion.text,
+                provider=provider,
+                model=model,
+            )
+            is_fallback = extraction.error_code is not None
+            _runtime_log(
+                "LLM结构化",
+                (
+                    f"function={function_name}, structured_source={extraction.source_field or 'none'}, "
+                    f"validation={extraction.validation}, fallback={'true' if is_fallback else 'false'}"
+                ),
+            )
+            if extraction.payload is not None and cache_key:
+                await self._structured_cache_set(key=cache_key, value=extraction.payload)
+            status = extraction.error_code or "success"
+            return extraction
+        except Exception as exc:
+            error_code = self._classify_structured_exception(exc)
+            status = error_code
+            result = StructuredExtractionResult(
+                payload=None,
+                source_field="",
+                provider=provider,
+                model=model,
+                error_code=error_code,
+                raw_text=str(exc)[:1000],
+                validation="invalid",
+            )
+            _runtime_log(
+                "LLM结构化",
+                f"function={function_name}, structured_source=none, validation=invalid, fallback=true",
+            )
+            return result
+        finally:
+            try:
+                log_payload = response_text or status
+                await self._log_ai_call(
+                    tenant_id=tenant_id,
+                    function_name=function_name,
+                    prompt=merged_prompt,
+                    response=log_payload,
+                    status=status,
+                    provider=provider,
+                    model=model,
+                )
+            except Exception as log_exc:
+                _runtime_log("LLM", f"结构化调用日志写入失败：{log_exc}")
+
+    async def _chat_completion(
+        self,
+        *,
+        provider: str,
+        prompt: str,
+        system_prompt: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        temperature: float,
+        max_tokens: int,
+        function_name: str = "",
+        response_mode: str = "complete",
+    ) -> str:
+        result = await self._chat_completion_raw(
+            provider=provider,
+            prompt=prompt,
+            system_prompt=system_prompt,
+            api_key=api_key,
+            base_url=base_url,
+            model=model,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            function_name=function_name,
+            response_mode=response_mode,
+        )
+        return result.text
+
+    async def _chat_completion_raw(
+        self,
+        *,
+        provider: str,
+        prompt: str,
+        system_prompt: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        temperature: float,
+        max_tokens: int,
+        function_name: str = "",
+        response_mode: str = "complete",
+    ) -> ChatCompletionResult:
+        provider_name = _normalize_provider(provider)
+        structured_output = self._structured_output_required(function_name)
+        request_timeout = self._request_timeout_for_function(function_name, response_mode=response_mode)
+        is_openclaw_gateway_request = self._is_openclaw_gateway_request(
+            provider_name=provider_name,
+            base_url=base_url,
+        )
+        openclaw_agents_timeout_seconds = self._openclaw_agents_timeout_seconds()
+        if is_openclaw_gateway_request:
+            # 网关内部 agent 超时需小于客户端超时，避免本地请求先被中止。
+            request_timeout = max(float(request_timeout), float(openclaw_agents_timeout_seconds) + 5.0)
+        use_provider_circuit_breaker = self._use_provider_circuit_breaker(function_name)
+        circuit_key = self._provider_circuit_key(
+            provider=provider_name,
+            base_url=base_url,
+            model=model,
+        )
+        if use_provider_circuit_breaker and self._is_provider_circuit_open(circuit_key):
+            raise ValueError(f"provider_circuit_open:{provider_name}:{model}")
+        key = api_key.strip()
+        messages: list[dict[str, str]] = []
+        cleaned_system_prompt = str(system_prompt or "").strip()
+        if cleaned_system_prompt:
+            messages.append({"role": "system", "content": cleaned_system_prompt})
+        messages.append({"role": "user", "content": prompt})
+
+        try:
+            if provider_name == "ollama":
+                endpoint = f"{base_url.rstrip('/')}/api/chat"
+                payload = {
+                    "model": model,
+                    "messages": messages,
+                    "stream": False,
+                    "options": {
+                        "temperature": temperature,
+                        "num_predict": max_tokens,
+                    },
+                }
+                async with httpx.AsyncClient(timeout=request_timeout, trust_env=False) as client:
+                    resp = await client.post(endpoint, json=payload)
+                    resp.raise_for_status()
+                    data = resp.json()
+                message = data.get("message") if isinstance(data, dict) else {}
+                content = message.get("content") if isinstance(message, dict) else None
+                text = self._sanitize_visible_answer(self._extract_text_from_value(content))
+                if not text:
+                    text = self._sanitize_visible_answer(self._extract_text_from_value(data.get("response")))
+                if not text and not isinstance(data, dict):
+                    raise ValueError("Ollama 返回空内容")
+                if use_provider_circuit_breaker:
+                    self._record_provider_success(circuit_key)
+                return ChatCompletionResult(text=text, raw_response=data)
+
+            if not key:
+                raise ValueError("未配置 LLM API Key")
+            endpoint = f"{base_url.rstrip('/')}/chat/completions"
+            payload = {
+                "model": model,
+                "messages": messages,
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+            }
+            if is_openclaw_gateway_request:
+                payload["agents"] = {
+                    "defaults": {
+                        "timeoutSeconds": int(openclaw_agents_timeout_seconds),
+                    }
+                }
+            headers = {
+                "Authorization": f"Bearer {key}",
+                "Content-Type": "application/json",
+            }
+            async with httpx.AsyncClient(timeout=request_timeout, trust_env=False) as client:
+                resp = await client.post(endpoint, headers=headers, json=payload)
+                resp.raise_for_status()
+                data = resp.json()
+            choices = data.get("choices") or []
+            if not choices:
+                raise ValueError("LLM 响应无 choices")
+            message = choices[0].get("message") if isinstance(choices[0], dict) else {}
+            text = self._extract_text_from_chat_response(
+                message=message,
+                choice=choices[0] if isinstance(choices[0], dict) else {},
+                response=data,
+                structured_output=structured_output,
+            )
+            if not text and not structured_output:
+                msg_keys = list(message.keys()) if isinstance(message, dict) else []
+                choice_keys = list(choices[0].keys()) if isinstance(choices[0], dict) else []
+                choice_preview = ""
+                try:
+                    choice_preview = json.dumps(choices[0], ensure_ascii=False)[:300]
+                except Exception:
+                    choice_preview = str(choices[0])[:300]
+                raise ValueError(
+                    "LLM 返回空内容: "
+                    f"message_keys={msg_keys}, choice_keys={choice_keys}, choice_preview={choice_preview}"
+                )
+            if use_provider_circuit_breaker:
+                self._record_provider_success(circuit_key)
+            return ChatCompletionResult(text=text, raw_response=data)
+        except Exception as exc:
+            if use_provider_circuit_breaker:
+                self._record_provider_failure(circuit_key, exc)
+            raise
+
+    async def generate_task_breakdown_async(self, description: str, **kwargs: Any) -> list[dict]:
+        if not description.strip():
+            return []
+        prompt = (
+            "请将以下需求拆解为 5-8 条执行任务，返回 JSON 数组，"
+            "每项格式: {\"task_name\":\"\",\"type\":\"backend|frontend|database|ai|devops\","
+            "\"estimated_hours\":4,\"priority\":\"high|medium|low\"}。\n"
+            f"需求：{description[:1200]}"
+        )
+        text = await self.generate(prompt, function_name="generate_task_breakdown", **kwargs)
+        rows = self._parse_json_array(text)
+        if rows:
+            return rows
+        return [
+            {
+                "task_name": "需求分析",
+                "type": "ai",
+                "priority": "medium",
+                "estimated_hours": 4,
+                "description": description[:200],
+            }
+        ]
+
+    async def get_embedding_async(self, text: str, **kwargs: Any) -> list[float]:
+        db = kwargs.get("db") or kwargs.get("db_session")
+        tenant_id = int(kwargs.get("tenant_id") or 1)
+        if db is None:
+            async with AsyncSessionLocal() as session:
+                return await self.get_embedding_async(text, db=session, tenant_id=tenant_id)
+        try:
+            engine = await self._resolve_active_engine(db, tenant_id=tenant_id)
+            provider = str(engine.get("provider") or "")
+            if provider == "openai":
+                endpoint = f"{str(engine.get('base_url') or '').rstrip('/')}/embeddings"
+                headers = {
+                    "Authorization": f"Bearer {str(engine.get('api_key') or '').strip()}",
+                    "Content-Type": "application/json",
+                }
+                payload = {
+                    "model": str(engine.get("model") or "text-embedding-3-small"),
+                    "input": text,
+                }
+                async with httpx.AsyncClient(timeout=self.request_timeout, trust_env=False) as client:
+                    resp = await client.post(endpoint, headers=headers, json=payload)
+                    resp.raise_for_status()
+                    data = resp.json()
+                return list(data.get("data", [{}])[0].get("embedding") or [])
+        except Exception:
+            pass
+        # DeepSeek/GLM 及失败场景：回退伪向量，保持功能可用。
+        seed = hashlib.md5(text.encode("utf-8")).digest()
+        return [((seed[i % 16] / 255.0) * 2 - 1) for i in range(64)]
+
+    async def extract_context_insights_async(self, text: str, **kwargs: Any) -> dict:
+        prompt = (
+            "请提炼以下技术文档上下文的关键结论，返回 JSON："
+            "{\"summary\":\"\",\"keywords\":[\"\",\"\"],\"risks\":[\"\",\"\"]}\n"
+            f"上下文：{text[:2000]}"
+        )
+        raw = await self.generate(prompt, function_name="extract_context_insights", **kwargs)
+        parsed = self._parse_json_object(raw)
+        if parsed:
+            return parsed
+        return {"summary": text[:120], "keywords": []}
+
+    async def find_relevant_context(self, **_: Any) -> list[dict]:
+        return []
+
+    async def evaluate_with_voting_async(self, signal_content: str, impact_score: float, **_: Any) -> dict:
+        is_strategic = impact_score >= 0.8
+        return {
+            "is_strategic": is_strategic,
+            "confidence": min(1.0, max(0.2, impact_score)),
+            "decision_proposal": {
+                "summary": signal_content[:120],
+                "action": "follow_up" if is_strategic else "observe",
+            },
+            "tasks": [],
+            "_voting": {
+                "enabled": False,
+                "models_used": [],
+                "strategic_votes": 0,
+                "total_votes": 0,
+                "unanimous": True,
+            },
+        }
+
+    async def audit_code_change_async(
+        self,
+        task_name: str,
+        commit_message: str,
+        code_diff: str,
+        **_: Any,
+    ) -> dict:
+        passed = bool(task_name and commit_message and code_diff is not None)
+        return {
+            "passed": passed,
+            "reason": "basic_stub_check_passed" if passed else "invalid_input",
+        }
+
+    # ---------------------------------------------------------------------
+    # internals
+    # ---------------------------------------------------------------------
+    def _stub_response(self, prompt: str) -> str:
+        _ = prompt  # 保留签名，避免调用方改动
+        return (
+            "我暂时无法生成可靠回答（模型服务异常）。"
+            "为避免误导，本次不直接给结论；请稍后重试，"
+            "或补充更具体的目标对象后继续提问。"
+        )
+
+    def _sanitize_engine_for_response(self, item: dict[str, Any]) -> dict[str, Any]:
+        return {
+            "id": str(item.get("id") or ""),
+            "name": str(item.get("name") or ""),
+            "provider": str(item.get("provider") or ""),
+            "model": str(item.get("model") or ""),
+            "base_url": str(item.get("base_url") or ""),
+            "api_key_masked": _mask_secret(str(item.get("api_key") or "")),
+            "updated_at": item.get("updated_at"),
+            "created_at": item.get("created_at"),
+        }
+
+    async def _get_or_init_bundle(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+    ) -> tuple[Integration, dict[str, Any]]:
+        result = await db.execute(
+            select(Integration).where(
+                and_(
+                    Integration.tenant_id == tenant_id,
+                    Integration.type == "llm_engine",
+                )
+            )
+        )
+        integration = result.scalar_one_or_none()
+        if not integration:
+            integration = Integration(
+                type="llm_engine",
+                config={"active_engine_id": None, "engines": []},
+                is_active=True,
+                tenant_id=tenant_id,
+            )
+            db.add(integration)
+            await db.flush()
+        bundle = _safe_json(integration.config)
+        engines = bundle.get("engines")
+        if not isinstance(engines, list):
+            engines = []
+        bundle["engines"] = engines
+        if "active_engine_id" not in bundle:
+            bundle["active_engine_id"] = None
+        return integration, bundle
+
+    async def _resolve_active_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+    ) -> dict[str, Any]:
+        _, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        return self._resolve_active_engine_from_bundle(bundle)
+
+    async def _resolve_engine_for_call(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        function_name: str,
+    ) -> dict[str, Any]:
+        _, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        active_engine = self._resolve_active_engine_from_bundle(bundle)
+        active_engine = self._ensure_openclaw_control_plane_engine(active_engine)
+        engines: list[dict[str, Any]] = [e for e in bundle.get("engines", []) if isinstance(e, dict)]
+
+        openclaw_routed = self._resolve_openclaw_engine(
+            function_name=function_name,
+            engines=engines,
+            active_engine=active_engine,
+        )
+        if openclaw_routed is not None:
+            openclaw_routed = self._ensure_openclaw_control_plane_engine(openclaw_routed)
+            provider = str(openclaw_routed.get("provider") or "")
+            api_key = str(openclaw_routed.get("api_key") or "").strip()
+            if provider != "ollama" and not api_key:
+                _runtime_log(
+                    "OpenClaw路由",
+                    (
+                        "命中 OpenClaw 路由但目标引擎缺少 API Key，已回退 active："
+                        f"function={function_name}, provider={provider}, model={openclaw_routed.get('model')}"
+                    ),
+                )
+            else:
+                return openclaw_routed
+
+        if not self._mixed_routing_enabled():
+            return active_engine
+
+        tier = self._resolve_function_tier(function_name)
+        if tier not in {"high", "low"}:
+            return active_engine
+
+        routed = self._resolve_tier_engine(
+            tier=tier,
+            engines=engines,
+            active_engine=active_engine,
+        )
+        if not routed:
+            return active_engine
+
+        provider = str(routed.get("provider") or "")
+        api_key = str(routed.get("api_key") or "").strip()
+        if provider != "ollama" and not api_key:
+            _runtime_log(
+                "LLM路由",
+                (
+                    "命中分层路由但目标引擎缺少 API Key，回退可用引擎："
+                    f"function={function_name}, tier={tier}, provider={provider}, model={routed.get('model')}"
+                ),
+            )
+            fallback_engine = self._pick_tier_fallback_engine(
+                tier=tier,
+                engines=engines,
+                active_engine=active_engine,
+            )
+            if fallback_engine:
+                return fallback_engine
+            return active_engine
+        _runtime_log(
+            "LLM路由",
+            (
+                "函数路由生效："
+                f"function={function_name}, tier={tier}, provider={provider}, model={routed.get('model')}"
+            ),
+        )
+        return self._ensure_openclaw_control_plane_engine(routed)
+
+    def _resolve_active_engine_from_bundle(self, bundle: dict[str, Any]) -> dict[str, Any]:
+        engines: list[dict[str, Any]] = [e for e in bundle.get("engines", []) if isinstance(e, dict)]
+        active_engine_id = str(bundle.get("active_engine_id") or "")
+        if active_engine_id:
+            for engine in engines:
+                if str(engine.get("id") or "") == active_engine_id:
+                    return self._normalize_engine_runtime(engine)
+        if engines:
+            return self._normalize_engine_runtime(engines[0])
+        return self._env_engine_runtime()
+
+    def _mixed_routing_enabled(self) -> bool:
+        return _env_bool("LLM_MIXED_ROUTING_ENABLED", False)
+
+    def _openclaw_routing_enabled(self) -> bool:
+        return _env_bool("OPENCLAW_ROUTING_ENABLED", True)
+
+    def _openclaw_control_plane_only(self) -> bool:
+        return self._openclaw_routing_enabled() and not self._direct_emergency_enabled()
+
+    def _ensure_openclaw_control_plane_engine(self, engine: dict[str, Any]) -> dict[str, Any]:
+        runtime = dict(engine or {})
+        if not self._openclaw_control_plane_only():
+            return runtime
+
+        openai_default = self._env_provider_runtime("openai")
+        runtime["provider"] = "openai"
+        runtime["base_url"] = (
+            str(os.getenv("OPENAI_BASE_URL", openai_default.get("base_url") or OPENCLAW_PRIMARY_GATEWAY_URL) or "")
+            .strip()
+            .rstrip("/")
+            or OPENCLAW_PRIMARY_GATEWAY_URL
+        )
+        runtime["api_key"] = (
+            str(os.getenv("OPENAI_API_KEY", "") or "").strip()
+            or str(runtime.get("api_key") or "").strip()
+        )
+        runtime["model"] = (
+            str(runtime.get("model") or "").strip()
+            or str(openai_default.get("model") or OPENCLAW_DEFAULT_LOW_MODEL).strip()
+        )
+        runtime["name"] = str(runtime.get("name") or "OpenClaw Gateway").strip() or "OpenClaw Gateway"
+        runtime["id"] = str(runtime.get("id") or "openclaw-gateway")
+        return runtime
+
+    def _openclaw_rules_path(self) -> str:
+        custom = str(os.getenv("OPENCLAW_ROUTING_RULES_PATH", "") or "").strip()
+        if custom:
+            return custom
+        if OPENCLAW_BACKEND_RULES_PATH.exists():
+            return str(OPENCLAW_BACKEND_RULES_PATH)
+        if OPENCLAW_PROJECT_RULES_PATH.exists():
+            return str(OPENCLAW_PROJECT_RULES_PATH)
+        return str(OPENCLAW_BACKEND_RULES_PATH)
+
+    def _load_openclaw_routing_bundle(self) -> dict[str, Any]:
+        now = time.time()
+        if self._openclaw_routing_cache and self._openclaw_routing_cache[0] > now:
+            return dict(self._openclaw_routing_cache[1])
+
+        path = Path(self._openclaw_rules_path())
+        bundle: dict[str, Any] = {}
+        if path.exists():
+            try:
+                parsed = json.loads(path.read_text(encoding="utf-8"))
+                if isinstance(parsed, dict):
+                    bundle = parsed
+            except Exception as exc:
+                _runtime_log("OpenClaw路由", f"规则文件解析失败：path={path} error={exc}")
+                bundle = {}
+        self._openclaw_routing_cache = (now + OPENCLAW_ROUTING_CACHE_TTL_SECONDS, dict(bundle))
+        return dict(bundle)
+
+    @staticmethod
+    def _rule_matches_function(rule: dict[str, Any], function_name: str) -> bool:
+        if bool(rule.get("match_all", False)):
+            return True
+        fn = str(function_name or "").strip().lower()
+        if not fn:
+            return False
+        exact = str(rule.get("function_exact") or "").strip().lower()
+        if exact and fn == exact:
+            return True
+        prefix = str(rule.get("function_prefix") or "").strip().lower()
+        if prefix == "*" or prefix == "all":
+            return True
+        if prefix and fn.startswith(prefix):
+            return True
+        match_any = rule.get("match_any")
+        if isinstance(match_any, list):
+            for hint in match_any:
+                token = str(hint or "").strip().lower()
+                if token and token in fn:
+                    return True
+        return False
+
+    def _match_openclaw_rule(
+        self,
+        function_name: str,
+        *,
+        tier_filter: Optional[set[str]] = None,
+    ) -> Optional[dict[str, Any]]:
+        if not self._openclaw_routing_enabled():
+            return None
+        bundle = self._load_openclaw_routing_bundle()
+        rules = bundle.get("rules") if isinstance(bundle.get("rules"), list) else []
+        if not rules:
+            return None
+
+        indexed: list[tuple[int, int, dict[str, Any]]] = []
+        for idx, row in enumerate(rules):
+            if not isinstance(row, dict):
+                continue
+            if not bool(row.get("enabled", True)):
+                continue
+            if tier_filter:
+                tier_name = str(row.get("tier") or "").strip().lower()
+                if tier_name not in tier_filter:
+                    continue
+            if not self._rule_matches_function(row, function_name):
+                continue
+            priority = int(row.get("priority") or 0)
+            indexed.append((priority, -idx, row))
+
+        if not indexed:
+            return None
+        indexed.sort(reverse=True)
+        return dict(indexed[0][2])
+
+    def _resolve_openclaw_engine(
+        self,
+        *,
+        function_name: str,
+        engines: list[dict[str, Any]],
+        active_engine: dict[str, Any],
+    ) -> Optional[dict[str, Any]]:
+        rule = self._match_openclaw_rule(function_name, tier_filter={"low", "high"})
+        if not rule:
+            return None
+
+        routed = self._resolve_engine_from_rule(
+            rule=rule,
+            engines=engines,
+            active_engine=active_engine,
+        )
+        if routed is None:
+            return None
+
+        _runtime_log(
+            "OpenClaw路由",
+            (
+                "规则命中："
+                f"function={function_name}, provider={routed.get('provider')}, "
+                f"model={routed.get('model')}, rule={rule.get('name') or 'unnamed'}"
+            ),
+        )
+        return routed
+
+    def _resolve_engine_from_rule(
+        self,
+        *,
+        rule: dict[str, Any],
+        engines: list[dict[str, Any]],
+        active_engine: dict[str, Any],
+    ) -> Optional[dict[str, Any]]:
+        tier = str(rule.get("tier") or "").strip().lower()
+        provider_raw = str(rule.get("provider") or "").strip().lower()
+        engine_id = str(rule.get("engine_id") or "").strip()
+        model_override = str(rule.get("model") or "").strip()
+        base_url_override = str(rule.get("base_url") or "").strip().rstrip("/")
+
+        routed: Optional[dict[str, Any]] = None
+        if engine_id:
+            for item in engines:
+                if str(item.get("id") or "").strip() == engine_id:
+                    routed = self._normalize_engine_runtime(item)
+                    break
+
+        if routed is None and provider_raw:
+            try:
+                provider = _normalize_provider(provider_raw)
+            except Exception:
+                provider = ""
+            if provider:
+                for item in engines:
+                    if str(item.get("provider") or "").strip().lower() == provider:
+                        routed = self._normalize_engine_runtime(item)
+                        break
+                if routed is None:
+                    try:
+                        routed = self._env_provider_runtime(provider)
+                    except Exception:
+                        routed = None
+
+        if routed is None and tier in {"high", "low", "safety"}:
+            routed = self._resolve_tier_engine(
+                tier=tier,
+                engines=engines,
+                active_engine=active_engine,
+            )
+
+        if routed is None:
+            return None
+
+        if model_override:
+            routed["model"] = model_override
+        if base_url_override:
+            routed["base_url"] = base_url_override
+        return routed
+
+    def _resolve_function_tier(self, function_name: str) -> str:
+        fn = str(function_name or "").strip().lower()
+        if not fn:
+            return str(os.getenv("LLM_ROUTING_DEFAULT_TIER", "active") or "active").strip().lower()
+
+        high_hints = _parse_csv_items(os.getenv("LLM_ROUTING_HIGH_FUNCTIONS", ""))
+        low_hints = _parse_csv_items(os.getenv("LLM_ROUTING_LOW_FUNCTIONS", ""))
+        high_pool = high_hints or DEFAULT_HIGH_TIER_HINTS
+        low_pool = low_hints or DEFAULT_LOW_TIER_HINTS
+
+        if any(hint and hint in fn for hint in high_pool):
+            return "high"
+        if any(hint and hint in fn for hint in low_pool):
+            return "low"
+        default_tier = str(os.getenv("LLM_ROUTING_DEFAULT_TIER", "active") or "active").strip().lower()
+        if default_tier != "active":
+            return default_tier
+        # mixed routing 开启时，未知函数默认走 high，避免被 active 引擎“全吞”。
+        return "high" if self._mixed_routing_enabled() else "active"
+
+    def _low_to_high_escalation_enabled(self) -> bool:
+        return _env_bool("LLM_LOW_TO_HIGH_ESCALATION_ENABLED", True)
+
+    def _is_low_tier_function(self, function_name: str) -> bool:
+        rule = self._match_openclaw_rule(function_name)
+        if rule is not None:
+            tier = str(rule.get("tier") or "").strip().lower()
+            if tier:
+                return tier == "low"
+        return self._resolve_function_tier(function_name) == "low"
+
+    @staticmethod
+    def _looks_like_unanswerable_response(text: str) -> bool:
+        cleaned = str(text or "").strip()
+        lowered = cleaned.lower()
+        if not lowered:
+            return True
+        if len(cleaned) <= 8:
+            return True
+        return any(marker in lowered for marker in LOW_TO_HIGH_ESCALATION_FALLBACK_MARKERS)
+
+    def _should_escalate_low_to_high(self, *, function_name: str, response_text: str) -> bool:
+        if not self._low_to_high_escalation_enabled():
+            return False
+        if not self._is_low_tier_function(function_name):
+            return False
+        return self._looks_like_unanswerable_response(response_text)
+
+    @staticmethod
+    def _merge_prompt_with_profile(*, prompt: str, prompt_profile: dict[str, str]) -> str:
+        system_prompt = str(prompt_profile.get("system_prompt") or "").strip()
+        if not system_prompt:
+            return prompt
+        role_name = str(prompt_profile.get("role_name") or "unknown").strip() or "unknown"
+        return f"[system_role:{role_name}]\n{system_prompt}\n\n{prompt}"
+
+    def _direct_emergency_enabled(self) -> bool:
+        return _env_bool("LLM_DIRECT_EMERGENCY_ENABLED", False)
+
+    @staticmethod
+    def _prefer_fast_local_fallback(function_name: str) -> bool:
+        fn = str(function_name or "").strip().lower()
+        if not fn:
+            return False
+        return any(hint and hint in fn for hint in FAST_LOCAL_FALLBACK_FUNCTION_HINTS)
+
+    @staticmethod
+    def _is_fast_local_planner_function(function_name: str) -> bool:
+        fn = str(function_name or "").strip().lower()
+        return any(hint and hint in fn for hint in FAST_LOCAL_PLANNER_FUNCTION_HINTS)
+
+    @staticmethod
+    def _is_fast_local_worker_function(function_name: str) -> bool:
+        fn = str(function_name or "").strip().lower()
+        return any(hint and hint in fn for hint in FAST_LOCAL_WORKER_FUNCTION_HINTS)
+
+    @staticmethod
+    def _structured_cache_enabled(function_name: str) -> bool:
+        fn = str(function_name or "").strip().lower()
+        return fn in {
+            "signal_source_onboarding_plan",
+            "openclaw_command_dispatch",
+        }
+
+    def _request_timeout_for_function(self, function_name: str, *, response_mode: str = "complete") -> float:
+        if not self._prefer_fast_local_fallback(function_name):
+            return self.request_timeout
+        env_key = "LLM_FAST_LOCAL_FALLBACK_TIMEOUT_COMPLETE"
+        default_timeout = FAST_LOCAL_FALLBACK_COMPLETE_TIMEOUT_SECONDS
+        if self._is_fast_local_planner_function(function_name) and str(response_mode or "").strip().lower() == "fast":
+            env_key = "LLM_FAST_LOCAL_FALLBACK_TIMEOUT_FAST"
+            default_timeout = FAST_LOCAL_FALLBACK_FAST_TIMEOUT_SECONDS
+        try:
+            configured = float(os.getenv(env_key, str(default_timeout)))
+        except Exception:
+            configured = default_timeout
+        if configured <= 0:
+            return self.request_timeout
+        return min(self.request_timeout, configured)
+
+    @staticmethod
+    def _is_openclaw_gateway_request(*, provider_name: str, base_url: str) -> bool:
+        if str(provider_name or "").strip().lower() != "openai":
+            return False
+        normalized_base_url = str(base_url or "").strip().lower()
+        if not normalized_base_url:
+            return False
+        return "openclaw" in normalized_base_url
+
+    @staticmethod
+    def _openclaw_agents_timeout_seconds() -> int:
+        raw = os.getenv(
+            "OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS",
+            str(OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS),
+        )
+        try:
+            configured = int(str(raw).strip())
+        except Exception:
+            configured = OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS
+        if configured <= 0:
+            configured = OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS
+        return min(configured, OPENCLAW_AGENTS_TIMEOUT_SECONDS_MAX)
+
+    def _use_provider_circuit_breaker(self, function_name: str) -> bool:
+        return not self._prefer_fast_local_fallback(function_name)
+
+    def _structured_output_required(self, function_name: str) -> bool:
+        fn = str(function_name or "").strip().lower()
+        if not fn:
+            return False
+        custom_hints = _parse_csv_items(os.getenv("LLM_STRUCTURED_OUTPUT_FUNCTIONS", ""))
+        hint_pool = custom_hints or STRUCTURED_OUTPUT_FUNCTION_HINTS
+        return any(hint and hint in fn for hint in hint_pool)
+
+    @staticmethod
+    def _extract_json_candidate(text: str) -> str:
+        raw = str(text or "").strip()
+        if not raw:
+            return ""
+        if raw.startswith("```"):
+            chunks = raw.split("```")
+            if len(chunks) >= 2:
+                raw = chunks[1].strip()
+            if raw.startswith("json"):
+                raw = raw[4:].strip()
+        if raw.startswith("{") or raw.startswith("["):
+            return raw
+
+        object_candidate = LLMService._extract_balanced_json_segment(raw, "{", "}")
+        array_candidate = LLMService._extract_balanced_json_segment(raw, "[", "]")
+        if object_candidate and array_candidate:
+            return object_candidate if raw.find("{") <= raw.find("[") else array_candidate
+        if object_candidate:
+            return object_candidate
+        if array_candidate:
+            return array_candidate
+        return raw
+
+    @staticmethod
+    def _looks_like_json_container(text: str) -> bool:
+        raw = str(text or "").strip()
+        if not raw:
+            return False
+        if raw.startswith("{") or raw.startswith("["):
+            return True
+        return "{" in raw or "[" in raw
+
+    @staticmethod
+    def _extract_balanced_json_segment(text: str, open_char: str, close_char: str) -> str:
+        raw = str(text or "")
+        start = raw.find(open_char)
+        if start < 0:
+            return ""
+        depth = 0
+        in_str = False
+        escaped = False
+        for idx in range(start, len(raw)):
+            ch = raw[idx]
+            if in_str:
+                if escaped:
+                    escaped = False
+                elif ch == "\\":
+                    escaped = True
+                elif ch == '"':
+                    in_str = False
+                continue
+            if ch == '"':
+                in_str = True
+                continue
+            if ch == open_char:
+                depth += 1
+            elif ch == close_char:
+                depth -= 1
+                if depth == 0:
+                    return raw[start : idx + 1].strip()
+        return ""
+
+    def _is_invalid_structured_output(self, *, function_name: str, response_text: str) -> bool:
+        if not self._structured_output_required(function_name):
+            return False
+        candidate = self._extract_json_candidate(response_text)
+        if not candidate:
+            return True
+        try:
+            parsed = json.loads(candidate)
+        except Exception:
+            return True
+        parsed = self._unwrap_structured_payload_once(parsed)
+        validation_error = self._validate_structured_payload(function_name=function_name, payload=parsed)
+        if validation_error is not None:
+            return True
+        return not isinstance(parsed, (dict, list))
+
+    @staticmethod
+    def _strip_code_fence(text: str) -> str:
+        raw = str(text or "").strip()
+        if not raw.startswith("```"):
+            return raw
+        chunks = raw.split("```")
+        if len(chunks) >= 2:
+            raw = chunks[1].strip()
+        if raw.lower().startswith("json"):
+            raw = raw[4:].strip()
+        return raw.strip()
+
+    def _extract_text_from_chat_response(
+        self,
+        *,
+        message: Any,
+        choice: Any,
+        response: Any,
+        structured_output: bool,
+    ) -> str:
+        if structured_output:
+            for _, raw_value in self._iter_structured_candidates(message=message, choice=choice, response=response):
+                text = self._sanitize_visible_answer(self._extract_text_from_value(raw_value))
+                normalized = self._strip_code_fence(text)
+                if not self._looks_like_json_container(normalized):
+                    continue
+                candidate = self._extract_json_candidate(normalized)
+                if candidate:
+                    return candidate
+            return ""
+        if isinstance(message, dict):
+            text = self._sanitize_visible_answer(self._extract_text_from_value(message.get("content")))
+            if text:
+                return text
+            tool_calls = message.get("tool_calls")
+            if isinstance(tool_calls, list) and tool_calls:
+                first = tool_calls[0] if isinstance(tool_calls[0], dict) else {}
+                fn = first.get("function") if isinstance(first, dict) else {}
+                if isinstance(fn, dict):
+                    text = self._sanitize_visible_answer(self._extract_text_from_value(fn.get("arguments")))
+                    if text:
+                        return text
+        if isinstance(choice, dict):
+            text = self._sanitize_visible_answer(self._extract_text_from_value(choice.get("text")))
+            if text:
+                return text
+        if isinstance(response, dict):
+            text = self._sanitize_visible_answer(self._extract_text_from_value(response.get("output_text")))
+            if text:
+                return text
+        return ""
+
+    def _iter_structured_candidates(
+        self,
+        *,
+        message: Any,
+        choice: Any,
+        response: Any,
+    ) -> list[tuple[str, Any]]:
+        candidates: list[tuple[str, Any]] = []
+        if isinstance(message, dict):
+            tool_calls = message.get("tool_calls")
+            if isinstance(tool_calls, list) and tool_calls:
+                first = tool_calls[0] if isinstance(tool_calls[0], dict) else {}
+                fn = first.get("function") if isinstance(first, dict) else {}
+                candidates.append(("tool_calls.function.arguments", fn.get("arguments") if isinstance(fn, dict) else None))
+            candidates.append(("message.content", message.get("content")))
+        if isinstance(choice, dict):
+            candidates.append(("choice.text", choice.get("text")))
+        if isinstance(response, dict):
+            candidates.append(("response.output_text", response.get("output_text")))
+        if isinstance(message, dict):
+            candidates.append(("message.reasoning_content", message.get("reasoning_content")))
+            candidates.append(("message.reasoning_content", message.get("reasoning")))
+        if isinstance(choice, dict):
+            candidates.append(("choice.reasoning_content", choice.get("reasoning_content")))
+            candidates.append(("choice.reasoning_content", choice.get("reasoning")))
+        if isinstance(response, dict):
+            candidates.append(("response.reasoning_content", response.get("reasoning_content")))
+            candidates.append(("response.reasoning_content", response.get("reasoning")))
+        return candidates
+
+    @classmethod
+    def _unwrap_structured_payload_once(cls, payload: Any) -> Any:
+        if not isinstance(payload, dict):
+            return payload
+        for key in ("json", "data", "result", "arguments"):
+            if key not in payload:
+                continue
+            inner = payload.get(key)
+            if isinstance(inner, str):
+                candidate = cls._extract_json_candidate(cls._strip_code_fence(inner))
+                if not candidate:
+                    continue
+                try:
+                    return json.loads(candidate)
+                except Exception:
+                    return inner
+            if isinstance(inner, (dict, list)):
+                return inner
+        return payload
+
+    @staticmethod
+    def _matches_expected_type(value: Any, expected: str) -> bool:
+        if expected == "bool":
+            return isinstance(value, bool)
+        if expected == "str":
+            return isinstance(value, str)
+        if expected == "list":
+            return isinstance(value, list)
+        if expected == "number":
+            return isinstance(value, (int, float)) and not isinstance(value, bool)
+        if expected == "dict_or_none":
+            return value is None or isinstance(value, dict)
+        return True
+
+    def _validate_structured_payload(self, *, function_name: str, payload: Any) -> Optional[str]:
+        fn = str(function_name or "").strip().lower()
+        schema = STRUCTURED_OBJECT_SCHEMAS.get(fn)
+        if schema is None:
+            return None if isinstance(payload, (dict, list)) else "structured_json_invalid"
+        if not isinstance(payload, dict):
+            return "structured_schema_invalid"
+        required_keys = schema.get("required_keys") or ()
+        for key in required_keys:
+            if key not in payload:
+                return "structured_schema_invalid"
+        type_checks = schema.get("type_checks") or {}
+        for key, expected in type_checks.items():
+            if not self._matches_expected_type(payload.get(key), str(expected)):
+                return "structured_schema_invalid"
+        optional_type_checks = schema.get("optional_type_checks") or {}
+        for key, expected_values in optional_type_checks.items():
+            if key not in payload:
+                continue
+            if not any(self._matches_expected_type(payload.get(key), str(expected)) for expected in expected_values):
+                return "structured_schema_invalid"
+        return None
+
+    def _extract_structured_object_from_response(
+        self,
+        *,
+        function_name: str,
+        response: Any,
+        fallback_text: str,
+        provider: str,
+        model: str,
+    ) -> StructuredExtractionResult:
+        message = {}
+        choice: dict[str, Any] = {}
+        if isinstance(response, dict):
+            choices = response.get("choices") if isinstance(response.get("choices"), list) else []
+            choice = choices[0] if choices and isinstance(choices[0], dict) else {}
+            message = choice.get("message") if isinstance(choice.get("message"), dict) else {}
+
+        last_nonempty_field = ""
+        last_nonempty_text = ""
+        for field, raw_value in self._iter_structured_candidates(message=message, choice=choice, response=response):
+            text = self._sanitize_visible_answer(self._extract_text_from_value(raw_value))
+            normalized = self._strip_code_fence(text)
+            if normalized:
+                last_nonempty_field = field
+                last_nonempty_text = normalized[:4000]
+            if not self._looks_like_json_container(normalized):
+                continue
+            candidate = self._extract_json_candidate(normalized)
+            if not candidate:
+                continue
+            try:
+                parsed = json.loads(candidate)
+            except Exception:
+                return StructuredExtractionResult(
+                    payload=None,
+                    source_field=field,
+                    provider=provider,
+                    model=model,
+                    error_code="structured_json_invalid",
+                    raw_text=candidate[:4000],
+                    validation="invalid",
+                )
+            parsed = self._unwrap_structured_payload_once(parsed)
+            validation_error = self._validate_structured_payload(function_name=function_name, payload=parsed)
+            if validation_error is not None:
+                return StructuredExtractionResult(
+                    payload=None,
+                    source_field=field,
+                    provider=provider,
+                    model=model,
+                    error_code=validation_error,
+                    raw_text=candidate[:4000],
+                    validation="invalid",
+                )
+            return StructuredExtractionResult(
+                payload=parsed if isinstance(parsed, dict) else None,
+                source_field=field,
+                provider=provider,
+                model=model,
+                raw_text=candidate[:4000],
+                validation="ok",
+            )
+
+        fallback_normalized = self._strip_code_fence(fallback_text)
+        fallback_candidate = (
+            self._extract_json_candidate(fallback_normalized)
+            if self._looks_like_json_container(fallback_normalized)
+            else ""
+        )
+        if fallback_candidate:
+            try:
+                parsed = json.loads(fallback_candidate)
+            except Exception:
+                return StructuredExtractionResult(
+                    payload=None,
+                    source_field="generated_text",
+                    provider=provider,
+                    model=model,
+                    error_code="structured_json_invalid",
+                    raw_text=fallback_candidate[:4000],
+                    validation="invalid",
+                )
+            parsed = self._unwrap_structured_payload_once(parsed)
+            validation_error = self._validate_structured_payload(function_name=function_name, payload=parsed)
+            if validation_error is None and isinstance(parsed, dict):
+                return StructuredExtractionResult(
+                    payload=parsed,
+                    source_field="generated_text",
+                    provider=provider,
+                    model=model,
+                    raw_text=fallback_candidate[:4000],
+                    validation="ok",
+                )
+
+        if not last_nonempty_field and str(fallback_text or "").strip():
+            last_nonempty_field = "generated_text"
+            last_nonempty_text = str(fallback_text or "").strip()[:4000]
+        error_code = "structured_json_not_found" if last_nonempty_field else "structured_empty"
+        return StructuredExtractionResult(
+            payload=None,
+            source_field=last_nonempty_field,
+            provider=provider,
+            model=model,
+            error_code=error_code,
+            raw_text=last_nonempty_text,
+            validation="invalid",
+        )
+
+    @staticmethod
+    def _classify_structured_exception(exc: Exception) -> str:
+        if isinstance(exc, (asyncio.TimeoutError, httpx.TimeoutException)):
+            return "structured_timeout"
+        if isinstance(exc, (httpx.HTTPError, httpx.RequestError)):
+            return "structured_http_error"
+        lowered = str(exc or "").strip().lower()
+        if "timeout" in lowered:
+            return "structured_timeout"
+        if "status" in lowered or "http" in lowered:
+            return "structured_http_error"
+        return "structured_empty"
+
+    @staticmethod
+    def _is_same_runtime_engine(a: dict[str, Any], b: dict[str, Any]) -> bool:
+        return (
+            str(a.get("provider") or "").strip().lower() == str(b.get("provider") or "").strip().lower()
+            and str(a.get("model") or "").strip().lower() == str(b.get("model") or "").strip().lower()
+            and str(a.get("base_url") or "").strip().rstrip("/").lower()
+            == str(b.get("base_url") or "").strip().rstrip("/").lower()
+        )
+
+    async def _resolve_high_tier_engine_for_escalation(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        primary_engine: dict[str, Any],
+    ) -> Optional[dict[str, Any]]:
+        _, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        active_engine = self._resolve_active_engine_from_bundle(bundle)
+        engines: list[dict[str, Any]] = [e for e in bundle.get("engines", []) if isinstance(e, dict)]
+        routed = self._resolve_tier_engine(
+            tier="high",
+            engines=engines,
+            active_engine=active_engine,
+        )
+        if not routed:
+            return None
+        if self._is_same_runtime_engine(routed, primary_engine):
+            return None
+
+        provider = str(routed.get("provider") or "").strip().lower()
+        api_key = str(routed.get("api_key") or "").strip()
+        if provider != "ollama" and not api_key:
+            return None
+        return routed
+
+    async def _resolve_safety_tier_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        function_name: str,
+        primary_engine: dict[str, Any],
+    ) -> Optional[dict[str, Any]]:
+        _, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        active_engine = self._resolve_active_engine_from_bundle(bundle)
+        engines: list[dict[str, Any]] = [e for e in bundle.get("engines", []) if isinstance(e, dict)]
+
+        safety_rule = self._match_openclaw_rule(function_name, tier_filter={"safety"})
+        routed: Optional[dict[str, Any]] = None
+        if safety_rule is not None:
+            routed = self._resolve_engine_from_rule(
+                rule=safety_rule,
+                engines=engines,
+                active_engine=active_engine,
+            )
+        if routed is None:
+            routed = self._resolve_tier_engine(
+                tier="safety",
+                engines=engines,
+                active_engine=active_engine,
+            )
+        if not routed:
+            return None
+        routed = self._ensure_openclaw_control_plane_engine(routed)
+        if self._is_same_runtime_engine(routed, primary_engine):
+            return None
+
+        provider = str(routed.get("provider") or "").strip().lower()
+        api_key = str(routed.get("api_key") or "").strip()
+        if provider != "ollama" and not api_key:
+            return None
+        return routed
+
+    async def _try_safety_engine(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        function_name: str,
+        prompt: str,
+        system_prompt: str,
+        temperature: float,
+        max_tokens: int,
+        primary_engine: dict[str, Any],
+        reason: str,
+    ) -> tuple[str, Optional[dict[str, Any]]]:
+        safety_engine = await self._resolve_safety_tier_engine(
+            db,
+            tenant_id=tenant_id,
+            function_name=function_name,
+            primary_engine=primary_engine,
+        )
+        if safety_engine is None:
+            _runtime_log(
+                "LLM路由",
+                f"Safety 引擎不可用，无法触发兜底：function={function_name}, reason={reason}",
+            )
+            return "", None
+        safety_provider = str(safety_engine.get("provider") or "")
+        safety_model = str(safety_engine.get("model") or "")
+        _runtime_log(
+            "LLM路由",
+            (
+                "触发 Safety 兜底："
+                f"function={function_name}, provider={safety_provider}, model={safety_model}, reason={reason}"
+            ),
+        )
+        try:
+            text = await self._chat_completion(
+                provider=safety_provider,
+                prompt=prompt,
+                system_prompt=system_prompt,
+                api_key=str(safety_engine.get("api_key") or ""),
+                base_url=str(safety_engine.get("base_url") or ""),
+                model=safety_model,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                function_name=function_name,
+            )
+            if not text.strip():
+                raise ValueError("safety_engine_empty_response")
+            if self._is_invalid_structured_output(function_name=function_name, response_text=text):
+                raise ValueError("safety_engine_invalid_structured_output")
+            return text, safety_engine
+        except Exception as exc:
+            _runtime_log("LLM路由", f"Safety 兜底失败：function={function_name}, error={exc}")
+            return "", None
+
+    def _pick_tier_fallback_engine(
+        self,
+        *,
+        tier: str,
+        engines: list[dict[str, Any]],
+        active_engine: dict[str, Any],
+    ) -> Optional[dict[str, Any]]:
+        for item in engines:
+            candidate = self._normalize_engine_runtime(item)
+            provider = str(candidate.get("provider") or "")
+            api_key = str(candidate.get("api_key") or "").strip()
+            if provider != "ollama" and not api_key:
+                continue
+            if tier == "high" and provider == "ollama":
+                continue
+            if str(candidate.get("id") or "") == str(active_engine.get("id") or ""):
+                continue
+            _runtime_log(
+                "LLM路由",
+                (
+                    "分层路由回退引擎生效："
+                    f"tier={tier}, provider={provider}, model={candidate.get('model')}"
+                ),
+            )
+            return candidate
+        return None
+
+    def _resolve_tier_engine(
+        self,
+        *,
+        tier: str,
+        engines: list[dict[str, Any]],
+        active_engine: dict[str, Any],
+    ) -> Optional[dict[str, Any]]:
+        tier = str(tier or "").strip().lower()
+        upper = tier.upper()
+        default_provider_by_tier = {
+            "low": "openai",
+            "high": "openai",
+            "safety": "openai",
+        }
+        provider_env = str(
+            os.getenv(
+                f"LLM_ROUTING_{upper}_PROVIDER",
+                default_provider_by_tier.get(
+                    tier,
+                    os.getenv("LLM_PROVIDER", "openai"),
+                ),
+            )
+            or ""
+        ).strip()
+        engine_id = str(os.getenv(f"LLM_ROUTING_{upper}_ENGINE_ID", "") or "").strip()
+        model_override = str(os.getenv(f"LLM_ROUTING_{upper}_MODEL", "") or "").strip()
+        base_url_override = str(os.getenv(f"LLM_ROUTING_{upper}_BASE_URL", "") or "").strip().rstrip("/")
+
+        routed: Optional[dict[str, Any]] = None
+        if engine_id:
+            for item in engines:
+                if str(item.get("id") or "") == engine_id:
+                    routed = self._normalize_engine_runtime(item)
+                    break
+
+        if routed is None and provider_env:
+            try:
+                provider = _normalize_provider(provider_env)
+            except Exception:
+                provider = ""
+            if provider:
+                for item in engines:
+                    if str(item.get("provider") or "").strip().lower() == provider:
+                        routed = self._normalize_engine_runtime(item)
+                        break
+                if routed is None:
+                    try:
+                        routed = self._env_provider_runtime(provider)
+                    except Exception:
+                        routed = None
+
+        if routed is None:
+            return None
+
+        provider = str(routed.get("provider") or "").strip().lower()
+        if model_override:
+            routed["model"] = model_override
+        elif self._openclaw_routing_enabled() and provider == "openai":
+            default_model = DEFAULT_TIER_MODEL_OVERRIDES.get(tier)
+            if default_model:
+                routed["model"] = default_model
+        if base_url_override:
+            routed["base_url"] = base_url_override
+        elif self._openclaw_routing_enabled() and provider == "openai":
+            routed["base_url"] = OPENCLAW_PRIMARY_GATEWAY_URL
+
+        if str(routed.get("id") or "") != str(active_engine.get("id") or ""):
+            _runtime_log(
+                "LLM路由",
+                (
+                    "混合路由已命中："
+                    f"层级={tier}, provider={routed.get('provider')}, model={routed.get('model')}"
+                ),
+            )
+        return self._ensure_openclaw_control_plane_engine(routed)
+
+    async def _resolve_engine_for_test(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        engine_id: Optional[str],
+    ) -> dict[str, Any]:
+        if not engine_id:
+            engine = await self._resolve_active_engine(db, tenant_id=tenant_id)
+            if str(engine.get("provider") or "") != "ollama" and not str(engine.get("api_key") or "").strip():
+                raise ValueError("当前激活引擎未配置 API Key")
+            return engine
+
+        _, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        engines: list[dict[str, Any]] = [e for e in bundle.get("engines", []) if isinstance(e, dict)]
+        for item in engines:
+            if str(item.get("id") or "") == str(engine_id):
+                engine = self._normalize_engine_runtime(item)
+                if str(engine.get("provider") or "") != "ollama" and not str(engine.get("api_key") or "").strip():
+                    raise ValueError("目标引擎未配置 API Key")
+                return engine
+        raise ValueError("engine 不存在")
+
+    async def _resolve_active_prompt_profile(self, db: AsyncSession, *, tenant_id: int) -> dict[str, str]:
+        cached = self._prompt_profile_cache.get(int(tenant_id))
+        now = time.time()
+        if cached and cached[0] > now:
+            return dict(cached[1])
+
+        role_name = ""
+        system_prompt = ""
+        role_key = ""
+        try:
+            row = (
+                await db.execute(
+                    select(Integration).where(
+                        and_(
+                            Integration.tenant_id == int(tenant_id),
+                            Integration.type == "ai_prompt_profiles",
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            bundle = _safe_json(row.config if row else {})
+            active_key = str(bundle.get("active_role_key") or "").strip()
+            roles = bundle.get("roles") if isinstance(bundle.get("roles"), list) else []
+            selected = None
+            for item in roles:
+                if not isinstance(item, dict):
+                    continue
+                item_key = str(item.get("role_key") or "").strip()
+                if item_key == active_key:
+                    selected = item
+                    break
+            if selected is None:
+                for item in roles:
+                    if isinstance(item, dict) and bool(item.get("enabled", True)):
+                        selected = item
+                        break
+            if isinstance(selected, dict):
+                role_name = str(selected.get("role_name") or "").strip()
+                system_prompt = str(selected.get("system_prompt") or "").strip()
+                role_key = str(selected.get("role_key") or "").strip()
+        except Exception:
+            role_name = ""
+            system_prompt = ""
+            role_key = ""
+
+        if len(system_prompt) < 20:
+            fallback = build_default_role_prompt_bundle()
+            fallback_key = str(fallback.get("active_role_key") or "").strip()
+            roles = fallback.get("roles") if isinstance(fallback.get("roles"), list) else []
+            role = None
+            for item in roles:
+                if not isinstance(item, dict):
+                    continue
+                if str(item.get("role_key") or "").strip() == fallback_key:
+                    role = item
+                    break
+            if role is None and roles:
+                role = roles[0] if isinstance(roles[0], dict) else None
+            if isinstance(role, dict):
+                role_name = str(role.get("role_name") or role_name or "默认角色").strip()
+                system_prompt = str(role.get("system_prompt") or "").strip()
+                role_key = str(role.get("role_key") or role_key or fallback_key).strip()
+
+        result = {
+            "role_key": role_key,
+            "role_name": role_name,
+            "system_prompt": system_prompt,
+        }
+        self._prompt_profile_cache[int(tenant_id)] = (now + PROMPT_PROFILE_CACHE_TTL_SECONDS, result)
+        return dict(result)
+
+    def _build_fallback_engines(self, primary_engine: dict[str, Any]) -> list[dict[str, Any]]:
+        candidates: list[dict[str, Any]] = []
+        provider = _normalize_provider(primary_engine.get("provider"))
+        current_model = str(primary_engine.get("model") or "").strip()
+
+        provider_runtime = self._env_provider_runtime(provider)
+        default_model = str(provider_runtime.get("model") or PROVIDER_DEFAULTS.get(provider, {}).get("model") or "").strip()
+        if default_model and default_model != current_model:
+            same_provider = dict(primary_engine)
+            same_provider["model"] = default_model
+            candidates.append(self._normalize_engine_runtime(same_provider))
+
+        fallback_provider_raw = str(os.getenv("LLM_FALLBACK_PROVIDER", "ollama")).strip()
+        if fallback_provider_raw:
+            fallback_provider = _normalize_provider(fallback_provider_raw)
+            if fallback_provider != provider:
+                candidates.append(self._env_provider_runtime(fallback_provider))
+
+        return candidates
+
+    def _env_provider_runtime(self, provider: str) -> dict[str, Any]:
+        provider_name = _normalize_provider(provider)
+        defaults = PROVIDER_DEFAULTS[provider_name]
+        api_env = str(defaults.get("api_env") or "").strip()
+        api_key = os.getenv(api_env, "") if api_env else ""
+        if provider_name == "glm":
+            api_key = os.getenv("GLM_API_KEY", os.getenv("ZHIPUAI_API_KEY", api_key))
+
+        base_url = os.getenv(f"{provider_name.upper()}_BASE_URL", defaults["base_url"])
+        model = os.getenv(f"{provider_name.upper()}_MODEL", defaults["model"])
+        if provider_name == "ollama":
+            base_url = os.getenv("OLLAMA_BASE_URL", os.getenv("OLLAMA_HOST", defaults["base_url"]))
+            model = os.getenv("OLLAMA_MODEL", os.getenv("OLLAMA_CHAT_MODEL", defaults["model"]))
+
+        return {
+            "id": f"env-{provider_name}",
+            "name": f"{defaults['name']} (ENV)",
+            "provider": provider_name,
+            "api_key": str(api_key or "").strip(),
+            "base_url": str(base_url or defaults["base_url"]).strip().rstrip("/"),
+            "model": str(model or defaults["model"]).strip(),
+        }
+
+    def _env_engine_runtime(self) -> dict[str, Any]:
+        provider = _normalize_provider(os.getenv("LLM_PROVIDER", "deepseek"))
+        runtime = self._env_provider_runtime(provider)
+        runtime["id"] = "env-default"
+        return runtime
+
+    def _normalize_engine_runtime(self, engine: dict[str, Any]) -> dict[str, Any]:
+        provider = _normalize_provider(engine.get("provider"))
+        defaults = PROVIDER_DEFAULTS[provider]
+        runtime = dict(engine)
+        runtime["provider"] = provider
+        runtime["base_url"] = str(runtime.get("base_url") or defaults["base_url"]).strip().rstrip("/")
+        runtime["model"] = str(runtime.get("model") or defaults["model"]).strip()
+        runtime["api_key"] = str(runtime.get("api_key") or "").strip()
+        runtime["name"] = str(runtime.get("name") or defaults["name"]).strip()
+        if not runtime.get("id"):
+            runtime["id"] = str(uuid4())
+        return runtime
+
+    async def _log_ai_call(
+        self,
+        *,
+        tenant_id: int,
+        function_name: str,
+        prompt: str,
+        response: str,
+        status: str,
+        provider: str = "",
+        model: str = "",
+    ) -> None:
+        tokens_used = max(len(prompt) // 4, 1) + max(len(response) // 4, 0)
+        cost = self._estimate_cost(provider=provider, model=model, tokens=tokens_used)
+        row_payload = {
+            "tenant_id": tenant_id,
+            "user_id": None,
+            "function_name": function_name,
+            "prompt": prompt[:8000],
+            "response": (response or "")[:8000],
+            "tokens_used": tokens_used,
+            "cost": cost,
+            "status": status,
+        }
+        async with AsyncSessionLocal() as session:
+            for attempt in range(2):
+                try:
+                    session.add(AICallLog(**row_payload))
+                    await session.commit()
+                    return
+                except OperationalError as exc:
+                    await session.rollback()
+                    if "database is locked" not in str(exc).lower() or attempt >= 1:
+                        return
+                    await asyncio.sleep(0.05 * (attempt + 1))
+                except Exception:
+                    await session.rollback()
+                    return
+
+    @staticmethod
+    def _estimate_cost(provider: str, model: str, tokens: int) -> float:
+        _ = model
+        per_1k = {
+            "deepseek": 0.0003,
+            "glm": 0.0006,
+            "openai": 0.0012,
+            "ollama": 0.0,
+        }.get(provider, 0.0005)
+        return round((tokens / 1000.0) * per_1k, 6)
+
+    @staticmethod
+    def _parse_json_object(text: str) -> dict[str, Any]:
+        raw = (text or "").strip()
+        if raw.startswith("```"):
+            parts = raw.split("```")
+            if len(parts) >= 2:
+                raw = parts[1].strip()
+            if raw.startswith("json"):
+                raw = raw[4:].strip()
+        try:
+            parsed = json.loads(raw)
+            return parsed if isinstance(parsed, dict) else {}
+        except Exception:
+            return {}
+
+    @staticmethod
+    def _parse_json_array(text: str) -> list[dict]:
+        raw = (text or "").strip()
+        if raw.startswith("```"):
+            parts = raw.split("```")
+            if len(parts) >= 2:
+                raw = parts[1].strip()
+            if raw.startswith("json"):
+                raw = raw[4:].strip()
+        try:
+            parsed = json.loads(raw)
+            if isinstance(parsed, list):
+                return [item for item in parsed if isinstance(item, dict)]
+        except Exception:
+            return []
+        return []
+
+    @staticmethod
+    def _extract_text_from_value(value: Any) -> str:
+        if value is None:
+            return ""
+        if isinstance(value, str):
+            return value
+        if isinstance(value, (int, float, bool)):
+            return str(value)
+        if isinstance(value, dict):
+            parts: list[str] = []
+            for key in ("text", "content", "value", "output_text", "input_text", "arguments", "json"):
+                if key in value:
+                    parts.append(LLMService._extract_text_from_value(value.get(key)))
+            return "".join(parts)
+        if isinstance(value, list):
+            return "".join(LLMService._extract_text_from_value(item) for item in value)
+        return str(value)
+
+    @staticmethod
+    def _sanitize_visible_answer(text: str) -> str:
+        cleaned = str(text or "")
+        if not cleaned:
+            return ""
+        cleaned = re.sub(r"<think>.*?</think>", "", cleaned, flags=re.IGNORECASE | re.DOTALL)
+        return cleaned.strip()
+
+
+llm_service = LLMService()
diff --git a/backend/app/services/llm_client.py b/backend/app/services/llm_client.py
new file mode 100644
index 00000000..72aa9e81
--- /dev/null
+++ b/backend/app/services/llm_client.py
@@ -0,0 +1,184 @@
+"""
+统一 LLM 客户端 - 支持多种提供商
+
+支持的提供商:
+- ollama: 本地运行 (免费，但占用 CPU/内存)
+- deepseek: 云端 API (便宜，¥1/百万 tokens)
+- openai: 云端 API (GPT-4)
+- qwen: 阿里云通义千问 (便宜，国内稳定)
+"""
+
+import httpx
+from typing import Dict, List, Optional, Any
+import os
+import json
+
+
+class LLMClient:
+    """统一 LLM 客户端"""
+
+    def __init__(self):
+        self.provider = os.getenv("LLM_PROVIDER", "deepseek").lower()
+        self.client = self._create_client()
+
+    def _create_client(self) -> 'LLMProvider':
+        """根据配置创建对应的提供商客户端"""
+        if self.provider == "ollama":
+            return OllamaClient()
+        elif self.provider == "deepseek":
+            return OpenAICompatibleClient(
+                api_key=os.getenv("DEEPSEEK_API_KEY"),
+                base_url=os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com/v1"),
+                model=os.getenv("DEEPSEEK_MODEL", "deepseek-chat")
+            )
+        elif self.provider == "openai":
+            return OpenAICompatibleClient(
+                api_key=os.getenv("OPENAI_API_KEY"),
+                base_url=os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1"),
+                model=os.getenv("OPENAI_MODEL", "gpt-4o-mini")
+            )
+        elif self.provider == "qwen":
+            return OpenAICompatibleClient(
+                api_key=os.getenv("QWEN_API_KEY"),
+                base_url=os.getenv("QWEN_BASE_URL", "https://dashscope.aliyuncs.com/compatible-mode/v1"),
+                model=os.getenv("QWEN_MODEL", "qwen-plus")
+            )
+        else:
+            raise ValueError(f"不支持的 LLM 提供商: {self.provider}")
+
+    def generate(self, prompt: str, **kwargs) -> str:
+        """生成文本"""
+        return self.client.generate(prompt, **kwargs)
+
+    def generate_json(self, prompt: str, **kwargs) -> Dict:
+        """生成 JSON 格式响应"""
+        return self.client.generate_json(prompt, **kwargs)
+
+    def get_embedding(self, text: str) -> List[float]:
+        """获取向量嵌入"""
+        return self.client.get_embedding(text)
+
+
+class LLMProvider:
+    """LLM 提供商基类"""
+    def generate(self, prompt: str, **kwargs) -> str:
+        raise NotImplementedError
+
+    def generate_json(self, prompt: str, **kwargs) -> Dict:
+        raise NotImplementedError
+
+    def get_embedding(self, text: str) -> List[float]:
+        raise NotImplementedError
+
+
+class OllamaClient(LLMProvider):
+    """Ollama 本地客户端"""
+
+    def __init__(self):
+        self.host = os.getenv("OLLAMA_HOST", "http://127.0.0.1:11434").rstrip('/')
+        self.chat_model = os.getenv("OLLAMA_CHAT_MODEL", "llama3:8b")
+        self.embedding_model = os.getenv("OLLAMA_EMBEDDING_MODEL", "nomic-embed-text")
+        self.client = httpx.Client(base_url=self.host, timeout=300.0, trust_env=False)
+
+    def generate(self, prompt: str, **kwargs) -> str:
+        response = self.client.post("/api/generate", json={
+            "model": self.chat_model,
+            "prompt": prompt,
+            "stream": False,
+            **kwargs
+        })
+        response.raise_for_status()
+        return response.json()["response"]
+
+    def generate_json(self, prompt: str, **kwargs) -> Dict:
+        response = self.client.post("/api/generate", json={
+            "model": self.chat_model,
+            "prompt": prompt,
+            "stream": False,
+            "format": "json",
+            **kwargs
+        })
+        response.raise_for_status()
+        content = response.json()["response"]
+        # 清理 markdown 代码块
+        if content.startswith("```json"):
+            content = content[7:]
+        if content.startswith("```"):
+            content = content[3:]
+        if content.endswith("```"):
+            content = content[:-3]
+        content = content.strip()
+        return json.loads(content)
+
+    def get_embedding(self, text: str) -> List[float]:
+        response = self.client.post("/api/embeddings", json={
+            "model": self.embedding_model,
+            "prompt": text
+        })
+        response.raise_for_status()
+        return response.json().get("embedding", [])
+
+
+class OpenAICompatibleClient(LLMProvider):
+    """OpenAI 兼容客户端 (适用于 DeepSeek, Qwen 等)"""
+
+    def __init__(self, api_key: str, base_url: str, model: str):
+        self.model = model
+        self.base_url = base_url
+        self.api_key = api_key
+        # 延迟初始化客户端，在实际调用时才检查 API_KEY
+        self._client = None
+
+    @property
+    def client(self):
+        """延迟初始化客户端"""
+        if self._client is None:
+            if not self.api_key:
+                raise ValueError(f"请设置 {self.base_url} 的 API_KEY")
+            self._client = httpx.Client(
+                base_url=self.base_url,
+                timeout=120.0,
+                headers={"Authorization": f"Bearer {self.api_key}"}
+            )
+        return self._client
+
+    def generate(self, prompt: str, **kwargs) -> str:
+        response = self.client.post("/chat/completions", json={
+            "model": self.model,
+            "messages": [{"role": "user", "content": prompt}],
+            **kwargs
+        })
+        response.raise_for_status()
+        return response.json()["choices"][0]["message"]["content"]
+
+    def generate_json(self, prompt: str, **kwargs) -> Dict:
+        response = self.client.post("/chat/completions", json={
+            "model": self.model,
+            "messages": [{"role": "user", "content": prompt}],
+            "response_format": {"type": "json_object"},
+            **kwargs
+        })
+        response.raise_for_status()
+        content = response.json()["choices"][0]["message"]["content"]
+        return json.loads(content)
+
+    def get_embedding(self, text: str) -> List[float]:
+        # DeepSeek 不支持 embedding，使用替代方案或返回空
+        if "deepseek" in self.base_url:
+            print("[WARNING] DeepSeek 不支持 embedding API，使用简化向量")
+            # 简化方案：基于文本内容生成伪向量（实际使用时应该用其他服务）
+            import hashlib
+            hash_obj = hashlib.md5(text.encode())
+            # 生成 768 维伪向量（与 nomic-embed-text 维度一致）
+            return [float((hash_obj.digest()[i % 16] / 255.0) * 2 - 1) for _ in range(768)]
+
+        response = self.client.post("/embeddings", json={
+            "model": self.model,
+            "input": text
+        })
+        response.raise_for_status()
+        return response.json()["data"][0]["embedding"]
+
+
+# 单例实例
+llm_client = LLMClient()
diff --git a/backend/app/services/model_capability_service.py b/backend/app/services/model_capability_service.py
new file mode 100644
index 00000000..64c9826b
--- /dev/null
+++ b/backend/app/services/model_capability_service.py
@@ -0,0 +1,701 @@
+"""模型能力矩阵核心服务
+
+AI 开发团队专属垂直能力矩阵的核心业务逻辑：
+- 模型-需求智能匹配校验
+- 算力成本跟踪与 ROI 核算
+- 能力差距检测与预警
+- 成本优化建议
+"""
+import json
+from datetime import datetime, timedelta
+from typing import Any, Dict, List, Optional, Union
+
+from sqlalchemy import select, and_, func, desc
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import (
+    ModelVersion,
+    ModelRequirementBinding,
+    TruthNode,
+    Task,
+    AICallLog,
+    ModelCapabilityGap,
+    CostTracking,
+    User,
+)
+from ..services.ai_service import AIService
+
+
+class ModelCapabilityService:
+    """模型能力矩阵核心服务"""
+
+    def __init__(self, db: AsyncSession, ai_service: AIService):
+        self.db = db
+        self.ai_service = ai_service
+
+    # ========================================================================
+    # 模型-需求智能匹配校验
+    # ========================================================================
+
+    async def validate_model_requirement_match(
+        self,
+        truth_node_id: int,
+        model_version_id: int,
+        tenant_id: int,
+        project_id: Optional[int] = None,
+    ) -> Dict[str, Any]:
+        """智能匹配校验：评估需求与模型的匹配度
+
+        Args:
+            truth_node_id: 宪章（需求）ID
+            model_version_id: 模型版本 ID
+            tenant_id: 租户 ID
+
+        Returns:
+            包含匹配度、成本预估、风险提示、能力差距等信息的字典
+        """
+        # 获取宪章和模型信息
+        truth_node = await self._get_truth_node(truth_node_id, tenant_id)
+        model_version = await self._get_model_version(model_version_id)
+
+        if not truth_node:
+            raise ValueError(f"宪章 {truth_node_id} 不存在")
+        if not model_version:
+            raise ValueError(f"模型版本 {model_version_id} 不存在")
+        if project_id is not None:
+            await self.validate_truth_node_in_project(
+                truth_node_id=truth_node_id,
+                tenant_id=tenant_id,
+                project_id=project_id,
+            )
+
+        # 格式化模型信息用于 AI 分析
+        model_info = self._format_model_info(model_version)
+        charter_content = str(truth_node.content or truth_node.title or "")
+
+        # AI 评估匹配度
+        evaluation = await self.ai_service.evaluate_model_requirement(
+            charter_content=charter_content,
+            model_version_id=model_version_id,
+            model_info=json.dumps(model_info, ensure_ascii=False),
+            tenant_id=tenant_id,
+            db=self.db,
+        )
+
+        # 创建或更新绑定记录
+        binding = await self._upsert_binding(
+            truth_node_id, model_version_id, evaluation, tenant_id
+        )
+
+        # 检测能力差距
+        gaps = await self._detect_capability_gaps(
+            truth_node_id, model_version_id, evaluation, tenant_id
+        )
+
+        return {
+            "binding_id": binding.id,
+            "compatibility_score": evaluation.get("compatibility_score", 0),
+            "cost_estimate_usd": evaluation.get("cost_estimate_usd", 0.0),
+            "cost_estimate_tokens_per_day": evaluation.get("cost_estimate_tokens_per_day", 0),
+            "risk_hints": evaluation.get("risk_hints", []),
+            "capability_gaps": gaps,
+            "recommendations": evaluation.get("recommendations", []),
+        }
+
+    async def validate_truth_node_in_project(
+        self, truth_node_id: int, tenant_id: int, project_id: int
+    ) -> TruthNode:
+        """校验宪章是否属于指定项目。"""
+        truth_node = await self._get_truth_node(truth_node_id, tenant_id)
+        if not truth_node:
+            raise ValueError(f"宪章 {truth_node_id} 不存在")
+        if truth_node.project_id != project_id:
+            raise ValueError(
+                f"宪章 {truth_node_id} 不属于项目 {project_id}，拒绝跨项目操作"
+            )
+        return truth_node
+
+    async def _get_truth_node(
+        self, truth_node_id: int, tenant_id: int
+    ) -> Optional[TruthNode]:
+        """获取宪章"""
+        result = await self.db.execute(
+            select(TruthNode).where(
+                and_(
+                    TruthNode.id == truth_node_id,
+                    TruthNode.tenant_id == tenant_id,
+                )
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def _get_model_version(
+        self, model_version_id: int
+    ) -> Optional[ModelVersion]:
+        """获取模型版本"""
+        result = await self.db.execute(
+            select(ModelVersion).where(ModelVersion.id == model_version_id)
+        )
+        return result.scalar_one_or_none()
+
+    def _format_model_info(self, model: ModelVersion) -> Dict[str, Any]:
+        """格式化模型信息用于 AI 分析"""
+        return {
+            "name": model.name,
+            "provider": model.provider,
+            "version": model.version,
+            "capability_boundaries": model.capability_boundaries or {},
+            "cost_per_1k_tokens": float(model.cost_per_1k_tokens) if model.cost_per_1k_tokens else 0.0,
+        }
+
+    async def _upsert_binding(
+        self,
+        truth_node_id: int,
+        model_version_id: int,
+        evaluation: Dict[str, Any],
+        tenant_id: int,
+    ) -> ModelRequirementBinding:
+        """创建或更新模型-需求绑定"""
+        # 检查是否已存在绑定
+        result = await self.db.execute(
+            select(ModelRequirementBinding).where(
+                and_(
+                    ModelRequirementBinding.truth_node_id == truth_node_id,
+                    ModelRequirementBinding.model_version_id == model_version_id,
+                )
+            )
+        )
+        binding = result.scalar_one_or_none()
+
+        compatibility_score = evaluation.get("compatibility_score", 0)
+        cost_estimate = evaluation.get("cost_estimate_usd", 0.0)
+        risk_notes = ", ".join(evaluation.get("risk_hints", []))
+
+        if binding:
+            # 更新现有绑定
+            binding.compatibility_score = int(compatibility_score)
+            binding.estimated_cost = float(cost_estimate)
+            binding.risk_notes = risk_notes[:1000] if risk_notes else None  # 限制长度
+        else:
+            # 创建新绑定
+            binding = ModelRequirementBinding(
+                tenant_id=tenant_id,
+                truth_node_id=truth_node_id,
+                model_version_id=model_version_id,
+                compatibility_score=int(compatibility_score),
+                estimated_cost=float(cost_estimate),
+                risk_notes=risk_notes[:1000] if risk_notes else None,
+            )
+            self.db.add(binding)
+
+        await self.db.commit()
+        await self.db.refresh(binding)
+        return binding
+
+    async def _detect_capability_gaps(
+        self,
+        truth_node_id: int,
+        model_version_id: int,
+        evaluation: Dict[str, Any],
+        tenant_id: int,
+    ) -> List[Dict[str, Any]]:
+        """检测能力差距并创建预警记录"""
+        gaps = []
+
+        # 检查兼容性评分
+        compatibility = evaluation.get("compatibility_score", 100)
+        if compatibility < 60:
+            gap = await self._create_capability_gap(
+                truth_node_id=truth_node_id,
+                model_version_id=model_version_id,
+                gap_type="capability",
+                severity="critical" if compatibility < 40 else "high",
+                description=f"模型兼容性评分仅 {compatibility}/100，可能无法满足需求",
+                ai_analysis={"compatibility_score": compatibility},
+                tenant_id=tenant_id,
+            )
+            gaps.append(gap)
+
+        # 检查成本风险
+        cost_estimate = evaluation.get("cost_estimate_usd", 0)
+        if cost_estimate > 100:  # 日成本超过 $100 视为高风险
+            gap = await self._create_capability_gap(
+                truth_node_id=truth_node_id,
+                model_version_id=model_version_id,
+                gap_type="cost",
+                severity="high" if cost_estimate > 500 else "medium",
+                description=f"预估日成本 ${cost_estimate:.2f}，超出常规范围",
+                ai_analysis={"cost_estimate_usd": cost_estimate},
+                tenant_id=tenant_id,
+            )
+            gaps.append(gap)
+
+        # 检查风险提示
+        risk_hints = evaluation.get("risk_hints", [])
+        if risk_hints:
+            gap = await self._create_capability_gap(
+                truth_node_id=truth_node_id,
+                model_version_id=model_version_id,
+                gap_type="risk",
+                severity="medium",
+                description="检测到潜在风险",
+                ai_analysis={"risk_hints": risk_hints},
+                tenant_id=tenant_id,
+            )
+            gaps.append(gap)
+
+        return gaps
+
+    async def _create_capability_gap(
+        self,
+        truth_node_id: int,
+        model_version_id: int,
+        gap_type: str,
+        severity: str,
+        description: str,
+        ai_analysis: Dict[str, Any],
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        """创建能力差距记录"""
+        # 检查是否已存在相同类型的未解决差距
+        existing = await self.db.execute(
+            select(ModelCapabilityGap).where(
+                and_(
+                    ModelCapabilityGap.truth_node_id == truth_node_id,
+                    ModelCapabilityGap.model_version_id == model_version_id,
+                    ModelCapabilityGap.gap_type == gap_type,
+                    ModelCapabilityGap.status.in_(["open", "acknowledged"]),
+                )
+            )
+        )
+        existing_gap = existing.scalar_one_or_none()
+
+        if existing_gap:
+            # 更新现有差距
+            existing_gap.description = description
+            existing_gap.ai_analysis = ai_analysis
+            existing_gap.severity = severity
+            await self.db.commit()
+            await self.db.refresh(existing_gap)
+            return {
+                "id": existing_gap.id,
+                "gap_type": existing_gap.gap_type,
+                "severity": existing_gap.severity,
+                "description": existing_gap.description,
+                "status": existing_gap.status,
+            }
+        else:
+            # 创建新差距
+            gap = ModelCapabilityGap(
+                tenant_id=tenant_id,
+                model_version_id=model_version_id,
+                truth_node_id=truth_node_id,
+                gap_type=gap_type,
+                severity=severity,
+                description=description,
+                ai_analysis=ai_analysis,
+                status="open",
+            )
+            self.db.add(gap)
+            await self.db.commit()
+            await self.db.refresh(gap)
+            return {
+                "id": gap.id,
+                "gap_type": gap.gap_type,
+                "severity": gap.severity,
+                "description": gap.description,
+                "status": gap.status,
+            }
+
+    # ========================================================================
+    # 算力成本跟踪与 ROI 核算
+    # ========================================================================
+
+    async def track_cost_and_roi(
+        self,
+        truth_node_id: int,
+        period: str = "weekly",
+        tenant_id: int = 1,
+    ) -> Dict[str, Any]:
+        """算力成本跟踪与 ROI 核算
+
+        Args:
+            truth_node_id: 宪章（需求）ID
+            period: 周期类型（daily, weekly, monthly, project）
+            tenant_id: 租户 ID
+
+        Returns:
+            包含成本明细、ROI 分析等信息的字典
+        """
+        # 获取相关任务和模型绑定
+        tasks = await self._get_truth_node_tasks(truth_node_id, tenant_id)
+        bindings = await self._get_truth_node_bindings(truth_node_id, tenant_id)
+
+        if not tasks:
+            return {
+                "total_tokens": 0,
+                "total_cost_usd": 0.0,
+                "task_breakdown": [],
+                "roi_analysis": {},
+                "cost_variance": 0.0,
+            }
+
+        # 汇总 token 使用量和成本
+        total_tokens = 0
+        total_cost = 0.0
+        task_costs = []
+
+        for task in tasks:
+            # 从 AI 调用日志统计 token
+            task_tokens = await self._get_task_token_usage(task.id)
+            task_cost = await self._calculate_task_cost(task, bindings)
+
+            total_tokens += task_tokens
+            total_cost += task_cost
+            task_costs.append({
+                "task_id": task.id,
+                "task_name": task.title,
+                "tokens": task_tokens,
+                "cost_usd": round(task_cost, 4),
+            })
+
+        # 获取或创建成本跟踪记录
+        cost_tracking = await self._get_or_create_cost_tracking(
+            truth_node_id=truth_node_id,
+            period=period,
+            tenant_id=tenant_id,
+            bindings=bindings,
+        )
+
+        # 更新实际数据
+        cost_tracking.actual_tokens = total_tokens
+        cost_tracking.actual_cost_usd = round(total_cost, 4)
+
+        # AI 分析 ROI
+        truth_node = await self._get_truth_node(truth_node_id, tenant_id)
+        roi_analysis = await self._analyze_cost_roi(
+            total_cost=total_cost,
+            total_tokens=total_tokens,
+            task_count=len(tasks),
+            truth_node_content=str(truth_node.content or truth_node.title or "") if truth_node else "",
+            tenant_id=tenant_id,
+        )
+
+        cost_tracking.roi_score = roi_analysis.get("roi_score")
+        cost_tracking.roi_analysis = roi_analysis
+
+        # 生成优化建议
+        optimization = await self._generate_cost_optimization(
+            truth_node_id=truth_node_id,
+            tenant_id=tenant_id,
+            total_cost=total_cost,
+        )
+        cost_tracking.optimization_suggestions = optimization
+
+        await self.db.commit()
+        await self.db.refresh(cost_tracking)
+
+        # 计算成本偏差
+        cost_variance = total_cost - (cost_tracking.estimated_cost_usd or 0)
+
+        return {
+            "total_tokens": total_tokens,
+            "total_cost_usd": round(total_cost, 4),
+            "task_breakdown": task_costs,
+            "roi_analysis": roi_analysis,
+            "cost_variance": round(cost_variance, 4),
+            "optimization_suggestions": optimization,
+        }
+
+    async def _get_truth_node_tasks(
+        self, truth_node_id: int, tenant_id: int
+    ) -> List[Task]:
+        """获取宪章下的所有任务"""
+        result = await self.db.execute(
+            select(Task).where(
+                and_(
+                    Task.truth_node_id == truth_node_id,
+                    Task.tenant_id == tenant_id,
+                )
+            )
+        )
+        return list(result.scalars().all())
+
+    async def _get_truth_node_bindings(
+        self, truth_node_id: int, tenant_id: int
+    ) -> List[ModelRequirementBinding]:
+        """获取宪章的模型绑定"""
+        result = await self.db.execute(
+            select(ModelRequirementBinding).where(
+                and_(
+                    ModelRequirementBinding.truth_node_id == truth_node_id,
+                    ModelRequirementBinding.tenant_id == tenant_id,
+                )
+            )
+        )
+        return list(result.scalars().all())
+
+    async def _get_task_token_usage(self, task_id: int) -> int:
+        """从 AI 调用日志统计任务的 token 使用量"""
+        # 注意：需要根据实际日志记录方式调整
+        # 这里假设日志中有 task_id 或相关标识
+        result = await self.db.execute(
+            select(func.sum(AICallLog.total_tokens)).where(
+                AICallLog.function_name.contains(f"task_{task_id}")
+            )
+        )
+        total = result.scalar() or 0
+        return int(total)
+
+    async def _calculate_task_cost(
+        self, task: Task, bindings: List[ModelRequirementBinding]
+    ) -> float:
+        """计算任务成本"""
+        # 如果有绑定记录，使用绑定的预估成本
+        if bindings:
+            return sum(float(b.estimated_cost or 0) for b in bindings) / max(len(bindings), 1)
+        return 0.0
+
+    async def _get_or_create_cost_tracking(
+        self,
+        truth_node_id: int,
+        period: str,
+        tenant_id: int,
+        bindings: List[ModelRequirementBinding],
+    ) -> CostTracking:
+        """获取或创建成本跟踪记录"""
+        # 查找现有记录
+        result = await self.db.execute(
+            select(CostTracking).where(
+                and_(
+                    CostTracking.truth_node_id == truth_node_id,
+                    CostTracking.period == period,
+                    CostTracking.tenant_id == tenant_id,
+                )
+            ).order_by(desc(CostTracking.created_at))
+        )
+        tracking = result.scalar_one_or_none()
+
+        if not tracking:
+            # 创建新记录
+            tracking = CostTracking(
+                tenant_id=tenant_id,
+                truth_node_id=truth_node_id,
+                model_version_id=bindings[0].model_version_id if bindings else None,
+                period=period,
+                estimated_tokens=0,
+                actual_tokens=0,
+                estimated_cost_usd=sum(float(b.estimated_cost or 0) for b in bindings),
+                actual_cost_usd=0.0,
+                roi_score=None,
+                roi_analysis={},
+                optimization_suggestions={},
+            )
+            self.db.add(tracking)
+            await self.db.commit()
+            await self.db.refresh(tracking)
+
+        return tracking
+
+    async def _analyze_cost_roi(
+        self,
+        total_cost: float,
+        total_tokens: int,
+        task_count: int,
+        truth_node_content: str,
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        """AI 分析成本效益"""
+        from ..config.ai_prompts import PROMPT_ANALYZE_COST_ROI
+
+        prompt = PROMPT_ANALYZE_COST_ROI.format(
+            total_cost=f"${total_cost:.2f}",
+            total_tokens=f"{total_tokens:,}",
+            task_count=task_count,
+            truth_node_content=truth_node_content[:2000],  # 限制长度
+        )
+
+        try:
+            text = await self.ai_service._invoke_with_retry(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="analyze_cost_roi",
+            )
+            return self.ai_service._parse_json(text)
+        except Exception:
+            # 失败时返回默认分析
+            return {
+                "roi_score": 5.0,
+                "cost_efficiency": "fair",
+                "cost_per_task": round(total_cost / max(task_count, 1), 2),
+                "optimization_suggestions": [],
+                "benchmark_comparison": "无法分析",
+            }
+
+    async def _generate_cost_optimization(
+        self, truth_node_id: int, tenant_id: int, total_cost: float
+    ) -> Dict[str, Any]:
+        """生成成本优化建议"""
+        # 简单的优化建议逻辑
+        suggestions = []
+
+        # 检查是否可以使用更便宜的模型
+        bindings = await self._get_truth_node_bindings(truth_node_id, tenant_id)
+        if bindings:
+            for binding in bindings:
+                model = await self._get_model_version(binding.model_version_id)
+                if model and model.cost_per_1k_tokens and model.cost_per_1k_tokens > 0.001:
+                    suggestions.append({
+                        "type": "model_downgrade",
+                        "priority": "medium",
+                        "description": f"考虑切换到更便宜的模型替代 {model.name}",
+                        "estimated_savings_percent": 30,
+                    })
+
+        # 检查是否可以启用缓存
+        suggestions.append({
+            "type": "cache_optimization",
+            "priority": "high",
+            "description": "启用 Redis 缓存可减少 20-40% 重复调用",
+            "estimated_savings_percent": 30,
+        })
+
+        return {
+            "suggestions": suggestions,
+            "total_potential_savings_percent": 30 if suggestions else 0,
+        }
+
+    # ========================================================================
+    # 能力差距管理
+    # ========================================================================
+
+    async def list_capability_gaps(
+        self,
+        tenant_id: int,
+        status: Optional[str] = None,
+        severity: Optional[str] = None,
+        limit: int = 100,
+    ) -> Dict[str, Any]:
+        """列出能力差距记录"""
+        query = select(ModelCapabilityGap).where(
+            ModelCapabilityGap.tenant_id == tenant_id
+        )
+
+        if status:
+            query = query.where(ModelCapabilityGap.status == status)
+        if severity:
+            query = query.where(ModelCapabilityGap.severity == severity)
+
+        query = query.order_by(desc(ModelCapabilityGap.created_at)).limit(limit)
+
+        result = await self.db.execute(query)
+        gaps = result.scalars().all()
+
+        return {
+            "gaps": [
+                {
+                    "id": g.id,
+                    "gap_type": g.gap_type,
+                    "severity": g.severity,
+                    "description": g.description,
+                    "ai_analysis": g.ai_analysis,
+                    "status": g.status,
+                    "created_at": g.created_at.isoformat() if g.created_at else None,
+                    "resolved_at": g.resolved_at.isoformat() if g.resolved_at else None,
+                }
+                for g in gaps
+            ],
+            "total": len(gaps),
+        }
+
+    async def acknowledge_gap(
+        self, gap_id: int, tenant_id: int
+    ) -> Dict[str, Any]:
+        """确认能力差距"""
+        result = await self.db.execute(
+            select(ModelCapabilityGap).where(
+                and_(
+                    ModelCapabilityGap.id == gap_id,
+                    ModelCapabilityGap.tenant_id == tenant_id,
+                )
+            )
+        )
+        gap = result.scalar_one_or_none()
+
+        if not gap:
+            raise ValueError(f"能力差距 {gap_id} 不存在")
+
+        gap.status = "acknowledged"
+        await self.db.commit()
+
+        return {"id": gap.id, "status": gap.status}
+
+    async def resolve_gap(
+        self, gap_id: int, tenant_id: int, resolution_notes: Optional[str] = None
+    ) -> Dict[str, Any]:
+        """解决能力差距"""
+        result = await self.db.execute(
+            select(ModelCapabilityGap).where(
+                and_(
+                    ModelCapabilityGap.id == gap_id,
+                    ModelCapabilityGap.tenant_id == tenant_id,
+                )
+            )
+        )
+        gap = result.scalar_one_or_none()
+
+        if not gap:
+            raise ValueError(f"能力差距 {gap_id} 不存在")
+
+        gap.status = "resolved"
+        gap.resolved_at = datetime.utcnow()
+        gap.resolution_notes = resolution_notes
+        await self.db.commit()
+
+        return {"id": gap.id, "status": gap.status}
+
+    # ========================================================================
+    # 成本跟踪查询
+    # ========================================================================
+
+    async def get_cost_tracking(
+        self, truth_node_id: int, period: str = "weekly", tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """获取成本跟踪记录"""
+        result = await self.db.execute(
+            select(CostTracking).where(
+                and_(
+                    CostTracking.truth_node_id == truth_node_id,
+                    CostTracking.period == period,
+                    CostTracking.tenant_id == tenant_id,
+                )
+            ).order_by(desc(CostTracking.created_at))
+        )
+        tracking = result.scalar_one_or_none()
+
+        if not tracking:
+            return {
+                "truth_node_id": truth_node_id,
+                "period": period,
+                "estimated_tokens": 0,
+                "actual_tokens": 0,
+                "estimated_cost_usd": 0.0,
+                "actual_cost_usd": 0.0,
+                "roi_score": None,
+                "roi_analysis": {},
+                "optimization_suggestions": {},
+            }
+
+        return {
+            "id": tracking.id,
+            "truth_node_id": tracking.truth_node_id,
+            "period": tracking.period,
+            "estimated_tokens": tracking.estimated_tokens,
+            "actual_tokens": tracking.actual_tokens,
+            "estimated_cost_usd": float(tracking.estimated_cost_usd),
+            "actual_cost_usd": float(tracking.actual_cost_usd),
+            "roi_score": tracking.roi_score,
+            "roi_analysis": tracking.roi_analysis,
+            "optimization_suggestions": tracking.optimization_suggestions,
+        }
diff --git a/backend/app/services/model_lifecycle_service.py b/backend/app/services/model_lifecycle_service.py
new file mode 100644
index 00000000..0bdb99ef
--- /dev/null
+++ b/backend/app/services/model_lifecycle_service.py
@@ -0,0 +1,211 @@
+"""模型生命周期管理服务（P1）"""
+
+from typing import Any, Dict, List, Optional
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import (
+    ModelRequirementBinding,
+    ModelUpdateImpact,
+    ModelVersion,
+)
+
+
+class ModelLifecycleService:
+    """模型更新影响评估与迁移任务生成。"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def assess_model_update_impact(
+        self,
+        old_model_version_id: int,
+        new_model_version_id: int,
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        old_model = await self._get_model(old_model_version_id)
+        new_model = await self._get_model(new_model_version_id)
+        if not old_model or not new_model:
+            raise ValueError("模型版本不存在")
+
+        affected = await self._get_affected_truth_nodes(old_model_version_id, tenant_id)
+        cost_change = self._calc_cost_change_percent(old_model, new_model)
+        performance_change = self._infer_performance_change(old_model, new_model)
+        risks = self._infer_compatibility_risks(old_model, new_model)
+        recommendation = self._recommend(cost_change, performance_change, risks)
+
+        impact = ModelUpdateImpact(
+            tenant_id=tenant_id,
+            old_model_version_id=old_model_version_id,
+            new_model_version_id=new_model_version_id,
+            affected_truth_nodes=affected,
+            cost_change_percentage=cost_change,
+            performance_change=performance_change,
+            compatibility_risks=risks,
+            migration_tasks={},
+            recommendation=recommendation,
+            status="pending_review",
+        )
+        self.db.add(impact)
+        await self.db.commit()
+        await self.db.refresh(impact)
+
+        return {
+            "impact_id": impact.id,
+            "affected_truth_nodes": affected,
+            "cost_change_percentage": cost_change,
+            "performance_change": performance_change,
+            "compatibility_risks": risks,
+            "recommendation": recommendation,
+            "status": impact.status,
+        }
+
+    async def generate_migration_tasks(
+        self,
+        impact_id: int,
+        tenant_id: int,
+        max_tasks: int = 5,
+    ) -> Dict[str, Any]:
+        impact = await self._get_impact(impact_id, tenant_id)
+        if not impact:
+            raise ValueError("影响评估记录不存在")
+
+        tasks: List[Dict[str, Any]] = []
+        affected_map = impact.affected_truth_nodes or {}
+        for truth_node_id, score in list(affected_map.items())[:max_tasks]:
+            tasks.append(
+                {
+                    "title": f"评估并迁移宪章 {truth_node_id} 到新模型",
+                    "type": "ai",
+                    "priority": "high" if float(score) >= 0.7 else "medium",
+                    "estimated_hours": 2.0 if float(score) >= 0.7 else 1.0,
+                    "notes": f"旧模型 -> 新模型影响分: {score}",
+                }
+            )
+
+        if impact.compatibility_risks:
+            tasks.append(
+                {
+                    "title": "执行回归验证（模型升级）",
+                    "type": "backend",
+                    "priority": "high",
+                    "estimated_hours": 3.0,
+                    "notes": "覆盖接口行为、输出一致性、预算消耗变化",
+                }
+            )
+
+        impact.migration_tasks = {"tasks": tasks[:max_tasks]}
+        impact.status = "in_progress"
+        await self.db.commit()
+        await self.db.refresh(impact)
+
+        return {
+            "impact_id": impact.id,
+            "status": impact.status,
+            "migration_tasks": impact.migration_tasks,
+        }
+
+    async def list_impacts(self, tenant_id: int, limit: int = 50) -> Dict[str, Any]:
+        result = await self.db.execute(
+            select(ModelUpdateImpact)
+            .where(ModelUpdateImpact.tenant_id == tenant_id)
+            .order_by(ModelUpdateImpact.created_at.desc())
+            .limit(limit)
+        )
+        impacts = result.scalars().all()
+        return {
+            "items": [
+                {
+                    "id": i.id,
+                    "old_model_version_id": i.old_model_version_id,
+                    "new_model_version_id": i.new_model_version_id,
+                    "cost_change_percentage": i.cost_change_percentage,
+                    "recommendation": i.recommendation,
+                    "status": i.status,
+                    "created_at": i.created_at.isoformat() if i.created_at else None,
+                }
+                for i in impacts
+            ]
+        }
+
+    async def _get_model(self, model_version_id: int) -> Optional[ModelVersion]:
+        result = await self.db.execute(
+            select(ModelVersion).where(ModelVersion.id == model_version_id)
+        )
+        return result.scalar_one_or_none()
+
+    async def _get_impact(self, impact_id: int, tenant_id: int) -> Optional[ModelUpdateImpact]:
+        result = await self.db.execute(
+            select(ModelUpdateImpact).where(
+                and_(ModelUpdateImpact.id == impact_id, ModelUpdateImpact.tenant_id == tenant_id)
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def _get_affected_truth_nodes(
+        self, old_model_version_id: int, tenant_id: int
+    ) -> Dict[str, float]:
+        result = await self.db.execute(
+            select(ModelRequirementBinding).where(
+                and_(
+                    ModelRequirementBinding.model_version_id == old_model_version_id,
+                    ModelRequirementBinding.tenant_id == tenant_id,
+                )
+            )
+        )
+        bindings = result.scalars().all()
+
+        affected: Dict[str, float] = {}
+        for b in bindings:
+            # 兼容度越低，升级影响越高。
+            impact_score = round(max(0.1, (100 - int(b.compatibility_score)) / 100), 2)
+            affected[str(b.truth_node_id)] = impact_score
+
+        # 如果当前没有绑定，返回空映射。
+        return affected
+
+    def _calc_cost_change_percent(self, old_model: ModelVersion, new_model: ModelVersion) -> float:
+        old_cost = float(old_model.cost_per_1k_tokens or 0.0)
+        new_cost = float(new_model.cost_per_1k_tokens or 0.0)
+        if old_cost <= 0:
+            return 0.0
+        return round(((new_cost - old_cost) / old_cost) * 100, 2)
+
+    def _infer_performance_change(
+        self, old_model: ModelVersion, new_model: ModelVersion
+    ) -> Dict[str, float]:
+        old_caps = old_model.capability_boundaries or {}
+        new_caps = new_model.capability_boundaries or {}
+        old_ctx = float(old_caps.get("context_window", 0) or 0)
+        new_ctx = float(new_caps.get("context_window", 0) or 0)
+        ctx_change = 0.0 if old_ctx <= 0 else round(((new_ctx - old_ctx) / old_ctx) * 100, 2)
+        quality_change = round(10.0 if len(new_caps.keys()) > len(old_caps.keys()) else 0.0, 2)
+        return {
+            "context_window_change_percent": ctx_change,
+            "quality_change_percent": quality_change,
+        }
+
+    def _infer_compatibility_risks(
+        self, old_model: ModelVersion, new_model: ModelVersion
+    ) -> Dict[str, Any]:
+        risks: List[Dict[str, str]] = []
+        if old_model.provider != new_model.provider:
+            risks.append({"type": "provider_change", "severity": "high"})
+        if old_model.name != new_model.name:
+            risks.append({"type": "model_family_change", "severity": "medium"})
+        return {"risks": risks, "count": len(risks)}
+
+    def _recommend(
+        self,
+        cost_change: float,
+        performance_change: Dict[str, float],
+        risks: Dict[str, Any],
+    ) -> str:
+        if risks.get("count", 0) > 0 and performance_change.get("quality_change_percent", 0) <= 0:
+            return "evaluate"
+        if cost_change <= 10 and performance_change.get("quality_change_percent", 0) >= 5:
+            return "immediate_upgrade"
+        if cost_change > 30 and performance_change.get("quality_change_percent", 0) < 5:
+            return "defer"
+        return "evaluate"
diff --git a/backend/app/services/notification_service.py b/backend/app/services/notification_service.py
new file mode 100644
index 00000000..c1e1babe
--- /dev/null
+++ b/backend/app/services/notification_service.py
@@ -0,0 +1,241 @@
+"""站内通知服务"""
+from __future__ import annotations
+
+from datetime import datetime, timedelta, timezone
+from typing import Any, Iterable, Optional
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import Notification, User
+
+
+class NotificationService:
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def create_notification(
+        self,
+        *,
+        tenant_id: int,
+        user_id: int,
+        title: str,
+        content: Optional[str] = None,
+        category: str = "system",
+        source: Optional[str] = None,
+        entity_type: Optional[str] = None,
+        entity_id: Optional[int] = None,
+        target_url: Optional[str] = None,
+        kind: Optional[str] = None,
+        priority: Optional[str] = None,
+        status: Optional[str] = None,
+        blocking_reason: Optional[str] = None,
+        recommended_action: Optional[str] = None,
+        dedupe_key: Optional[str] = None,
+        occurrence_count: int = 1,
+        payload: Optional[dict[str, Any]] = None,
+        commit: bool = False,
+    ) -> Notification:
+        normalized_dedupe_key = (dedupe_key or "").strip()[:255] or None
+        if normalized_dedupe_key:
+            existing = (
+                await self.db.execute(
+                    select(Notification)
+                    .where(
+                        and_(
+                            Notification.tenant_id == int(tenant_id),
+                            Notification.user_id == int(user_id),
+                            Notification.dedupe_key == normalized_dedupe_key,
+                            Notification.is_read.is_(False),
+                        )
+                    )
+                    .order_by(Notification.updated_at.desc(), Notification.id.desc())
+                    .limit(1)
+                )
+            ).scalar_one_or_none()
+            if existing is not None:
+                existing.category = (category or existing.category or "system")[:50]
+                existing.title = (title or existing.title or "系统通知")[:200]
+                existing.content = (content or existing.content or "")[:4000] or None
+                existing.source = (source or existing.source or "")[:50] or None
+                existing.entity_type = (entity_type or existing.entity_type or "")[:50] or None
+                existing.entity_id = int(entity_id) if entity_id is not None else existing.entity_id
+                existing.target_url = (target_url or existing.target_url or "")[:500] or None
+                existing.kind = (kind or existing.kind or "")[:30] or None
+                existing.priority = (priority or existing.priority or "")[:20] or None
+                existing.status = (status or existing.status or "")[:30] or None
+                existing.blocking_reason = (blocking_reason or existing.blocking_reason or "")[:2000] or None
+                existing.recommended_action = (recommended_action or existing.recommended_action or "")[:80] or None
+                existing.payload = payload or existing.payload
+                existing.occurrence_count = max(1, int(existing.occurrence_count or 1)) + max(1, int(occurrence_count or 1))
+                await self.db.flush()
+                if commit:
+                    await self.db.commit()
+                return existing
+
+        row = Notification(
+            tenant_id=int(tenant_id),
+            user_id=int(user_id),
+            category=(category or "system")[:50],
+            title=(title or "系统通知")[:200],
+            content=(content or "")[:4000] or None,
+            source=(source or "")[:50] or None,
+            entity_type=(entity_type or "")[:50] or None,
+            entity_id=int(entity_id) if entity_id is not None else None,
+            target_url=(target_url or "")[:500] or None,
+            kind=(kind or "")[:30] or None,
+            priority=(priority or "")[:20] or None,
+            status=(status or "")[:30] or None,
+            blocking_reason=(blocking_reason or "")[:2000] or None,
+            recommended_action=(recommended_action or "")[:80] or None,
+            dedupe_key=normalized_dedupe_key,
+            occurrence_count=max(1, int(occurrence_count or 1)),
+            payload=payload or None,
+            is_read=False,
+        )
+        self.db.add(row)
+        await self.db.flush()
+        if commit:
+            await self.db.commit()
+        return row
+
+    async def create_notifications(
+        self,
+        *,
+        tenant_id: int,
+        user_ids: Iterable[int],
+        title: str,
+        content: Optional[str] = None,
+        category: str = "system",
+        source: Optional[str] = None,
+        entity_type: Optional[str] = None,
+        entity_id: Optional[int] = None,
+        target_url: Optional[str] = None,
+        kind: Optional[str] = None,
+        priority: Optional[str] = None,
+        status: Optional[str] = None,
+        blocking_reason: Optional[str] = None,
+        recommended_action: Optional[str] = None,
+        dedupe_key: Optional[str] = None,
+        occurrence_count: int = 1,
+        payload: Optional[dict[str, Any]] = None,
+        commit: bool = False,
+    ) -> list[Notification]:
+        rows: list[Notification] = []
+        for user_id in user_ids:
+            if user_id is None:
+                continue
+            rows.append(
+                await self.create_notification(
+                    tenant_id=tenant_id,
+                    user_id=int(user_id),
+                    title=title,
+                    content=content,
+                    category=category,
+                    source=source,
+                    entity_type=entity_type,
+                    entity_id=entity_id,
+                    target_url=target_url,
+                    kind=kind,
+                    priority=priority,
+                    status=status,
+                    blocking_reason=blocking_reason,
+                    recommended_action=recommended_action,
+                    dedupe_key=dedupe_key,
+                    occurrence_count=occurrence_count,
+                    payload=payload,
+                    commit=False,
+                )
+            )
+        if commit and rows:
+            await self.db.commit()
+        return rows
+
+    async def notify_signal_created(
+        self,
+        *,
+        tenant_id: int,
+        signal_id: int,
+        signal_content: str,
+        source: str = "feishu",
+        target_url: Optional[str] = None,
+        recipient_user_ids: Optional[Iterable[int]] = None,
+        proposer: Optional[str] = None,
+        occurred_at: Optional[datetime] = None,
+    ) -> int:
+        """向指定用户（或租户内全部活跃用户）发送“新需求已生成”通知。"""
+        user_ids = await self._resolve_recipients(
+            tenant_id=tenant_id,
+            recipient_user_ids=recipient_user_ids,
+        )
+        if not user_ids:
+            return 0
+
+        snippet = (signal_content or "").strip()
+        if len(snippet) > 80:
+            snippet = f"{snippet[:80]}..."
+        if not snippet:
+            snippet = "飞书消息触发了新的需求信号。"
+
+        proposer_name = (proposer or "").strip() or "未知用户"
+        occurred_text = self._format_cn_time(occurred_at)
+        title = "AI 已识别新需求"
+        content = (
+            f"提出人：{proposer_name}\n"
+            f"时间：{occurred_text}\n"
+            f"内容：{snippet}\n"
+            f"信号ID：{signal_id}"
+        )
+
+        rows = await self.create_notifications(
+            tenant_id=tenant_id,
+            user_ids=user_ids,
+            category="requirement",
+            title=title,
+            content=content,
+            source=source,
+            entity_type="signal",
+            entity_id=signal_id,
+            target_url=target_url,
+            kind="info",
+            priority="medium",
+            status="notified",
+            recommended_action="view_result",
+            dedupe_key=f"signal-created:{signal_id}",
+            payload={
+                "message_class": "info",
+                "project_id": None,
+                "project_name": None,
+                "summary": snippet,
+            },
+            commit=True,
+        )
+        return len(rows)
+
+    @staticmethod
+    def _format_cn_time(dt: Optional[datetime]) -> str:
+        if not dt:
+            dt = datetime.now(timezone.utc)
+        if dt.tzinfo is None:
+            dt = dt.replace(tzinfo=timezone.utc)
+        cn_tz = timezone(timedelta(hours=8))
+        return dt.astimezone(cn_tz).strftime("%Y-%m-%d %H:%M:%S")
+
+    async def _resolve_recipients(
+        self,
+        *,
+        tenant_id: int,
+        recipient_user_ids: Optional[Iterable[int]] = None,
+    ) -> list[int]:
+        if recipient_user_ids:
+            return [int(uid) for uid in recipient_user_ids if uid is not None]
+
+        result = await self.db.execute(
+            select(User.id).where(
+                and_(
+                    User.tenant_id == tenant_id,
+                    User.is_active.is_(True),
+                )
+            )
+        )
+        return [int(uid) for uid in result.scalars().all()]
diff --git a/backend/app/services/openclaw_commander_service.py b/backend/app/services/openclaw_commander_service.py
new file mode 100644
index 00000000..2b4273f4
--- /dev/null
+++ b/backend/app/services/openclaw_commander_service.py
@@ -0,0 +1,2417 @@
+"""OpenClaw 指挥层与操作执行服务。"""
+
+from __future__ import annotations
+
+import asyncio
+from datetime import datetime, timezone
+import hashlib
+import json
+import re
+import time
+from types import SimpleNamespace
+from typing import Any, Optional
+from uuid import uuid4
+
+from sqlalchemy import and_, func, or_, select
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import PROMPT_OPENCLAW_COMMAND_DISPATCH
+from ..models import (
+    OperationExecutionLedger,
+    Project,
+    SkillTag,
+    Task,
+    TruthNode,
+    User,
+    UserProject,
+    UserSkillBinding,
+    UserWorkload,
+)
+from . import workload_service
+from .adaptive_rag_service import AdaptiveRAGService
+from .ai_orchestration_run_service import AIOrchestrationRunService
+from .context_cache import context_cache_service
+from .llm import llm_service
+from .operation_catalog_service import operation_catalog_service
+from .operation_log_service import log_operation
+from .notification_service import NotificationService
+from .skill_score_service import process_task_completion
+from .sstc_service import ALIGNED_STATUSES
+from .task_persistence import build_task_create_draft, persist_task
+from ..utils.project_members import ensure_project_default_members
+from ..utils.sequence import ensure_project_id_sequence
+from ..utils.title import compact_title
+
+SUPPORTED_COMMANDER_OPERATION_CODES = (
+    "project.create",
+    "project.update",
+    "project.archive",
+    "project.reopen",
+    "task.create",
+    "task.assign",
+    "task.prioritize",
+    "task.status_transition",
+    "notification.create",
+    "personnel.workload_update",
+)
+PERSISTENT_IDEMPOTENCY_OPERATION_CODES = {"notification.create", "project.create"}
+SUPPORTED_PROJECT_STATUSES = {"active", "archived", "done"}
+SUPPORTED_TASK_STATUSES = {"todo", "in_progress", "review", "done", "blocked", "closed"}
+DONE_TASK_STATUSES = {"done", "completed", "closed"}
+SOURCE_ONBOARDING_FALLBACK_RE = re.compile(
+    r"(新增|增加|添加|接入|配置|注册|建立).{0,10}(信号源|来源|监控源|github|飞书|webhook|rss|url|source|channel)",
+    re.IGNORECASE,
+)
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+class OpenClawCommanderService:
+    """将自然语言指令转成可执行 operation plan。"""
+
+    def __init__(self, db: AsyncSession) -> None:
+        self.db = db
+
+    async def dispatch_command(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project_id: int,
+        instruction: str,
+        dry_run: bool,
+        max_operations: int,
+        response_mode: str = "fast",
+        trace_id: Optional[str] = None,
+        persist_run: bool = True,
+    ) -> dict[str, Any]:
+        project_row = await self._load_project(tenant_id=tenant_id, project_id=project_id)
+        if project_row is None:
+            raise ValueError("project not found")
+        project = self._build_project_snapshot(project_row)
+
+        members = await self._load_project_team_context(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            owner_id=int(project.owner_id) if project.owner_id else None,
+        )
+        tasks, task_summary = await self._load_project_tasks(tenant_id=tenant_id, project_id=project_id, members=members)
+        planner_role = await llm_service._resolve_active_prompt_profile(self.db, tenant_id=tenant_id)
+        trace_id = str(trace_id or f"openclaw-{tenant_id}-{project_id}-{int(time.time() * 1000)}-{uuid4().hex[:8]}")
+
+        plan, planner_status, planner_diagnostics = await self._plan_operations(
+            tenant_id=tenant_id,
+            operator_user=operator_user,
+            project=project,
+            tasks=tasks,
+            task_summary=task_summary,
+            members=members,
+            instruction=instruction,
+            dry_run=dry_run,
+            max_operations=max_operations,
+            response_mode=response_mode,
+            project_directory=[self._serialize_project_directory_item(project_row)],
+        )
+
+        executor = OperationExecutorService(self.db)
+        execution = await executor.execute_operations(
+            tenant_id=tenant_id,
+            operator_user=operator_user,
+            project=project,
+            operations=plan.get("operations") or [],
+            dry_run=dry_run,
+            trace_id=trace_id,
+        )
+        openclaw_status = self._resolve_openclaw_status(execution=execution, dry_run=dry_run)
+        run_status = openclaw_status if openclaw_status in {"completed", "partial", "failed"} else "completed"
+
+        result = {
+            "trace_id": trace_id,
+            "response_mode": response_mode,
+            "run_status": run_status,
+            "planner_status": planner_status,
+            "planner_diagnostics": planner_diagnostics,
+            "openclaw_status": openclaw_status,
+            "project": {
+                "id": str(project.id),
+                "title": project.title or project.name,
+                "status": project.status,
+            },
+            "prompt_role": {
+                "role_key": str(planner_role.get("role_key") or ""),
+                "role_name": str(planner_role.get("role_name") or ""),
+            },
+            "instruction": instruction,
+            "dry_run": bool(dry_run),
+            "summary": str(plan.get("summary") or "").strip(),
+            "confidence": self._to_float(plan.get("confidence"), default=0.0),
+            "task_summary": task_summary,
+            "member_count": len(members),
+            "planned_operations": len(plan.get("operations") or []),
+            "execution": execution,
+            "pending_workers": [],
+        }
+        if persist_run:
+            run_service = AIOrchestrationRunService(self.db)
+            await run_service.create_run(
+                tenant_id=tenant_id,
+                trace_id=trace_id,
+                run_type="openclaw_dispatch",
+                project_id=int(project.id),
+                signal_id=None,
+                response_mode=response_mode,
+                status=run_status,
+                planner_status=planner_status,
+                openclaw_status=openclaw_status,
+                worker_total=0,
+                worker_completed=0,
+                worker_failed=0,
+                summary=str(result.get("summary") or "").strip(),
+                planner_diagnostics=planner_diagnostics,
+                result_payload=dict(result),
+                created_by_user_id=int(operator_user.id) if getattr(operator_user, "id", None) else None,
+                completed_at=_utc_now_naive() if run_status in {"completed", "partial", "failed"} else None,
+            )
+            await self.db.commit()
+        return result
+
+    async def dispatch_global_command(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        instruction: str,
+        dry_run: bool,
+        max_operations: int,
+        response_mode: str = "fast",
+        trace_id: Optional[str] = None,
+        persist_run: bool = False,
+    ) -> dict[str, Any]:
+        project = self._build_global_project_snapshot()
+        project_directory = await self._load_project_directory(tenant_id=tenant_id)
+        planner_role = await llm_service._resolve_active_prompt_profile(self.db, tenant_id=tenant_id)
+        trace_id = str(trace_id or f"openclaw-{tenant_id}-global-{int(time.time() * 1000)}-{uuid4().hex[:8]}")
+
+        plan, planner_status, planner_diagnostics = await self._plan_operations(
+            tenant_id=tenant_id,
+            operator_user=operator_user,
+            project=project,
+            tasks=[],
+            task_summary={"total": 0, "active": 0, "done": 0, "blocked": 0, "sampled": 0},
+            members=[],
+            instruction=instruction,
+            dry_run=dry_run,
+            max_operations=max_operations,
+            response_mode=response_mode,
+            project_directory=project_directory,
+        )
+
+        executor = OperationExecutorService(self.db)
+        execution = await executor.execute_operations(
+            tenant_id=tenant_id,
+            operator_user=operator_user,
+            project=None,
+            operations=plan.get("operations") or [],
+            dry_run=dry_run,
+            trace_id=trace_id,
+        )
+        openclaw_status = self._resolve_openclaw_status(execution=execution, dry_run=dry_run)
+        run_status = openclaw_status if openclaw_status in {"completed", "partial", "failed"} else "completed"
+        project_ref = self._extract_project_reference_from_execution(execution) or {
+            "id": None,
+            "title": "",
+            "status": "",
+        }
+        result = {
+            "trace_id": trace_id,
+            "response_mode": response_mode,
+            "run_status": run_status,
+            "planner_status": planner_status,
+            "planner_diagnostics": planner_diagnostics,
+            "openclaw_status": openclaw_status,
+            "project": project_ref,
+            "prompt_role": {
+                "role_key": str(planner_role.get("role_key") or ""),
+                "role_name": str(planner_role.get("role_name") or ""),
+            },
+            "instruction": instruction,
+            "dry_run": bool(dry_run),
+            "summary": str(plan.get("summary") or "").strip(),
+            "confidence": self._to_float(plan.get("confidence"), default=0.0),
+            "task_summary": {"total": 0, "active": 0, "done": 0, "blocked": 0, "sampled": 0},
+            "member_count": 0,
+            "planned_operations": len(plan.get("operations") or []),
+            "execution": execution,
+            "pending_workers": [],
+        }
+        resolved_project_id = self._to_int(project_ref.get("id")) if isinstance(project_ref, dict) else None
+        if persist_run and resolved_project_id is not None and resolved_project_id > 0:
+            run_service = AIOrchestrationRunService(self.db)
+            await run_service.create_run(
+                tenant_id=tenant_id,
+                trace_id=trace_id,
+                run_type="openclaw_dispatch",
+                project_id=int(resolved_project_id),
+                signal_id=None,
+                response_mode=response_mode,
+                status=run_status,
+                planner_status=planner_status,
+                openclaw_status=openclaw_status,
+                worker_total=0,
+                worker_completed=0,
+                worker_failed=0,
+                summary=str(result.get("summary") or "").strip(),
+                planner_diagnostics=planner_diagnostics,
+                result_payload=dict(result),
+                created_by_user_id=int(operator_user.id) if getattr(operator_user, "id", None) else None,
+                completed_at=_utc_now_naive() if run_status in {"completed", "partial", "failed"} else None,
+            )
+            await self.db.commit()
+        return result
+
+    @staticmethod
+    def _build_project_snapshot(project: Project) -> SimpleNamespace:
+        return SimpleNamespace(
+            id=int(project.id),
+            owner_id=int(project.owner_id) if getattr(project, "owner_id", None) else None,
+            title=str(project.title or "").strip(),
+            name=str(project.name or "").strip(),
+            status=str(project.status or "").strip(),
+            assigned_role=str(project.assigned_role or "").strip(),
+            description=str(project.description or "").strip(),
+        )
+
+    @staticmethod
+    def _build_global_project_snapshot() -> SimpleNamespace:
+        return SimpleNamespace(
+            id=0,
+            owner_id=None,
+            title="",
+            name="",
+            status="",
+            assigned_role="",
+            description="",
+        )
+
+    @staticmethod
+    def _serialize_project_directory_item(project: Project) -> dict[str, Any]:
+        return {
+            "project_id": int(project.id),
+            "title": str(project.title or project.name or "").strip(),
+            "status": str(project.status or "").strip(),
+        }
+
+    async def _plan_operations(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Project,
+        tasks: list[dict[str, Any]],
+        task_summary: dict[str, Any],
+        members: list[dict[str, Any]],
+        instruction: str,
+        dry_run: bool,
+        max_operations: int,
+        response_mode: str,
+        project_directory: Optional[list[dict[str, Any]]] = None,
+    ) -> tuple[dict[str, Any], str, dict[str, Any]]:
+        command_payload = {
+            "instruction": instruction,
+            "operator_user_id": int(operator_user.id),
+            "operator_name": str(operator_user.name or operator_user.username or operator_user.id),
+            "operator_role": str(operator_user.role or ""),
+            "dry_run": bool(dry_run),
+            "max_operations": int(max_operations),
+        }
+        project_payload = {
+            "scope": "project" if self._to_int(project.id) and int(project.id) > 0 else "global",
+            "project_id": int(project.id) if self._to_int(project.id) and int(project.id) > 0 else None,
+            "title": str(project.title or project.name or "").strip(),
+            "description": str(project.description or "").strip()[:400],
+            "status": str(project.status or "").strip(),
+            "assigned_role": str(project.assigned_role or "").strip(),
+            "task_summary": task_summary,
+            "candidate_projects": project_directory or [],
+        }
+        operation_catalog_payload = [
+            {
+                "operation_code": item["operation_code"],
+                "description": item["description"],
+                "allowed_fields": sorted(
+                    list(
+                        (
+                            (item.get("input_schema") or {}).get("properties") or {}
+                        ).keys()
+                    )
+                ),
+            }
+            for item in self._commander_operation_catalog()
+        ]
+
+        prompt = PROMPT_OPENCLAW_COMMAND_DISPATCH.format(
+            command_json=json.dumps(command_payload, ensure_ascii=False),
+            project_json=json.dumps(project_payload, ensure_ascii=False),
+            tasks_json=json.dumps(tasks, ensure_ascii=False),
+            members_json=json.dumps(members, ensure_ascii=False),
+            operation_catalog_json=json.dumps(operation_catalog_payload, ensure_ascii=False),
+            max_operations=int(max_operations),
+        )
+
+        try:
+            extraction = await llm_service.generate_structured_object(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="openclaw_command_dispatch",
+                use_prompt_profile=False,
+                temperature=0.0,
+                max_tokens=520,
+                response_mode=response_mode,
+            )
+            diagnostics = self._build_planner_diagnostics(extraction)
+            if extraction.payload is None:
+                fallback = await self._build_fallback_plan(
+                    tenant_id=tenant_id,
+                    project_id=self._to_int(project.id),
+                    instruction=instruction,
+                    max_operations=max_operations,
+                    error=str(extraction.error_code or "structured_unknown"),
+                    project_directory=project_directory,
+                )
+                return fallback, "fallback", diagnostics
+            plan = extraction.payload
+            normalized_ops = self._normalize_operations(plan.get("operations"), max_operations=max_operations)
+            return (
+                {
+                    "summary": str(plan.get("summary") or "").strip() or "OpenClaw 已生成操作计划。",
+                    "confidence": self._to_float(plan.get("confidence"), default=0.0),
+                    "operations": normalized_ops,
+                },
+                "planned",
+                diagnostics,
+            )
+        except Exception as exc:
+            fallback = await self._build_fallback_plan(
+                tenant_id=tenant_id,
+                project_id=self._to_int(project.id),
+                instruction=instruction,
+                max_operations=max_operations,
+                error=str(exc),
+                project_directory=project_directory,
+            )
+            return (
+                fallback,
+                "fallback",
+                {
+                    "used_fallback": True,
+                    "error_code": "structured_empty",
+                    "source_field": None,
+                    "provider": "",
+                    "model": "",
+                },
+            )
+
+    async def _build_fallback_plan(
+        self,
+        *,
+        tenant_id: int,
+        project_id: Optional[int],
+        instruction: str,
+        max_operations: int,
+        error: str,
+        project_directory: Optional[list[dict[str, Any]]] = None,
+    ) -> dict[str, Any]:
+        instruction_text = str(instruction or "").strip()
+        text = instruction_text.lower()
+        project_ops = await self._build_project_fallback_operations(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            instruction=instruction_text,
+            max_operations=max_operations,
+            project_directory=project_directory,
+        )
+        if project_ops:
+            return {
+                "summary": "OpenClaw 规划失败，已回退到项目命令最小执行闭环。",
+                "confidence": 0.58,
+                "operations": project_ops,
+            }
+        if self._looks_like_signal_source_instruction(instruction_text):
+            if project_id is None:
+                return {
+                    "summary": "当前信号源指令缺少项目上下文，系统未自动执行。",
+                    "confidence": 0.0,
+                    "operations": [],
+                }
+            operations = await self._build_signal_source_fallback_operations(
+                tenant_id=tenant_id,
+                project_id=int(project_id),
+                instruction=instruction_text,
+                max_operations=max_operations,
+            )
+            if operations:
+                return {
+                    "summary": "OpenClaw 规划失败，已回退为信号源接入最小闭环。",
+                    "confidence": 0.28,
+                    "operations": operations,
+                }
+            return {
+                "summary": "OpenClaw 未生成可安全执行的信号源接入动作，等待人工补充上下文。",
+                "confidence": 0.0,
+                "operations": [],
+            }
+
+        unassigned_only = any(token in text for token in ("未分配", "没人负责", "unassigned"))
+        rebalance = any(token in text for token in ("重平衡", "优化负载", "重新分配", "调度"))
+        assign_like = rebalance or any(token in text for token in ("分配", "负责人", "指派", "接管"))
+
+        if not assign_like:
+            return {
+                "summary": "当前指令缺少足够上下文，系统未执行任何动作。",
+                "confidence": 0.0,
+                "operations": [],
+            }
+        if project_id is None:
+            return {
+                "summary": "当前指令缺少项目上下文，系统未自动执行分配动作。",
+                "confidence": 0.0,
+                "operations": [],
+            }
+
+        proposals = await workload_service.build_ai_optimize_preview(
+            self.db,
+            tenant_id=tenant_id,
+            project_id=int(project_id),
+            max_moves=max_operations,
+            unassigned_only=unassigned_only and not rebalance,
+        )
+        operations = [
+            {
+                "operation_code": "task.assign",
+                "payload": {
+                    "task_id": int(item.get("task_id")),
+                    "assignee_id": int(item.get("new_assignee_id")),
+                },
+                "reason": str(item.get("reason") or "fallback_workload_optimizer"),
+            }
+            for item in proposals[:max_operations]
+            if self._to_int(item.get("task_id")) and self._to_int(item.get("new_assignee_id"))
+        ]
+
+        if operations:
+            return {
+                "summary": "OpenClaw 规划失败，已回退到内置负载优化器生成分配计划。",
+                "confidence": 0.35,
+                "operations": operations,
+            }
+        return {
+            "summary": "系统未生成可安全执行的分配动作，等待补充更明确的负责人或任务上下文。",
+            "confidence": 0.0,
+            "operations": [],
+        }
+
+    @staticmethod
+    def _build_planner_diagnostics(extraction: Any) -> dict[str, Any]:
+        return {
+            "used_fallback": extraction.error_code is not None,
+            "error_code": extraction.error_code,
+            "source_field": extraction.source_field or None,
+            "provider": str(extraction.provider or ""),
+            "model": str(extraction.model or ""),
+        }
+
+    @staticmethod
+    def _resolve_openclaw_status(*, execution: dict[str, Any], dry_run: bool) -> str:
+        if dry_run:
+            return "planned"
+        failed_count = int(execution.get("failed_count") or 0)
+        applied_count = int(execution.get("applied_count") or 0)
+        skipped_count = int(execution.get("skipped_count") or 0)
+        if failed_count > 0 and applied_count == 0 and skipped_count == 0:
+            return "failed"
+        if failed_count > 0:
+            return "partial"
+        return "completed"
+
+    @staticmethod
+    def _looks_like_signal_source_instruction(text: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(text or ""))
+        if len(normalized) < 6:
+            return False
+        if "信号源" in normalized and any(token in normalized for token in ("新增", "增加", "添加", "接入", "配置", "注册", "建立")):
+            return True
+        return bool(SOURCE_ONBOARDING_FALLBACK_RE.search(normalized))
+
+    async def _build_signal_source_fallback_operations(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        instruction: str,
+        max_operations: int,
+    ) -> list[dict[str, Any]]:
+        if max_operations <= 0:
+            return []
+        project = await self._load_project(tenant_id=tenant_id, project_id=project_id)
+        if project is None:
+            return []
+
+        source_name = self._extract_signal_source_name(instruction)
+        charter = await self._load_latest_aligned_charter(tenant_id=tenant_id, project_id=project_id)
+        recipient_id = await self._resolve_signal_source_notification_recipient(
+            tenant_id=tenant_id,
+            owner_id=int(project.owner_id) if project.owner_id else None,
+        )
+        operations: list[dict[str, Any]] = []
+
+        if charter is not None and len(operations) < max_operations:
+            operations.append(
+                {
+                    "operation_code": "task.create",
+                    "payload": {
+                        "project_id": int(project_id),
+                        "title": f"接入{source_name}信号源"[:120],
+                        "description": self._build_signal_source_task_description(
+                            source_name=source_name,
+                            instruction=instruction,
+                        ),
+                        "priority": "high",
+                    },
+                    "reason": "fallback_signal_source_onboarding",
+                }
+            )
+
+        if recipient_id is not None and len(operations) < max_operations:
+            operations.append(
+                {
+                    "operation_code": "notification.create",
+                    "payload": {
+                        "user_id": int(recipient_id),
+                        "title": f"{source_name}信号源接入待确认"[:120],
+                        "content": self._build_signal_source_notification_content(
+                            source_name=source_name,
+                            instruction=instruction,
+                            charter_ready=charter is not None,
+                        ),
+                    },
+                    "reason": "fallback_signal_source_onboarding",
+                }
+            )
+
+        return operations[:max_operations]
+
+    @staticmethod
+    def _extract_signal_source_name(text: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip()
+        if not cleaned:
+            return "未命名"
+        repo_match = re.search(r"\b([A-Za-z0-9_.-]+/[A-Za-z0-9_.-]+)\b", cleaned)
+        if repo_match:
+            return repo_match.group(1)[:80]
+        quoted_match = re.search(r"[\"'“](.+?)[\"'”]", cleaned)
+        if quoted_match:
+            candidate = quoted_match.group(1).strip(" []【】()（）:：\"'“”")
+            if candidate:
+                return candidate[:80]
+        for pattern in (
+            r"(?:新增|增加|添加|接入|配置|注册|建立)(?:一个|1个|一条|一组|新的|一个新的)?\s*(.+?)(?:信号源|来源|监控源)",
+            r"(?:信号源|来源|监控源)(?:为|是|叫|命名为)?\s*(.+)",
+        ):
+            match = re.search(pattern, cleaned, re.IGNORECASE)
+            if not match:
+                continue
+            candidate = re.sub(r"[，。,.;；:：].*$", "", match.group(1)).strip(" []【】()（）:：\"'“”")
+            candidate = re.sub(
+                r"^(?:一个|1个|一条|一组|一个新的|新的|新增的)\s+",
+                "",
+                candidate,
+                flags=re.IGNORECASE,
+            )
+            candidate = re.sub(r"(?:信号源|来源|监控源|source|channel)$", "", candidate, flags=re.IGNORECASE).strip()
+            if candidate:
+                return candidate[:80]
+        lower = cleaned.lower()
+        if "github" in lower:
+            return "GitHub"
+        if "飞书" in cleaned or "lark" in lower:
+            return "Feishu"
+        if "webhook" in lower:
+            return "Webhook"
+        return cleaned[:80]
+
+    @staticmethod
+    def _build_signal_source_task_description(*, source_name: str, instruction: str) -> str:
+        return (
+            f"为信号源“{source_name}”完成最小接入任务。"
+            "包含接入配置、权限校验、数据字段确认和回流验证。"
+            f"原始指令：{instruction[:400]}"
+        )[:5000]
+
+    @staticmethod
+    def _build_signal_source_notification_content(*, source_name: str, instruction: str, charter_ready: bool) -> str:
+        suffix = (
+            "已自动生成接入任务，请确认负责人和接入边界。"
+            if charter_ready
+            else "当前项目尚未有 aligned charter，系统只先发出确认通知，避免无宪章自动建任务。"
+        )
+        return f"OpenClaw 已识别到新增信号源请求：{source_name}。{suffix} 原始指令：{instruction[:300]}".strip()
+
+    async def _resolve_signal_source_notification_recipient(
+        self,
+        *,
+        tenant_id: int,
+        owner_id: Optional[int],
+    ) -> Optional[int]:
+        if owner_id is not None:
+            owner = (
+                await self.db.execute(
+                    select(User.id).where(
+                        and_(
+                            User.tenant_id == tenant_id,
+                            User.id == int(owner_id),
+                            User.is_active.is_(True),
+                            User.status == "active",
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if owner is not None:
+                return int(owner)
+
+        manager = (
+            await self.db.execute(
+                select(User.id).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.is_active.is_(True),
+                        User.status == "active",
+                        User.role.in_(["BOSS", "ADMIN"]),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+            )
+        ).scalar_one_or_none()
+        return int(manager) if manager is not None else None
+
+    def _commander_operation_catalog(self) -> list[dict[str, Any]]:
+        items: list[dict[str, Any]] = []
+        for code in SUPPORTED_COMMANDER_OPERATION_CODES:
+            row = operation_catalog_service.get_operation(code)
+            if row is not None:
+                items.append(row)
+        return items
+
+    async def _load_project(self, *, tenant_id: int, project_id: int) -> Optional[Project]:
+        return (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == tenant_id,
+                        Project.id == project_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _load_project_directory(self, *, tenant_id: int, limit: int = 20) -> list[dict[str, Any]]:
+        rows = (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == tenant_id,
+                        Project.status != "inactive",
+                    )
+                ).order_by(Project.updated_at.desc(), Project.id.desc()).limit(limit)
+            )
+        ).scalars().all()
+        return [self._serialize_project_directory_item(row) for row in rows]
+
+    async def _build_project_fallback_operations(
+        self,
+        *,
+        tenant_id: int,
+        project_id: Optional[int],
+        instruction: str,
+        max_operations: int,
+        project_directory: Optional[list[dict[str, Any]]],
+    ) -> list[dict[str, Any]]:
+        if max_operations <= 0:
+            return []
+        title = self._extract_project_create_title(instruction)
+        if title:
+            payload: dict[str, Any] = {"title": title}
+            description = self._extract_project_create_description(instruction=instruction, title=title)
+            if description:
+                payload["description"] = description
+            return [
+                {
+                    "operation_code": "project.create",
+                    "payload": payload,
+                    "reason": "explicit_project_create",
+                }
+            ]
+
+        resolved_project_id = project_id or await self._resolve_project_id_from_instruction(
+            tenant_id=tenant_id,
+            instruction=instruction,
+            project_directory=project_directory,
+        )
+        if resolved_project_id is None:
+            return []
+
+        status_operation = self._build_project_status_fallback_operation(
+            project_id=int(resolved_project_id),
+            instruction=instruction,
+        )
+        return [status_operation] if status_operation else []
+
+    async def _resolve_project_id_from_instruction(
+        self,
+        *,
+        tenant_id: int,
+        instruction: str,
+        project_directory: Optional[list[dict[str, Any]]],
+    ) -> Optional[int]:
+        directory = project_directory or await self._load_project_directory(tenant_id=tenant_id, limit=50)
+        normalized = re.sub(r"\s+", "", str(instruction or "")).lower()
+        for item in directory:
+            title = re.sub(r"\s+", "", str(item.get("title") or "")).lower()
+            project_id = self._to_int(item.get("project_id"))
+            if project_id is None or not title or len(title) < 2:
+                continue
+            if title in normalized:
+                return int(project_id)
+        return None
+
+    @staticmethod
+    def _normalize_project_title_candidate(candidate: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(candidate or "")).strip(" []【】()（）:：\"'“”")
+        if not cleaned:
+            return ""
+        normalized = re.sub(r"\s+", "", cleaned).lower()
+        generic_tokens = {
+            "项目",
+            "一个项目",
+            "新项目",
+            "项目列表",
+            "项目列表里面",
+            "项目库",
+            "项目台账",
+            "项目主数据",
+            "pmflow",
+            "pmflow项目列表",
+        }
+        if normalized in generic_tokens:
+            return ""
+        if normalized.startswith(("在项目列表", "到项目列表", "进项目列表", "在pmflow", "到pmflow", "进pmflow")):
+            return ""
+        return compact_title(cleaned, max_len=24)
+
+    @staticmethod
+    def _extract_project_create_title(text: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip()
+        if not cleaned:
+            return ""
+        quoted_match = re.search(r"(?:项目|新项目)[^\"“”']{0,10}[\"“”']([^\"“”']{2,40})[\"“”']", cleaned)
+        if quoted_match:
+            return OpenClawCommanderService._normalize_project_title_candidate(quoted_match.group(1))
+        for pattern in (
+            r"(?:帮我|请|麻烦)?(?:创建|新建|建立|开个|开一个|建个|建一个|做个|做一个)(?:一个|个)?项目(?:叫|名为|名称是|名称为|是|为)?[：:，,\s]*(.+)",
+            r"(?:创建|新建|建立)项目[：:，,\s]*(.+)",
+            r"项目(?:叫|名为|名称是|名称为)[：:，,\s]*(.+)",
+        ):
+            match = re.search(pattern, cleaned, re.IGNORECASE)
+            if not match:
+                continue
+            candidate = re.sub(r"[，。,.;；!?？].*$", "", match.group(1)).strip(" []【】()（）:：\"'“”")
+            candidate = re.sub(r"^(?:一个|个|新项目)\s*", "", candidate, flags=re.IGNORECASE)
+            if candidate:
+                return OpenClawCommanderService._normalize_project_title_candidate(candidate)
+        return ""
+
+    @staticmethod
+    def _extract_project_create_description(*, instruction: str, title: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(instruction or "")).strip()
+        if not cleaned:
+            return ""
+        if title:
+            cleaned = cleaned.replace(title, "", 1).strip(" ，,:：")
+        cleaned = re.sub(
+            r"^(?:帮我|请|麻烦)?(?:创建|新建|建立|开个|开一个|建个|建一个|做个|做一个)(?:一个|个)?项目(?:叫|名为|名称是|名称为|是|为)?",
+            "",
+            cleaned,
+            flags=re.IGNORECASE,
+        ).strip(" ，,:：")
+        if not cleaned:
+            return ""
+        return cleaned[:4000]
+
+    @staticmethod
+    def _build_project_status_fallback_operation(*, project_id: int, instruction: str) -> Optional[dict[str, Any]]:
+        normalized = re.sub(r"\s+", "", str(instruction or "")).lower()
+        if any(token in normalized for token in ("归档", "archive", "停用项目", "项目归档", "状态改成archived", "状态改为archived")):
+            return {
+                "operation_code": "project.archive",
+                "payload": {"project_id": int(project_id)},
+                "reason": "explicit_project_archive",
+            }
+        if any(
+            token in normalized
+            for token in (
+                "重新打开",
+                "重开",
+                "恢复active",
+                "恢复为active",
+                "恢复项目",
+                "激活项目",
+                "设为active",
+                "设成active",
+                "改成active",
+                "改为active",
+                "状态改成active",
+                "状态改为active",
+                "状态设为active",
+            )
+        ):
+            return {
+                "operation_code": "project.reopen",
+                "payload": {"project_id": int(project_id)},
+                "reason": "explicit_project_reopen",
+            }
+        if any(
+            token in normalized
+            for token in (
+                "设为done",
+                "设成done",
+                "改成done",
+                "改为done",
+                "状态改成done",
+                "状态改为done",
+                "状态设为done",
+                "标记完成",
+            )
+        ):
+            return {
+                "operation_code": "project.update",
+                "payload": {"project_id": int(project_id), "status": "done"},
+                "reason": "explicit_project_done",
+            }
+        return None
+
+    @staticmethod
+    def _extract_project_reference_from_execution(execution: dict[str, Any]) -> Optional[dict[str, Any]]:
+        results = execution.get("results") if isinstance(execution, dict) else None
+        if not isinstance(results, list):
+            return None
+        for row in results:
+            if not isinstance(row, dict):
+                continue
+            if str(row.get("entity_type") or "").strip() != "project":
+                continue
+            payload = row.get("result") if isinstance(row.get("result"), dict) else {}
+            project_id = payload.get("project_id") or row.get("entity_id")
+            title = payload.get("title") or payload.get("project_title") or ""
+            status = payload.get("status") or ""
+            if project_id is None:
+                continue
+            return {
+                "id": str(project_id),
+                "title": str(title or "").strip(),
+                "status": str(status or "").strip(),
+            }
+        return None
+
+    async def _load_latest_aligned_charter(self, *, tenant_id: int, project_id: int) -> Optional[TruthNode]:
+        return (
+            await self.db.execute(
+                select(TruthNode).where(
+                    and_(
+                        TruthNode.tenant_id == tenant_id,
+                        TruthNode.project_id == int(project_id),
+                        TruthNode.node_type == "charter",
+                        TruthNode.status.in_(tuple(ALIGNED_STATUSES)),
+                    )
+                ).order_by(TruthNode.updated_at.desc(), TruthNode.id.desc())
+            )
+        ).scalars().first()
+
+    async def _load_project_tasks(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        members: list[dict[str, Any]],
+    ) -> tuple[list[dict[str, Any]], dict[str, Any]]:
+        task_rows = (
+            await self.db.execute(
+                select(Task).where(
+                    and_(
+                        Task.tenant_id == tenant_id,
+                        Task.project_id == project_id,
+                    )
+                ).order_by(Task.id.asc())
+            )
+        ).scalars().all()
+
+        member_name_map = {int(item["user_id"]): str(item.get("name") or "") for item in members if self._to_int(item.get("user_id"))}
+        active_tasks = [task for task in task_rows if str(task.status or "").strip().lower() not in DONE_TASK_STATUSES]
+        selected_tasks = active_tasks[:40] if active_tasks else task_rows[:20]
+        serialized = [
+            {
+                "task_id": int(task.id),
+                "title": str(task.title or "").strip(),
+                "status": str(task.status or "").strip().lower(),
+                "priority": str(task.priority or "").strip().lower(),
+                "assignee_id": int(task.assignee_id) if task.assignee_id else None,
+                "assignee_name": member_name_map.get(int(task.assignee_id), "") if task.assignee_id else "",
+                "required_skill_tag_ids": [int(item) for item in (task.required_skill_tag_ids or []) if self._to_int(item)],
+                "estimated_hours": float(task.estimated_hours or 0.0),
+            }
+            for task in selected_tasks
+        ]
+        summary = {
+            "total": len(task_rows),
+            "active": len(active_tasks),
+            "done": len([task for task in task_rows if str(task.status or "").strip().lower() in DONE_TASK_STATUSES]),
+            "blocked": len([task for task in task_rows if str(task.status or "").strip().lower() == "blocked"]),
+            "sampled": len(serialized),
+        }
+        return serialized, summary
+
+    async def _load_project_team_context(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        owner_id: Optional[int],
+    ) -> list[dict[str, Any]]:
+        member_ids_result = await self.db.execute(
+            select(UserProject.user_id).where(
+                and_(
+                    UserProject.tenant_id == tenant_id,
+                    UserProject.project_id == project_id,
+                )
+            )
+        )
+        member_ids = {int(uid) for uid in member_ids_result.scalars().all()}
+        if owner_id:
+            member_ids.add(int(owner_id))
+        if not member_ids:
+            manager_result = await self.db.execute(
+                select(User.id).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.status == "active",
+                        User.is_active.is_(True),
+                        User.role.in_(["BOSS", "ADMIN"]),
+                    )
+                ).limit(5)
+            )
+            member_ids = {int(uid) for uid in manager_result.scalars().all()}
+        if not member_ids:
+            return []
+
+        users = (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id.in_(sorted(member_ids)),
+                        User.status == "active",
+                        User.is_active.is_(True),
+                    )
+                )
+            )
+        ).scalars().all()
+        if not users:
+            return []
+
+        user_ids = [int(user.id) for user in users]
+        workload_rows = (
+            await self.db.execute(
+                select(UserWorkload).where(
+                    and_(
+                        UserWorkload.tenant_id == tenant_id,
+                        UserWorkload.user_id.in_(user_ids),
+                    )
+                )
+            )
+        ).scalars().all()
+        workload_map = {int(item.user_id): item for item in workload_rows}
+
+        skill_rows = (
+            await self.db.execute(
+                select(
+                    UserSkillBinding.user_id,
+                    SkillTag.id,
+                    SkillTag.name,
+                    SkillTag.slug,
+                    UserSkillBinding.proficiency,
+                )
+                .join(SkillTag, SkillTag.id == UserSkillBinding.skill_tag_id)
+                .where(
+                    and_(
+                        UserSkillBinding.tenant_id == tenant_id,
+                        UserSkillBinding.user_id.in_(user_ids),
+                        SkillTag.tenant_id == tenant_id,
+                    )
+                )
+            )
+        ).all()
+
+        skill_map: dict[int, list[dict[str, Any]]] = {}
+        for uid, sid, name, slug, proficiency in skill_rows:
+            skill_map.setdefault(int(uid), []).append(
+                {
+                    "skill_id": int(sid),
+                    "name": str(name or "").strip(),
+                    "slug": str(slug or "").strip(),
+                    "proficiency": int(proficiency or 0),
+                }
+            )
+
+        members: list[dict[str, Any]] = []
+        for user in users:
+            uid = int(user.id)
+            workload_row = workload_map.get(uid)
+            member_skills = skill_map.get(uid, [])
+            members.append(
+                {
+                    "user_id": uid,
+                    "name": str(user.name or user.username or uid),
+                    "role": str(user.role or ""),
+                    "active": bool(user.is_active and str(user.status or "").strip().lower() == "active"),
+                    "workload_ratio": round(
+                        float(workload_row.workload_ratio if workload_row is not None else float(user.workload or 0.0)),
+                        2,
+                    ),
+                    "current_assigned_hours": round(
+                        float(workload_row.current_assigned_hours or 0.0) if workload_row else 0.0,
+                        2,
+                    ),
+                    "max_available_hours_per_week": round(
+                        float(workload_row.max_available_hours_per_week or 40.0) if workload_row else 40.0,
+                        2,
+                    ),
+                    "skill_tag_ids": [int(item["skill_id"]) for item in member_skills],
+                    "skill_names": [str(item["name"]) for item in member_skills],
+                }
+            )
+        members.sort(key=lambda item: (float(item.get("workload_ratio") or 0.0), int(item.get("user_id") or 0)))
+        return members
+
+    @classmethod
+    def _parse_plan_response(cls, raw_response: str) -> dict[str, Any]:
+        content = (raw_response or "").strip()
+        if not content:
+            raise ValueError("empty llm response")
+        if content.startswith("```"):
+            parts = content.split("```")
+            if len(parts) >= 2:
+                content = parts[1].strip()
+            if content.startswith("json"):
+                content = content[4:].strip()
+        try:
+            parsed = json.loads(content)
+            if isinstance(parsed, dict):
+                return parsed
+        except Exception:
+            pass
+        object_text = cls._extract_first_json_object(content)
+        if not object_text:
+            raise ValueError("llm json not found")
+        parsed = json.loads(object_text)
+        if not isinstance(parsed, dict):
+            raise ValueError("llm json is not object")
+        return parsed
+
+    @classmethod
+    def _extract_first_json_object(cls, text: str) -> Optional[str]:
+        raw = str(text or "")
+        start = raw.find("{")
+        if start < 0:
+            return None
+        depth = 0
+        in_str = False
+        escaped = False
+        for idx in range(start, len(raw)):
+            ch = raw[idx]
+            if in_str:
+                if escaped:
+                    escaped = False
+                elif ch == "\\":
+                    escaped = True
+                elif ch == '"':
+                    in_str = False
+                continue
+            if ch == '"':
+                in_str = True
+                continue
+            if ch == "{":
+                depth += 1
+            elif ch == "}":
+                depth -= 1
+                if depth == 0:
+                    return raw[start : idx + 1]
+        return None
+
+    def _normalize_operations(self, raw_operations: Any, *, max_operations: int) -> list[dict[str, Any]]:
+        if not isinstance(raw_operations, list):
+            return []
+        normalized: list[dict[str, Any]] = []
+        for item in raw_operations:
+            if not isinstance(item, dict):
+                continue
+            operation_code = str(item.get("operation_code") or "").strip().lower()
+            payload = item.get("payload")
+            if not operation_code or not isinstance(payload, dict):
+                continue
+            normalized.append(
+                {
+                    "operation_code": operation_code,
+                    "payload": dict(payload),
+                    "reason": str(item.get("reason") or "").strip()[:300],
+                }
+            )
+            if len(normalized) >= max_operations:
+                break
+        return normalized
+
+    @staticmethod
+    def _to_int(value: Any) -> Optional[int]:
+        if isinstance(value, bool):
+            return None
+        try:
+            return int(str(value).strip())
+        except Exception:
+            return None
+
+    @staticmethod
+    def _to_float(value: Any, *, default: float = 0.0) -> float:
+        try:
+            return float(value)
+        except Exception:
+            return float(default)
+
+
+class OperationExecutorService:
+    """对白名单操作做 schema 校验并执行。"""
+
+    def __init__(self, db: AsyncSession) -> None:
+        self.db = db
+        self._project_member_cache: dict[int, set[int]] = {}
+
+    async def execute_operations(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        operations: list[dict[str, Any]],
+        dry_run: bool,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        seen_idempotency_keys: set[str] = set()
+        results: list[dict[str, Any]] = []
+        applied_count = 0
+        skipped_count = 0
+        failed_count = 0
+        context_project_id = self._project_context_id(project)
+
+        for index, item in enumerate(operations, start=1):
+            operation_code = str(item.get("operation_code") or "").strip().lower()
+            raw_payload = item.get("payload")
+            reason = str(item.get("reason") or "").strip()
+            persistent_claim: OperationExecutionLedger | None = None
+
+            try:
+                payload = self._validate_and_normalize_payload(operation_code=operation_code, payload=raw_payload)
+                idempotency_key = self._build_idempotency_key(
+                    operation_code=operation_code,
+                    payload=payload,
+                    tenant_id=tenant_id,
+                    project_id=context_project_id,
+                )
+                if idempotency_key in seen_idempotency_keys:
+                    skipped_count += 1
+                    results.append(
+                        {
+                            "index": index,
+                            "operation_code": operation_code,
+                            "payload": payload,
+                            "idempotency_key": idempotency_key,
+                            "status": "skipped",
+                            "message": "duplicate_in_request",
+                        }
+                    )
+                    continue
+                seen_idempotency_keys.add(idempotency_key)
+
+                if dry_run:
+                    results.append(
+                        {
+                            "index": index,
+                            "operation_code": operation_code,
+                            "payload": payload,
+                            "idempotency_key": idempotency_key,
+                            "status": "planned",
+                            "message": reason or "dry_run",
+                        }
+                    )
+                    continue
+
+                persistent_claim, duplicate_result = await self._claim_persistent_idempotency(
+                    tenant_id=tenant_id,
+                    project_id=context_project_id if context_project_id > 0 else None,
+                    operation_code=operation_code,
+                    payload=payload,
+                    idempotency_key=idempotency_key,
+                    trace_id=trace_id,
+                )
+                if duplicate_result is not None:
+                    skipped_count += 1
+                    results.append(
+                        {
+                            "index": index,
+                            "operation_code": operation_code,
+                            "payload": payload,
+                            "idempotency_key": idempotency_key,
+                            **duplicate_result,
+                        }
+                    )
+                    continue
+
+                result = await self._execute_single_operation(
+                    tenant_id=tenant_id,
+                    operator_user=operator_user,
+                    project=project,
+                    operation_code=operation_code,
+                    payload=payload,
+                    reason=reason,
+                    trace_id=trace_id,
+                    idempotency_key=idempotency_key,
+                    persistent_claim=persistent_claim,
+                )
+                status = str(result.get("status") or "applied")
+                if status == "applied":
+                    applied_count += 1
+                elif status == "skipped":
+                    skipped_count += 1
+                results.append(
+                    {
+                        "index": index,
+                        "operation_code": operation_code,
+                        "payload": payload,
+                        "idempotency_key": idempotency_key,
+                        **result,
+                    }
+                )
+            except Exception as exc:
+                await self._safe_rollback()
+                failed_count += 1
+                results.append(
+                    {
+                        "index": index,
+                        "operation_code": operation_code,
+                        "payload": raw_payload if isinstance(raw_payload, dict) else {},
+                        "status": "failed",
+                        "message": str(exc),
+                    }
+                )
+
+        return {
+            "dry_run": bool(dry_run),
+            "requested_count": len(operations),
+            "applied_count": applied_count,
+            "skipped_count": skipped_count,
+            "failed_count": failed_count,
+            "results": results,
+        }
+
+    async def _claim_persistent_idempotency(
+        self,
+        *,
+        tenant_id: int,
+        project_id: Optional[int],
+        operation_code: str,
+        payload: dict[str, Any],
+        idempotency_key: str,
+        trace_id: str,
+    ) -> tuple[Optional[OperationExecutionLedger], Optional[dict[str, Any]]]:
+        if operation_code not in PERSISTENT_IDEMPOTENCY_OPERATION_CODES:
+            return None, None
+
+        claim = OperationExecutionLedger(
+            tenant_id=tenant_id,
+            operation_code=operation_code,
+            project_id=int(project_id) if project_id is not None and int(project_id) > 0 else None,
+            idempotency_key=idempotency_key,
+            trace_id=trace_id,
+            status="processing",
+            payload=dict(payload),
+        )
+        self.db.add(claim)
+        try:
+            await self.db.flush()
+            return claim, None
+        except IntegrityError:
+            await self._safe_rollback()
+            existing = await self._load_execution_claim(
+                tenant_id=tenant_id,
+                operation_code=operation_code,
+                idempotency_key=idempotency_key,
+            )
+            if existing is None:
+                raise
+            if str(existing.status or "").strip().lower() == "failed":
+                await self.db.delete(existing)
+                await self.db.flush()
+                return await self._claim_persistent_idempotency(
+                    tenant_id=tenant_id,
+                    project_id=project_id,
+                    operation_code=operation_code,
+                    payload=payload,
+                    idempotency_key=idempotency_key,
+                    trace_id=trace_id,
+                )
+            return (
+                None,
+                {
+                    "status": "skipped",
+                    "message": "duplicate_persistent"
+                    if str(existing.status or "").strip().lower() == "applied"
+                    else "duplicate_in_progress",
+                    "entity_type": str(existing.entity_type or "").strip() or None,
+                    "entity_id": str(existing.entity_id or "").strip() or None,
+                    "trace_id": str(existing.trace_id or "").strip() or None,
+                },
+            )
+
+    async def _load_execution_claim(
+        self,
+        *,
+        tenant_id: int,
+        operation_code: str,
+        idempotency_key: str,
+    ) -> Optional[OperationExecutionLedger]:
+        return (
+            await self.db.execute(
+                select(OperationExecutionLedger).where(
+                    and_(
+                        OperationExecutionLedger.tenant_id == tenant_id,
+                        OperationExecutionLedger.operation_code == operation_code,
+                        OperationExecutionLedger.idempotency_key == idempotency_key,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _safe_rollback(self) -> None:
+        try:
+            await self.db.rollback()
+        except Exception:
+            return
+
+    def _validate_and_normalize_payload(self, *, operation_code: str, payload: Any) -> dict[str, Any]:
+        row = operation_catalog_service.get_operation(operation_code)
+        if row is None or operation_code not in SUPPORTED_COMMANDER_OPERATION_CODES:
+            raise ValueError(f"unsupported operation_code: {operation_code}")
+        if not isinstance(payload, dict):
+            raise ValueError(f"{operation_code} payload must be object")
+
+        input_schema = row.get("input_schema") if isinstance(row.get("input_schema"), dict) else {}
+        properties = input_schema.get("properties") if isinstance(input_schema.get("properties"), dict) else {}
+        required = [str(item) for item in (input_schema.get("required") or [])]
+        normalized: dict[str, Any] = {}
+
+        if input_schema.get("additionalProperties") is False:
+            extra_keys = sorted(set(payload.keys()) - set(properties.keys()))
+            if extra_keys:
+                raise ValueError(f"{operation_code} payload has unsupported fields: {', '.join(extra_keys)}")
+
+        for field, spec in properties.items():
+            if field not in payload:
+                continue
+            normalized[field] = self._coerce_value(field=field, value=payload[field], spec=spec if isinstance(spec, dict) else {})
+
+        missing = [field for field in required if field not in normalized]
+        if missing:
+            raise ValueError(f"{operation_code} missing required fields: {', '.join(missing)}")
+        return normalized
+
+    @classmethod
+    def _coerce_value(cls, *, field: str, value: Any, spec: dict[str, Any]) -> Any:
+        value_type = str(spec.get("type") or "").strip().lower()
+        if value_type == "integer":
+            parsed = cls._to_int(value)
+            if parsed is None:
+                raise ValueError(f"{field} must be integer")
+            return parsed
+        if value_type == "number":
+            if isinstance(value, bool):
+                raise ValueError(f"{field} must be number")
+            try:
+                parsed_float = float(value)
+            except Exception as exc:
+                raise ValueError(f"{field} must be number") from exc
+            minimum = spec.get("minimum")
+            maximum = spec.get("maximum")
+            if minimum is not None and parsed_float < float(minimum):
+                raise ValueError(f"{field} below minimum")
+            if maximum is not None and parsed_float > float(maximum):
+                raise ValueError(f"{field} above maximum")
+            return parsed_float
+        if value_type == "string":
+            text = str(value or "").strip()
+            if not text:
+                raise ValueError(f"{field} must be non-empty string")
+            enum_values = spec.get("enum")
+            if isinstance(enum_values, list) and enum_values:
+                enum_map = {str(item).strip().lower(): str(item) for item in enum_values}
+                matched = enum_map.get(text.lower())
+                if matched is None:
+                    raise ValueError(f"{field} must be one of {', '.join(str(item) for item in enum_values)}")
+                text = matched
+            min_length = spec.get("minLength")
+            max_length = spec.get("maxLength")
+            if min_length is not None and len(text) < int(min_length):
+                raise ValueError(f"{field} shorter than minLength")
+            if max_length is not None and len(text) > int(max_length):
+                raise ValueError(f"{field} longer than maxLength")
+            return text
+        if value_type == "array":
+            if not isinstance(value, list):
+                raise ValueError(f"{field} must be array")
+            max_items = spec.get("maxItems")
+            if max_items is not None and len(value) > int(max_items):
+                raise ValueError(f"{field} too many items")
+            item_spec = spec.get("items") if isinstance(spec.get("items"), dict) else {}
+            return [
+                cls._coerce_value(field=field, value=item, spec=item_spec)
+                for item in value
+            ]
+        return value
+
+    async def _execute_single_operation(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        operation_code: str,
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+        idempotency_key: str,
+        persistent_claim: Optional[OperationExecutionLedger],
+    ) -> dict[str, Any]:
+        if operation_code == "project.create":
+            return await self._execute_project_create(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "project.update":
+            return await self._execute_project_update(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "project.archive":
+            return await self._execute_project_archive(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "project.reopen":
+            return await self._execute_project_reopen(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "task.assign":
+            return await self._execute_task_assign(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "task.create":
+            return await self._execute_task_create(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "task.prioritize":
+            return await self._execute_task_prioritize(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "task.status_transition":
+            return await self._execute_task_status_transition(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        if operation_code == "notification.create":
+            return await self._execute_notification_create(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+                idempotency_key=idempotency_key,
+                persistent_claim=persistent_claim,
+            )
+        if operation_code == "personnel.workload_update":
+            return await self._execute_personnel_workload_update(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project=project,
+                payload=payload,
+                reason=reason,
+                trace_id=trace_id,
+            )
+        raise ValueError(f"unsupported operation_code: {operation_code}")
+
+    async def _execute_project_create(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        title = compact_title(str(payload["title"]), max_len=24)
+        if not title:
+            raise ValueError("project title required")
+        existing = await self._load_project_by_title(tenant_id=tenant_id, title=title)
+        if existing is not None:
+            return {
+                "status": "skipped",
+                "message": "project_already_exists",
+                "entity_type": "project",
+                "entity_id": str(existing.id),
+                "result": {
+                    "project_id": int(existing.id),
+                    "title": str(existing.title or existing.name or "").strip(),
+                    "status": str(existing.status or "").strip(),
+                },
+            }
+
+        owner_id = self._to_int(payload.get("owner_id")) or int(operator_user.id)
+        if owner_id is not None:
+            owner = await self._load_active_user(tenant_id=tenant_id, user_id=int(owner_id))
+            if owner is None:
+                owner_id = int(operator_user.id)
+
+        await ensure_project_id_sequence(self.db)
+        project = Project(
+            tenant_id=tenant_id,
+            name=title,
+            title=title,
+            description=str(payload.get("description") or "").strip()[:4000] or None,
+            owner_id=int(owner_id) if owner_id is not None else None,
+            status="active",
+            domain_id=None,
+            assigned_role="ai_engineer",
+            signal_count=0,
+        )
+        self.db.add(project)
+        await self.db.flush()
+        await ensure_project_default_members(
+            self.db,
+            tenant_id=tenant_id,
+            project_id=int(project.id),
+            owner_id=int(project.owner_id) if project.owner_id is not None else None,
+            assigned_role=project.assigned_role,
+        )
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(project.id),
+            entity_type="project",
+            entity_id=str(project.id),
+            action="project.create",
+            before_data=None,
+            after_data={
+                "title": project.title,
+                "status": project.status,
+                "owner_id": int(project.owner_id) if project.owner_id is not None else None,
+            },
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+        await self.db.commit()
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        return {
+            "status": "applied",
+            "message": f"project created #{int(project.id)}",
+            "entity_type": "project",
+            "entity_id": str(project.id),
+            "result": {
+                "project_id": int(project.id),
+                "title": str(project.title or project.name or "").strip(),
+                "status": str(project.status or "").strip(),
+            },
+        }
+
+    async def _execute_project_update(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        target_project = await self._load_project_for_operation(
+            tenant_id=tenant_id,
+            project=project,
+            payload_project_id=int(payload["project_id"]),
+            operation_code="project.update",
+        )
+        before = {
+            "title": str(target_project.title or target_project.name or "").strip(),
+            "description": str(target_project.description or "").strip(),
+            "status": str(target_project.status or "").strip(),
+        }
+        changed = False
+        if "title" in payload:
+            next_title = compact_title(str(payload.get("title") or ""), max_len=24)
+            if not next_title:
+                raise ValueError("project title required")
+            if next_title != str(target_project.title or target_project.name or "").strip():
+                target_project.title = next_title
+                target_project.name = next_title
+                changed = True
+        if "description" in payload:
+            next_description = str(payload.get("description") or "").strip()[:4000] or None
+            if next_description != (target_project.description or None):
+                target_project.description = next_description
+                changed = True
+        if "status" in payload:
+            next_status = self._normalize_project_status(payload.get("status"))
+            if next_status != str(target_project.status or "").strip().lower():
+                target_project.status = next_status
+                changed = True
+        if not changed:
+            return {
+                "status": "skipped",
+                "message": "already_in_target_state",
+                "entity_type": "project",
+                "entity_id": str(target_project.id),
+                "result": {
+                    "project_id": int(target_project.id),
+                    "title": str(target_project.title or target_project.name or "").strip(),
+                    "status": str(target_project.status or "").strip(),
+                },
+            }
+
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(target_project.id),
+            entity_type="project",
+            entity_id=str(target_project.id),
+            action="project.update",
+            before_data=before,
+            after_data={
+                "title": str(target_project.title or target_project.name or "").strip(),
+                "description": str(target_project.description or "").strip(),
+                "status": str(target_project.status or "").strip(),
+            },
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+        await self.db.commit()
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        return {
+            "status": "applied",
+            "message": f"project updated #{int(target_project.id)}",
+            "entity_type": "project",
+            "entity_id": str(target_project.id),
+            "result": {
+                "project_id": int(target_project.id),
+                "title": str(target_project.title or target_project.name or "").strip(),
+                "status": str(target_project.status or "").strip(),
+            },
+        }
+
+    async def _execute_project_archive(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        return await self._execute_project_update(
+            tenant_id=tenant_id,
+            operator_user=operator_user,
+            project=project,
+            payload={"project_id": int(payload["project_id"]), "status": "archived"},
+            reason=reason,
+            trace_id=trace_id,
+        )
+
+    async def _execute_project_reopen(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        return await self._execute_project_update(
+            tenant_id=tenant_id,
+            operator_user=operator_user,
+            project=project,
+            payload={"project_id": int(payload["project_id"]), "status": "active"},
+            reason=reason,
+            trace_id=trace_id,
+        )
+
+    async def _execute_task_create(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        project = self._require_project_context(project=project, operation_code="task.create")
+        project_id = int(payload["project_id"])
+        if project_id != int(project.id):
+            raise ValueError("task.create project_id mismatch")
+
+        charter = await self._load_latest_aligned_charter(tenant_id=tenant_id, project_id=project_id)
+        if charter is None:
+            raise ValueError("aligned charter required before task.create")
+
+        assignee_id = self._to_int(payload.get("assignee_id"))
+        touched_user_ids: list[int] = []
+        if assignee_id is not None:
+            assignee = await self._load_active_user(tenant_id=tenant_id, user_id=assignee_id)
+            if assignee is None:
+                raise ValueError("assignee not found or inactive")
+            member_ids = await self._get_project_member_ids(
+                tenant_id=tenant_id,
+                project_id=project_id,
+                owner_id=int(project.owner_id) if project.owner_id else None,
+            )
+            if int(assignee.id) not in member_ids:
+                raise ValueError("assignee is not project member")
+            touched_user_ids.append(int(assignee.id))
+
+        due_date = self._parse_due_date_value(payload.get("due_date"))
+        required_skill_tag_ids = self._safe_int_list(payload.get("required_skill_tag_ids"))
+        task = await persist_task(
+            self.db,
+            tenant_id=tenant_id,
+            truth_node_id=int(charter.id),
+            project_id=project_id,
+            draft=build_task_create_draft(
+                title=str(payload["title"]),
+                description=str(payload.get("description") or "").strip() or None,
+                assignee_id=assignee_id,
+                required_skill_tag_ids=required_skill_tag_ids,
+                priority=payload.get("priority"),
+                status="todo",
+                estimated_hours=self._to_float(payload.get("estimated_hours"), default=0.0) if payload.get("estimated_hours") is not None else None,
+                due_date=due_date,
+            ),
+            actor_user_id=int(operator_user.id),
+            action="task.create",
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+            commit=not touched_user_ids,
+            refresh_workload=bool(touched_user_ids),
+        )
+
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        for user_id in touched_user_ids:
+            await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=user_id)
+        self._sync_task_rag(task=task, tenant_id=tenant_id)
+
+        return {
+            "status": "applied",
+            "message": f"task created #{int(task.id)}",
+            "entity_type": "task",
+            "entity_id": str(task.id),
+        }
+
+    async def _execute_task_assign(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        project = self._require_project_context(project=project, operation_code="task.assign")
+        task = await self._load_task(tenant_id=tenant_id, task_id=int(payload["task_id"]))
+        if task is None or int(task.project_id) != int(project.id):
+            raise ValueError("task not found in target project")
+
+        assignee = await self._load_active_user(tenant_id=tenant_id, user_id=int(payload["assignee_id"]))
+        if assignee is None:
+            raise ValueError("assignee not found or inactive")
+
+        member_ids = await self._get_project_member_ids(tenant_id=tenant_id, project_id=int(project.id), owner_id=int(project.owner_id) if project.owner_id else None)
+        if int(assignee.id) not in member_ids:
+            raise ValueError("assignee is not project member")
+
+        current_assignee_id = int(task.assignee_id) if task.assignee_id else None
+        if current_assignee_id == int(assignee.id):
+            return {
+                "status": "skipped",
+                "message": "already_assigned",
+                "entity_type": "task",
+                "entity_id": str(task.id),
+            }
+
+        before = {
+            "assignee_id": current_assignee_id,
+            "status": task.status,
+        }
+        task.assignee_id = int(assignee.id)
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(project.id),
+            entity_type="task",
+            entity_id=str(task.id),
+            action="task.assign",
+            before_data=before,
+            after_data={"assignee_id": int(assignee.id), "status": task.status},
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+
+        touched_user_ids: list[int] = []
+        if current_assignee_id:
+            touched_user_ids.append(current_assignee_id)
+        touched_user_ids.append(int(assignee.id))
+        for uid in touched_user_ids:
+            await workload_service.refresh_user_assigned_hours_from_tasks(self.db, tenant_id=tenant_id, user_id=uid)
+
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        for uid in sorted(set(touched_user_ids)):
+            await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=uid)
+        self._sync_task_rag(task=task, tenant_id=tenant_id)
+
+        return {
+            "status": "applied",
+            "message": f"task assigned to user#{int(assignee.id)}",
+            "entity_type": "task",
+            "entity_id": str(task.id),
+        }
+
+    async def _execute_task_prioritize(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        project = self._require_project_context(project=project, operation_code="task.prioritize")
+        task = await self._load_task(tenant_id=tenant_id, task_id=int(payload["task_id"]))
+        if task is None or int(task.project_id) != int(project.id):
+            raise ValueError("task not found in target project")
+
+        before_priority = self._normalize_task_priority(task.priority)
+        next_priority = self._normalize_task_priority(payload.get("priority"))
+        if before_priority == next_priority:
+            return {
+                "status": "skipped",
+                "message": "already_in_target_priority",
+                "entity_type": "task",
+                "entity_id": str(task.id),
+            }
+
+        task.priority = next_priority
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(project.id),
+            entity_type="task",
+            entity_id=str(task.id),
+            action="task.prioritize",
+            before_data={"priority": before_priority},
+            after_data={"priority": next_priority},
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+        await self.db.commit()
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        if task.assignee_id:
+            await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=int(task.assignee_id))
+        self._sync_task_rag(task=task, tenant_id=tenant_id)
+
+        return {
+            "status": "applied",
+            "message": f"task priority updated to {next_priority}",
+            "entity_type": "task",
+            "entity_id": str(task.id),
+        }
+
+    async def _execute_task_status_transition(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        project = self._require_project_context(project=project, operation_code="task.status_transition")
+        task = await self._load_task(tenant_id=tenant_id, task_id=int(payload["task_id"]))
+        if task is None or int(task.project_id) != int(project.id):
+            raise ValueError("task not found in target project")
+
+        status = str(payload["status"]).strip().lower()
+        if status not in SUPPORTED_TASK_STATUSES:
+            raise ValueError("unsupported task status")
+        before_status = str(task.status or "").strip().lower()
+        if before_status == status:
+            return {
+                "status": "skipped",
+                "message": "already_in_target_status",
+                "entity_type": "task",
+                "entity_id": str(task.id),
+            }
+
+        task.status = status
+        if status in DONE_TASK_STATUSES:
+            task.completed_at = datetime.now(timezone.utc).replace(tzinfo=None)
+
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(project.id),
+            entity_type="task",
+            entity_id=str(task.id),
+            action="task.status_transition",
+            before_data={"status": before_status},
+            after_data={"status": status},
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+
+        if task.assignee_id:
+            await workload_service.refresh_user_assigned_hours_from_tasks(
+                self.db,
+                tenant_id=tenant_id,
+                user_id=int(task.assignee_id),
+            )
+        else:
+            await self.db.commit()
+
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        if task.assignee_id:
+            await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=int(task.assignee_id))
+        self._sync_task_rag(task=task, tenant_id=tenant_id)
+
+        if status in DONE_TASK_STATUSES and before_status not in DONE_TASK_STATUSES and task.assignee_id:
+            asyncio.create_task(
+                process_task_completion(
+                    task_id=int(task.id),
+                    tenant_id=tenant_id,
+                    operator_user_id=int(operator_user.id),
+                )
+            )
+
+        return {
+            "status": "applied",
+            "message": f"task status updated to {status}",
+            "entity_type": "task",
+            "entity_id": str(task.id),
+        }
+
+    async def _execute_notification_create(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+        idempotency_key: str,
+        persistent_claim: Optional[OperationExecutionLedger],
+    ) -> dict[str, Any]:
+        project = self._require_project_context(project=project, operation_code="notification.create")
+        user = await self._load_active_user(tenant_id=tenant_id, user_id=int(payload["user_id"]))
+        if user is None:
+            raise ValueError("notification target user not found or inactive")
+
+        notification = await NotificationService(self.db).create_notification(
+            tenant_id=tenant_id,
+            user_id=int(user.id),
+            category="system",
+            title=str(payload["title"]).strip(),
+            content=str(payload["content"]).strip(),
+            source="openclaw_commander",
+            entity_type="project",
+            entity_id=int(project.id),
+            target_url=f"/projects/{int(project.id)}/board",
+            commit=False,
+        )
+        if persistent_claim is not None:
+            persistent_claim.status = "applied"
+            persistent_claim.entity_type = "notification"
+            persistent_claim.entity_id = str(notification.id)
+            persistent_claim.result_payload = {
+                "status": "applied",
+                "notification_id": int(notification.id),
+                "idempotency_key": idempotency_key,
+            }
+            persistent_claim.error_message = None
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(project.id),
+            entity_type="notification",
+            entity_id=str(notification.id),
+            action="notification.create",
+            before_data=None,
+            after_data={
+                "user_id": int(user.id),
+                "title": notification.title,
+                "source": notification.source,
+            },
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+        await self.db.commit()
+
+        return {
+            "status": "applied",
+            "message": f"notification created for user#{int(user.id)}",
+            "entity_type": "notification",
+            "entity_id": str(notification.id),
+        }
+
+    async def _execute_personnel_workload_update(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: User,
+        project: Optional[Project | SimpleNamespace],
+        payload: dict[str, Any],
+        reason: str,
+        trace_id: str,
+    ) -> dict[str, Any]:
+        project = self._require_project_context(project=project, operation_code="personnel.workload_update")
+        user = await self._load_active_user(tenant_id=tenant_id, user_id=int(payload["user_id"]))
+        if user is None:
+            raise ValueError("target user not found or inactive")
+
+        member_ids = await self._get_project_member_ids(
+            tenant_id=tenant_id,
+            project_id=int(project.id),
+            owner_id=int(project.owner_id) if project.owner_id else None,
+        )
+        if int(user.id) not in member_ids:
+            raise ValueError("target user is not project member")
+
+        before_workload = await self._load_user_workload(tenant_id=tenant_id, user_id=int(user.id))
+        result = await workload_service.update_user_max_weekly_hours(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(user.id),
+            max_weekly_hours=float(payload["max_weekly_hours"]),
+        )
+        if result is None:
+            raise ValueError("target user workload not found")
+
+        await log_operation(
+            self.db,
+            tenant_id=tenant_id,
+            user_id=int(operator_user.id),
+            project_id=int(project.id),
+            entity_type="user_workload",
+            entity_id=str(user.id),
+            action="personnel.workload_update",
+            before_data=before_workload,
+            after_data=result,
+            remark=f"{trace_id}:{reason}"[:1000] if reason else trace_id,
+        )
+        await self.db.commit()
+        await context_cache_service.invalidate_system_context(tenant_id=tenant_id)
+        await context_cache_service.invalidate_user_context(tenant_id=tenant_id, user_id=int(user.id))
+
+        return {
+            "status": "applied",
+            "message": f"user#{int(user.id)} max weekly hours updated",
+            "entity_type": "user_workload",
+            "entity_id": str(user.id),
+        }
+
+    async def _load_task(self, *, tenant_id: int, task_id: int) -> Optional[Task]:
+        return (
+            await self.db.execute(
+                select(Task).where(
+                    and_(
+                        Task.tenant_id == tenant_id,
+                        Task.id == int(task_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _load_project_by_title(self, *, tenant_id: int, title: str) -> Optional[Project]:
+        normalized = compact_title(str(title or ""), max_len=24)
+        if not normalized:
+            return None
+        return (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == tenant_id,
+                        or_(
+                            func.lower(Project.title) == normalized.lower(),
+                            func.lower(Project.name) == normalized.lower(),
+                        ),
+                        Project.status != "inactive",
+                    )
+                ).limit(1)
+            )
+        ).scalar_one_or_none()
+
+    async def _load_project_for_operation(
+        self,
+        *,
+        tenant_id: int,
+        project: Optional[Project | SimpleNamespace],
+        payload_project_id: int,
+        operation_code: str,
+    ) -> Project:
+        context_project_id = self._project_context_id(project)
+        if context_project_id > 0 and context_project_id != int(payload_project_id):
+            raise ValueError(f"{operation_code} project_id mismatch")
+        row = await self._load_project(tenant_id=tenant_id, project_id=int(payload_project_id))
+        if row is None:
+            raise ValueError("project not found")
+        return row
+
+    async def _load_active_user(self, *, tenant_id: int, user_id: int) -> Optional[User]:
+        return (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id == int(user_id),
+                        User.is_active.is_(True),
+                        User.status == "active",
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _load_user_workload(self, *, tenant_id: int, user_id: int) -> Optional[dict[str, Any]]:
+        row = (
+            await self.db.execute(
+                select(UserWorkload).where(
+                    and_(
+                        UserWorkload.tenant_id == tenant_id,
+                        UserWorkload.user_id == int(user_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return {
+            "user_id": str(user_id),
+            "current_assigned_hours": float(row.current_assigned_hours or 0.0),
+            "max_weekly_hours": float(row.max_available_hours_per_week or 40.0),
+            "workload_ratio": float(row.workload_ratio or 0.0),
+        }
+
+    async def _get_project_member_ids(self, *, tenant_id: int, project_id: int, owner_id: Optional[int]) -> set[int]:
+        cached = self._project_member_cache.get(int(project_id))
+        if cached is not None:
+            return set(cached)
+        rows = await self.db.execute(
+            select(UserProject.user_id).where(
+                and_(
+                    UserProject.tenant_id == tenant_id,
+                    UserProject.project_id == int(project_id),
+                )
+            )
+        )
+        member_ids = {int(uid) for uid in rows.scalars().all()}
+        if owner_id:
+            member_ids.add(int(owner_id))
+        self._project_member_cache[int(project_id)] = set(member_ids)
+        return member_ids
+
+    async def _load_latest_aligned_charter(self, *, tenant_id: int, project_id: int) -> Optional[TruthNode]:
+        return (
+            await self.db.execute(
+                select(TruthNode).where(
+                    and_(
+                        TruthNode.tenant_id == tenant_id,
+                        TruthNode.project_id == int(project_id),
+                        TruthNode.node_type == "charter",
+                        TruthNode.status.in_(tuple(ALIGNED_STATUSES)),
+                    )
+                ).order_by(TruthNode.updated_at.desc(), TruthNode.id.desc())
+            )
+        ).scalars().first()
+
+    @staticmethod
+    def _project_context_id(project: Optional[Project | SimpleNamespace]) -> int:
+        if project is None:
+            return 0
+        try:
+            return max(0, int(getattr(project, "id", 0) or 0))
+        except Exception:
+            return 0
+
+    def _require_project_context(
+        self,
+        *,
+        project: Optional[Project | SimpleNamespace],
+        operation_code: str,
+    ) -> Project | SimpleNamespace:
+        if self._project_context_id(project) <= 0:
+            raise ValueError(f"{operation_code} requires project context")
+        if project is None:
+            raise ValueError(f"{operation_code} requires project context")
+        return project
+
+    def _build_idempotency_key(
+        self,
+        *,
+        operation_code: str,
+        payload: dict[str, Any],
+        tenant_id: int,
+        project_id: int,
+    ) -> str:
+        row = operation_catalog_service.get_operation(operation_code)
+        fields = []
+        if isinstance(row, dict):
+            fields = [str(item) for item in ((row.get("idempotency_key") or {}).get("fields") or [])]
+
+        parts: dict[str, Any] = {}
+        for field in fields:
+            if field == "tenant_id":
+                parts[field] = int(tenant_id)
+            elif field == "project_id":
+                parts[field] = int(project_id)
+            elif field == "content_hash":
+                parts[field] = hashlib.sha256(str(payload.get("content") or "").strip().encode("utf-8")).hexdigest()[:16]
+            else:
+                parts[field] = payload.get(field)
+        encoded = json.dumps(parts, ensure_ascii=False, sort_keys=True, default=str)
+        return hashlib.sha256(f"{operation_code}|{encoded}".encode("utf-8")).hexdigest()[:24]
+
+    def _sync_task_rag(self, *, task: Task, tenant_id: int) -> None:
+        try:
+            content = (
+                (task.description or "")
+                + f"\n状态: {task.status}"
+                + f"\n负责人: {int(task.assignee_id) if task.assignee_id else '未分配'}"
+            )
+            asyncio.create_task(
+                AdaptiveRAGService.upsert_pmflow_record_async(
+                    tenant_id=tenant_id,
+                    record_type="task",
+                    record_id=int(task.id),
+                    project_id=int(task.project_id),
+                    title=str(task.title or "任务"),
+                    content=content,
+                )
+            )
+        except Exception:
+            return
+
+    @staticmethod
+    def _to_int(value: Any) -> Optional[int]:
+        if isinstance(value, bool):
+            return None
+        try:
+            return int(str(value).strip())
+        except Exception:
+            return None
+
+    @staticmethod
+    def _safe_int_list(value: Any) -> list[int]:
+        if not isinstance(value, list):
+            return []
+        items: list[int] = []
+        for raw in value:
+            try:
+                items.append(int(raw))
+            except Exception:
+                continue
+        return items
+
+    @classmethod
+    def _normalize_task_priority(cls, value: Any) -> str:
+        text = str(value or "medium").strip().lower()
+        if text in {"p0", "critical", "high"}:
+            return "high"
+        if text in {"p2", "p3", "low"}:
+            return "low"
+        return "medium"
+
+    @staticmethod
+    def _normalize_project_status(value: Any) -> str:
+        text = str(value or "active").strip().lower()
+        if text not in SUPPORTED_PROJECT_STATUSES:
+            raise ValueError("unsupported project status")
+        return text
+
+    @staticmethod
+    def _parse_due_date_value(value: Any) -> Optional[datetime]:
+        text = str(value or "").strip()
+        if not text:
+            return None
+        try:
+            return datetime.fromisoformat(text.replace("Z", "+00:00")).replace(tzinfo=None)
+        except ValueError as exc:
+            raise ValueError("invalid due_date") from exc
diff --git a/backend/app/services/operation_catalog_service.py b/backend/app/services/operation_catalog_service.py
new file mode 100644
index 00000000..ab51847b
--- /dev/null
+++ b/backend/app/services/operation_catalog_service.py
@@ -0,0 +1,89 @@
+"""Operation Catalog 服务。"""
+
+from __future__ import annotations
+
+from collections import Counter
+from typing import Any, Optional
+
+from ..config.operation_catalog import OPERATION_CATALOG, OperationCatalogEntry
+
+
+REQUIRED_OPERATION_CODES = {
+    "project.create",
+    "project.update",
+    "project.archive",
+    "project.reopen",
+    "task.create",
+    "task.update",
+    "task.assign",
+    "task.prioritize",
+    "task.status_transition",
+    "task.delete",
+    "personnel.role_update",
+    "personnel.workload_update",
+    "personnel.availability_update",
+    "knowledge.doc_upsert",
+    "knowledge.chunk_upsert",
+    "knowledge.link_upsert",
+    "knowledge.reindex",
+    "notification.create",
+    "notification.ack",
+    "notification.close",
+    "file.write_whitelist",
+}
+
+
+class OperationCatalogService:
+    def __init__(self) -> None:
+        self._catalog: tuple[OperationCatalogEntry, ...] = OPERATION_CATALOG
+
+    def list_operations(self, *, domain: Optional[str] = None) -> list[dict[str, Any]]:
+        rows = self._catalog
+        if domain:
+            domain_key = str(domain).strip().lower()
+            rows = tuple(
+                row for row in rows
+                if str(row.domain).strip().lower() == domain_key
+            )
+        return [row.model_dump() for row in rows]
+
+    def get_operation(self, operation_code: str) -> Optional[dict[str, Any]]:
+        target = str(operation_code or "").strip().lower()
+        for row in self._catalog:
+            if str(row.operation_code).strip().lower() == target:
+                return row.model_dump()
+        return None
+
+    def lint_catalog(self) -> dict[str, Any]:
+        codes = [row.operation_code for row in self._catalog]
+        domains = [str(row.domain) for row in self._catalog]
+        code_counter = Counter(codes)
+        duplicate_codes = sorted([code for code, count in code_counter.items() if count > 1])
+        missing_required = sorted(REQUIRED_OPERATION_CODES - set(codes))
+
+        invalid_entries: list[str] = []
+        for row in self._catalog:
+            if not row.operation_code:
+                invalid_entries.append("empty_operation_code")
+            if not row.input_schema:
+                invalid_entries.append(f"{row.operation_code}:missing_input_schema")
+            if not row.rbac_policy:
+                invalid_entries.append(f"{row.operation_code}:missing_rbac_policy")
+            if not row.idempotency_key.fields:
+                invalid_entries.append(f"{row.operation_code}:missing_idempotency_fields")
+            if not row.audit_fields:
+                invalid_entries.append(f"{row.operation_code}:missing_audit_fields")
+
+        ok = not duplicate_codes and not missing_required and not invalid_entries
+        return {
+            "ok": ok,
+            "operation_count": len(self._catalog),
+            "domain_count": len(set(domains)),
+            "domains": sorted(set(domains)),
+            "duplicate_codes": duplicate_codes,
+            "missing_required_codes": missing_required,
+            "invalid_entries": invalid_entries,
+        }
+
+
+operation_catalog_service = OperationCatalogService()
diff --git a/backend/app/services/operation_log_service.py b/backend/app/services/operation_log_service.py
new file mode 100644
index 00000000..75819850
--- /dev/null
+++ b/backend/app/services/operation_log_service.py
@@ -0,0 +1,40 @@
+"""操作日志服务"""
+
+from typing import Optional
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import OperationLog
+
+
+async def log_operation(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    user_id: int,
+    project_id: Optional[int],
+    entity_type: str,
+    entity_id: str,
+    action: str,
+    before_data: Optional[dict] = None,
+    after_data: Optional[dict] = None,
+    remark: Optional[str] = None,
+) -> None:
+    """记录操作日志，失败时不影响主流程。"""
+    try:
+        db.add(
+            OperationLog(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                project_id=project_id,
+                entity_type=entity_type,
+                entity_id=entity_id,
+                action=action,
+                before_data=before_data,
+                after_data=after_data,
+                remark=remark,
+            )
+        )
+    except Exception:
+        # 日志失败不阻断主交易
+        return
diff --git a/backend/app/services/plane.py b/backend/app/services/plane.py
new file mode 100644
index 00000000..be9d565f
--- /dev/null
+++ b/backend/app/services/plane.py
@@ -0,0 +1,14 @@
+"""Plane 服务兼容层（最小可运行实现）。"""
+
+from __future__ import annotations
+
+
+class PlaneService:
+    async def sync_tasks(self, tasks: list[dict], is_urgent: bool = False) -> dict:
+        return {"synced_count": len(tasks or []), "urgent": is_urgent}
+
+    async def update_issue_status(self, issue_key: str, status: str) -> dict:
+        return {"issue_key": issue_key, "status": status}
+
+
+plane_service = PlaneService()
diff --git a/backend/app/services/project_knowledge_graph.py b/backend/app/services/project_knowledge_graph.py
new file mode 100644
index 00000000..2c6ec012
--- /dev/null
+++ b/backend/app/services/project_knowledge_graph.py
@@ -0,0 +1,1769 @@
+"""Project Knowledge Graph service.
+
+第一阶段目标：
+1. 构建统一实体/关系/证据写入能力
+2. 支持结构化数据（project/task/signal/user）同步
+3. 支持代码目录扫描快照
+4. 提供图谱概览统计
+"""
+
+from __future__ import annotations
+
+import hashlib
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from typing import Any, Optional
+
+from sqlalchemy import and_, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import (
+    IngestEvent,
+    IngestOffset,
+    KGEvidence,
+    KGEntity,
+    KGRelation,
+    KGSyncRun,
+    Project,
+    FeishuItem,
+    Signal,
+    SymbolDefinition,
+    SymbolReference,
+    Task,
+    User,
+)
+from .context_cache import context_cache_service
+from .symbol_index_service import SymbolIndexService
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+class ProjectKnowledgeGraphService:
+    def __init__(self, db: AsyncSession, *, tenant_id: int) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+
+    async def upsert_entity(
+        self,
+        *,
+        entity_type: str,
+        entity_key: str,
+        display_name: str,
+        payload: Optional[dict[str, Any]] = None,
+        source_type: str = "system",
+        source_ref: Optional[str] = None,
+    ) -> KGEntity:
+        row = (
+            await self.db.execute(
+                select(KGEntity).where(
+                    and_(
+                        KGEntity.tenant_id == self.tenant_id,
+                        KGEntity.entity_type == str(entity_type).strip()[:40],
+                        KGEntity.entity_key == str(entity_key).strip()[:255],
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+        if row is None:
+            row = KGEntity(
+                tenant_id=self.tenant_id,
+                entity_type=str(entity_type).strip()[:40],
+                entity_key=str(entity_key).strip()[:255],
+                display_name=str(display_name or entity_key).strip()[:255],
+                payload=payload or {},
+                source_type=str(source_type or "system").strip()[:32],
+                source_ref=(str(source_ref).strip()[:255] if source_ref else None),
+                last_seen_at=_utc_now_naive(),
+            )
+            self.db.add(row)
+            await self.db.flush()
+            return row
+
+        row.display_name = str(display_name or row.display_name).strip()[:255]
+        row.payload = payload or {}
+        row.source_type = str(source_type or row.source_type or "system").strip()[:32]
+        row.source_ref = str(source_ref).strip()[:255] if source_ref else row.source_ref
+        row.last_seen_at = _utc_now_naive()
+        await self.db.flush()
+        return row
+
+    async def upsert_relation(
+        self,
+        *,
+        relation_type: str,
+        subject_entity_id: int,
+        object_entity_id: int,
+        weight: float = 1.0,
+        payload: Optional[dict[str, Any]] = None,
+        source_type: str = "system",
+        source_ref: Optional[str] = None,
+    ) -> KGRelation:
+        row = (
+            await self.db.execute(
+                select(KGRelation).where(
+                    and_(
+                        KGRelation.tenant_id == self.tenant_id,
+                        KGRelation.relation_type == str(relation_type).strip()[:40],
+                        KGRelation.subject_entity_id == int(subject_entity_id),
+                        KGRelation.object_entity_id == int(object_entity_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+        if row is None:
+            row = KGRelation(
+                tenant_id=self.tenant_id,
+                relation_type=str(relation_type).strip()[:40],
+                subject_entity_id=int(subject_entity_id),
+                object_entity_id=int(object_entity_id),
+                weight=float(weight),
+                payload=payload or {},
+                source_type=str(source_type or "system").strip()[:32],
+                source_ref=(str(source_ref).strip()[:255] if source_ref else None),
+            )
+            self.db.add(row)
+            await self.db.flush()
+            return row
+
+        row.weight = float(weight)
+        row.payload = payload or {}
+        row.source_type = str(source_type or row.source_type or "system").strip()[:32]
+        row.source_ref = str(source_ref).strip()[:255] if source_ref else row.source_ref
+        await self.db.flush()
+        return row
+
+    async def add_evidence(
+        self,
+        *,
+        evidence_type: str,
+        content_excerpt: str,
+        entity_id: Optional[int] = None,
+        relation_id: Optional[int] = None,
+        source_type: str = "system",
+        source_ref: Optional[str] = None,
+        metadata: Optional[dict[str, Any]] = None,
+    ) -> KGEvidence:
+        clean = str(content_excerpt or "").strip()
+        digest = hashlib.md5(clean.encode("utf-8")).hexdigest()
+
+        row = KGEvidence(
+            tenant_id=self.tenant_id,
+            entity_id=int(entity_id) if entity_id is not None else None,
+            relation_id=int(relation_id) if relation_id is not None else None,
+            evidence_type=str(evidence_type or "snapshot").strip()[:40],
+            content_hash=digest,
+            content_excerpt=clean[:4000] if clean else "(empty)",
+            source_type=str(source_type or "system").strip()[:32],
+            source_ref=(str(source_ref).strip()[:255] if source_ref else None),
+            meta_payload=metadata or {},
+            occurred_at=_utc_now_naive(),
+        )
+        self.db.add(row)
+        await self.db.flush()
+        return row
+
+    async def enqueue_ingest_event(
+        self,
+        *,
+        channel: str,
+        event_key: str,
+        payload: Optional[dict[str, Any]] = None,
+        status: str = "pending",
+    ) -> IngestEvent:
+        row = (
+            await self.db.execute(
+                select(IngestEvent).where(
+                    and_(
+                        IngestEvent.tenant_id == self.tenant_id,
+                        IngestEvent.channel == str(channel).strip()[:32],
+                        IngestEvent.event_key == str(event_key).strip()[:128],
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            row = IngestEvent(
+                tenant_id=self.tenant_id,
+                channel=str(channel).strip()[:32],
+                event_key=str(event_key).strip()[:128],
+                status=str(status or "pending").strip()[:20],
+                payload=payload or {},
+            )
+            self.db.add(row)
+        else:
+            row.status = str(status or row.status or "pending").strip()[:20]
+            row.payload = payload or row.payload or {}
+            row.error_message = None
+        await self.db.flush()
+        return row
+
+    async def set_ingest_offset(
+        self,
+        *,
+        channel: str,
+        offset_key: str,
+        offset_value: str,
+        payload: Optional[dict[str, Any]] = None,
+    ) -> IngestOffset:
+        row = (
+            await self.db.execute(
+                select(IngestOffset).where(
+                    and_(
+                        IngestOffset.tenant_id == self.tenant_id,
+                        IngestOffset.channel == str(channel).strip()[:32],
+                        IngestOffset.offset_key == str(offset_key).strip()[:128],
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+        if row is None:
+            row = IngestOffset(
+                tenant_id=self.tenant_id,
+                channel=str(channel).strip()[:32],
+                offset_key=str(offset_key).strip()[:128],
+                offset_value=str(offset_value or "").strip()[:255],
+                payload=payload or {},
+            )
+            self.db.add(row)
+        else:
+            row.offset_value = str(offset_value or row.offset_value).strip()[:255]
+            row.payload = payload or row.payload or {}
+        await self.db.flush()
+        return row
+
+    async def get_ingest_offset(self, *, channel: str, offset_key: str) -> Optional[str]:
+        row = (
+            await self.db.execute(
+                select(IngestOffset).where(
+                    and_(
+                        IngestOffset.tenant_id == self.tenant_id,
+                        IngestOffset.channel == str(channel).strip()[:32],
+                        IngestOffset.offset_key == str(offset_key).strip()[:128],
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return str(row.offset_value or "").strip() or None
+
+    @staticmethod
+    def _parse_iso_dt(value: Optional[str]) -> Optional[datetime]:
+        raw = str(value or "").strip()
+        if not raw:
+            return None
+        try:
+            dt = datetime.fromisoformat(raw)
+            return dt.replace(tzinfo=None) if dt.tzinfo else dt
+        except Exception:
+            return None
+
+    async def _open_sync_run(
+        self,
+        *,
+        run_type: str,
+        trigger_source: str,
+        input_payload: Optional[dict[str, Any]] = None,
+    ) -> KGSyncRun:
+        now = _utc_now_naive()
+        row = KGSyncRun(
+            tenant_id=self.tenant_id,
+            run_type=str(run_type or "incremental").strip()[:40],
+            trigger_source=str(trigger_source or "system").strip()[:40],
+            status="running",
+            started_at=now,
+            input_payload=input_payload or {},
+            result_payload={},
+        )
+        self.db.add(row)
+        await self.db.flush()
+        await self.db.commit()
+        await self.db.refresh(row)
+        return row
+
+    async def _close_sync_run(
+        self,
+        *,
+        row: Optional[KGSyncRun],
+        status: str,
+        result_payload: Optional[dict[str, Any]] = None,
+        error_message: Optional[str] = None,
+        started_at_hint: Optional[datetime] = None,
+    ) -> None:
+        if row is None:
+            return
+        now = _utc_now_naive()
+        started = started_at_hint or row.started_at or now
+        row.status = str(status or "unknown").strip()[:20]
+        row.ended_at = now
+        row.duration_ms = max(int((now - started).total_seconds() * 1000), 0)
+        row.result_payload = result_payload or {}
+        row.error_message = (str(error_message).strip()[:2000] if error_message else None)
+        await self.db.commit()
+
+    async def list_sync_runs(
+        self,
+        *,
+        limit: int = 50,
+        run_type: Optional[str] = None,
+        status: Optional[str] = None,
+    ) -> dict[str, Any]:
+        max_rows = max(1, min(int(limit), 200))
+        stmt = select(KGSyncRun).where(KGSyncRun.tenant_id == self.tenant_id)
+        if run_type:
+            stmt = stmt.where(KGSyncRun.run_type == str(run_type).strip()[:40])
+        if status:
+            stmt = stmt.where(KGSyncRun.status == str(status).strip()[:20])
+        rows = (
+            await self.db.execute(stmt.order_by(KGSyncRun.started_at.desc(), KGSyncRun.id.desc()).limit(max_rows))
+        ).scalars().all()
+
+        items = []
+        for row in rows:
+            items.append(
+                {
+                    "id": int(row.id),
+                    "run_type": str(row.run_type or ""),
+                    "trigger_source": str(row.trigger_source or ""),
+                    "status": str(row.status or ""),
+                    "started_at": row.started_at.isoformat() if row.started_at else None,
+                    "ended_at": row.ended_at.isoformat() if row.ended_at else None,
+                    "duration_ms": int(row.duration_ms or 0),
+                    "input_payload": row.input_payload or {},
+                    "result_payload": row.result_payload or {},
+                    "error_message": str(row.error_message or ""),
+                }
+            )
+        return {
+            "tenant_id": self.tenant_id,
+            "count": len(items),
+            "items": items,
+        }
+
+    async def get_sync_slo(self, *, lookback_hours: int = 24) -> dict[str, Any]:
+        hours = max(1, min(int(lookback_hours), 24 * 30))
+        now = _utc_now_naive()
+        since = now - timedelta(hours=hours)
+
+        runs = (
+            await self.db.execute(
+                select(KGSyncRun).where(
+                    and_(
+                        KGSyncRun.tenant_id == self.tenant_id,
+                        KGSyncRun.started_at >= since,
+                        KGSyncRun.run_type == "incremental",
+                    )
+                )
+            )
+        ).scalars().all()
+
+        total = len(runs)
+        success = sum(1 for x in runs if str(x.status or "") == "success")
+        partial = sum(1 for x in runs if str(x.status or "") == "partial")
+        failed = sum(1 for x in runs if str(x.status or "") == "failed")
+        avg_duration_ms = (
+            int(sum(int(x.duration_ms or 0) for x in runs) / total)
+            if total > 0
+            else 0
+        )
+
+        latest_success = (
+            await self.db.execute(
+                select(KGSyncRun)
+                .where(
+                    and_(
+                        KGSyncRun.tenant_id == self.tenant_id,
+                        KGSyncRun.run_type == "incremental",
+                        KGSyncRun.status == "success",
+                    )
+                )
+                .order_by(KGSyncRun.started_at.desc(), KGSyncRun.id.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+
+        dead_letter = (
+            await self.db.execute(
+                select(func.count(IngestEvent.id)).where(
+                    and_(
+                        IngestEvent.tenant_id == self.tenant_id,
+                        IngestEvent.status == "dead_letter",
+                    )
+                )
+            )
+        ).scalar_one()
+        pending_events = (
+            await self.db.execute(
+                select(func.count(IngestEvent.id)).where(
+                    and_(
+                        IngestEvent.tenant_id == self.tenant_id,
+                        IngestEvent.status.in_(["pending", "retry"]),
+                    )
+                )
+            )
+        ).scalar_one()
+
+        async def _freshness_minutes(channel: str, offset_key: str) -> Optional[int]:
+            raw = await self.get_ingest_offset(channel=channel, offset_key=offset_key)
+            dt = self._parse_iso_dt(raw)
+            if dt is None:
+                return None
+            return max(int((now - dt).total_seconds() // 60), 0)
+
+        return {
+            "tenant_id": self.tenant_id,
+            "lookback_hours": hours,
+            "window_started_at": since.isoformat(),
+            "window_ended_at": now.isoformat(),
+            "runs_total": total,
+            "runs_success": success,
+            "runs_partial": partial,
+            "runs_failed": failed,
+            "success_rate": round((success / total), 4) if total > 0 else 0.0,
+            "avg_duration_ms": avg_duration_ms,
+            "latest_success_at": latest_success.started_at.isoformat() if latest_success and latest_success.started_at else None,
+            "dead_letter_events": int(dead_letter or 0),
+            "pending_events": int(pending_events or 0),
+            "freshness_minutes": {
+                "db_structured": await _freshness_minutes("db_delta", "last_structured_sync_at"),
+                "feishu_delta": await _freshness_minutes("feishu_delta", "last_feishu_item_updated_at"),
+                "code_scan": await _freshness_minutes("code_scan", "last_code_scan_at"),
+                "symbol_scan": await _freshness_minutes("symbol_scan", "last_symbol_scan_at"),
+                "feature_catalog": await _freshness_minutes("feature_catalog", "last_feature_catalog_sync_at"),
+            },
+        }
+
+    async def sync_structured_delta(self, *, limit: int = 1000) -> dict[str, Any]:
+        max_rows = max(1, min(int(limit), 5000))
+        last_raw = await self.get_ingest_offset(channel="db_delta", offset_key="last_structured_sync_at")
+        last_dt = self._parse_iso_dt(last_raw) or (datetime(1970, 1, 1))
+        counters = {
+            "projects": 0,
+            "tasks": 0,
+            "signals": 0,
+            "users": 0,
+            "relations": 0,
+            "evidences": 0,
+            "from_offset": last_dt.isoformat(),
+        }
+        newest_seen = last_dt
+
+        project_rows = (
+            await self.db.execute(
+                select(Project)
+                .where(
+                    and_(
+                        Project.tenant_id == self.tenant_id,
+                        Project.updated_at >= last_dt,
+                    )
+                )
+                .order_by(Project.updated_at.asc(), Project.id.asc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        project_map: dict[int, int] = {}
+        for p in project_rows:
+            entity = await self.upsert_entity(
+                entity_type="project",
+                entity_key=f"project:{int(p.id)}",
+                display_name=str(p.title or p.name or f"Project#{int(p.id)}"),
+                payload={
+                    "project_id": int(p.id),
+                    "status": str(p.status or "unknown"),
+                    "owner_id": int(p.owner_id) if p.owner_id is not None else None,
+                },
+                source_type="db",
+                source_ref=f"projects:{int(p.id)}",
+            )
+            project_map[int(p.id)] = int(entity.id)
+            counters["projects"] += 1
+            if p.updated_at and p.updated_at > newest_seen:
+                newest_seen = p.updated_at
+
+        user_rows = (
+            await self.db.execute(
+                select(User)
+                .where(
+                    and_(
+                        User.tenant_id == self.tenant_id,
+                        User.updated_at >= last_dt,
+                    )
+                )
+                .order_by(User.updated_at.asc(), User.id.asc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        user_map: dict[int, int] = {}
+        for u in user_rows:
+            entity = await self.upsert_entity(
+                entity_type="user",
+                entity_key=f"user:{int(u.id)}",
+                display_name=str(u.name or f"User#{int(u.id)}"),
+                payload={
+                    "user_id": int(u.id),
+                    "role": str(u.role or "unknown"),
+                    "is_active": bool(u.is_active),
+                },
+                source_type="db",
+                source_ref=f"users:{int(u.id)}",
+            )
+            user_map[int(u.id)] = int(entity.id)
+            counters["users"] += 1
+            if u.updated_at and u.updated_at > newest_seen:
+                newest_seen = u.updated_at
+
+        task_rows = (
+            await self.db.execute(
+                select(Task)
+                .where(
+                    and_(
+                        Task.tenant_id == self.tenant_id,
+                        Task.updated_at >= last_dt,
+                    )
+                )
+                .order_by(Task.updated_at.asc(), Task.id.asc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        for t in task_rows:
+            task_entity = await self.upsert_entity(
+                entity_type="task",
+                entity_key=f"task:{int(t.id)}",
+                display_name=str(t.title or f"Task#{int(t.id)}"),
+                payload={
+                    "task_id": int(t.id),
+                    "project_id": int(t.project_id) if t.project_id is not None else None,
+                    "assignee_id": int(t.assignee_id) if t.assignee_id is not None else None,
+                    "status": str(t.status or "unknown"),
+                },
+                source_type="db",
+                source_ref=f"tasks:{int(t.id)}",
+            )
+            counters["tasks"] += 1
+            if t.updated_at and t.updated_at > newest_seen:
+                newest_seen = t.updated_at
+
+            if t.project_id is not None:
+                target_pid = int(t.project_id)
+                if target_pid not in project_map:
+                    p = (
+                        await self.db.execute(
+                            select(Project).where(
+                                and_(
+                                    Project.tenant_id == self.tenant_id,
+                                    Project.id == target_pid,
+                                )
+                            )
+                        )
+                    ).scalar_one_or_none()
+                    if p is not None:
+                        p_entity = await self.upsert_entity(
+                            entity_type="project",
+                            entity_key=f"project:{int(p.id)}",
+                            display_name=str(p.title or p.name or f"Project#{int(p.id)}"),
+                            payload={
+                                "project_id": int(p.id),
+                                "status": str(p.status or "unknown"),
+                                "owner_id": int(p.owner_id) if p.owner_id is not None else None,
+                            },
+                            source_type="db",
+                            source_ref=f"projects:{int(p.id)}",
+                        )
+                        project_map[int(p.id)] = int(p_entity.id)
+                if target_pid in project_map:
+                    rel = await self.upsert_relation(
+                        relation_type="belongs_to",
+                        subject_entity_id=int(task_entity.id),
+                        object_entity_id=int(project_map[target_pid]),
+                        source_type="db",
+                        source_ref=f"tasks:{int(t.id)}",
+                    )
+                    counters["relations"] += 1
+                    await self.add_evidence(
+                        evidence_type="db_delta",
+                        content_excerpt=f"task#{int(t.id)} belongs_to project#{target_pid}",
+                        relation_id=int(rel.id),
+                        source_type="db",
+                        source_ref=f"tasks:{int(t.id)}",
+                        metadata={"table": "tasks"},
+                    )
+                    counters["evidences"] += 1
+
+            if t.assignee_id is not None:
+                uid = int(t.assignee_id)
+                if uid not in user_map:
+                    u = (
+                        await self.db.execute(
+                            select(User).where(and_(User.tenant_id == self.tenant_id, User.id == uid))
+                        )
+                    ).scalar_one_or_none()
+                    if u is not None:
+                        u_entity = await self.upsert_entity(
+                            entity_type="user",
+                            entity_key=f"user:{uid}",
+                            display_name=str(u.name or f"User#{uid}"),
+                            payload={
+                                "user_id": uid,
+                                "role": str(u.role or "unknown"),
+                                "is_active": bool(u.is_active),
+                            },
+                            source_type="db",
+                            source_ref=f"users:{uid}",
+                        )
+                        user_map[uid] = int(u_entity.id)
+                if uid in user_map:
+                    rel = await self.upsert_relation(
+                        relation_type="assigned_to",
+                        subject_entity_id=int(task_entity.id),
+                        object_entity_id=int(user_map[uid]),
+                        source_type="db",
+                        source_ref=f"tasks:{int(t.id)}",
+                    )
+                    counters["relations"] += 1
+                    await self.add_evidence(
+                        evidence_type="db_delta",
+                        content_excerpt=f"task#{int(t.id)} assigned_to user#{uid}",
+                        relation_id=int(rel.id),
+                        source_type="db",
+                        source_ref=f"tasks:{int(t.id)}",
+                        metadata={"table": "tasks"},
+                    )
+                    counters["evidences"] += 1
+
+        signal_rows = (
+            await self.db.execute(
+                select(Signal)
+                .where(
+                    and_(
+                        Signal.tenant_id == self.tenant_id,
+                        Signal.updated_at >= last_dt,
+                    )
+                )
+                .order_by(Signal.updated_at.asc(), Signal.id.asc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        for s in signal_rows:
+            signal_entity = await self.upsert_entity(
+                entity_type="signal",
+                entity_key=f"signal:{int(s.id)}",
+                display_name=str(s.title or s.content or f"Signal#{int(s.id)}")[:255],
+                payload={
+                    "signal_id": int(s.id),
+                    "priority": str(s.priority or ""),
+                    "status": str(s.status or ""),
+                    "matched_project_id": (
+                        int(s.matched_project_id) if getattr(s, "matched_project_id", None) is not None else None
+                    ),
+                },
+                source_type="db",
+                source_ref=f"signals:{int(s.id)}",
+            )
+            counters["signals"] += 1
+            if s.updated_at and s.updated_at > newest_seen:
+                newest_seen = s.updated_at
+
+            pid = int(s.matched_project_id) if getattr(s, "matched_project_id", None) is not None else None
+            if pid is not None:
+                if pid not in project_map:
+                    p = (
+                        await self.db.execute(
+                            select(Project).where(and_(Project.tenant_id == self.tenant_id, Project.id == pid))
+                        )
+                    ).scalar_one_or_none()
+                    if p is not None:
+                        p_entity = await self.upsert_entity(
+                            entity_type="project",
+                            entity_key=f"project:{pid}",
+                            display_name=str(p.title or p.name or f"Project#{pid}"),
+                            payload={
+                                "project_id": pid,
+                                "status": str(p.status or "unknown"),
+                                "owner_id": int(p.owner_id) if p.owner_id is not None else None,
+                            },
+                            source_type="db",
+                            source_ref=f"projects:{pid}",
+                        )
+                        project_map[pid] = int(p_entity.id)
+                if pid in project_map:
+                    rel = await self.upsert_relation(
+                        relation_type="matches_project",
+                        subject_entity_id=int(signal_entity.id),
+                        object_entity_id=int(project_map[pid]),
+                        source_type="db",
+                        source_ref=f"signals:{int(s.id)}",
+                    )
+                    counters["relations"] += 1
+                    await self.add_evidence(
+                        evidence_type="db_delta",
+                        content_excerpt=f"signal#{int(s.id)} matches project#{pid}",
+                        relation_id=int(rel.id),
+                        source_type="db",
+                        source_ref=f"signals:{int(s.id)}",
+                        metadata={"table": "signals"},
+                    )
+                    counters["evidences"] += 1
+
+        await self.enqueue_ingest_event(
+            channel="db_delta",
+            event_key=f"db_delta:{_utc_now_naive().strftime('%Y%m%d%H%M%S')}",
+            payload=counters,
+            status="done",
+        )
+        await self.set_ingest_offset(
+            channel="db_delta",
+            offset_key="last_structured_sync_at",
+            offset_value=newest_seen.isoformat(),
+            payload={"counters": counters},
+        )
+        await self.db.commit()
+        counters["to_offset"] = newest_seen.isoformat()
+        return counters
+
+    async def enqueue_feishu_delta_events(self, *, limit: int = 500) -> dict[str, Any]:
+        max_rows = max(1, min(int(limit), 2000))
+        last_raw = await self.get_ingest_offset(channel="feishu_delta", offset_key="last_feishu_item_updated_at")
+        last_dt = self._parse_iso_dt(last_raw) or (datetime(1970, 1, 1))
+
+        rows = (
+            await self.db.execute(
+                select(FeishuItem)
+                .where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.updated_at >= last_dt,
+                    )
+                )
+                .order_by(FeishuItem.updated_at.asc(), FeishuItem.id.asc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        enqueued = 0
+        newest_seen = last_dt
+        for item in rows:
+            event_key = str(item.feishu_message_id or f"feishu_item:{int(item.id)}")
+            await self.enqueue_ingest_event(
+                channel="feishu_delta",
+                event_key=event_key[:128],
+                payload={
+                    "feishu_item_id": int(item.id),
+                    "matched_project_id": int(item.matched_project_id) if item.matched_project_id is not None else None,
+                    "level": str(item.level or ""),
+                    "route_status": str(item.route_status or ""),
+                },
+                status="pending",
+            )
+            enqueued += 1
+            if item.updated_at and item.updated_at > newest_seen:
+                newest_seen = item.updated_at
+
+        await self.set_ingest_offset(
+            channel="feishu_delta",
+            offset_key="last_feishu_item_updated_at",
+            offset_value=newest_seen.isoformat(),
+            payload={"batch_size": len(rows)},
+        )
+        await self.db.commit()
+        return {
+            "enqueued": enqueued,
+            "from_offset": last_dt.isoformat(),
+            "to_offset": newest_seen.isoformat(),
+            "candidate_count": len(rows),
+        }
+
+    async def process_pending_events(self, *, max_batch: int = 200, max_retry: int = 3) -> dict[str, Any]:
+        batch = max(1, min(int(max_batch), 1000))
+        retries = max(0, min(int(max_retry), 10))
+
+        rows = (
+            await self.db.execute(
+                select(IngestEvent)
+                .where(
+                    and_(
+                        IngestEvent.tenant_id == self.tenant_id,
+                        IngestEvent.status.in_(["pending", "retry"]),
+                    )
+                )
+                .order_by(IngestEvent.created_at.asc(), IngestEvent.id.asc())
+                .limit(batch)
+            )
+        ).scalars().all()
+
+        stats = {"processed": 0, "done": 0, "retry": 0, "dead_letter": 0}
+        for ev in rows:
+            stats["processed"] += 1
+            try:
+                if str(ev.channel) == "feishu_delta":
+                    await self._process_feishu_event(ev)
+                ev.status = "done"
+                ev.error_message = None
+                ev.processed_at = _utc_now_naive()
+                stats["done"] += 1
+            except Exception as exc:
+                ev.retry_count = int(ev.retry_count or 0) + 1
+                ev.error_message = str(exc)[:1000]
+                if ev.retry_count > retries:
+                    ev.status = "dead_letter"
+                    ev.processed_at = _utc_now_naive()
+                    stats["dead_letter"] += 1
+                else:
+                    ev.status = "retry"
+                    stats["retry"] += 1
+
+        await self.db.commit()
+        return stats
+
+    async def _process_feishu_event(self, event: IngestEvent) -> None:
+        payload = event.payload if isinstance(event.payload, dict) else {}
+        item_id = payload.get("feishu_item_id")
+        if item_id is None:
+            return
+        item = (
+            await self.db.execute(
+                select(FeishuItem).where(
+                    and_(
+                        FeishuItem.tenant_id == self.tenant_id,
+                        FeishuItem.id == int(item_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if item is None:
+            return
+
+        signal_entity = await self.upsert_entity(
+            entity_type="signal",
+            entity_key=f"feishu_item:{int(item.id)}",
+            display_name=str(item.raw_message or f"FeishuItem#{int(item.id)}")[:255],
+            payload={
+                "feishu_item_id": int(item.id),
+                "level": str(item.level or ""),
+                "route_status": str(item.route_status or ""),
+                "matched_project_id": int(item.matched_project_id) if item.matched_project_id is not None else None,
+            },
+            source_type="feishu",
+            source_ref=str(item.feishu_message_id or f"item:{int(item.id)}")[:255],
+        )
+        await self.add_evidence(
+            evidence_type="feishu_delta",
+            content_excerpt=str(item.raw_message or "")[:1000] or f"feishu_item#{int(item.id)}",
+            entity_id=int(signal_entity.id),
+            source_type="feishu",
+            source_ref=str(item.feishu_message_id or f"item:{int(item.id)}")[:255],
+            metadata={"feishu_item_id": int(item.id)},
+        )
+
+        if item.matched_project_id is not None:
+            project_entity = await self.upsert_entity(
+                entity_type="project",
+                entity_key=f"project:{int(item.matched_project_id)}",
+                display_name=f"Project#{int(item.matched_project_id)}",
+                payload={"project_id": int(item.matched_project_id)},
+                source_type="db",
+                source_ref=f"projects:{int(item.matched_project_id)}",
+            )
+            rel = await self.upsert_relation(
+                relation_type="matches_project",
+                subject_entity_id=int(signal_entity.id),
+                object_entity_id=int(project_entity.id),
+                source_type="feishu",
+                source_ref=str(item.feishu_message_id or f"item:{int(item.id)}")[:255],
+            )
+            await self.add_evidence(
+                evidence_type="feishu_delta",
+                content_excerpt=f"feishu_item#{int(item.id)} matches project#{int(item.matched_project_id)}",
+                relation_id=int(rel.id),
+                source_type="feishu",
+                source_ref=str(item.feishu_message_id or f"item:{int(item.id)}")[:255],
+                metadata={"feishu_item_id": int(item.id)},
+            )
+
+    async def sync_codebase_if_due(
+        self,
+        *,
+        base_dir: Optional[str] = None,
+        interval_hours: int = 6,
+    ) -> dict[str, Any]:
+        hours = max(1, int(interval_hours))
+        offset_raw = await self.get_ingest_offset(channel="code_scan", offset_key="last_code_scan_at")
+        last_scan_at = self._parse_iso_dt(offset_raw)
+        now = _utc_now_naive()
+        if last_scan_at is not None and now - last_scan_at < timedelta(hours=hours):
+            return {
+                "executed": False,
+                "reason": "not_due",
+                "last_scan_at": last_scan_at.isoformat(),
+                "next_due_at": (last_scan_at + timedelta(hours=hours)).isoformat(),
+            }
+        payload = await self.sync_codebase_snapshot(base_dir=base_dir)
+        return {
+            "executed": True,
+            "reason": "due",
+            "payload": payload,
+        }
+
+    async def sync_symbol_index_if_due(
+        self,
+        *,
+        base_dir: Optional[str] = None,
+        interval_hours: int = 6,
+        max_files: int = 500,
+    ) -> dict[str, Any]:
+        hours = max(1, int(interval_hours))
+        offset_raw = await self.get_ingest_offset(channel="symbol_scan", offset_key="last_symbol_scan_at")
+        last_scan_at = self._parse_iso_dt(offset_raw)
+        now = _utc_now_naive()
+        if last_scan_at is not None and now - last_scan_at < timedelta(hours=hours):
+            return {
+                "executed": False,
+                "reason": "not_due",
+                "last_scan_at": last_scan_at.isoformat(),
+                "next_due_at": (last_scan_at + timedelta(hours=hours)).isoformat(),
+            }
+
+        symbol_service = SymbolIndexService(self.db, tenant_id=self.tenant_id)
+        payload = await symbol_service.sync_codebase_symbols(
+            base_dir=base_dir,
+            max_files=max_files,
+            force_reindex=False,
+        )
+        symbol_graph = await self._sync_symbol_entities(limit=800)
+        payload["kg_symbol_sync"] = symbol_graph
+        await self.enqueue_ingest_event(
+            channel="symbol_scan",
+            event_key=f"symbol_scan:{_utc_now_naive().strftime('%Y%m%d%H%M%S')}",
+            payload=payload,
+            status="done",
+        )
+        await self.set_ingest_offset(
+            channel="symbol_scan",
+            offset_key="last_symbol_scan_at",
+            offset_value=_utc_now_naive().isoformat(),
+            payload=payload,
+        )
+        await self.db.commit()
+        return {
+            "executed": True,
+            "reason": "due",
+            "payload": payload,
+        }
+
+    async def _sync_symbol_entities(self, *, limit: int = 800) -> dict[str, Any]:
+        max_rows = max(10, min(int(limit), 3000))
+        defs = (
+            await self.db.execute(
+                select(SymbolDefinition)
+                .where(SymbolDefinition.tenant_id == self.tenant_id)
+                .order_by(SymbolDefinition.updated_at.desc(), SymbolDefinition.id.desc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        if not defs:
+            return {"symbols": 0, "modules": 0, "relations": 0, "evidences": 0}
+
+        counters = {"symbols": 0, "modules": 0, "relations": 0, "evidences": 0}
+        module_entity_map: dict[str, int] = {}
+        symbol_entity_map: dict[int, int] = {}
+        name_entity_map: dict[str, int] = {}
+        qn_entity_map: dict[str, int] = {}
+
+        for row in defs:
+            qn = str(row.qualified_name or row.symbol_name or f"symbol#{int(row.id)}")
+            symbol_entity = await self.upsert_entity(
+                entity_type="symbol",
+                entity_key=f"symbol:{int(row.id)}",
+                display_name=qn[:255],
+                payload={
+                    "symbol_id": int(row.id),
+                    "symbol_name": str(row.symbol_name or ""),
+                    "qualified_name": str(row.qualified_name or ""),
+                    "symbol_type": str(row.symbol_type or ""),
+                    "language": str(row.language or ""),
+                    "file_path": str(row.file_path or ""),
+                    "start_line": int(row.start_line) if row.start_line is not None else None,
+                    "end_line": int(row.end_line) if row.end_line is not None else None,
+                },
+                source_type="symbol_index",
+                source_ref=str(row.file_path or "")[:255],
+            )
+            counters["symbols"] += 1
+            symbol_entity_map[int(row.id)] = int(symbol_entity.id)
+            if str(row.symbol_name or "").strip():
+                name_entity_map[str(row.symbol_name).strip()] = int(symbol_entity.id)
+            if str(row.qualified_name or "").strip():
+                qn_entity_map[str(row.qualified_name).strip()] = int(symbol_entity.id)
+
+            file_key = str(row.file_path or "").strip()
+            if not file_key:
+                continue
+            module_entity_id = module_entity_map.get(file_key)
+            if module_entity_id is None:
+                module_entity = await self.upsert_entity(
+                    entity_type="module",
+                    entity_key=f"module:{file_key}",
+                    display_name=file_key[:255],
+                    payload={"file_path": file_key},
+                    source_type="symbol_index",
+                    source_ref=file_key[:255],
+                )
+                module_entity_id = int(module_entity.id)
+                module_entity_map[file_key] = module_entity_id
+                counters["modules"] += 1
+            rel = await self.upsert_relation(
+                relation_type="defined_in",
+                subject_entity_id=int(symbol_entity.id),
+                object_entity_id=int(module_entity_id),
+                source_type="symbol_index",
+                source_ref=file_key[:255],
+                payload={
+                    "symbol_id": int(row.id),
+                    "symbol_type": str(row.symbol_type or ""),
+                    "language": str(row.language or ""),
+                },
+            )
+            counters["relations"] += 1
+            await self.add_evidence(
+                evidence_type="symbol_index",
+                content_excerpt=(
+                    f"{str(row.symbol_name or '')} defined_in {file_key}:"
+                    f"{int(row.start_line) if row.start_line is not None else 0}"
+                ),
+                relation_id=int(rel.id),
+                source_type="symbol_index",
+                source_ref=file_key[:255],
+                metadata={"symbol_id": int(row.id), "file_path": file_key},
+            )
+            counters["evidences"] += 1
+
+        refs = (
+            await self.db.execute(
+                select(SymbolReference)
+                .where(SymbolReference.tenant_id == self.tenant_id)
+                .order_by(SymbolReference.updated_at.desc(), SymbolReference.id.desc())
+                .limit(max_rows * 3)
+            )
+        ).scalars().all()
+        for row in refs:
+            source_entity_id = None
+            if row.source_symbol_id is not None:
+                source_entity_id = symbol_entity_map.get(int(row.source_symbol_id))
+            if source_entity_id is None:
+                continue
+            target_entity_id = None
+            target_qn = str(row.target_qualified_name or "").strip()
+            target_name = str(row.target_symbol_name or "").strip()
+            if target_qn:
+                target_entity_id = qn_entity_map.get(target_qn)
+            if target_entity_id is None and target_name:
+                target_entity_id = name_entity_map.get(target_name)
+            if target_entity_id is None:
+                continue
+            rel_type = str(row.relation_type or "references").strip()
+            rel = await self.upsert_relation(
+                relation_type=rel_type[:40],
+                subject_entity_id=int(source_entity_id),
+                object_entity_id=int(target_entity_id),
+                source_type="symbol_index",
+                source_ref=str(row.file_path or "")[:255],
+                payload={
+                    "symbol_reference_id": int(row.id),
+                    "target_symbol_name": target_name,
+                    "target_qualified_name": target_qn,
+                },
+            )
+            counters["relations"] += 1
+            await self.add_evidence(
+                evidence_type="symbol_reference",
+                content_excerpt=f"{int(source_entity_id)} {rel_type} {target_name or target_qn}",
+                relation_id=int(rel.id),
+                source_type="symbol_index",
+                source_ref=str(row.file_path or "")[:255],
+                metadata={
+                    "symbol_reference_id": int(row.id),
+                    "target_symbol_name": target_name,
+                    "target_qualified_name": target_qn,
+                },
+            )
+            counters["evidences"] += 1
+        return counters
+
+    async def sync_feature_catalog_if_due(
+        self,
+        *,
+        interval_hours: int = 6,
+    ) -> dict[str, Any]:
+        hours = max(1, int(interval_hours))
+        offset_raw = await self.get_ingest_offset(channel="feature_catalog", offset_key="last_feature_catalog_sync_at")
+        last_sync_at = self._parse_iso_dt(offset_raw)
+        now = _utc_now_naive()
+        if last_sync_at is not None and now - last_sync_at < timedelta(hours=hours):
+            return {
+                "executed": False,
+                "reason": "not_due",
+                "last_sync_at": last_sync_at.isoformat(),
+                "next_due_at": (last_sync_at + timedelta(hours=hours)).isoformat(),
+            }
+
+        payload = await self._sync_feature_catalog_entities(limit=300)
+        await self.enqueue_ingest_event(
+            channel="feature_catalog",
+            event_key=f"feature_catalog:{_utc_now_naive().strftime('%Y%m%d%H%M%S')}",
+            payload=payload,
+            status="done",
+        )
+        await self.set_ingest_offset(
+            channel="feature_catalog",
+            offset_key="last_feature_catalog_sync_at",
+            offset_value=_utc_now_naive().isoformat(),
+            payload=payload,
+        )
+        await self.db.commit()
+        return {
+            "executed": True,
+            "reason": "due",
+            "payload": payload,
+        }
+
+    async def sync_feature_catalog_now(self) -> dict[str, Any]:
+        payload = await self._sync_feature_catalog_entities(limit=300)
+        await self.enqueue_ingest_event(
+            channel="feature_catalog",
+            event_key=f"feature_catalog:{_utc_now_naive().strftime('%Y%m%d%H%M%S')}",
+            payload=payload,
+            status="done",
+        )
+        await self.set_ingest_offset(
+            channel="feature_catalog",
+            offset_key="last_feature_catalog_sync_at",
+            offset_value=_utc_now_naive().isoformat(),
+            payload=payload,
+        )
+        await self.db.commit()
+        return payload
+
+    async def _sync_feature_catalog_entities(self, *, limit: int = 300) -> dict[str, Any]:
+        snapshot = await context_cache_service.get_system_snapshot(
+            db=self.db,
+            tenant_id=self.tenant_id,
+        )
+        platform = snapshot.get("platform_profile") if isinstance(snapshot.get("platform_profile"), dict) else {}
+        catalog = platform.get("feature_catalog") if isinstance(platform.get("feature_catalog"), list) else []
+        rows = [row for row in catalog if isinstance(row, dict)][: max(1, min(int(limit), 1000))]
+        if not rows:
+            return {"features": 0, "routes": 0, "pages": 0, "apis": 0, "relations": 0, "evidences": 0}
+
+        counters = {"features": 0, "routes": 0, "pages": 0, "apis": 0, "relations": 0, "evidences": 0}
+        route_entity_map: dict[str, int] = {}
+        page_entity_map: dict[str, int] = {}
+        api_entity_map: dict[str, int] = {}
+
+        for row in rows:
+            name = str(row.get("name") or "").strip() or "unknown_feature"
+            route = str(row.get("route") or "").strip()
+            summary = str(row.get("summary") or "").strip()
+            page_file = str(row.get("page_file") or "").strip()
+            page_component = str(row.get("page_component") or "").strip()
+            apis = row.get("apis") if isinstance(row.get("apis"), list) else []
+            aliases = row.get("aliases") if isinstance(row.get("aliases"), list) else []
+
+            feature_slug = (route or name).strip().lower()[:200]
+            feature_entity = await self.upsert_entity(
+                entity_type="feature",
+                entity_key=f"feature:{feature_slug}",
+                display_name=name[:255],
+                payload={
+                    "name": name,
+                    "route": route,
+                    "summary": summary,
+                    "page_file": page_file,
+                    "page_component": page_component,
+                    "apis": [str(x) for x in apis[:20] if str(x).strip()],
+                    "aliases": [str(x) for x in aliases[:20] if str(x).strip()],
+                },
+                source_type="feature_catalog",
+                source_ref=(route or page_file or name)[:255],
+            )
+            counters["features"] += 1
+            await self.add_evidence(
+                evidence_type="feature_catalog",
+                content_excerpt=f"{name} | route={route or 'n/a'} | page={page_file or 'n/a'}",
+                entity_id=int(feature_entity.id),
+                source_type="feature_catalog",
+                source_ref=(route or page_file or name)[:255],
+                metadata={"feature_name": name, "route": route, "page_file": page_file},
+            )
+            counters["evidences"] += 1
+
+            route_entity_id: Optional[int] = None
+            if route:
+                if route not in route_entity_map:
+                    route_entity = await self.upsert_entity(
+                        entity_type="route",
+                        entity_key=f"route:{route[:200]}",
+                        display_name=route[:255],
+                        payload={"route": route},
+                        source_type="feature_catalog",
+                        source_ref=route[:255],
+                    )
+                    route_entity_map[route] = int(route_entity.id)
+                    counters["routes"] += 1
+                route_entity_id = route_entity_map[route]
+                rel = await self.upsert_relation(
+                    relation_type="feature_has_route",
+                    subject_entity_id=int(feature_entity.id),
+                    object_entity_id=int(route_entity_id),
+                    source_type="feature_catalog",
+                    source_ref=route[:255],
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="feature_catalog",
+                    content_excerpt=f"{name} has route {route}",
+                    relation_id=int(rel.id),
+                    source_type="feature_catalog",
+                    source_ref=route[:255],
+                    metadata={"feature_name": name, "route": route},
+                )
+                counters["evidences"] += 1
+
+            page_entity_id: Optional[int] = None
+            if page_file:
+                if page_file not in page_entity_map:
+                    page_entity = await self.upsert_entity(
+                        entity_type="page",
+                        entity_key=f"page:{page_file[:220]}",
+                        display_name=(page_component or page_file)[:255],
+                        payload={"page_file": page_file, "page_component": page_component},
+                        source_type="feature_catalog",
+                        source_ref=page_file[:255],
+                    )
+                    page_entity_map[page_file] = int(page_entity.id)
+                    counters["pages"] += 1
+                page_entity_id = page_entity_map[page_file]
+                rel = await self.upsert_relation(
+                    relation_type="feature_rendered_by",
+                    subject_entity_id=int(feature_entity.id),
+                    object_entity_id=int(page_entity_id),
+                    source_type="feature_catalog",
+                    source_ref=page_file[:255],
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="feature_catalog",
+                    content_excerpt=f"{name} rendered_by {page_file}",
+                    relation_id=int(rel.id),
+                    source_type="feature_catalog",
+                    source_ref=page_file[:255],
+                    metadata={"feature_name": name, "page_file": page_file},
+                )
+                counters["evidences"] += 1
+
+            if route_entity_id is not None and page_entity_id is not None:
+                rel = await self.upsert_relation(
+                    relation_type="route_to_page",
+                    subject_entity_id=int(route_entity_id),
+                    object_entity_id=int(page_entity_id),
+                    source_type="feature_catalog",
+                    source_ref=(route or page_file)[:255],
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="feature_catalog",
+                    content_excerpt=f"{route} maps_to {page_file}",
+                    relation_id=int(rel.id),
+                    source_type="feature_catalog",
+                    source_ref=(route or page_file)[:255],
+                    metadata={"route": route, "page_file": page_file},
+                )
+                counters["evidences"] += 1
+
+            for raw_api in apis[:20]:
+                api_path = str(raw_api or "").strip()
+                if not api_path:
+                    continue
+                if api_path not in api_entity_map:
+                    api_entity = await self.upsert_entity(
+                        entity_type="api_endpoint",
+                        entity_key=f"api:{api_path[:220]}",
+                        display_name=api_path[:255],
+                        payload={"path": api_path},
+                        source_type="feature_catalog",
+                        source_ref=api_path[:255],
+                    )
+                    api_entity_map[api_path] = int(api_entity.id)
+                    counters["apis"] += 1
+                rel = await self.upsert_relation(
+                    relation_type="feature_calls_api",
+                    subject_entity_id=int(feature_entity.id),
+                    object_entity_id=int(api_entity_map[api_path]),
+                    source_type="feature_catalog",
+                    source_ref=api_path[:255],
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="feature_catalog",
+                    content_excerpt=f"{name} calls {api_path}",
+                    relation_id=int(rel.id),
+                    source_type="feature_catalog",
+                    source_ref=api_path[:255],
+                    metadata={"feature_name": name, "api_path": api_path},
+                )
+                counters["evidences"] += 1
+
+        return counters
+
+    async def run_incremental_cycle(
+        self,
+        *,
+        delta_limit: int = 1000,
+        event_batch: int = 200,
+        code_scan_hours: int = 6,
+        feature_scan_hours: int = 6,
+        code_scan_base_dir: Optional[str] = None,
+        run_source: str = "system",
+    ) -> dict[str, Any]:
+        started = _utc_now_naive()
+        sync_run = await self._open_sync_run(
+            run_type="incremental",
+            trigger_source=run_source,
+            input_payload={
+                "delta_limit": int(delta_limit),
+                "event_batch": int(event_batch),
+                "code_scan_hours": int(code_scan_hours),
+                "feature_scan_hours": int(feature_scan_hours),
+                "code_scan_base_dir": str(code_scan_base_dir or ""),
+            },
+        )
+        run_id = int(sync_run.id)
+        run_started_at = sync_run.started_at or started
+        stage_results: dict[str, Any] = {}
+        stage_errors: dict[str, str] = {}
+        success_count = 0
+
+        try:
+            try:
+                stage_results["structured_delta"] = await self.sync_structured_delta(limit=delta_limit)
+                success_count += 1
+            except Exception as exc:
+                await self.db.rollback()
+                stage_errors["structured_delta"] = str(exc)[:500]
+                stage_results["structured_delta"] = {"error": stage_errors["structured_delta"]}
+
+            try:
+                stage_results["feishu_enqueue"] = await self.enqueue_feishu_delta_events(limit=min(delta_limit, 1000))
+                success_count += 1
+            except Exception as exc:
+                await self.db.rollback()
+                stage_errors["feishu_enqueue"] = str(exc)[:500]
+                stage_results["feishu_enqueue"] = {"error": stage_errors["feishu_enqueue"]}
+
+            try:
+                stage_results["event_processing"] = await self.process_pending_events(max_batch=event_batch)
+                success_count += 1
+            except Exception as exc:
+                await self.db.rollback()
+                stage_errors["event_processing"] = str(exc)[:500]
+                stage_results["event_processing"] = {"error": stage_errors["event_processing"]}
+
+            try:
+                stage_results["code_scan"] = await self.sync_codebase_if_due(
+                    base_dir=code_scan_base_dir,
+                    interval_hours=code_scan_hours,
+                )
+                success_count += 1
+            except Exception as exc:
+                await self.db.rollback()
+                stage_errors["code_scan"] = str(exc)[:500]
+                stage_results["code_scan"] = {"error": stage_errors["code_scan"]}
+
+            try:
+                stage_results["symbol_index"] = await self.sync_symbol_index_if_due(
+                    base_dir=code_scan_base_dir,
+                    interval_hours=code_scan_hours,
+                )
+                success_count += 1
+            except Exception as exc:
+                await self.db.rollback()
+                stage_errors["symbol_index"] = str(exc)[:500]
+                stage_results["symbol_index"] = {"error": stage_errors["symbol_index"]}
+
+            try:
+                stage_results["feature_catalog"] = await self.sync_feature_catalog_if_due(
+                    interval_hours=feature_scan_hours,
+                )
+                success_count += 1
+            except Exception as exc:
+                await self.db.rollback()
+                stage_errors["feature_catalog"] = str(exc)[:500]
+                stage_results["feature_catalog"] = {"error": stage_errors["feature_catalog"]}
+
+            ended = _utc_now_naive()
+            if success_count == 6:
+                final_status = "success"
+            elif success_count == 0:
+                final_status = "failed"
+            else:
+                final_status = "partial"
+
+            payload = {
+                "tenant_id": self.tenant_id,
+                "started_at": started.isoformat(),
+                "ended_at": ended.isoformat(),
+                "run_id": run_id,
+                "run_status": final_status,
+                "success_stage_count": success_count,
+                "failed_stage_count": len(stage_errors),
+                "stage_errors": stage_errors,
+                **stage_results,
+            }
+            await self._close_sync_run(
+                row=sync_run,
+                status=final_status,
+                result_payload=payload,
+                error_message=("; ".join(f"{k}:{v}" for k, v in stage_errors.items()) if stage_errors else None),
+                started_at_hint=run_started_at,
+            )
+            return payload
+        except Exception as exc:
+            await self.db.rollback()
+            ended = _utc_now_naive()
+            payload = {
+                "tenant_id": self.tenant_id,
+                "started_at": started.isoformat(),
+                "ended_at": ended.isoformat(),
+                "run_id": run_id,
+                "run_status": "failed",
+                "success_stage_count": success_count,
+                "failed_stage_count": max(1, len(stage_errors)),
+                "stage_errors": stage_errors,
+                **stage_results,
+            }
+            await self._close_sync_run(
+                row=sync_run,
+                status="failed",
+                result_payload=payload,
+                error_message=str(exc)[:1000],
+                started_at_hint=run_started_at,
+            )
+            return payload
+
+    async def sync_structured_entities(self, *, limit: int = 1000) -> dict[str, Any]:
+        max_rows = max(1, min(int(limit), 5000))
+        counters = {
+            "projects": 0,
+            "tasks": 0,
+            "signals": 0,
+            "users": 0,
+            "relations": 0,
+            "evidences": 0,
+        }
+
+        project_map: dict[int, int] = {}
+        projects = (
+            await self.db.execute(
+                select(Project)
+                .where(Project.tenant_id == self.tenant_id)
+                .order_by(Project.updated_at.desc(), Project.id.desc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        for p in projects:
+            entity = await self.upsert_entity(
+                entity_type="project",
+                entity_key=f"project:{int(p.id)}",
+                display_name=str(p.title or p.name or f"Project#{int(p.id)}"),
+                payload={
+                    "project_id": int(p.id),
+                    "status": str(p.status or "unknown"),
+                    "owner_id": int(p.owner_id) if p.owner_id is not None else None,
+                },
+                source_type="db",
+                source_ref=f"projects:{int(p.id)}",
+            )
+            project_map[int(p.id)] = int(entity.id)
+            counters["projects"] += 1
+
+        user_map: dict[int, int] = {}
+        users = (
+            await self.db.execute(
+                select(User)
+                .where(User.tenant_id == self.tenant_id)
+                .order_by(User.updated_at.desc(), User.id.desc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        for u in users:
+            entity = await self.upsert_entity(
+                entity_type="user",
+                entity_key=f"user:{int(u.id)}",
+                display_name=str(u.name or f"User#{int(u.id)}"),
+                payload={
+                    "user_id": int(u.id),
+                    "role": str(u.role or "unknown"),
+                    "is_active": bool(u.is_active),
+                },
+                source_type="db",
+                source_ref=f"users:{int(u.id)}",
+            )
+            user_map[int(u.id)] = int(entity.id)
+            counters["users"] += 1
+
+        tasks = (
+            await self.db.execute(
+                select(Task)
+                .where(Task.tenant_id == self.tenant_id)
+                .order_by(Task.updated_at.desc(), Task.id.desc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        for t in tasks:
+            task_entity = await self.upsert_entity(
+                entity_type="task",
+                entity_key=f"task:{int(t.id)}",
+                display_name=str(t.title or f"Task#{int(t.id)}"),
+                payload={
+                    "task_id": int(t.id),
+                    "project_id": int(t.project_id) if t.project_id is not None else None,
+                    "assignee_id": int(t.assignee_id) if t.assignee_id is not None else None,
+                    "status": str(t.status or "unknown"),
+                },
+                source_type="db",
+                source_ref=f"tasks:{int(t.id)}",
+            )
+            counters["tasks"] += 1
+
+            if t.project_id is not None and int(t.project_id) in project_map:
+                rel = await self.upsert_relation(
+                    relation_type="belongs_to",
+                    subject_entity_id=int(task_entity.id),
+                    object_entity_id=int(project_map[int(t.project_id)]),
+                    source_type="db",
+                    source_ref=f"tasks:{int(t.id)}",
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="db_row",
+                    content_excerpt=f"task#{int(t.id)} belongs_to project#{int(t.project_id)}",
+                    relation_id=int(rel.id),
+                    source_type="db",
+                    source_ref=f"tasks:{int(t.id)}",
+                    metadata={"table": "tasks"},
+                )
+                counters["evidences"] += 1
+
+            if t.assignee_id is not None and int(t.assignee_id) in user_map:
+                rel = await self.upsert_relation(
+                    relation_type="assigned_to",
+                    subject_entity_id=int(task_entity.id),
+                    object_entity_id=int(user_map[int(t.assignee_id)]),
+                    source_type="db",
+                    source_ref=f"tasks:{int(t.id)}",
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="db_row",
+                    content_excerpt=f"task#{int(t.id)} assigned_to user#{int(t.assignee_id)}",
+                    relation_id=int(rel.id),
+                    source_type="db",
+                    source_ref=f"tasks:{int(t.id)}",
+                    metadata={"table": "tasks"},
+                )
+                counters["evidences"] += 1
+
+        signals = (
+            await self.db.execute(
+                select(Signal)
+                .where(Signal.tenant_id == self.tenant_id)
+                .order_by(Signal.updated_at.desc(), Signal.id.desc())
+                .limit(max_rows)
+            )
+        ).scalars().all()
+        for s in signals:
+            signal_entity = await self.upsert_entity(
+                entity_type="signal",
+                entity_key=f"signal:{int(s.id)}",
+                display_name=str(s.title or s.content or f"Signal#{int(s.id)}")[:255],
+                payload={
+                    "signal_id": int(s.id),
+                    "priority": str(s.priority or ""),
+                    "status": str(s.status or ""),
+                    "matched_project_id": (
+                        int(s.matched_project_id) if getattr(s, "matched_project_id", None) is not None else None
+                    ),
+                },
+                source_type="db",
+                source_ref=f"signals:{int(s.id)}",
+            )
+            counters["signals"] += 1
+
+            target_pid = int(s.matched_project_id) if getattr(s, "matched_project_id", None) is not None else None
+            if target_pid is not None and target_pid in project_map:
+                rel = await self.upsert_relation(
+                    relation_type="matches_project",
+                    subject_entity_id=int(signal_entity.id),
+                    object_entity_id=int(project_map[target_pid]),
+                    source_type="db",
+                    source_ref=f"signals:{int(s.id)}",
+                )
+                counters["relations"] += 1
+                await self.add_evidence(
+                    evidence_type="db_row",
+                    content_excerpt=f"signal#{int(s.id)} matches project#{target_pid}",
+                    relation_id=int(rel.id),
+                    source_type="db",
+                    source_ref=f"signals:{int(s.id)}",
+                    metadata={"table": "signals"},
+                )
+                counters["evidences"] += 1
+
+        await self.enqueue_ingest_event(
+            channel="db_delta",
+            event_key=f"structured_sync:{_utc_now_naive().strftime('%Y%m%d%H%M%S')}",
+            payload={"limit": max_rows, "counters": counters},
+            status="done",
+        )
+        await self.set_ingest_offset(
+            channel="db_delta",
+            offset_key="last_structured_sync_at",
+            offset_value=_utc_now_naive().isoformat(),
+            payload={"counters": counters},
+        )
+
+        await self.db.commit()
+        return counters
+
+    async def sync_codebase_snapshot(self, *, base_dir: Optional[str] = None) -> dict[str, Any]:
+        root = Path(base_dir or "/app/backend").resolve()
+        if not root.exists() or not root.is_dir():
+            raise ValueError(f"codebase path not found: {root}")
+
+        py_files = [p for p in root.rglob("*.py") if p.is_file()]
+        route_markers = ("@router.get(", "@router.post(", "@router.put(", "@router.patch(", "@router.delete(")
+        route_count = 0
+        service_file_count = 0
+        model_file_count = 0
+
+        for py_file in py_files:
+            rel = str(py_file.relative_to(root))
+            if rel.startswith("app/services/"):
+                service_file_count += 1
+            if rel.startswith("app/models/"):
+                model_file_count += 1
+            try:
+                text = py_file.read_text(encoding="utf-8", errors="ignore")
+            except Exception:
+                continue
+            for marker in route_markers:
+                route_count += text.count(marker)
+
+        payload = {
+            "root": str(root),
+            "python_files": len(py_files),
+            "service_files": service_file_count,
+            "model_files": model_file_count,
+            "api_route_markers": route_count,
+            "scanned_at": _utc_now_naive().isoformat(),
+        }
+        entity = await self.upsert_entity(
+            entity_type="service",
+            entity_key="service:backend_codebase",
+            display_name="Backend Codebase Snapshot",
+            payload=payload,
+            source_type="code_scan",
+            source_ref="filesystem",
+        )
+        await self.add_evidence(
+            evidence_type="code_scan",
+            content_excerpt=f"python_files={len(py_files)}, routes={route_count}",
+            entity_id=int(entity.id),
+            source_type="code_scan",
+            source_ref=str(root),
+            metadata=payload,
+        )
+        await self.enqueue_ingest_event(
+            channel="code_scan",
+            event_key=f"code_scan:{_utc_now_naive().strftime('%Y%m%d%H%M%S')}",
+            payload=payload,
+            status="done",
+        )
+        await self.set_ingest_offset(
+            channel="code_scan",
+            offset_key="last_code_scan_at",
+            offset_value=_utc_now_naive().isoformat(),
+            payload=payload,
+        )
+        await self.db.commit()
+        return payload
+
+    async def get_overview(self) -> dict[str, Any]:
+        entity_total = (
+            await self.db.execute(select(func.count(KGEntity.id)).where(KGEntity.tenant_id == self.tenant_id))
+        ).scalar_one()
+        relation_total = (
+            await self.db.execute(select(func.count(KGRelation.id)).where(KGRelation.tenant_id == self.tenant_id))
+        ).scalar_one()
+        evidence_total = (
+            await self.db.execute(select(func.count(KGEvidence.id)).where(KGEvidence.tenant_id == self.tenant_id))
+        ).scalar_one()
+        pending_events = (
+            await self.db.execute(
+                select(func.count(IngestEvent.id)).where(
+                    and_(
+                        IngestEvent.tenant_id == self.tenant_id,
+                        IngestEvent.status.in_(["pending", "retry"]),
+                    )
+                )
+            )
+        ).scalar_one()
+        latest_run = (
+            await self.db.execute(
+                select(KGSyncRun)
+                .where(KGSyncRun.tenant_id == self.tenant_id)
+                .order_by(KGSyncRun.started_at.desc(), KGSyncRun.id.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+
+        entity_groups = (
+            await self.db.execute(
+                select(KGEntity.entity_type, func.count(KGEntity.id))
+                .where(KGEntity.tenant_id == self.tenant_id)
+                .group_by(KGEntity.entity_type)
+            )
+        ).all()
+        relation_groups = (
+            await self.db.execute(
+                select(KGRelation.relation_type, func.count(KGRelation.id))
+                .where(KGRelation.tenant_id == self.tenant_id)
+                .group_by(KGRelation.relation_type)
+            )
+        ).all()
+
+        return {
+            "tenant_id": self.tenant_id,
+            "entity_total": int(entity_total or 0),
+            "relation_total": int(relation_total or 0),
+            "evidence_total": int(evidence_total or 0),
+            "pending_events": int(pending_events or 0),
+            "entities_by_type": {str(k): int(v or 0) for k, v in entity_groups},
+            "relations_by_type": {str(k): int(v or 0) for k, v in relation_groups},
+            "latest_sync_run": (
+                {
+                    "id": int(latest_run.id),
+                    "run_type": str(latest_run.run_type or ""),
+                    "trigger_source": str(latest_run.trigger_source or ""),
+                    "status": str(latest_run.status or ""),
+                    "started_at": latest_run.started_at.isoformat() if latest_run.started_at else None,
+                    "ended_at": latest_run.ended_at.isoformat() if latest_run.ended_at else None,
+                    "duration_ms": int(latest_run.duration_ms or 0),
+                }
+                if latest_run is not None
+                else None
+            ),
+            "snapshot_at": _utc_now_naive().isoformat(),
+        }
diff --git a/backend/app/services/prompt_profile_service.py b/backend/app/services/prompt_profile_service.py
new file mode 100644
index 00000000..73b6f15c
--- /dev/null
+++ b/backend/app/services/prompt_profile_service.py
@@ -0,0 +1,241 @@
+"""AI 角色 Prompt 配置服务。"""
+
+from __future__ import annotations
+
+import json
+import re
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import build_default_role_prompt_bundle
+from ..models import Integration
+
+INTEGRATION_TYPE = "ai_prompt_profiles"
+ROLE_KEY_RE = re.compile(r"^[a-z0-9][a-z0-9_-]{1,31}$")
+MAX_ROLES = 12
+
+
+def _safe_json(value: Any) -> dict[str, Any]:
+    if isinstance(value, dict):
+        try:
+            return json.loads(json.dumps(value))
+        except Exception:
+            return dict(value)
+    if isinstance(value, str):
+        try:
+            parsed = json.loads(value)
+            if isinstance(parsed, dict):
+                return parsed
+        except Exception:
+            return {}
+    return {}
+
+
+def _now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+class PromptProfileService:
+    """管理租户级 AI 角色 Prompt 配置。"""
+
+    async def get_profiles(self, db: AsyncSession, *, tenant_id: int) -> dict[str, Any]:
+        integration, bundle = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        normalized = self._normalize_bundle(bundle)
+        if normalized != bundle:
+            integration.config = json.loads(json.dumps(normalized, ensure_ascii=False))
+            integration.is_active = True
+            await db.commit()
+        return normalized
+
+    async def update_profiles(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+        roles: list[dict[str, Any]],
+        active_role_key: Optional[str] = None,
+    ) -> dict[str, Any]:
+        if not roles:
+            raise ValueError("至少需要保留一个角色")
+        if len(roles) > MAX_ROLES:
+            raise ValueError(f"角色数量不能超过 {MAX_ROLES} 个")
+
+        dedup: dict[str, dict[str, Any]] = {}
+        for idx, raw in enumerate(roles):
+            key = str(raw.get("role_key") or "").strip().lower()
+            if not key:
+                raise ValueError(f"第 {idx + 1} 个角色缺少 role_key")
+            if not ROLE_KEY_RE.match(key):
+                raise ValueError(f"角色标识 {key} 格式不正确，仅支持小写字母/数字/_/-")
+            if key in dedup:
+                raise ValueError(f"角色标识重复: {key}")
+
+            role_name = str(raw.get("role_name") or "").strip()
+            system_prompt = str(raw.get("system_prompt") or "").strip()
+            if not role_name:
+                raise ValueError(f"角色 {key} 缺少角色名称")
+            if len(role_name) > 32:
+                raise ValueError(f"角色 {key} 名称过长（最多 32 字）")
+            if len(system_prompt) < 20:
+                raise ValueError(f"角色 {key} 的提示词过短（至少 20 字）")
+            if len(system_prompt) > 8000:
+                raise ValueError(f"角色 {key} 的提示词过长（最多 8000 字）")
+
+            role = {
+                "role_key": key,
+                "role_name": role_name,
+                "role_description": str(raw.get("role_description") or "").strip()[:300],
+                "system_prompt": system_prompt,
+                "enabled": bool(raw.get("enabled", True)),
+                "updated_at": str(raw.get("updated_at") or _now_iso()),
+            }
+            dedup[key] = role
+
+        ordered_roles = list(dedup.values())
+        enabled_keys = [str(r["role_key"]) for r in ordered_roles if bool(r.get("enabled"))]
+        if not enabled_keys:
+            raise ValueError("至少需要启用一个角色")
+
+        requested_active = str(active_role_key or "").strip().lower()
+        if requested_active and requested_active in enabled_keys:
+            active = requested_active
+        else:
+            active = enabled_keys[0]
+
+        bundle = {
+            "version": 1,
+            "active_role_key": active,
+            "roles": ordered_roles,
+            "updated_at": _now_iso(),
+        }
+
+        integration, _ = await self._get_or_init_bundle(db, tenant_id=tenant_id)
+        integration.config = json.loads(json.dumps(bundle, ensure_ascii=False))
+        integration.is_active = True
+        await db.commit()
+        return self._normalize_bundle(bundle)
+
+    async def get_active_role(self, db: AsyncSession, *, tenant_id: int) -> dict[str, Any]:
+        data = await self.get_profiles(db, tenant_id=tenant_id)
+        active_key = str(data.get("active_role_key") or "")
+        roles = data.get("roles", [])
+        if not isinstance(roles, list):
+            roles = []
+        active_role = next((r for r in roles if str(r.get("role_key")) == active_key), None)
+        if active_role is None and roles:
+            active_role = roles[0]
+        return {
+            "active_role_key": active_key,
+            "role": active_role,
+        }
+
+    async def _get_or_init_bundle(
+        self,
+        db: AsyncSession,
+        *,
+        tenant_id: int,
+    ) -> tuple[Integration, dict[str, Any]]:
+        result = await db.execute(
+            select(Integration).where(
+                and_(
+                    Integration.tenant_id == int(tenant_id),
+                    Integration.type == INTEGRATION_TYPE,
+                )
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration is None:
+            bundle = self._normalize_bundle(build_default_role_prompt_bundle())
+            integration = Integration(
+                tenant_id=int(tenant_id),
+                type=INTEGRATION_TYPE,
+                config=json.loads(json.dumps(bundle, ensure_ascii=False)),
+                is_active=True,
+            )
+            db.add(integration)
+            await db.commit()
+            return integration, bundle
+
+        bundle = self._normalize_bundle(_safe_json(integration.config))
+        return integration, bundle
+
+    def _normalize_bundle(self, raw: dict[str, Any]) -> dict[str, Any]:
+        defaults = build_default_role_prompt_bundle()
+        default_roles_raw = defaults.get("roles", [])
+        default_roles: dict[str, dict[str, Any]] = {}
+        if isinstance(default_roles_raw, list):
+            for row in default_roles_raw:
+                if not isinstance(row, dict):
+                    continue
+                key = str(row.get("role_key") or "").strip().lower()
+                if key:
+                    default_roles[key] = {
+                        "role_key": key,
+                        "role_name": str(row.get("role_name") or key)[:32],
+                        "role_description": str(row.get("role_description") or "")[:300],
+                        "system_prompt": str(row.get("system_prompt") or "").strip(),
+                        "enabled": bool(row.get("enabled", True)),
+                        "updated_at": str(row.get("updated_at") or _now_iso()),
+                    }
+
+        incoming_roles = raw.get("roles") if isinstance(raw.get("roles"), list) else []
+        roles: list[dict[str, Any]] = []
+        seen: set[str] = set()
+
+        for item in incoming_roles:
+            if not isinstance(item, dict):
+                continue
+            key = str(item.get("role_key") or "").strip().lower()
+            if not key or key in seen or not ROLE_KEY_RE.match(key):
+                continue
+            seen.add(key)
+
+            fallback = default_roles.get(key, {})
+            prompt_text = str(item.get("system_prompt") or fallback.get("system_prompt") or "").strip()
+            if len(prompt_text) < 20:
+                prompt_text = str(fallback.get("system_prompt") or "").strip()
+            if len(prompt_text) < 20:
+                continue
+
+            roles.append(
+                {
+                    "role_key": key,
+                    "role_name": str(item.get("role_name") or fallback.get("role_name") or key).strip()[:32] or key,
+                    "role_description": str(item.get("role_description") or fallback.get("role_description") or "").strip()[:300],
+                    "system_prompt": prompt_text[:8000],
+                    "enabled": bool(item.get("enabled", fallback.get("enabled", True))),
+                    "updated_at": str(item.get("updated_at") or _now_iso()),
+                }
+            )
+            if len(roles) >= MAX_ROLES:
+                break
+
+        if not roles:
+            roles = list(default_roles.values())[:MAX_ROLES]
+
+        enabled_keys = [str(r["role_key"]) for r in roles if bool(r.get("enabled"))]
+        if not enabled_keys and roles:
+            roles[0]["enabled"] = True
+            enabled_keys = [str(roles[0]["role_key"])]
+
+        active_raw = str(raw.get("active_role_key") or "").strip().lower()
+        active_role_key = active_raw if active_raw in enabled_keys else (enabled_keys[0] if enabled_keys else str(roles[0]["role_key"]))
+
+        return {
+            "version": 1,
+            "active_role_key": active_role_key,
+            "roles": roles,
+            "updated_at": str(raw.get("updated_at") or _now_iso()),
+            "tips": {
+                "active_role_key": "当前生效角色，影响 AI 回复语气与任务分派风格。",
+                "system_prompt": "建议写清角色职责、回复风格、输出边界；支持后续扩展多个角色。",
+                "role_key": "角色标识用于系统路由，保存后建议保持稳定。",
+            },
+        }
+
+
+prompt_profile_service = PromptProfileService()
diff --git a/backend/app/services/recipe_service.py b/backend/app/services/recipe_service.py
new file mode 100644
index 00000000..f437ba73
--- /dev/null
+++ b/backend/app/services/recipe_service.py
@@ -0,0 +1,266 @@
+"""食谱服务 - 包含热量识别与营养分析"""
+
+import base64
+import io
+from typing import Optional
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import Recipe
+from ..schemas import RecipeCreate, RecipeUpdate, RecipeResponse
+from .ai_service import AIService
+from ..config.ai_prompts import PROMPT_ANALYZE_FOOD_CALORIES
+
+
+class RecipeService:
+    """食谱服务"""
+
+    def __init__(self, ai_service: AIService):
+        self.ai_service = ai_service
+
+    async def create_recipe(
+        self,
+        db: AsyncSession,
+        recipe_data: RecipeCreate,
+        tenant_id: int,
+    ) -> Recipe:
+        """创建食谱"""
+        recipe = Recipe(
+            **recipe_data.model_dump(),
+            tenant_id=tenant_id,
+            status="draft",
+        )
+        db.add(recipe)
+        await db.commit()
+        await db.refresh(recipe)
+        return recipe
+
+    async def get_recipe(
+        self,
+        db: AsyncSession,
+        recipe_id: int,
+        tenant_id: int,
+    ) -> Optional[Recipe]:
+        """获取单个食谱"""
+        result = await db.execute(
+            select(Recipe).where(
+                Recipe.id == recipe_id,
+                Recipe.tenant_id == tenant_id,
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def list_recipes(
+        self,
+        db: AsyncSession,
+        tenant_id: int,
+        skip: int = 0,
+        limit: int = 20,
+        category: Optional[str] = None,
+        cuisine: Optional[str] = None,
+        status: Optional[str] = None,
+        is_favorite: Optional[bool] = None,
+    ) -> list[Recipe]:
+        """获取食谱列表"""
+        query = select(Recipe).where(Recipe.tenant_id == tenant_id)
+
+        if category:
+            query = query.where(Recipe.category == category)
+        if cuisine:
+            query = query.where(Recipe.cuisine == cuisine)
+        if status:
+            query = query.where(Recipe.status == status)
+        if is_favorite is not None:
+            query = query.where(Recipe.is_favorite == is_favorite)
+
+        query = query.order_by(Recipe.created_at.desc()).offset(skip).limit(limit)
+        result = await db.execute(query)
+        return list(result.scalars().all())
+
+    async def update_recipe(
+        self,
+        db: AsyncSession,
+        recipe_id: int,
+        recipe_data: RecipeUpdate,
+        tenant_id: int,
+    ) -> Optional[Recipe]:
+        """更新食谱"""
+        recipe = await self.get_recipe(db, recipe_id, tenant_id)
+        if not recipe:
+            return None
+
+        update_data = recipe_data.model_dump(exclude_unset=True)
+        for field, value in update_data.items():
+            setattr(recipe, field, value)
+
+        await db.commit()
+        await db.refresh(recipe)
+        return recipe
+
+    async def delete_recipe(
+        self,
+        db: AsyncSession,
+        recipe_id: int,
+        tenant_id: int,
+    ) -> bool:
+        """删除食谱"""
+        recipe = await self.get_recipe(db, recipe_id, tenant_id)
+        if not recipe:
+            return False
+
+        await db.delete(recipe)
+        await db.commit()
+        return True
+
+    async def analyze_food_from_description(
+        self,
+        db: AsyncSession,
+        description: str,
+        tenant_id: int,
+    ) -> dict:
+        """
+        基于食物描述分析热量和营养成分
+
+        注意：此方法使用纯文本分析。对于图片识别，
+        建议先使用图像识别服务将图片转为描述文本，再调用此方法。
+        """
+        tid = str(tenant_id)
+        prompt = PROMPT_ANALYZE_FOOD_CALORIES.format(food_description=description)
+
+        try:
+            response_text = await self.ai_service._invoke_with_retry(
+                prompt,
+                db=db,
+                tenant_id=tid,
+                function_name="analyze_food_calories",
+                use_cache=True,
+            )
+            result = self.ai_service._parse_json(response_text)
+            return {
+                "success": True,
+                "data": result,
+                "message": "分析成功",
+            }
+        except Exception as e:
+            return {
+                "success": False,
+                "data": None,
+                "message": f"分析失败: {str(e)}",
+            }
+
+    async def create_recipe_with_ai_analysis(
+        self,
+        db: AsyncSession,
+        recipe_data: RecipeCreate,
+        food_description: str,
+        tenant_id: int,
+    ) -> tuple[Recipe, dict]:
+        """
+        创建食谱并进行 AI 热量分析
+
+        Args:
+            db: 数据库会话
+            recipe_data: 食谱基础数据
+            food_description: 食物描述（可以是图片识别后的描述）
+            tenant_id: 租户ID
+
+        Returns:
+            (Recipe, analysis_result): 创建的食谱对象和分析结果
+        """
+        # 先创建食谱
+        recipe = await self.create_recipe(db, recipe_data, tenant_id)
+
+        # 进行 AI 分析
+        analysis = await self.analyze_food_from_description(
+            db, food_description, tenant_id
+        )
+
+        if analysis["success"] and analysis["data"]:
+            # 更新食谱的营养信息
+            data = analysis["data"]
+            recipe.calories = data.get("calories")
+            recipe.protein = data.get("protein")
+            recipe.fat = data.get("fat")
+            recipe.carbs = data.get("carbs")
+            recipe.fiber = data.get("fiber")
+            recipe.ai_recognition_confidence = data.get("confidence")
+            recipe.ai_analysis_result = data
+            recipe.category = data.get("category") or recipe_data.category
+            recipe.cuisine = data.get("cuisine") or recipe_data.cuisine
+
+            await db.commit()
+            await db.refresh(recipe)
+
+        return recipe, analysis
+
+    def decode_base64_image(self, image_base64: str) -> tuple[bytes, str]:
+        """
+        解码 Base64 图片数据
+
+        Args:
+            image_base64: Base64 编码的图片数据（可能包含 data:image/xxx;base64, 前缀）
+
+        Returns:
+            (image_bytes, mime_type): 图片字节数据和 MIME 类型
+        """
+        # 移除可能的数据 URI 前缀
+        if "," in image_base64:
+            header, image_base64 = image_base64.split(",", 1)
+            # 提取 MIME 类型
+            mime_type = header.split(";")[0].split(":")[-1] if ":" in header else "image/jpeg"
+        else:
+            mime_type = "image/jpeg"
+
+        # 解码 Base64
+        image_bytes = base64.b64decode(image_base64)
+        return image_bytes, mime_type
+
+    async def upload_image(
+        self,
+        image_base64: str,
+        filename: Optional[str] = None,
+    ) -> str:
+        """
+        上传图片并返回 URL
+
+        注意：这是一个示例实现，实际项目中应该集成：
+        - AWS S3
+        - 阿里云 OSS
+        - 腾讯云 COS
+        - MinIO
+        等对象存储服务
+
+        Args:
+            image_base64: Base64 编码的图片数据
+            filename: 原始文件名
+
+        Returns:
+            图片 URL
+        """
+        # TODO: 集成实际的文件存储服务
+        # 这里只是示例，实际应该上传到对象存储
+        image_bytes, mime_type = self.decode_base64_image(image_base64)
+
+        # 示例：生成一个模拟 URL
+        # 实际应该上传后返回真实的 URL
+        import hashlib
+        import time
+
+        file_hash = hashlib.md5(image_bytes).hexdigest()
+        timestamp = int(time.time())
+        mock_url = f"https://example.com/recipe-images/{file_hash}_{timestamp}.jpg"
+
+        return mock_url
+
+
+# 单例
+_recipe_service: Optional[RecipeService] = None
+
+
+def get_recipe_service() -> RecipeService:
+    """获取食谱服务单例"""
+    global _recipe_service
+    if _recipe_service is None:
+        from .ai_service import get_ai_service
+        _recipe_service = RecipeService(ai_service=get_ai_service())
+    return _recipe_service
diff --git a/backend/app/services/scheduler_center.py b/backend/app/services/scheduler_center.py
new file mode 100644
index 00000000..ff4ea6bd
--- /dev/null
+++ b/backend/app/services/scheduler_center.py
@@ -0,0 +1,448 @@
+"""轻量级调度中心。"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from datetime import datetime, timedelta
+from typing import Optional
+
+from sqlalchemy import and_, or_, select
+from sqlalchemy.ext.asyncio import async_sessionmaker
+
+from ..core.logging import get_logger
+from ..models import KnowledgeBaseDoc, PortfolioSnapshot, Project
+from .adaptive_rag_service import AdaptiveRAGService
+from .ai_pmo_cadence_service import AIPMOCadenceService
+from .project_knowledge_graph import ProjectKnowledgeGraphService
+from . import workload_service
+from . import workload_alert_service
+
+logger = get_logger(__name__)
+
+
+class WorkloadRecalcScheduler:
+    """每日 00:00 重算 workload_ratio。"""
+
+    def __init__(self, session_factory: async_sessionmaker):
+        self._session_factory = session_factory
+        self._task: Optional[asyncio.Task] = None
+        self._stopped = asyncio.Event()
+
+    def start(self) -> None:
+        if self._task and not self._task.done():
+            return
+        self._stopped.clear()
+        self._task = asyncio.create_task(self._run_loop(), name="workload-recalc-daily")
+        logger.info("workload_scheduler_started")
+
+    async def stop(self) -> None:
+        self._stopped.set()
+        if not self._task:
+            return
+        self._task.cancel()
+        try:
+            await self._task
+        except asyncio.CancelledError:
+            pass
+        logger.info("workload_scheduler_stopped")
+
+    async def run_once(self) -> dict:
+        """手动执行一次，便于调试与测试。"""
+        return await self._run_recalc()
+
+    async def _run_loop(self) -> None:
+        while not self._stopped.is_set():
+            seconds = self._seconds_until_next_midnight()
+            try:
+                await asyncio.wait_for(self._stopped.wait(), timeout=seconds)
+                break
+            except TimeoutError:
+                pass
+
+            try:
+                await self._run_recalc()
+            except Exception as exc:  # pragma: no cover
+                logger.warning("workload_scheduler_recalc_failed", error=str(exc))
+
+    @staticmethod
+    def _seconds_until_next_midnight() -> float:
+        now = datetime.now()
+        tomorrow = (now + timedelta(days=1)).date()
+        next_midnight = datetime.combine(tomorrow, datetime.min.time())
+        return max((next_midnight - now).total_seconds(), 1.0)
+
+    async def _run_recalc(self) -> dict:
+        async with self._session_factory() as db:
+            tenant_ids = await workload_service.list_tenant_ids(db)
+            total_users = 0
+            total_suggestions = 0
+            for tenant_id in tenant_ids or [1]:
+                total_users += await workload_service.recalculate_all_users_workload_ratio(
+                    db,
+                    tenant_id=tenant_id,
+                )
+                summary = await workload_alert_service.process_tenant_unresolved_overloads(
+                    db,
+                    tenant_id=tenant_id,
+                )
+                total_suggestions += int(summary.get("suggestion_count", 0))
+            logger.info(
+                "workload_scheduler_recalc_done",
+                tenant_count=len(tenant_ids or [1]),
+                user_count=total_users,
+                suggestion_count=total_suggestions,
+            )
+            return {
+                "tenant_count": len(tenant_ids or [1]),
+                "user_count": total_users,
+                "suggestion_count": total_suggestions,
+            }
+
+
+class KnowledgeUrlRefreshScheduler:
+    """按周期回抓 URL 知识源（默认每 168 小时）。"""
+
+    def __init__(self, session_factory: async_sessionmaker):
+        self._session_factory = session_factory
+        self._task: Optional[asyncio.Task] = None
+        self._stopped = asyncio.Event()
+        try:
+            hours = int(os.getenv("KB_URL_REFRESH_HOURS", "168"))
+        except Exception:
+            hours = 168
+        self._interval_seconds = max(hours, 1) * 3600
+
+    def start(self) -> None:
+        if self._task and not self._task.done():
+            return
+        self._stopped.clear()
+        self._task = asyncio.create_task(self._run_loop(), name="knowledge-url-refresh")
+        logger.info("knowledge_url_refresh_scheduler_started", interval_seconds=self._interval_seconds)
+
+    async def stop(self) -> None:
+        self._stopped.set()
+        if not self._task:
+            return
+        self._task.cancel()
+        try:
+            await self._task
+        except asyncio.CancelledError:
+            pass
+        logger.info("knowledge_url_refresh_scheduler_stopped")
+
+    async def run_once(self) -> dict:
+        return await self._run_refresh()
+
+    async def _run_loop(self) -> None:
+        while not self._stopped.is_set():
+            try:
+                await asyncio.wait_for(self._stopped.wait(), timeout=self._interval_seconds)
+                break
+            except TimeoutError:
+                pass
+
+            try:
+                await self._run_refresh()
+            except Exception as exc:  # pragma: no cover
+                logger.warning("knowledge_url_refresh_scheduler_failed", error=str(exc))
+
+    async def _run_refresh(self) -> dict:
+        async with self._session_factory() as db:
+            now = datetime.now()
+            cutoff = now - timedelta(days=7)
+            docs = (
+                await db.execute(
+                    select(KnowledgeBaseDoc).where(
+                        and_(
+                            KnowledgeBaseDoc.source_type == "url",
+                            or_(
+                                KnowledgeBaseDoc.last_fetched_at.is_(None),
+                                KnowledgeBaseDoc.last_fetched_at <= cutoff,
+                            ),
+                        )
+                    )
+                )
+            ).scalars().all()
+
+            refreshed = 0
+            failed = 0
+            for doc in docs:
+                tenant_id = int(getattr(doc, "tenant_id", 1) or 1)
+                service = AdaptiveRAGService(db, tenant_id=tenant_id)
+                try:
+                    await service.refresh_url_doc(doc_id=int(doc.id), project_id=None)
+                    refreshed += 1
+                except Exception as exc:  # pragma: no cover
+                    failed += 1
+                    logger.warning(
+                        "knowledge_url_refresh_doc_failed",
+                        doc_id=int(doc.id),
+                        tenant_id=tenant_id,
+                        error=str(exc),
+                    )
+
+            logger.info(
+                "knowledge_url_refresh_done",
+                candidate_count=len(docs),
+                refreshed=refreshed,
+                failed=failed,
+            )
+            return {
+                "candidate_count": len(docs),
+                "refreshed": refreshed,
+                "failed": failed,
+            }
+
+
+class KnowledgeGraphIncrementalScheduler:
+    """按周期执行知识图谱增量学习流水线。"""
+
+    def __init__(self, session_factory: async_sessionmaker):
+        self._session_factory = session_factory
+        self._task: Optional[asyncio.Task] = None
+        self._stopped = asyncio.Event()
+        try:
+            minutes = int(os.getenv("KG_INCREMENTAL_INTERVAL_MINUTES", "5"))
+        except Exception:
+            minutes = 5
+        self._interval_seconds = max(minutes, 1) * 60
+
+        try:
+            self._delta_limit = max(100, int(os.getenv("KG_INCREMENTAL_DELTA_LIMIT", "1000")))
+        except Exception:
+            self._delta_limit = 1000
+
+        try:
+            self._event_batch = max(10, int(os.getenv("KG_INCREMENTAL_EVENT_BATCH", "200")))
+        except Exception:
+            self._event_batch = 200
+
+        try:
+            self._code_scan_hours = max(1, int(os.getenv("KG_CODE_SCAN_INTERVAL_HOURS", "6")))
+        except Exception:
+            self._code_scan_hours = 6
+
+    def start(self) -> None:
+        if self._task and not self._task.done():
+            return
+        self._stopped.clear()
+        self._task = asyncio.create_task(self._run_loop(), name="knowledge-graph-incremental")
+        logger.info(
+            "knowledge_graph_incremental_scheduler_started",
+            interval_seconds=self._interval_seconds,
+            delta_limit=self._delta_limit,
+            event_batch=self._event_batch,
+            code_scan_hours=self._code_scan_hours,
+        )
+
+    async def stop(self) -> None:
+        self._stopped.set()
+        if not self._task:
+            return
+        self._task.cancel()
+        try:
+            await self._task
+        except asyncio.CancelledError:
+            pass
+        logger.info("knowledge_graph_incremental_scheduler_stopped")
+
+    async def run_once(self) -> dict:
+        return await self._run_cycle()
+
+    async def _run_loop(self) -> None:
+        while not self._stopped.is_set():
+            try:
+                await asyncio.wait_for(self._stopped.wait(), timeout=self._interval_seconds)
+                break
+            except TimeoutError:
+                pass
+
+            try:
+                await self._run_cycle()
+            except Exception as exc:  # pragma: no cover
+                logger.warning("knowledge_graph_incremental_scheduler_failed", error=str(exc))
+
+    async def _run_cycle(self) -> dict:
+        async with self._session_factory() as db:
+            tenant_ids = await workload_service.list_tenant_ids(db)
+            effective_tenants = tenant_ids or [1]
+            tenant_results: list[dict] = []
+            for tenant_id in effective_tenants:
+                service = ProjectKnowledgeGraphService(db, tenant_id=int(tenant_id))
+                result = await service.run_incremental_cycle(
+                    delta_limit=self._delta_limit,
+                    event_batch=self._event_batch,
+                    code_scan_hours=self._code_scan_hours,
+                    run_source="scheduler",
+                )
+                tenant_results.append(result)
+
+            logger.info(
+                "knowledge_graph_incremental_scheduler_done",
+                tenant_count=len(effective_tenants),
+                results=tenant_results,
+            )
+            return {
+                "tenant_count": len(effective_tenants),
+                "results": tenant_results,
+            }
+
+
+class AIPMOCadenceScheduler:
+    """AI PMO cadence 自动调度。"""
+
+    def __init__(self, session_factory: async_sessionmaker):
+        self._session_factory = session_factory
+        self._task: Optional[asyncio.Task] = None
+        self._stopped = asyncio.Event()
+        try:
+            minutes = int(os.getenv("AI_PMO_CADENCE_INTERVAL_MINUTES", "60"))
+        except Exception:
+            minutes = 60
+        self._interval_seconds = max(minutes, 5) * 60
+
+    def start(self) -> None:
+        if self._task and not self._task.done():
+            return
+        self._stopped.clear()
+        self._task = asyncio.create_task(self._run_loop(), name="ai-pmo-cadence")
+        logger.info("ai_pmo_cadence_scheduler_started", interval_seconds=self._interval_seconds)
+
+    async def stop(self) -> None:
+        self._stopped.set()
+        if not self._task:
+            return
+        self._task.cancel()
+        try:
+            await self._task
+        except asyncio.CancelledError:
+            pass
+        logger.info("ai_pmo_cadence_scheduler_stopped")
+
+    async def run_once(self) -> dict:
+        return await self._run_cycle()
+
+    async def _run_loop(self) -> None:
+        while not self._stopped.is_set():
+            try:
+                await asyncio.wait_for(self._stopped.wait(), timeout=self._interval_seconds)
+                break
+            except TimeoutError:
+                pass
+
+            try:
+                await self._run_cycle()
+            except Exception as exc:  # pragma: no cover
+                logger.warning("ai_pmo_cadence_scheduler_failed", error=str(exc))
+
+    async def _run_cycle(self) -> dict:
+        async with self._session_factory() as db:
+            now = datetime.now()
+            tenant_ids = await workload_service.list_tenant_ids(db)
+            effective_tenants = tenant_ids or [1]
+            emitted = 0
+            for tenant_id in effective_tenants:
+                service = AIPMOCadenceService(db, tenant_id=int(tenant_id))
+                project_ids = (
+                    await db.execute(
+                        select(Project.id).where(
+                            and_(
+                                Project.tenant_id == int(tenant_id),
+                                Project.status.in_(["active", "planning", "doing", "todo"]),
+                            )
+                        )
+                    )
+                ).scalars().all()
+
+                if await self._snapshot_due(
+                    db=db,
+                    tenant_id=int(tenant_id),
+                    scope_type="risk_review",
+                    scope_ref=f"tenant:{tenant_id}",
+                    freshness=timedelta(minutes=50),
+                ):
+                    await service.generate_snapshot(cadence_type="risk_review", project_id=None)
+                    emitted += 1
+
+                for project_id in project_ids:
+                    if await self._snapshot_due(
+                        db=db,
+                        tenant_id=int(tenant_id),
+                        scope_type="risk_review",
+                        scope_ref=f"project:{int(project_id)}",
+                        freshness=timedelta(minutes=50),
+                    ):
+                        await service.generate_snapshot(cadence_type="risk_review", project_id=int(project_id))
+                        emitted += 1
+
+                if 8 <= now.hour <= 10:
+                    if await self._snapshot_due(
+                        db=db,
+                        tenant_id=int(tenant_id),
+                        scope_type="daily_standup",
+                        scope_ref=f"tenant:{tenant_id}",
+                        freshness=timedelta(hours=20),
+                    ):
+                        await service.generate_snapshot(cadence_type="daily_standup", project_id=None)
+                        emitted += 1
+                    for project_id in project_ids:
+                        if await self._snapshot_due(
+                            db=db,
+                            tenant_id=int(tenant_id),
+                            scope_type="daily_standup",
+                            scope_ref=f"project:{int(project_id)}",
+                            freshness=timedelta(hours=20),
+                        ):
+                            await service.generate_snapshot(cadence_type="daily_standup", project_id=int(project_id))
+                            emitted += 1
+
+                if 18 <= now.hour <= 20:
+                    if await self._snapshot_due(
+                        db=db,
+                        tenant_id=int(tenant_id),
+                        scope_type="milestone_check",
+                        scope_ref=f"tenant:{tenant_id}",
+                        freshness=timedelta(hours=20),
+                    ):
+                        await service.generate_snapshot(cadence_type="milestone_check", project_id=None)
+                        emitted += 1
+
+                if now.weekday() == 0 and 8 <= now.hour <= 11:
+                    if await self._snapshot_due(
+                        db=db,
+                        tenant_id=int(tenant_id),
+                        scope_type="weekly_review",
+                        scope_ref=f"tenant:{tenant_id}",
+                        freshness=timedelta(days=6),
+                    ):
+                        await service.generate_snapshot(cadence_type="weekly_review", project_id=None)
+                        emitted += 1
+
+            await db.commit()
+            logger.info("ai_pmo_cadence_scheduler_done", tenant_count=len(effective_tenants), snapshot_count=emitted)
+            return {"tenant_count": len(effective_tenants), "snapshot_count": emitted}
+
+    @staticmethod
+    async def _snapshot_due(
+        *,
+        db,
+        tenant_id: int,
+        scope_type: str,
+        scope_ref: str,
+        freshness: timedelta,
+    ) -> bool:
+        row = (
+            await db.execute(
+                select(PortfolioSnapshot.created_at).where(
+                    and_(
+                        PortfolioSnapshot.tenant_id == int(tenant_id),
+                        PortfolioSnapshot.scope_type == scope_type,
+                        PortfolioSnapshot.scope_ref == scope_ref[:64],
+                    )
+                ).order_by(PortfolioSnapshot.id.desc()).limit(1)
+            )
+        ).scalars().first()
+        if row is None:
+            return True
+        return row <= (datetime.now() - freshness)
diff --git a/backend/app/services/signal_capture.py b/backend/app/services/signal_capture.py
new file mode 100644
index 00000000..b53e7db5
--- /dev/null
+++ b/backend/app/services/signal_capture.py
@@ -0,0 +1,2278 @@
+"""
+Proactive Signal Capture Service - 主动式信号捕获服务
+
+核心功能：
+1. 飞书群聊监控 - 自动捕获群内讨论的产品/技术相关内容
+2. 浏览器标签监控 - 检测用户访问的技术文档、竞品页面
+3. GitHub 仓库监控 - 跟踪上游仓库更新、竞品开源项目
+4. HackerNews 自动抓取 - 定时抓取技术前沿信息
+
+每个信号都会自动：
+- 评估战略重要性
+- 提取关键技术栈
+- 生成任务建议
+- 推送给相关角色
+"""
+import asyncio
+import inspect
+import json
+import re
+from uuid import uuid4
+from datetime import datetime, timedelta, timezone
+from typing import Any, Dict, List, Optional
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from sqlalchemy import and_, text
+
+from ..core.logging import get_logger
+from ..database import AsyncSessionLocal
+from ..models import (
+    Signal, Project, ProjectDomain, User,
+    ProjectSignalLink, Task, TruthNode,
+    SkillTag, UserProject, UserWorkload, UserSkillBinding,
+)
+from ..config.ai_prompts import (
+    PROMPT_FEISHU_PROJECT_SIGNAL_CLASSIFICATION,
+    PROMPT_FEISHU_TASK_ORCHESTRATION,
+)
+from .llm import LLMService
+from .notification_service import NotificationService
+from . import workload_service
+from .adaptive_rag_service import AdaptiveRAGService
+from .context_cache import context_cache_service
+from .signal_source_orchestration_service import SignalSourceOrchestrationService
+from .ai_pmo_kernel_service import AIPMOKernelService
+from ..utils.project_members import ensure_project_default_members
+from ..utils.sequence import ensure_project_id_sequence, ensure_signal_id_sequence
+
+
+EMOJI_UNICODE_RE = re.compile(
+    r"[\U0001F300-\U0001F6FF\U0001F700-\U0001F77F\U0001F780-\U0001F7FF"
+    r"\U0001F800-\U0001F8FF\U0001F900-\U0001F9FF\U0001FA00-\U0001FAFF"
+    r"\U00002600-\U000026FF\U00002700-\U000027BF]+"
+)
+ONLY_PUNCT_OR_SPACE_RE = re.compile(r"^[\s\W_]+$", re.UNICODE)
+MENTION_TAG_RE = re.compile(r"</?at[^>]*>", re.IGNORECASE)
+CONTROL_CHAR_RE = re.compile(r"[\x00-\x08\x0B\x0C\x0E-\x1F\x7F]+")
+NON_REQUIREMENT_EXACT = {
+    "收到",
+    "收到啦",
+    "好的",
+    "ok",
+    "okk",
+    "okay",
+    "yes",
+    "no",
+    "1",
+    "2",
+    "3",
+    "thanks",
+    "thx",
+    "赞",
+    "辛苦",
+    "你是谁",
+    "你是誰",
+    "你是谁？",
+    "你是誰？",
+    "说话",
+    "在吗",
+    "在不在",
+    "今天什么日期",
+    "今天几号",
+    "今天星期几",
+    "现在几点",
+    "当前日期",
+}
+NON_REQUIREMENT_CONTAINS = [
+    "你是谁",
+    "什么身份",
+    "身份确认",
+    "今天什么日期",
+    "当前日期",
+    "日期疑问",
+    "询问日期",
+    "现在几点",
+    "几点了",
+    "今天周几",
+    "今天星期几",
+    "说话",
+    "在吗",
+    "在不在",
+]
+REQUIREMENT_HINTS = [
+    "需求",
+    "优化",
+    "改一下",
+    "调整",
+    "修复",
+    "bug",
+    "故障",
+    "报错",
+    "项目",
+    "页面",
+    "接口",
+    "功能",
+    "排版",
+    "样式",
+    "任务",
+    "立项",
+    "上线",
+    "配置",
+]
+
+logger = get_logger(__name__)
+AI_THINK_STAGE_ZH = {
+    "capture_start": "开始处理飞书消息",
+    "capture_skip_empty": "消息为空，忽略",
+    "capture_skip_noise": "识别为噪音消息，忽略",
+    "capture_done_ai_pmo": "AI PMO kernel 已完成决策闭环",
+    "capture_kernel_failed": "AI PMO kernel 执行异常",
+    "capture_kernel_failed_rollback_failed": "kernel 异常后回滚失败",
+    "classification_done": "语义分类完成",
+    "classification_parse_failed": "语义分类结果解析失败",
+    "capture_skip_not_requirement": "非需求消息，忽略",
+    "capture_skip_non_requirement": "判定为非需求，忽略",
+    "signal_created": "信号已创建",
+    "project_link_done": "项目关联完成",
+    "project_link_failed": "项目关联失败",
+    "task_orchestrate_failed": "任务编排失败",
+    "capture_done": "飞书消息处理完成",
+    "pipeline_post_orchestrate_failed": "分流后置任务编排失败",
+    "pipeline_post_update_signal_failed": "分流后置信号结果回写失败",
+    "pipeline_post_source_onboard_failed": "分流后置信号源接入编排失败",
+    "task_plan_llm_done": "任务编排 LLM 返回成功",
+    "task_plan_llm_failed": "任务编排 LLM 调用失败",
+    "task_created": "任务已创建",
+    "charter_auto_created": "自动创建宪章节点",
+    "hn_story_captured": "HackerNews 信号已捕获",
+    "hn_monitor_error": "HackerNews 监控异常",
+    "new_project_created_from_signal": "已根据信号自动创建项目",
+    "draft_new_project_brief_failed": "新项目摘要生成失败，已走兜底",
+}
+
+
+def _build_feishu_rag_conversation_key(
+    *,
+    source_chat_id: Optional[str],
+    source_sender_open_id: Optional[str],
+    sender: Optional[str],
+) -> str:
+    chat = str(source_chat_id or "").strip()
+    if not chat:
+        return ""
+    actor = str(source_sender_open_id or "").strip()
+    if not actor:
+        actor = str(sender or "").strip().lower()
+        actor = re.sub(r"\s+", "", actor)
+    actor = actor[:96] if actor else "unknown"
+    return f"feishu:{chat}:{actor}"
+
+
+class ProactiveSignalCaptureService:
+    """主动式信号捕获服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+        self.llm = LLMService()
+
+    async def capture_from_feishu(
+        self,
+        group_chat_id: str,
+        message_content: str,
+        sender: str,
+        tenant_id: int = 1,
+        proposer_open_id: Optional[str] = None,
+        captured_at: Optional[datetime] = None,
+        mentioned_bot_names: Optional[List[str]] = None,
+        feishu_message_id: Optional[str] = None,
+        chat_type: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """
+        从飞书群聊捕获信号
+
+        触发条件：
+        - 提到产品、竞品、技术关键词
+        - 讨论架构、技术选型
+        - 分享技术文章
+        """
+        trace_id = f"feishu-{uuid4().hex[:12]}"
+        captured_dt = captured_at or datetime.now(timezone.utc)
+        if captured_dt.tzinfo is None:
+            captured_dt = captured_dt.replace(tzinfo=timezone.utc)
+
+        proposer_name = self._normalize_single_line_text(sender or "未知用户", max_len=80) or "未知用户"
+        proposer_open_id = self._normalize_single_line_text(proposer_open_id or "", max_len=128) or None
+        self._emit_ai_think(
+            "capture_start",
+            trace_id=trace_id,
+            tenant_id=tenant_id,
+            group_chat_id=group_chat_id,
+            sender=proposer_name,
+        )
+
+        cleaned_message = self._sanitize_feishu_message(message_content)
+        if not cleaned_message:
+            self._emit_ai_think("capture_skip_empty", trace_id=trace_id)
+            return {"status": "ignored", "reason": "empty_or_emoji_only"}
+        if self._is_non_requirement_message(cleaned_message) and not self._is_question_like(cleaned_message):
+            self._emit_ai_think(
+                "capture_skip_noise",
+                trace_id=trace_id,
+                message_preview=cleaned_message[:120],
+            )
+            return {"status": "ignored", "reason": "non_requirement_noise"}
+
+        kernel = AIPMOKernelService(self.db, tenant_id=tenant_id)
+        try:
+            pipeline_result = await kernel.process_feishu_message(
+                group_chat_id=group_chat_id,
+                message_content=cleaned_message,
+                sender=proposer_name,
+                source_message_id=feishu_message_id,
+                source_sender_open_id=proposer_open_id,
+                captured_at=captured_dt,
+                chat_type=chat_type,
+            )
+            if not bool(pipeline_result.get("handled")):
+                self._emit_ai_think(
+                    "capture_kernel_failed",
+                    trace_id=trace_id,
+                    reason=pipeline_result.get("reason"),
+                    error=pipeline_result.get("error"),
+                )
+                return {
+                    "status": "error",
+                    "reason": str(pipeline_result.get("reason") or "kernel_not_handled"),
+                    "classification": "error",
+                    "project_scope": "error",
+                    "linked_project_ids": [],
+                    "signal_id": None,
+                    "created_project_id": None,
+                    "auto_tasks": [],
+                    "ai_trace_id": trace_id,
+                    "direct_reply_required": True,
+                    "direct_reply_text": "系统处理中出现异常，本次未执行入库。请稍后重试，或补充更明确的目标/对象/动作。",
+                }
+            self._emit_ai_think(
+                "capture_done_ai_pmo",
+                trace_id=trace_id,
+                classification=pipeline_result.get("classification"),
+                linked_project_ids=pipeline_result.get("linked_project_ids", []),
+                work_item_id=pipeline_result.get("work_item_id"),
+            )
+            pipeline_result = await self._apply_pipeline_post_actions(
+                trace_id=trace_id,
+                tenant_id=tenant_id,
+                cleaned_message=cleaned_message,
+                proposer_name=proposer_name,
+                captured_at=captured_dt,
+                pipeline_result=pipeline_result,
+            )
+            try:
+                await self._sync_feishu_recent_cache(
+                    tenant_id=tenant_id,
+                    group_chat_id=group_chat_id,
+                    sender=proposer_name,
+                    sender_open_id=proposer_open_id,
+                    message=cleaned_message,
+                    pipeline_result=pipeline_result,
+                )
+            except Exception as cache_exc:
+                self._emit_ai_think(
+                    "capture_recent_cache_sync_failed",
+                    trace_id=trace_id,
+                    error=str(cache_exc),
+                )
+            try:
+                asyncio.create_task(
+                    AdaptiveRAGService.ingest_feishu_message_async(
+                        tenant_id=tenant_id,
+                        content=cleaned_message,
+                        sender=proposer_name,
+                        chat_id=group_chat_id,
+                        message_id=feishu_message_id,
+                        captured_at=captured_dt,
+                        project_id=(
+                            int(pipeline_result["linked_project_ids"][0])
+                            if isinstance(pipeline_result.get("linked_project_ids"), list)
+                            and pipeline_result.get("linked_project_ids")
+                            else None
+                        ),
+                    )
+                )
+            except Exception:
+                pass
+            return pipeline_result
+        except Exception as exc:
+            self._emit_ai_think("capture_kernel_failed", trace_id=trace_id, error=str(exc))
+            try:
+                await self.db.rollback()
+            except Exception as rollback_exc:
+                self._emit_ai_think(
+                    "capture_kernel_failed_rollback_failed",
+                    trace_id=trace_id,
+                    error=str(rollback_exc),
+                )
+            return {
+                "status": "error",
+                "reason": "kernel_failed",
+                "classification": "error",
+                "project_scope": "error",
+                "linked_project_ids": [],
+                "signal_id": None,
+                "created_project_id": None,
+                "auto_tasks": [],
+                "ai_trace_id": trace_id,
+                "direct_reply_required": True,
+                "direct_reply_text": "系统处理中出现异常，本次未执行入库。请稍后重试，或补充更明确的目标/对象/动作。",
+            }
+
+        projects_context = await self._load_active_projects_context(
+            tenant_id=tenant_id,
+        )
+        classification = await self._analyze_chat_message(
+            message=cleaned_message,
+            projects_context=projects_context,
+            tenant_id=tenant_id,
+        )
+        self._emit_ai_think(
+            "classification_done",
+            trace_id=trace_id,
+            classification=classification.get("classification"),
+            confidence=classification.get("confidence"),
+            matched_project_ids=classification.get("matched_project_ids", []),
+            reason=classification.get("reason"),
+        )
+        if not classification.get("is_requirement", False):
+            self._emit_ai_think(
+                "capture_skip_not_requirement",
+                trace_id=trace_id,
+                reason=classification.get("reason", "non_requirement"),
+            )
+            return {"status": "ignored", "reason": classification.get("reason", "non_requirement")}
+
+        project_scope = str(classification.get("classification") or "pending_classification")
+        if project_scope == "non_requirement":
+            self._emit_ai_think(
+                "capture_skip_non_requirement",
+                trace_id=trace_id,
+                reason=classification.get("reason", "non_requirement"),
+            )
+            return {"status": "ignored", "reason": classification.get("reason", "non_requirement")}
+
+        # 创建信号
+        await ensure_signal_id_sequence(self.db)
+        strategic_payload = {
+            "captured_at": captured_dt.isoformat(),
+            "source_context": {
+                "group_id": group_chat_id,
+                "sender": proposer_name,
+                "sender_open_id": proposer_open_id,
+                "mentioned_bot_names": [
+                    self._normalize_single_line_text(n, max_len=80)
+                    for n in (mentioned_bot_names or [])
+                    if self._normalize_single_line_text(n, max_len=80)
+                ],
+            },
+            "original_message": (message_content or "")[:500],
+            "cleaned_message": cleaned_message[:500],
+            "ai_analysis": {
+                "requirement_project_scope": project_scope,
+                "reason": classification.get("reason", ""),
+                "confidence": classification.get("confidence", 0.0),
+                "matched_project_ids": classification.get("matched_project_ids", []),
+                "matched_project_titles": classification.get("matched_project_titles", []),
+                "summary": classification.get("summary", ""),
+            },
+        }
+
+        signal = Signal(
+            source="Feishu",
+            content=f"[{proposer_name}]: {self._normalize_single_line_text(cleaned_message, max_len=200)}",
+            impact_score=float(classification.get("impact_score") or 0.7),
+            priority=(classification.get("priority") or "medium"),
+            processed=False,
+            strategic_analysis=json.dumps(strategic_payload, ensure_ascii=False),
+            tasks_generated=json.dumps([], ensure_ascii=False),
+            tenant_id=tenant_id
+        )
+        self.db.add(signal)
+        await self.db.commit()
+        await self.db.refresh(signal)
+        self._emit_ai_think("signal_created", trace_id=trace_id, signal_id=signal.id)
+
+        # 自动关联到相关项目（失败不阻断后续通知）
+        linked_project_ids: List[int] = []
+        try:
+            if project_scope == "new_project":
+                new_project_id = await self._auto_create_project_from_new_signal(
+                    signal_id=int(signal.id),
+                    message=cleaned_message,
+                    sender=proposer_name,
+                    tenant_id=tenant_id,
+                    requested_at=captured_dt,
+                )
+                linked_project_ids = [new_project_id] if new_project_id else []
+            else:
+                linked_project_ids = await self._auto_link_to_projects(
+                    signal.id,
+                    classification,
+                    tenant_id,
+                    message_content=cleaned_message,
+                )
+            self._emit_ai_think(
+                "project_link_done",
+                trace_id=trace_id,
+                signal_id=signal.id,
+                linked_project_ids=linked_project_ids,
+            )
+        except Exception as exc:
+            self._emit_ai_think("project_link_failed", trace_id=trace_id, error=str(exc))
+
+        auto_tasks: List[Dict[str, Any]] = []
+        ai_execution: Dict[str, Any] = {
+            "trace_id": trace_id,
+            "status": "skipped",
+            "reason": "no_linked_project",
+            "project_id": None,
+            "summary": "",
+            "tasks_created": 0,
+        }
+        if linked_project_ids:
+            primary_project_id = int(linked_project_ids[0])
+            try:
+                auto_tasks, ai_execution = await self._auto_orchestrate_project_tasks(
+                    trace_id=trace_id,
+                    tenant_id=tenant_id,
+                    project_id=primary_project_id,
+                    signal_id=int(signal.id),
+                    cleaned_message=cleaned_message,
+                    classification=classification,
+                    proposer_name=proposer_name,
+                    captured_at=captured_dt,
+                )
+            except Exception as exc:
+                self._emit_ai_think(
+                    "task_orchestrate_failed",
+                    trace_id=trace_id,
+                    project_id=primary_project_id,
+                    error=str(exc),
+                )
+                ai_execution = {
+                    "trace_id": trace_id,
+                    "status": "failed",
+                    "reason": str(exc),
+                    "project_id": primary_project_id,
+                    "summary": "",
+                    "tasks_created": 0,
+                }
+
+        strategic_payload["ai_execution"] = ai_execution
+        signal.tasks_generated = json.dumps(auto_tasks, ensure_ascii=False)
+        signal.strategic_analysis = json.dumps(strategic_payload, ensure_ascii=False)
+        signal.processed = bool(linked_project_ids or auto_tasks)
+        await self.db.commit()
+
+        # 站内通知：右上角消息提醒
+        target_url = "/ai-steward"
+        if linked_project_ids:
+            target_url = f"/projects/{linked_project_ids[0]}/board"
+        try:
+            notification_service = NotificationService(self.db)
+            await notification_service.notify_signal_created(
+                tenant_id=tenant_id,
+                signal_id=int(signal.id),
+                signal_content=cleaned_message,
+                source="feishu",
+                target_url=target_url,
+                proposer=proposer_name,
+                occurred_at=captured_dt,
+            )
+        except Exception as exc:
+            self._emit_ai_think("notify_failed", trace_id=trace_id, error=str(exc))
+
+        self._emit_ai_think(
+            "capture_done",
+            trace_id=trace_id,
+            signal_id=signal.id,
+            linked_project_ids=linked_project_ids,
+            auto_tasks=len(auto_tasks),
+        )
+        return {
+            "signal_id": str(signal.id),
+            "status": "captured",
+            "impact_score": signal.impact_score,
+            "auto_tasks": auto_tasks,
+            "project_scope": project_scope,
+            "classification": project_scope,
+            "linked_project_ids": linked_project_ids,
+            "created_project_id": linked_project_ids[0] if project_scope == "new_project" and linked_project_ids else None,
+            "proposer": proposer_name,
+            "ai_trace_id": trace_id,
+        }
+
+    async def _sync_feishu_recent_cache(
+        self,
+        *,
+        tenant_id: int,
+        group_chat_id: str,
+        sender: str,
+        sender_open_id: Optional[str],
+        message: str,
+        pipeline_result: Dict[str, Any],
+    ) -> None:
+        if not isinstance(pipeline_result, dict):
+            return
+
+        conversation_key = _build_feishu_rag_conversation_key(
+            source_chat_id=group_chat_id,
+            source_sender_open_id=sender_open_id,
+            sender=sender,
+        )
+        external_session_id = AdaptiveRAGService._external_session_id_from_key(conversation_key)
+        if external_session_id is None:
+            return
+
+        answer_text = str(pipeline_result.get("direct_reply_text") or "").strip()
+        if not answer_text:
+            answer_text = self._build_recent_cache_ack_text(pipeline_result)
+        if not answer_text:
+            return
+
+        recent_messages = await context_cache_service.get_session_messages(
+            tenant_id=int(tenant_id),
+            session_id=int(external_session_id),
+        )
+        recent_turns = context_cache_service.messages_to_turns(recent_messages)
+        if recent_turns:
+            latest_turn = recent_turns[-1]
+            latest_question = str(latest_turn.get("question") or "").strip()
+            latest_answer = str(latest_turn.get("answer") or "").strip()
+            if latest_question == str(message or "").strip() and latest_answer == answer_text:
+                return
+
+        await context_cache_service.append_session_messages(
+            tenant_id=int(tenant_id),
+            session_id=int(external_session_id),
+            question=str(message or "")[:3000],
+            answer=answer_text[:4000],
+        )
+
+    @staticmethod
+    def _build_recent_cache_ack_text(pipeline_result: Dict[str, Any]) -> str:
+        status = str(pipeline_result.get("status") or "").strip().lower()
+        if status == "captured":
+            return "已收到并记录该需求，已进入信号处理队列。"
+        if status == "ignored":
+            return "已收到该消息，本次未进入需求处理队列。"
+        if status == "error":
+            return str(pipeline_result.get("direct_reply_text") or "系统处理中出现异常，请稍后重试。").strip()
+        return "已收到该消息。"
+
+    async def _apply_pipeline_post_actions(
+        self,
+        *,
+        trace_id: str,
+        tenant_id: int,
+        cleaned_message: str,
+        proposer_name: str,
+        captured_at: datetime,
+        pipeline_result: Dict[str, Any],
+    ) -> Dict[str, Any]:
+        if not isinstance(pipeline_result, dict):
+            return pipeline_result
+        if str(pipeline_result.get("status") or "").lower() != "captured":
+            return pipeline_result
+
+        existing_tasks = pipeline_result.get("auto_tasks")
+        if isinstance(existing_tasks, list) and existing_tasks:
+            return pipeline_result
+
+        linked_project_ids = pipeline_result.get("linked_project_ids")
+        if not isinstance(linked_project_ids, list) or not linked_project_ids:
+            return pipeline_result
+
+        signal_id = self._to_int(pipeline_result.get("signal_id"))
+        project_id = self._to_int(linked_project_ids[0])
+        if signal_id is None or project_id is None:
+            return pipeline_result
+
+        signal = (
+            await self.db.execute(
+                select(Signal).where(
+                    and_(
+                        Signal.id == signal_id,
+                        Signal.tenant_id == tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if signal is None:
+            return pipeline_result
+
+        if self._is_question_like(cleaned_message):
+            return pipeline_result
+
+        source_orchestration_result: Optional[Dict[str, Any]] = None
+        if SignalSourceOrchestrationService.looks_like_signal_source_request(cleaned_message):
+            try:
+                source_orchestration_result = await SignalSourceOrchestrationService(self.db).orchestrate_from_signal(
+                    tenant_id=tenant_id,
+                    project_id=project_id,
+                    signal_id=signal_id,
+                    initiating_text=cleaned_message,
+                    requested_via="feishu",
+                    operator_user_id=None,
+                    dry_run=False,
+                )
+            except Exception as exc:
+                self._emit_ai_think(
+                    "pipeline_post_source_onboard_failed",
+                    trace_id=trace_id,
+                    signal_id=signal_id,
+                    project_id=project_id,
+                    error=str(exc),
+                )
+            else:
+                orchestration_status = str(source_orchestration_result.get("status") or "").lower()
+                if orchestration_status in {"completed", "partial", "planned", "pending_manual"}:
+                    auto_tasks = self._build_openclaw_auto_tasks(
+                        source_orchestration_result.get("openclaw"),
+                    )
+                    try:
+                        strategic_payload: Dict[str, Any] = {}
+                        raw_strategic = signal.strategic_analysis
+                        if isinstance(raw_strategic, dict):
+                            strategic_payload = dict(raw_strategic)
+                        elif isinstance(raw_strategic, str) and raw_strategic.strip():
+                            parsed = json.loads(raw_strategic)
+                            if isinstance(parsed, dict):
+                                strategic_payload = parsed
+                        strategic_payload["signal_source_orchestration"] = source_orchestration_result
+                        strategic_payload["ai_execution"] = source_orchestration_result
+                        signal.strategic_analysis = json.dumps(strategic_payload, ensure_ascii=False)
+                        signal.tasks_generated = json.dumps(auto_tasks, ensure_ascii=False)
+                        signal.processed = orchestration_status in {"completed", "partial"}
+                        await self.db.commit()
+                    except Exception as exc:
+                        self._emit_ai_think(
+                            "pipeline_post_update_signal_failed",
+                            trace_id=trace_id,
+                            signal_id=signal_id,
+                            project_id=project_id,
+                            error=str(exc),
+                        )
+                    pipeline_result["auto_tasks"] = auto_tasks
+                    pipeline_result["ai_execution"] = source_orchestration_result
+                    pipeline_result["signal_source_orchestration"] = source_orchestration_result
+                    return pipeline_result
+
+        signal_type = str(signal.signal_type or "").strip()
+        recommended_action = str(signal.recommended_action or "").strip()
+        should_orchestrate = (
+            recommended_action in {"create_charter", "trigger_alert"}
+            or signal_type == "新需求"
+        )
+        if not should_orchestrate:
+            return pipeline_result
+
+        classification_payload: Dict[str, Any] = {
+            "classification": str(pipeline_result.get("classification") or "signal"),
+            "impact_score": float(signal.impact_score or 0.0),
+            "priority": str(signal.priority or "medium"),
+            "confidence": float(signal.confidence or 0.0),
+            "reason": str(signal.summary or "")[:300],
+            "matched_project_ids": [project_id],
+            "matched_project_titles": [],
+        }
+
+        try:
+            auto_tasks, ai_execution = await self._auto_orchestrate_project_tasks(
+                trace_id=f"{trace_id}-pipeline",
+                tenant_id=tenant_id,
+                project_id=project_id,
+                signal_id=signal_id,
+                cleaned_message=cleaned_message,
+                classification=classification_payload,
+                proposer_name=proposer_name,
+                captured_at=captured_at,
+            )
+        except Exception as exc:
+            self._emit_ai_think(
+                "pipeline_post_orchestrate_failed",
+                trace_id=trace_id,
+                signal_id=signal_id,
+                project_id=project_id,
+                error=str(exc),
+            )
+            return pipeline_result
+
+        try:
+            strategic_payload: Dict[str, Any] = {}
+            raw_strategic = signal.strategic_analysis
+            if isinstance(raw_strategic, dict):
+                strategic_payload = dict(raw_strategic)
+            elif isinstance(raw_strategic, str) and raw_strategic.strip():
+                parsed = json.loads(raw_strategic)
+                if isinstance(parsed, dict):
+                    strategic_payload = parsed
+            strategic_payload["ai_execution"] = ai_execution
+            signal.strategic_analysis = json.dumps(strategic_payload, ensure_ascii=False)
+            signal.tasks_generated = json.dumps(auto_tasks, ensure_ascii=False)
+            signal.processed = bool(signal.processed or auto_tasks)
+            await self.db.commit()
+        except Exception as exc:
+            self._emit_ai_think(
+                "pipeline_post_update_signal_failed",
+                trace_id=trace_id,
+                signal_id=signal_id,
+                project_id=project_id,
+                error=str(exc),
+            )
+
+        pipeline_result["auto_tasks"] = auto_tasks
+        pipeline_result["ai_execution"] = ai_execution
+        return pipeline_result
+
+    @staticmethod
+    def _build_openclaw_auto_tasks(openclaw_payload: Any) -> List[Dict[str, Any]]:
+        if not isinstance(openclaw_payload, dict):
+            return []
+        result = openclaw_payload.get("result") if isinstance(openclaw_payload.get("result"), dict) else openclaw_payload
+        execution = result.get("execution") if isinstance(result, dict) and isinstance(result.get("execution"), dict) else {}
+        rows = execution.get("results") if isinstance(execution.get("results"), list) else []
+        items: List[Dict[str, Any]] = []
+        for row in rows:
+            if not isinstance(row, dict):
+                continue
+            operation_code = str(row.get("operation_code") or "").strip()
+            payload = row.get("payload") if isinstance(row.get("payload"), dict) else {}
+            status = str(row.get("status") or "").strip()
+            if operation_code not in {"task.create", "task.assign", "task.prioritize", "task.status_transition"}:
+                continue
+            if status not in {"applied", "planned", "skipped"}:
+                continue
+            items.append(
+                {
+                    "operation_code": operation_code,
+                    "status": status,
+                    "entity_id": row.get("entity_id"),
+                    "task_id": payload.get("task_id"),
+                    "title": payload.get("title"),
+                    "summary": str(row.get("message") or "").strip(),
+                }
+            )
+        return items
+
+    async def capture_from_browser(
+        self,
+        user_id: int,
+        url: str,
+        title: str,
+        content_preview: str,
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        从浏览器标签页捕获信号
+
+        触发条件：
+        - 访问技术文档、API 文档
+        - 浏览竞品网站、产品页面
+        - 查看 GitHub issues/PRs
+        - 阅读技术博客
+        """
+        # AI 分析页面价值
+        analysis = await self._analyze_webpage(
+            url,
+            title,
+            content_preview,
+            tenant_id=tenant_id,
+        )
+
+        if not analysis['is_valuable']:
+            return {'status': 'ignored', 'reason': analysis['reason']}
+
+        # 创建信号
+        await ensure_signal_id_sequence(self.db)
+
+        signal = Signal(
+            source='Browser',
+            content=f"[浏览] {title}: {content_preview[:200]}",
+            impact_score=analysis['impact_score'],
+            priority=analysis['priority'],
+            processed=True,
+            strategic_analysis=json.dumps({
+                'url': url,
+                'captured_at': datetime.now(timezone.utc).isoformat(),
+                'user_id': user_id,
+                'ai_analysis': analysis.get('analysis', {})
+            }),
+            tasks_generated=json.dumps([]),
+            tenant_id=tenant_id
+        )
+        self.db.add(signal)
+        await self.db.commit()
+        await self.db.refresh(signal)
+
+        # 自动关联到相关项目
+        await self._auto_link_to_projects(signal.id, analysis, tenant_id)
+
+        return {
+            'signal_id': str(signal.id),
+            'status': 'captured',
+            'impact_score': signal.impact_score
+        }
+
+    async def capture_from_github(
+        self,
+        repo_url: str,
+        event_type: str,
+        event_data: Dict[str, Any],
+        tenant_id: int = 1
+    ) -> Dict[str, Any]:
+        """
+        从 GitHub 事件捕获信号
+
+        监控事件：
+        - 竞品仓库更新
+        - 上游依赖更新
+        - 技术趋势项目发布
+        """
+        # AI 分析 GitHub 事件
+        analysis = await self._analyze_github_event(
+            repo_url,
+            event_type,
+            event_data,
+            tenant_id=tenant_id,
+        )
+
+        if not analysis['is_valuable']:
+            return {'status': 'ignored', 'reason': analysis['reason']}
+
+        # 创建信号
+        await ensure_signal_id_sequence(self.db)
+
+        signal = Signal(
+            source='GitHub',
+            content=f"[{event_type}] {repo_url}: {event_data.get('title', event_data.get('message', ''))[:200]}",
+            impact_score=analysis['impact_score'],
+            priority=analysis['priority'],
+            processed=True,
+            strategic_analysis=json.dumps({
+                'repo_url': repo_url,
+                'event_type': event_type,
+                'event_data': event_data,
+                'captured_at': datetime.now(timezone.utc).isoformat(),
+                'ai_analysis': analysis.get('analysis', {})
+            }),
+            tasks_generated=json.dumps([]),
+            tenant_id=tenant_id
+        )
+        self.db.add(signal)
+        await self.db.commit()
+        await self.db.refresh(signal)
+
+        return {
+            'signal_id': str(signal.id),
+            'status': 'captured',
+            'impact_score': signal.impact_score
+        }
+
+    async def start_hackernews_monitor(
+        self,
+        interval_minutes: int = 30,
+        tenant_id: int = 1
+    ):
+        """
+        启动 HackerNews 定时监控
+
+        定期抓取 HackerNews 首页，分析技术趋势
+        """
+        while True:
+            try:
+                # 抓取 HackerNews 首页
+                from .fetcher import fetch_hn_stories
+                # 兼容三种实现：
+                # 1) async def -> 直接 await
+                # 2) sync def -> to_thread 防阻塞
+                # 3) sync def 返回 coroutine -> 对返回值再 await
+                if inspect.iscoroutinefunction(fetch_hn_stories):
+                    stories = await fetch_hn_stories(limit=30)
+                else:
+                    result = await asyncio.to_thread(fetch_hn_stories, 30)
+                    stories = await result if inspect.isawaitable(result) else result
+
+                for story in stories[:10]:  # 只分析前10条
+                    # AI 分析每条故事
+                    analysis = await self._analyze_hackernews_story(
+                        story,
+                        tenant_id=tenant_id,
+                    )
+
+                    if analysis.get('is_valuable'):
+                        # 创建信号
+                        await ensure_signal_id_sequence(self.db)
+
+                        signal = Signal(
+                            source='HackerNews',
+                            content=f"HN: {story.get('title', '')[:200]}",
+                            impact_score=analysis['impact_score'],
+                            priority=analysis['priority'],
+                            processed=True,
+                            strategic_analysis=json.dumps({
+                                'hn_url': story.get('url'),
+                                'hn_points': story.get('points'),
+                                'captured_at': datetime.now(timezone.utc).isoformat(),
+                                'ai_analysis': analysis.get('analysis', {})
+                            }),
+                            tasks_generated=json.dumps(analysis.get('suggested_tasks', [])),
+                            tenant_id=tenant_id
+                        )
+                        self.db.add(signal)
+                        await self.db.commit()
+
+                        self._emit_ai_think(
+                            "hn_story_captured",
+                            title=str(story.get("title") or "")[:50],
+                        )
+
+                # 等待下次抓取
+                await asyncio.sleep(interval_minutes * 60)
+
+            except Exception as e:
+                self._emit_ai_think("hn_monitor_error", error=str(e))
+                await asyncio.sleep(interval_minutes * 60)
+
+    async def _load_active_projects_context(self, tenant_id: int) -> List[Dict[str, Any]]:
+        result = await self.db.execute(
+            select(Project).where(
+                Project.status == "active",
+                Project.tenant_id == tenant_id,
+            )
+        )
+        projects = result.scalars().all()
+        rows: List[Dict[str, Any]] = []
+        for project in projects:
+            rows.append(
+                {
+                    "id": int(project.id),
+                    "title": (project.title or project.name or "").strip(),
+                    "description": (project.description or "").strip()[:300],
+                    "status": project.status,
+                }
+            )
+        return rows
+
+    def _emit_ai_think(self, stage: str, **payload: Any) -> None:
+        safe_payload: Dict[str, Any] = {}
+        for key, value in payload.items():
+            safe_payload[key] = self._safe_log_value(value)
+        logger.info("signal_capture_ai_think", stage=stage, **safe_payload)
+        try:
+            stage_zh = AI_THINK_STAGE_ZH.get(stage, stage)
+            now_text = self._format_cn_time(datetime.now(timezone.utc))
+            print(
+                f"[{now_text}][AI思考][SignalCapture] {stage_zh} "
+                + json.dumps({"stage": stage, **safe_payload}, ensure_ascii=False)
+            )
+        except Exception:
+            pass
+
+    @staticmethod
+    def _safe_log_value(value: Any, max_len: int = 600) -> Any:
+        if isinstance(value, (str, int, float, bool)) or value is None:
+            text = str(value) if value is not None else None
+            if text is None:
+                return None
+            return text if len(text) <= max_len else f"{text[:max_len]}...(truncated)"
+        try:
+            text = json.dumps(value, ensure_ascii=False, default=str)
+        except Exception:
+            text = str(value)
+        return text if len(text) <= max_len else f"{text[:max_len]}...(truncated)"
+
+    async def _auto_orchestrate_project_tasks(
+        self,
+        *,
+        trace_id: str,
+        tenant_id: int,
+        project_id: int,
+        signal_id: int,
+        cleaned_message: str,
+        classification: Dict[str, Any],
+        proposer_name: str,
+        captured_at: datetime,
+    ) -> tuple[List[Dict[str, Any]], Dict[str, Any]]:
+        project = (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.id == project_id,
+                        Project.tenant_id == tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not project:
+            return [], {
+                "trace_id": trace_id,
+                "status": "skipped",
+                "reason": "project_not_found",
+                "project_id": project_id,
+                "summary": "",
+                "tasks_created": 0,
+            }
+
+        charter_id = await self._ensure_project_charter_node(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            proposer_name=proposer_name,
+            cleaned_message=cleaned_message,
+            captured_at=captured_at,
+        )
+
+        team_members = await self._load_project_team_context(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            owner_id=int(project.owner_id) if project.owner_id else None,
+        )
+        if not team_members:
+            return [], {
+                "trace_id": trace_id,
+                "status": "skipped",
+                "reason": "no_active_members",
+                "project_id": project_id,
+                "summary": "",
+                "tasks_created": 0,
+            }
+
+        project_payload = {
+            "id": int(project.id),
+            "title": (project.title or project.name or "").strip(),
+            "description": (project.description or "").strip()[:600],
+            "assigned_role": (project.assigned_role or "").strip(),
+        }
+        signal_payload = {
+            "signal_id": signal_id,
+            "message": cleaned_message[:1200],
+            "classification": classification,
+            "proposer": proposer_name,
+            "captured_at": captured_at.isoformat(),
+        }
+        prompt = PROMPT_FEISHU_TASK_ORCHESTRATION.format(
+            project_json=json.dumps(project_payload, ensure_ascii=False),
+            signal_json=json.dumps(signal_payload, ensure_ascii=False),
+            members_json=json.dumps(team_members, ensure_ascii=False),
+        )
+
+        llm_status = "success"
+        llm_error = ""
+        parsed: Dict[str, Any] = {}
+        try:
+            llm_raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="feishu_task_orchestration",
+            )
+            parsed = self._parse_llm_json_response(llm_raw)
+            self._emit_ai_think(
+                "task_plan_llm_done",
+                trace_id=trace_id,
+                project_id=project_id,
+                llm_summary=parsed.get("overall_summary", ""),
+            )
+        except Exception as exc:
+            llm_status = "failed"
+            llm_error = str(exc)
+            self._emit_ai_think(
+                "task_plan_llm_failed",
+                trace_id=trace_id,
+                project_id=project_id,
+                error=llm_error,
+            )
+
+        tasks_payload = parsed.get("tasks") if isinstance(parsed.get("tasks"), list) else []
+        if not tasks_payload:
+            tasks_payload = self._build_task_orchestration_fallback(
+                cleaned_message=cleaned_message,
+                classification=classification,
+            )
+
+        created_items: List[Dict[str, Any]] = []
+        touched_user_ids: set[int] = set()
+        created_count = 0
+        for idx, raw_task in enumerate(tasks_payload[:8]):
+            if not isinstance(raw_task, dict):
+                continue
+            title = self._normalize_single_line_text(
+                str(raw_task.get("title") or raw_task.get("task_name") or ""),
+                max_len=64,
+            )
+            if not title:
+                continue
+            description = self._normalize_multiline_text(
+                str(raw_task.get("description") or cleaned_message),
+                max_len=800,
+            )
+            priority = self._normalize_priority(raw_task.get("priority"))
+            estimated_hours = self._normalize_estimated_hours(raw_task.get("estimated_hours"))
+            required_skill_names = self._normalize_skill_name_list(raw_task.get("required_skill_names"))
+            required_skill_tag_ids = await self._resolve_skill_tag_ids(
+                tenant_id=tenant_id,
+                skill_names=required_skill_names,
+            )
+            ai_recommended_assignee_id = self._to_int(raw_task.get("recommended_assignee_id"))
+            assignee_id, decision = self._pick_task_assignee(
+                team_members=team_members,
+                required_skill_tag_ids=required_skill_tag_ids,
+                ai_recommended_assignee_id=ai_recommended_assignee_id,
+            )
+
+            task = Task(
+                tenant_id=tenant_id,
+                truth_node_id=charter_id,
+                project_id=project_id,
+                assignee_id=assignee_id,
+                title=title,
+                description=description,
+                required_skill_tag_ids=required_skill_tag_ids,
+                status="todo",
+                priority=priority,
+                estimated_hours=estimated_hours,
+                ai_recommended_assignees={
+                    "trace_id": trace_id,
+                    "ai_recommended_assignee_id": ai_recommended_assignee_id,
+                    "final_assignee_id": assignee_id,
+                    "assignee_reason": self._normalize_single_line_text(
+                        str(raw_task.get("assignee_reason") or ""),
+                        max_len=300,
+                    ),
+                    "required_skill_names": required_skill_names,
+                    "decision": decision,
+                    "confidence": float(raw_task.get("confidence") or 0.0),
+                },
+            )
+            self.db.add(task)
+            await self.db.flush()
+            created_count += 1
+
+            assignee_name = ""
+            if assignee_id:
+                touched_user_ids.add(int(assignee_id))
+                match_member = next((m for m in team_members if int(m.get("user_id", 0)) == int(assignee_id)), None)
+                if isinstance(match_member, dict):
+                    assignee_name = str(match_member.get("name") or "")
+
+            item = {
+                "task_id": int(task.id),
+                "task_name": title,
+                "priority": priority,
+                "estimated_hours": float(task.estimated_hours or 0),
+                "required_skill_names": required_skill_names,
+                "required_skill_tag_ids": required_skill_tag_ids,
+                "assignee_id": int(assignee_id) if assignee_id else None,
+                "assignee_name": assignee_name,
+                "reason": decision.get("picked_reason", ""),
+            }
+            created_items.append(item)
+            self._emit_ai_think(
+                "task_created",
+                trace_id=trace_id,
+                project_id=project_id,
+                rank=idx + 1,
+                task_id=task.id,
+                task_name=title,
+                assignee_id=item["assignee_id"],
+                decision=item["reason"],
+            )
+
+        await self.db.commit()
+        for uid in sorted(touched_user_ids):
+            await workload_service.refresh_user_assigned_hours_from_tasks(
+                self.db,
+                tenant_id=tenant_id,
+                user_id=uid,
+            )
+
+        summary = self._normalize_single_line_text(
+            str(parsed.get("overall_summary") or ""),
+            max_len=200,
+        )
+        if not summary:
+            summary = "AI 自动拆解并分配飞书需求任务"
+
+        execution_meta = {
+            "trace_id": trace_id,
+            "status": "success" if created_count > 0 else "skipped",
+            "reason": "ok" if created_count > 0 else "no_valid_tasks",
+            "project_id": project_id,
+            "charter_id": charter_id,
+            "summary": summary,
+            "tasks_created": created_count,
+            "llm_status": llm_status,
+            "llm_error": llm_error[:300] if llm_error else "",
+        }
+        return created_items, execution_meta
+
+    async def _ensure_project_charter_node(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        proposer_name: str,
+        cleaned_message: str,
+        captured_at: datetime,
+    ) -> int:
+        existing = (
+            await self.db.execute(
+                select(TruthNode).where(
+                    and_(
+                        TruthNode.tenant_id == tenant_id,
+                        TruthNode.project_id == project_id,
+                        TruthNode.node_type == "charter",
+                    )
+                ).order_by(TruthNode.updated_at.desc(), TruthNode.id.desc())
+            )
+        ).scalars().first()
+        if existing:
+            return int(existing.id)
+
+        project = (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.id == project_id,
+                        Project.tenant_id == tenant_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not project:
+            raise ValueError("project_not_found_for_charter")
+
+        creator_id = await self._pick_truth_node_creator(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            owner_id=int(project.owner_id) if project.owner_id else None,
+        )
+        node = TruthNode(
+            tenant_id=tenant_id,
+            node_type="charter",
+            project_id=project_id,
+            created_by=creator_id,
+            title=self._normalize_single_line_text(
+                f"{project.title or project.name or '项目'} 自动宪章",
+                max_len=120,
+            ),
+            content={
+                "rationale": "由飞书信号自动触发的初版执行宪章",
+                "source": "feishu_auto_orchestration",
+                "proposer": proposer_name,
+                "captured_at": captured_at.isoformat(),
+                "seed_message": cleaned_message[:500],
+                "success_criteria": ["建立可执行任务链", "明确负责人", "控制团队负载"],
+                "constraints": ["后续可由管理员更新细化"],
+            },
+            version=1,
+            status="active",
+        )
+        self.db.add(node)
+        await self.db.flush()
+        self._emit_ai_think(
+            "charter_auto_created",
+            project_id=project_id,
+            charter_id=node.id,
+            created_by=creator_id,
+        )
+        return int(node.id)
+
+    async def _pick_truth_node_creator(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        owner_id: Optional[int],
+    ) -> int:
+        if owner_id:
+            owner = (
+                await self.db.execute(
+                    select(User.id).where(
+                        and_(
+                            User.tenant_id == tenant_id,
+                            User.id == owner_id,
+                            User.status == "active",
+                        )
+                    ).limit(1)
+                )
+            ).scalars().first()
+            if owner:
+                return int(owner)
+
+        project_member = (
+            await self.db.execute(
+                select(UserProject.user_id)
+                .join(
+                    User,
+                    and_(
+                        User.id == UserProject.user_id,
+                        User.tenant_id == tenant_id,
+                        User.status == "active",
+                        User.is_active.is_(True),
+                    ),
+                )
+                .where(
+                    and_(
+                        UserProject.tenant_id == tenant_id,
+                        UserProject.project_id == project_id,
+                    )
+                )
+                .order_by(UserProject.user_id.asc())
+                .limit(1)
+            )
+        ).scalars().first()
+        if project_member:
+            return int(project_member)
+
+        admin_user = (
+            await self.db.execute(
+                select(User.id).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.status == "active",
+                        User.is_active.is_(True),
+                        User.role.in_(["BOSS", "ADMIN"]),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+            )
+        ).scalars().first()
+        if admin_user:
+            return int(admin_user)
+
+        any_active = (
+            await self.db.execute(
+                select(User.id).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.status == "active",
+                        User.is_active.is_(True),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+            )
+        ).scalars().first()
+        if any_active:
+            return int(any_active)
+        raise ValueError("no_active_user_for_charter_creator")
+
+    async def _load_project_team_context(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        owner_id: Optional[int],
+    ) -> List[Dict[str, Any]]:
+        member_ids_result = await self.db.execute(
+            select(UserProject.user_id).where(
+                and_(
+                    UserProject.tenant_id == tenant_id,
+                    UserProject.project_id == project_id,
+                )
+            )
+        )
+        member_ids = {int(uid) for uid in member_ids_result.scalars().all()}
+        if owner_id:
+            member_ids.add(int(owner_id))
+        if not member_ids:
+            manager_result = await self.db.execute(
+                select(User.id).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.status == "active",
+                        User.is_active.is_(True),
+                        User.role.in_(["BOSS", "ADMIN"]),
+                    )
+                ).limit(5)
+            )
+            member_ids = {int(uid) for uid in manager_result.scalars().all()}
+        if not member_ids:
+            return []
+
+        users = (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.id.in_(sorted(member_ids)),
+                        User.status == "active",
+                        User.is_active.is_(True),
+                    )
+                )
+            )
+        ).scalars().all()
+        if not users:
+            return []
+
+        user_ids = [int(u.id) for u in users]
+        workload_rows = (
+            await self.db.execute(
+                select(UserWorkload).where(
+                    and_(
+                        UserWorkload.tenant_id == tenant_id,
+                        UserWorkload.user_id.in_(user_ids),
+                    )
+                )
+            )
+        ).scalars().all()
+        workload_map = {int(item.user_id): item for item in workload_rows}
+
+        skill_rows = (
+            await self.db.execute(
+                select(
+                    UserSkillBinding.user_id,
+                    SkillTag.id,
+                    SkillTag.name,
+                    SkillTag.slug,
+                    UserSkillBinding.proficiency,
+                )
+                .join(SkillTag, SkillTag.id == UserSkillBinding.skill_tag_id)
+                .where(
+                    and_(
+                        UserSkillBinding.tenant_id == tenant_id,
+                        UserSkillBinding.user_id.in_(user_ids),
+                        SkillTag.tenant_id == tenant_id,
+                    )
+                )
+            )
+        ).all()
+
+        skill_map: Dict[int, List[Dict[str, Any]]] = {}
+        for uid, sid, name, slug, proficiency in skill_rows:
+            skill_map.setdefault(int(uid), []).append(
+                {
+                    "skill_id": int(sid),
+                    "name": str(name or "").strip(),
+                    "slug": str(slug or "").strip(),
+                    "proficiency": int(proficiency or 0),
+                }
+            )
+
+        members: List[Dict[str, Any]] = []
+        for user in users:
+            uid = int(user.id)
+            workload_row = workload_map.get(uid)
+            member_skills = skill_map.get(uid, [])
+            members.append(
+                {
+                    "user_id": uid,
+                    "name": str(user.name or user.username or uid),
+                    "role": str(user.role or ""),
+                    "workload_ratio": round(
+                        float(
+                            workload_row.workload_ratio
+                            if workload_row is not None
+                            else float(user.workload or 0.0)
+                        ),
+                        2,
+                    ),
+                    "current_assigned_hours": round(
+                        float(workload_row.current_assigned_hours or 0.0) if workload_row else 0.0,
+                        2,
+                    ),
+                    "max_available_hours_per_week": round(
+                        float(workload_row.max_available_hours_per_week or 40.0) if workload_row else 40.0,
+                        2,
+                    ),
+                    "skill_tag_ids": [int(item["skill_id"]) for item in member_skills],
+                    "skill_names": [str(item["name"]) for item in member_skills],
+                    "skills": member_skills,
+                }
+            )
+        members.sort(key=lambda x: (float(x.get("workload_ratio") or 0.0), int(x.get("user_id") or 0)))
+        return members
+
+    async def _resolve_skill_tag_ids(
+        self,
+        *,
+        tenant_id: int,
+        skill_names: List[str],
+    ) -> List[int]:
+        if not skill_names:
+            return []
+        result = await self.db.execute(
+            select(SkillTag).where(SkillTag.tenant_id == tenant_id)
+        )
+        tags = result.scalars().all()
+        if not tags:
+            return []
+
+        exact: Dict[str, int] = {}
+        fuzzy: List[tuple[str, int]] = []
+        for tag in tags:
+            name = str(tag.name or "").strip().lower()
+            slug = str(tag.slug or "").strip().lower()
+            if name:
+                exact[name] = int(tag.id)
+                fuzzy.append((name, int(tag.id)))
+            if slug:
+                exact[slug] = int(tag.id)
+                fuzzy.append((slug, int(tag.id)))
+
+        resolved_ids: List[int] = []
+        for raw in skill_names:
+            key = str(raw or "").strip().lower()
+            if not key:
+                continue
+            if key in exact:
+                resolved_ids.append(exact[key])
+                continue
+            matched = None
+            for candidate, sid in fuzzy:
+                if key in candidate or candidate in key:
+                    matched = sid
+                    break
+            if matched:
+                resolved_ids.append(int(matched))
+        return sorted(set(resolved_ids))
+
+    def _pick_task_assignee(
+        self,
+        *,
+        team_members: List[Dict[str, Any]],
+        required_skill_tag_ids: List[int],
+        ai_recommended_assignee_id: Optional[int],
+    ) -> tuple[Optional[int], Dict[str, Any]]:
+        if not team_members:
+            return None, {"picked_reason": "no_team_members", "candidate_scores": []}
+
+        candidates: List[Dict[str, Any]] = []
+        for member in team_members:
+            uid = self._to_int(member.get("user_id"))
+            if not uid:
+                continue
+            skill_ids = {int(v) for v in (member.get("skill_tag_ids") or []) if self._to_int(v)}
+            workload_ratio = float(member.get("workload_ratio") or 0.0)
+            if required_skill_tag_ids:
+                matched = sum(1 for sid in required_skill_tag_ids if sid in skill_ids)
+                skill_score = matched / len(required_skill_tag_ids)
+            else:
+                skill_score = 0.4
+            workload_score = max(0.0, 140.0 - workload_ratio) / 140.0
+            ai_bonus = 0.15 if ai_recommended_assignee_id and uid == int(ai_recommended_assignee_id) else 0.0
+            overload_penalty = 0.2 if workload_ratio >= 150 else 0.0
+            total_score = (skill_score * 0.7) + (workload_score * 0.3) + ai_bonus - overload_penalty
+            candidates.append(
+                {
+                    "user_id": uid,
+                    "name": member.get("name"),
+                    "skill_score": round(skill_score, 4),
+                    "workload_ratio": round(workload_ratio, 2),
+                    "total_score": round(total_score, 4),
+                }
+            )
+
+        candidates.sort(key=lambda x: (-x["total_score"], x["workload_ratio"], x["user_id"]))
+        if not candidates:
+            return None, {"picked_reason": "no_valid_candidate", "candidate_scores": []}
+
+        top = candidates[0]
+        if float(top.get("workload_ratio") or 0.0) > 160 and float(top.get("skill_score") or 0.0) < 0.5:
+            return None, {
+                "picked_reason": "all_candidates_overloaded",
+                "candidate_scores": candidates[:3],
+            }
+        picked_uid = int(top["user_id"])
+        return picked_uid, {
+            "picked_reason": (
+                f"score={top['total_score']},skill={top['skill_score']},"
+                f"workload={top['workload_ratio']}"
+            ),
+            "candidate_scores": candidates[:3],
+        }
+
+    def _build_task_orchestration_fallback(
+        self,
+        *,
+        cleaned_message: str,
+        classification: Dict[str, Any],
+    ) -> List[Dict[str, Any]]:
+        summary = self._normalize_single_line_text(
+            str(classification.get("summary") or cleaned_message),
+            max_len=120,
+        )
+        return [
+            {
+                "title": "需求澄清与范围确认",
+                "description": f"基于飞书信号补全需求边界、验收口径与优先级。上下文：{summary}",
+                "priority": "high",
+                "estimated_hours": 4,
+                "required_skill_names": ["产品管理", "项目管理"],
+                "recommended_assignee_id": None,
+                "assignee_reason": "fallback_without_llm",
+                "confidence": 0.0,
+            },
+            {
+                "title": "核心方案落地",
+                "description": f"围绕需求信号推进可交付实现，拆分后端/前端/AI 子任务并排期。上下文：{summary}",
+                "priority": "medium",
+                "estimated_hours": 12,
+                "required_skill_names": ["FastAPI", "React", "LangChain"],
+                "recommended_assignee_id": None,
+                "assignee_reason": "fallback_without_llm",
+                "confidence": 0.0,
+            },
+        ]
+
+    @staticmethod
+    def _normalize_priority(value: Any) -> str:
+        priority = str(value or "medium").strip().lower()
+        if priority in {"p0", "critical", "high"}:
+            return "high"
+        if priority in {"p2", "low"}:
+            return "low"
+        return "medium"
+
+    @staticmethod
+    def _normalize_estimated_hours(value: Any) -> Optional[float]:
+        try:
+            hours = float(value)
+        except Exception:
+            return 8.0
+        if hours <= 0:
+            return 1.0
+        return round(min(hours, 80.0), 2)
+
+    def _normalize_skill_name_list(self, value: Any) -> List[str]:
+        if isinstance(value, list):
+            names = [self._normalize_single_line_text(str(item or ""), max_len=60) for item in value]
+            return [name for name in names if name]
+        if isinstance(value, str):
+            split = re.split(r"[,，/|]", value)
+            names = [self._normalize_single_line_text(part, max_len=60) for part in split]
+            return [name for name in names if name]
+        return []
+
+    @staticmethod
+    def _to_int(value: Any) -> Optional[int]:
+        try:
+            if value is None or value == "":
+                return None
+            return int(value)
+        except Exception:
+            return None
+
+    @staticmethod
+    def _parse_llm_json_response(raw_response: str) -> Dict[str, Any]:
+        content = (raw_response or "").strip()
+        if not content:
+            raise ValueError("empty llm response")
+        if content.startswith("```"):
+            parts = content.split("```")
+            if len(parts) >= 2:
+                content = parts[1].strip()
+            if content.startswith("json"):
+                content = content[4:].strip()
+        try:
+            parsed = json.loads(content.strip())
+            if isinstance(parsed, dict):
+                return parsed
+        except Exception:
+            pass
+
+        # 兼容模型返回“解释文本 + JSON”场景：抽取首个 JSON 对象再解析
+        obj_text = ProactiveSignalCaptureService._extract_first_json_object(content)
+        if not obj_text:
+            raise ValueError("empty llm response")
+        parsed = json.loads(obj_text)
+        if isinstance(parsed, dict):
+            return parsed
+        raise ValueError("llm json is not object")
+
+    @staticmethod
+    def _extract_first_json_object(text: str) -> Optional[str]:
+        raw = str(text or "")
+        start = raw.find("{")
+        if start < 0:
+            return None
+        depth = 0
+        in_str = False
+        escaped = False
+        for i in range(start, len(raw)):
+            ch = raw[i]
+            if in_str:
+                if escaped:
+                    escaped = False
+                elif ch == "\\":
+                    escaped = True
+                elif ch == '"':
+                    in_str = False
+                continue
+            if ch == '"':
+                in_str = True
+                continue
+            if ch == "{":
+                depth += 1
+            elif ch == "}":
+                depth -= 1
+                if depth == 0:
+                    return raw[start : i + 1]
+        return None
+
+    async def _analyze_chat_message(
+        self,
+        *,
+        message: str,
+        projects_context: List[Dict[str, Any]],
+        tenant_id: int,
+    ) -> Dict[str, Any]:
+        """AI 语义分类：existing/new/pending/non_requirement。"""
+        prompt = PROMPT_FEISHU_PROJECT_SIGNAL_CLASSIFICATION.format(
+            projects_json=json.dumps(projects_context, ensure_ascii=False),
+            message=message[:800],
+        )
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="feishu_project_signal_classification",
+            )
+            try:
+                parsed = self._parse_llm_json_response(response)
+            except Exception as parse_exc:
+                self._emit_ai_think(
+                    "classification_parse_failed",
+                    error=str(parse_exc),
+                    response_preview=(response or "")[:500],
+                )
+                raise
+            classification = str(parsed.get("classification") or "pending_classification").lower()
+            if classification not in {
+                "existing_project",
+                "new_project",
+                "pending_classification",
+                "non_requirement",
+            }:
+                classification = "pending_classification"
+
+            matched_project_ids: List[int] = []
+            for item in parsed.get("matched_project_ids", []) or []:
+                try:
+                    matched_project_ids.append(int(item))
+                except Exception:
+                    continue
+
+            matched_titles = [str(x) for x in (parsed.get("matched_project_titles") or []) if str(x).strip()]
+            if classification == "existing_project" and not matched_project_ids:
+                classification = "pending_classification"
+            # 对“明确新建项目意图”进行升级，避免模型过度保守地给 pending。
+            if classification == "pending_classification" and self._has_explicit_new_project_intent(message):
+                classification = "new_project"
+            raw_is_requirement = parsed.get("is_requirement")
+            if isinstance(raw_is_requirement, bool):
+                is_requirement = raw_is_requirement
+            elif isinstance(raw_is_requirement, str):
+                is_requirement = raw_is_requirement.strip().lower() in {"1", "true", "yes", "y"}
+            else:
+                is_requirement = classification != "non_requirement"
+            if classification in {"existing_project", "new_project", "pending_classification"}:
+                is_requirement = True
+            if classification == "non_requirement":
+                is_requirement = False
+            return {
+                "is_requirement": is_requirement,
+                "reason": str(parsed.get("reason", "") or "AI 判定"),
+                "confidence": float(parsed.get("confidence") or 0.0),
+                "classification": classification,
+                "matched_project_ids": matched_project_ids,
+                "matched_project_titles": matched_titles,
+                "impact_score": float(parsed.get("impact_score") or 0.7),
+                "priority": str(parsed.get("priority", "medium") or "medium"),
+                "summary": str(parsed.get("summary") or "")[:200],
+            }
+        except Exception as exc:
+            self._emit_ai_think("classification_failed", error=str(exc), message_preview=message[:120])
+            fallback_new_project = self._has_explicit_new_project_intent(message)
+            return {
+                "is_requirement": bool(fallback_new_project),
+                "reason": "AI 分析失败，启用保守策略",
+                "confidence": 0.0,
+                "classification": "new_project" if fallback_new_project else "non_requirement",
+                "matched_project_ids": [],
+                "matched_project_titles": [],
+                "impact_score": 0.6,
+                "priority": "medium",
+                "summary": "",
+            }
+
+    @staticmethod
+    def _sanitize_feishu_message(message: str) -> str:
+        text = ProactiveSignalCaptureService._normalize_single_line_text(message or "", max_len=2000)
+        if not text:
+            return ""
+
+        text = MENTION_TAG_RE.sub(" ", text)
+        text = text.replace("@所有人", " ").replace("@everyone", " ")
+        text = re.sub(r"@[_a-zA-Z0-9\-]+", " ", text)
+        text = EMOJI_UNICODE_RE.sub(" ", text)
+        text = re.sub(r"\s+", " ", text).strip()
+
+        if not text or ONLY_PUNCT_OR_SPACE_RE.match(text):
+            return ""
+        return text
+
+    @staticmethod
+    def _normalize_single_line_text(text: str, *, max_len: int = 500) -> str:
+        content = str(text or "")
+        if not content:
+            return ""
+        content = content.replace("\ufffd", "")
+        content = CONTROL_CHAR_RE.sub(" ", content)
+        content = re.sub(r"\s+", " ", content).strip()
+        if max_len > 0:
+            content = content[:max_len]
+        return content
+
+    @staticmethod
+    def _normalize_multiline_text(text: str, *, max_len: int = 2000) -> str:
+        content = str(text or "")
+        if not content:
+            return ""
+        content = content.replace("\ufffd", "")
+        content = content.replace("\r\n", "\n").replace("\r", "\n")
+        content = CONTROL_CHAR_RE.sub(" ", content)
+        content = re.sub(r"[ \t]+", " ", content)
+        content = re.sub(r"\n{3,}", "\n\n", content).strip()
+        if max_len > 0:
+            content = content[:max_len]
+        return content
+
+    @staticmethod
+    def _format_cn_time(dt: datetime) -> str:
+        safe_dt = dt
+        if safe_dt.tzinfo is None:
+            safe_dt = safe_dt.replace(tzinfo=timezone.utc)
+        cn_tz = timezone(timedelta(hours=8))
+        return safe_dt.astimezone(cn_tz).strftime("%Y-%m-%d %H:%M:%S")
+
+    @staticmethod
+    def _is_non_requirement_message(message: str) -> bool:
+        content = (message or "").strip().lower()
+        if not content:
+            return True
+        normalized = re.sub(r"\s+", "", content)
+        if content in NON_REQUIREMENT_EXACT:
+            return True
+        if normalized in NON_REQUIREMENT_EXACT:
+            return True
+        if any(k in normalized for k in NON_REQUIREMENT_CONTAINS) and not any(
+            hint in normalized for hint in REQUIREMENT_HINTS
+        ):
+            return True
+        if len(content) <= 2 and content.isdigit():
+            return True
+        non_requirement_keywords = [
+            "收到",
+            "辛苦了",
+            "赞",
+            "哈哈",
+            "good",
+            "nice",
+            "ok",
+        ]
+        if len(content) <= 8 and any(k in content for k in non_requirement_keywords):
+            return True
+        if ONLY_PUNCT_OR_SPACE_RE.match(content):
+            return True
+        return False
+
+    @staticmethod
+    def _has_explicit_new_project_intent(message: str) -> bool:
+        content = (message or "").strip().lower()
+        if not content:
+            return False
+        explicit_phrases = [
+            "新建项目",
+            "新增项目",
+            "新开项目",
+            "独立立项",
+            "重新立项",
+            "单独立项",
+            "开个新项目",
+            "做个新项目",
+            "从0到1",
+        ]
+        if any(p in content for p in explicit_phrases):
+            return True
+        # 兼容常见中文表达：“新增个需求，关于 xxx 的项目”
+        if ("新增" in content or "新增个需求" in content) and "项目" in content:
+            return True
+        return False
+
+    @staticmethod
+    def _is_question_like(message: str) -> bool:
+        text = str(message or "").strip().lower()
+        if not text:
+            return False
+        if "?" in text or "？" in text:
+            return True
+        hints = ("吗", "怎么", "如何", "为什么", "谁", "多少", "是否", "有没有", "几号", "几点", "周几")
+        return any(k in text for k in hints)
+
+    def _rule_based_chat_signal(self, message: str) -> Optional[Dict[str, Any]]:
+        _ = message
+        return None
+
+    async def _analyze_webpage(self, url: str, title: str, content: str, tenant_id: int) -> Dict[str, Any]:
+        """AI 分析网页价值"""
+        prompt = f"""分析以下网页内容是否有价值：
+
+标题：{title}
+URL: {url}
+内容摘要：{content[:500]}
+
+请判断：
+1. 是否是技术文档、竞品信息、行业趋势？
+2. 是否值得团队关注？
+3. 影响程度（0-1）？
+
+以 JSON 格式返回：
+{{
+    "is_valuable": boolean,
+    "reason": "原因",
+    "impact_score": 0.8,
+    "priority": "high/medium/low",
+    "analysis": {{}},
+    "suggested_tasks": [{{"task_name": "任务", "type": "backend", "estimated_hours": 8}}]
+}}
+
+如果不是有价值内容，is_valuable 设为 false。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except:
+            return {'is_valuable': False, 'reason': 'AI 分析失败'}
+
+    async def _analyze_github_event(self, repo_url: str, event_type: str, event_data: Dict, tenant_id: int) -> Dict[str, Any]:
+        """AI 分析 GitHub 事件"""
+        description = f"{event_type}: {event_data.get('title', event_data.get('message', ''))}"
+
+        prompt = f"""分析以下 GitHub 事件是否有价值：
+
+仓库：{repo_url}
+事件：{description}
+
+请判断：
+1. 是否是技术栈更新、安全修复、重要功能？
+2. 是否影响我们的项目？
+3. 影响程度（0-1）？
+
+以 JSON 格式返回：
+{{
+    "is_valuable": boolean,
+    "reason": "原因",
+    "impact_score": 0.8,
+    "priority": "high/medium/low",
+    "analysis": {{}},
+    "suggested_tasks": [{{"task_name": "任务", "type": "backend", "estimated_hours": 8}}]
+}}
+
+如果不是有价值内容，is_valuable 设为 false。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            return json.loads(response.strip())
+        except:
+            return {'is_valuable': False, 'reason': 'AI 分析失败'}
+
+    async def _analyze_hackernews_story(self, story: Dict, tenant_id: int) -> Dict[str, Any]:
+        """AI 分析 HackerNews 故事"""
+        prompt = f"""分析以下 HackerNews 故事是否是重要技术趋势：
+
+标题：{story.get('title', '')}
+链接：{story.get('url', '')}
+得分：{story.get('points', 0)}
+
+请判断：
+1. 是否是行业趋势、技术突破？
+2. 是否值得我们关注？
+3. 影响程度（0-1）？
+
+以 JSON 格式返回：
+{{
+    "is_strategic": boolean,
+    "impact_score": 0.8,
+    "priority": "high/medium/low",
+    "analysis": {{}},
+    "suggested_tasks": [{{"task_name": "任务", "type": "backend", "estimated_hours": 8}}]
+}}
+
+如果不是战略性内容，is_strategic 设为 false。"""
+
+        try:
+            response = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+            )
+            response = response.strip()
+            if response.startswith('```'):
+                response = response.split('```')[1]
+                if response.startswith('json'):
+                    response = response[4:]
+            result = json.loads(response.strip())
+            # 重命名字段以匹配其他方法
+            return {
+                'is_valuable': result.get('is_strategic', False),
+                'reason': result.get('reason', ''),
+                'impact_score': result.get('impact_score', 0.5),
+                'priority': result.get('priority', 'low'),
+                'analysis': result.get('analysis', {}),
+                'suggested_tasks': result.get('suggested_tasks', [])
+            }
+        except:
+            return {'is_valuable': False, 'reason': 'AI 分析失败'}
+
+    async def _auto_link_to_projects(
+        self,
+        signal_id: int,
+        analysis: Dict[str, Any],
+        tenant_id: int,
+        message_content: Optional[str] = None,
+    ) -> List[int]:
+        """自动关联信号到相关项目"""
+        _ = message_content
+        classification = str(analysis.get("classification") or "").lower()
+        if classification in {"new_project", "pending_classification", "non_requirement"}:
+            return []
+
+        if classification == "existing_project":
+            matched_ids: List[int] = []
+            for pid in analysis.get("matched_project_ids", []) or []:
+                try:
+                    matched_ids.append(int(pid))
+                except Exception:
+                    continue
+            matched_ids = sorted(set(matched_ids))
+            if not matched_ids:
+                return []
+
+            result = await self.db.execute(
+                select(Project.id).where(
+                    and_(
+                        Project.id.in_(matched_ids),
+                        Project.status == "active",
+                        Project.tenant_id == tenant_id,
+                    )
+                )
+            )
+            valid_ids = {int(v) for v in result.scalars().all()}
+            if not valid_ids:
+                return []
+            confidence = float(analysis.get("confidence") or 0.0)
+            relevance = max(0.5, min(0.98, 0.55 + confidence * 0.4))
+            link_type = "ai_existing_project_signal"
+        else:
+            result = await self.db.execute(
+                select(Project.id).where(
+                    and_(Project.status == "active", Project.tenant_id == tenant_id)
+                )
+            )
+            valid_ids = {int(v) for v in result.scalars().all()}
+            if not valid_ids:
+                return []
+            relevance = 0.7
+            link_type = "auto_captured"
+
+        linked_project_ids: List[int] = []
+
+        for project_id in sorted(valid_ids):
+            existing = await self.db.execute(
+                text("SELECT id FROM project_signal_links WHERE project_id = :project_id AND signal_id = :signal_id"),
+                {"project_id": project_id, "signal_id": signal_id},
+            )
+            if existing.first() is None:
+                from ..models import ProjectSignalLink
+                link = ProjectSignalLink(
+                    project_id=project_id,
+                    signal_id=signal_id,
+                    link_type=link_type,
+                    relevance_score=relevance,
+                    tenant_id=tenant_id,
+                )
+                self.db.add(link)
+                linked_project_ids.append(project_id)
+
+        await self.db.commit()
+        return linked_project_ids
+
+    async def _auto_create_project_from_new_signal(
+        self,
+        *,
+        signal_id: int,
+        message: str,
+        sender: str,
+        tenant_id: int,
+        requested_at: Optional[datetime] = None,
+    ) -> Optional[int]:
+        """当分类为 new_project 时，自动创建项目并关联当前信号。"""
+        requested_dt = requested_at or datetime.now(timezone.utc)
+        if requested_dt.tzinfo is None:
+            requested_dt = requested_dt.replace(tzinfo=timezone.utc)
+        project_brief = await self._draft_new_project_brief(
+            message=message,
+            tenant_id=tenant_id,
+            sender=sender,
+            requested_at=requested_dt,
+        )
+        title = self._normalize_single_line_text(project_brief.get("title") or "", max_len=120)
+        if not title:
+            title = self._normalize_single_line_text(message or "飞书新需求立项", max_len=40) or "飞书新需求立项"
+        raw_desc = self._normalize_multiline_text(project_brief.get("description") or message or "", max_len=1200)
+        requested_at_text = self._format_cn_time(requested_dt)
+        description = self._normalize_multiline_text(
+            f"需求提出人：{sender}\n"
+            f"提出时间：{requested_at_text}\n"
+            f"需求来源：飞书群聊\n\n"
+            f"需求说明：{raw_desc or self._normalize_multiline_text(message, max_len=500)}",
+            max_len=1800,
+        )
+        assigned_role = str(project_brief.get("primary_role") or "ai_engineer").strip()[:50] or "ai_engineer"
+
+        owner_id: Optional[int] = None
+        owner_result = await self.db.execute(
+            select(User.id).where(
+                and_(
+                    User.tenant_id == tenant_id,
+                    User.status == "active",
+                    User.role.in_(["BOSS", "ADMIN"]),
+                )
+            ).order_by(User.id.asc()).limit(1)
+        )
+        owner_id = owner_result.scalar_one_or_none()
+
+        await ensure_project_id_sequence(self.db)
+        project = Project(
+            name=title[:80] or "新项目",
+            title=title,
+            description=description,
+            owner_id=owner_id,
+            status="active",
+            assigned_role=assigned_role,
+            signal_count=1,
+            tenant_id=tenant_id,
+        )
+        self.db.add(project)
+        await self.db.flush()
+        await ensure_project_default_members(
+            self.db,
+            tenant_id=tenant_id,
+            project_id=int(project.id),
+            owner_id=int(project.owner_id) if project.owner_id is not None else None,
+            assigned_role=project.assigned_role,
+        )
+
+        existing = await self.db.execute(
+            text("SELECT id FROM project_signal_links WHERE project_id = :project_id AND signal_id = :signal_id"),
+            {"project_id": int(project.id), "signal_id": int(signal_id)},
+        )
+        if existing.first() is None:
+            link = ProjectSignalLink(
+                project_id=int(project.id),
+                signal_id=int(signal_id),
+                link_type="ai_new_project_signal",
+                relevance_score=0.98,
+                tenant_id=tenant_id,
+            )
+            self.db.add(link)
+
+        await self.db.execute(
+            text("UPDATE signals SET processed = :processed WHERE id = :signal_id"),
+            {"processed": True, "signal_id": int(signal_id)},
+        )
+        await self.db.commit()
+
+        self._emit_ai_think(
+            "new_project_created_from_signal",
+            signal_id=signal_id,
+            project_id=int(project.id),
+            sender=sender,
+        )
+        return int(project.id)
+
+    async def _draft_new_project_brief(
+        self,
+        *,
+        message: str,
+        tenant_id: int,
+        sender: str,
+        requested_at: datetime,
+    ) -> Dict[str, Any]:
+        requested_at_text = self._format_cn_time(requested_at)
+        prompt = f"""你是 PMFlow 的立项助手。基于下面需求消息，提炼新项目基础信息。
+
+需求提出人：{sender}
+提出时间：{requested_at_text}
+
+需求消息：
+{message[:1000]}
+
+输出 JSON：
+{{
+  "title": "项目标题（15-30字）",
+  "description": "项目简介（80-200字）",
+  "primary_role": "ai_engineer|backend|frontend|devops"
+}}
+
+只输出 JSON，不要 markdown，不要解释，不要乱码。"""
+        try:
+            raw = await self.llm.generate(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="feishu_new_project_brief",
+            )
+            parsed = self._parse_llm_json_response(raw)
+            role = str(parsed.get("primary_role") or "ai_engineer").strip().lower()
+            if role not in {"ai_engineer", "backend", "frontend", "devops"}:
+                role = "ai_engineer"
+            return {
+                "title": self._normalize_single_line_text(str(parsed.get("title") or ""), max_len=120),
+                "description": self._normalize_multiline_text(str(parsed.get("description") or ""), max_len=1200),
+                "primary_role": role,
+            }
+        except Exception as exc:
+            self._emit_ai_think("draft_new_project_brief_failed", error=str(exc))
+            return {
+                "title": self._normalize_single_line_text(message or "飞书新需求立项", max_len=40),
+                "description": self._normalize_multiline_text(message, max_len=300),
+                "primary_role": "ai_engineer",
+            }
+
+
+# 全局服务实例
+_signal_capture_service: Optional[ProactiveSignalCaptureService] = None
+
+
+def get_signal_capture_service() -> ProactiveSignalCaptureService:
+    """获取信号捕获服务实例（需要在 async 上下文中使用）"""
+    return ProactiveSignalCaptureService(None)  # Placeholder
diff --git a/backend/app/services/signal_source_orchestration_service.py b/backend/app/services/signal_source_orchestration_service.py
new file mode 100644
index 00000000..5b926d2d
--- /dev/null
+++ b/backend/app/services/signal_source_orchestration_service.py
@@ -0,0 +1,1927 @@
+"""信号源接入编排服务。"""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import json
+import re
+import time
+from datetime import datetime, timezone
+from types import SimpleNamespace
+from typing import Any, Optional
+from uuid import uuid4
+
+from sqlalchemy import and_, select
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import (
+    PROMPT_AI_WORKER_SIGNAL_SOURCE_EXECUTION,
+    PROMPT_SIGNAL_SOURCE_ONBOARDING,
+)
+from ..models import (
+    AIWorkerExecution,
+    Project,
+    Signal,
+    SignalSourceRegistry,
+    User,
+)
+from ..database import AsyncSessionLocal
+from .ai_orchestration_run_service import AIOrchestrationRunService
+from .prompt_profile_service import prompt_profile_service
+from .llm import llm_service
+
+SOURCE_ONBOARDING_ACTION_RE = re.compile(
+    r"(新增|增加|添加|接入|配置|注册|建立).{0,8}(信号源|来源|监控源|channel|source|webhook|飞书|github|url|rss)",
+    re.IGNORECASE,
+)
+SOURCE_KIND_HINTS: tuple[tuple[str, tuple[str, ...]], ...] = (
+    ("feishu", ("飞书", "lark", "群聊", "chat")),
+    ("github", ("github", "repo", "仓库", "pull request", "issue")),
+    ("webhook", ("webhook", "回调", "回传")),
+    ("url", ("url", "链接", "网页", "站点", "rss")),
+    ("browser", ("浏览器", "标签页", "页面")),
+    ("pmflow", ("pmflow", "内置ai", "ai管家", "steward")),
+)
+DEFAULT_WORKER_ROLE_KEYS = ("pm_dispatcher", "tech_architect", "delivery_coach")
+MAX_SOURCE_NAME_LEN = 120
+MAX_SOURCE_IDENTIFIER_LEN = 180
+WORKER_MAX_CONCURRENCY = 3
+COMPLETE_WORKER_WAIT_TIMEOUT_SECONDS = 6.0
+
+
+def _utc_now_naive() -> datetime:
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+
+class SignalSourceOrchestrationService:
+    """把“新增信号源”转换为 OpenClaw 指挥与下级 AI worker 执行链。"""
+
+    session_factory = AsyncSessionLocal
+
+    def __init__(self, db: AsyncSession) -> None:
+        self.db = db
+
+    @classmethod
+    def looks_like_signal_source_request(cls, text: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(text or ""))
+        if len(normalized) < 6:
+            return False
+        if "信号源" in normalized and any(token in normalized for token in ("新增", "增加", "添加", "接入", "配置", "注册", "建立")):
+            return True
+        return bool(SOURCE_ONBOARDING_ACTION_RE.search(normalized))
+
+    async def orchestrate_from_signal(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        signal_id: Optional[int],
+        initiating_text: str,
+        requested_via: str,
+        operator_user_id: Optional[int] = None,
+        dry_run: bool = False,
+        response_mode: str = "fast",
+    ) -> dict[str, Any]:
+        response_mode = "complete" if str(response_mode or "").strip().lower() == "complete" else "fast"
+        trace_id = f"source-onboard-{tenant_id}-{project_id}-{int(time.time() * 1000)}-{uuid4().hex[:8]}"
+        project = await self._load_project(tenant_id=tenant_id, project_id=project_id)
+        signal = await self._load_signal(tenant_id=tenant_id, signal_id=signal_id) if signal_id is not None else None
+        run_service = AIOrchestrationRunService(self.db) if self.db is not None else None
+        if project is None:
+            return {
+                "trace_id": trace_id,
+                "status": "skipped",
+                "reason": "project_or_signal_not_found",
+            }
+        if signal is None:
+            if dry_run:
+                signal = self._build_ephemeral_signal(initiating_text=initiating_text)
+            else:
+                return {
+                    "trace_id": trace_id,
+                "status": "skipped",
+                "reason": "project_or_signal_not_found",
+                }
+        project_ref = SimpleNamespace(
+            id=int(project.id),
+            title=str(project.title or "").strip(),
+            name=str(project.name or "").strip(),
+            description=str(project.description or "").strip(),
+            owner_id=int(project.owner_id) if getattr(project, "owner_id", None) else None,
+        )
+        signal_ref = SimpleNamespace(
+            id=int(signal.id) if getattr(signal, "id", None) is not None else None,
+            title=str(signal.title or "").strip(),
+            summary=str(signal.summary or "").strip(),
+            signal_type=str(signal.signal_type or "").strip(),
+            recommended_action=str(signal.recommended_action or "").strip(),
+            content=str(signal.content or "").strip(),
+        )
+        real_signal_id = signal_ref.id
+        persisted_signal_id = int(real_signal_id) if real_signal_id and int(real_signal_id) > 0 else None
+
+        if not self.looks_like_signal_source_request(initiating_text):
+            result = {
+                "trace_id": trace_id,
+                "status": "skipped",
+                "run_status": "completed",
+                "response_mode": response_mode,
+                "planner_status": "skipped",
+                "planner_diagnostics": None,
+                "summary": "当前输入不属于信号源接入请求。",
+                "pending_workers": [],
+                "worker_executions": [],
+                "openclaw": {"status": "skipped", "reason": "not_signal_source_request"},
+                "reason": "not_signal_source_request",
+            }
+            if run_service is not None:
+                await run_service.create_run(
+                    tenant_id=tenant_id,
+                    trace_id=trace_id,
+                    run_type="signal_source_onboarding",
+                    project_id=project_id,
+                    signal_id=persisted_signal_id,
+                    response_mode=response_mode,
+                    status="completed",
+                    planner_status="skipped",
+                    openclaw_status="skipped",
+                    worker_total=0,
+                    worker_completed=0,
+                    worker_failed=0,
+                    summary=str(result.get("summary") or "").strip(),
+                    planner_diagnostics=None,
+                    result_payload=dict(result),
+                    created_by_user_id=operator_user_id,
+                    completed_at=_utc_now_naive(),
+                )
+                await self.db.commit()
+            return result
+
+        plan, planner_status, planner_diagnostics = await self._plan_onboarding(
+            tenant_id=tenant_id,
+            project=project_ref,
+            signal=signal_ref,
+            initiating_text=initiating_text,
+            requested_via=requested_via,
+            response_mode=response_mode,
+        )
+        if not bool(plan.get("should_onboard")):
+            result = {
+                "trace_id": trace_id,
+                "status": "skipped",
+                "run_status": "completed",
+                "response_mode": response_mode,
+                "reason": "planner_rejected",
+                "planner_status": planner_status,
+                "planner_diagnostics": planner_diagnostics,
+                "summary": str(plan.get("summary") or "").strip(),
+                "pending_workers": [],
+                "worker_executions": [],
+                "openclaw": {"status": "skipped", "reason": "planner_rejected"},
+            }
+            if run_service is not None:
+                await run_service.create_run(
+                    tenant_id=tenant_id,
+                    trace_id=trace_id,
+                    run_type="signal_source_onboarding",
+                    project_id=project_id,
+                    signal_id=persisted_signal_id,
+                    response_mode=response_mode,
+                    status="completed",
+                    planner_status=planner_status,
+                    openclaw_status="skipped",
+                    worker_total=0,
+                    worker_completed=0,
+                    worker_failed=0,
+                    summary=str(result.get("summary") or "").strip(),
+                    planner_diagnostics=planner_diagnostics,
+                    result_payload=dict(result),
+                    created_by_user_id=operator_user_id,
+                    completed_at=_utc_now_naive(),
+                )
+                await self.db.commit()
+            return result
+
+        source_payload = plan.get("source") if isinstance(plan.get("source"), dict) else {}
+        worker_role_keys = self._normalize_worker_roles(plan.get("worker_roles"))
+        registry_result = await self._create_or_reuse_registry(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            signal_id=signal_id,
+            operator_user_id=operator_user_id,
+            requested_via=requested_via,
+            source_payload=source_payload,
+            initiating_text=initiating_text,
+            dry_run=dry_run,
+        )
+
+        operator_user = await self._resolve_operator_user(
+            tenant_id=tenant_id,
+            project=project_ref,
+            operator_user_id=operator_user_id,
+        )
+        operator_ref = self._build_user_snapshot(operator_user)
+        openclaw_result = await self._dispatch_openclaw(
+            tenant_id=tenant_id,
+            operator_user=operator_ref,
+            project=project_ref,
+            instruction=str(plan.get("openclaw_instruction") or "").strip(),
+            dry_run=dry_run,
+            response_mode=response_mode,
+        )
+        openclaw_status = str(openclaw_result.get("status") or "").strip().lower()
+        base_result = self._build_orchestration_result(
+            trace_id=trace_id,
+            response_mode=response_mode,
+            planner_status=planner_status,
+            planner_diagnostics=planner_diagnostics,
+            summary=str(plan.get("summary") or "").strip(),
+            registry_result=registry_result,
+            source_payload=source_payload,
+            openclaw_result=openclaw_result,
+            worker_results=[],
+            pending_workers=list(worker_role_keys),
+            status="planned" if dry_run else "partial",
+            run_status="completed" if dry_run or not worker_role_keys else "running",
+        )
+
+        if dry_run:
+            planned_results = self._build_planned_worker_results(worker_role_keys)
+            final_result = self._build_orchestration_result(
+                trace_id=trace_id,
+                response_mode=response_mode,
+                planner_status=planner_status,
+                planner_diagnostics=planner_diagnostics,
+                summary=str(plan.get("summary") or "").strip(),
+                registry_result=registry_result,
+                source_payload=source_payload,
+                openclaw_result=openclaw_result,
+                worker_results=planned_results,
+                pending_workers=[],
+                status="planned",
+                run_status="completed",
+            )
+            if run_service is not None:
+                await run_service.create_run(
+                    tenant_id=tenant_id,
+                    trace_id=trace_id,
+                    run_type="signal_source_onboarding",
+                    project_id=project_id,
+                    signal_id=persisted_signal_id,
+                    response_mode=response_mode,
+                    status="completed",
+                    planner_status=planner_status,
+                    openclaw_status=openclaw_status,
+                    worker_total=len(worker_role_keys),
+                    worker_completed=len(planned_results),
+                    worker_failed=0,
+                    summary=str(final_result.get("summary") or "").strip(),
+                    planner_diagnostics=planner_diagnostics,
+                    result_payload=dict(final_result),
+                    created_by_user_id=operator_user_id,
+                    completed_at=_utc_now_naive(),
+                )
+                await self.db.commit()
+            return final_result
+
+        if not worker_role_keys:
+            registry_status, onboarding_status = self._resolve_registry_activation(
+                source_payload=source_payload,
+                openclaw_result=openclaw_result,
+                worker_results=[],
+            )
+            final_status = self._resolve_orchestration_status(openclaw_status=openclaw_status, worker_results=[])
+            if final_status == "completed" and onboarding_status != "completed":
+                final_status = "pending_manual"
+            final_result = self._build_orchestration_result(
+                trace_id=trace_id,
+                response_mode=response_mode,
+                planner_status=planner_status,
+                planner_diagnostics=planner_diagnostics,
+                summary=str(plan.get("summary") or "").strip(),
+                registry_result={
+                    "registry_id": registry_result.get("registry_id"),
+                    "status": registry_status,
+                    "onboarding_status": onboarding_status,
+                },
+                source_payload=source_payload,
+                openclaw_result=openclaw_result,
+                worker_results=[],
+                pending_workers=[],
+                status=final_status,
+                run_status="completed",
+            )
+            if run_service is not None:
+                await run_service.create_run(
+                    tenant_id=tenant_id,
+                    trace_id=trace_id,
+                    run_type="signal_source_onboarding",
+                    project_id=project_id,
+                    signal_id=persisted_signal_id,
+                    response_mode=response_mode,
+                    status=final_status,
+                    planner_status=planner_status,
+                    openclaw_status=openclaw_status,
+                    worker_total=0,
+                    worker_completed=0,
+                    worker_failed=0,
+                    summary=str(final_result.get("summary") or "").strip(),
+                    planner_diagnostics=planner_diagnostics,
+                    result_payload=dict(final_result),
+                    created_by_user_id=operator_user_id,
+                    completed_at=_utc_now_naive(),
+                )
+                if registry_result.get("registry_id") is not None:
+                    registry_row = await self._load_signal_source_registry(
+                        tenant_id=tenant_id,
+                        registry_id=int(registry_result["registry_id"]),
+                    )
+                    if registry_row is not None:
+                        registry_row.last_signal_at = _utc_now_naive()
+                        registry_row.last_worker_run_at = _utc_now_naive()
+                        registry_row.status = registry_status
+                        registry_row.onboarding_status = onboarding_status
+                await self.db.commit()
+            if real_signal_id:
+                persisted_signal = await self._load_signal(tenant_id=tenant_id, signal_id=int(real_signal_id))
+                if persisted_signal is not None:
+                    await self._persist_signal_orchestration_result(signal=persisted_signal, orchestration_result=final_result)
+            return final_result
+
+        if run_service is not None:
+            await run_service.create_run(
+                tenant_id=tenant_id,
+                trace_id=trace_id,
+                run_type="signal_source_onboarding",
+                project_id=project_id,
+                signal_id=persisted_signal_id,
+                response_mode=response_mode,
+                status="running",
+                planner_status=planner_status,
+                openclaw_status=openclaw_status,
+                worker_total=len(worker_role_keys),
+                worker_completed=0,
+                worker_failed=0,
+                summary=str(base_result.get("summary") or "").strip(),
+                planner_diagnostics=planner_diagnostics,
+                result_payload=dict(base_result),
+                created_by_user_id=operator_user_id,
+            )
+            await self.db.commit()
+
+        task = asyncio.create_task(
+            type(self)._background_finalize_with_guard(
+                tenant_id=tenant_id,
+                trace_id=trace_id,
+                project_payload=self._serialize_snapshot(project_ref),
+                signal_payload=self._serialize_snapshot(signal_ref),
+                registry_result=dict(registry_result),
+                openclaw_result=dict(openclaw_result),
+                worker_role_keys=list(worker_role_keys),
+                operator_user_payload=self._serialize_snapshot(operator_ref),
+                planner_status=planner_status,
+                planner_diagnostics=planner_diagnostics,
+                summary=str(plan.get("summary") or "").strip(),
+                signal_id=persisted_signal_id,
+            ),
+            name=f"signal-source-workers-{trace_id}",
+        )
+        if response_mode == "fast":
+            return base_result
+        try:
+            return await asyncio.wait_for(asyncio.shield(task), timeout=COMPLETE_WORKER_WAIT_TIMEOUT_SECONDS)
+        except asyncio.TimeoutError:
+            if run_service is None:
+                return base_result
+            run_row = await run_service.get_run(tenant_id=tenant_id, trace_id=trace_id)
+            if run_row is None:
+                return base_result
+            snapshot = AIOrchestrationRunService.serialize_run(run_row)
+            payload = dict(run_row.result_payload or base_result)
+            payload["run_status"] = str(snapshot.get("run_status") or "running")
+            payload["status"] = str(payload.get("status") or "partial")
+            payload["pending_workers"] = snapshot.get("pending_workers") or payload.get("pending_workers") or []
+            return payload
+
+    async def list_project_sources(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        limit: int = 20,
+    ) -> list[dict[str, Any]]:
+        rows = (
+            await self.db.execute(
+                select(SignalSourceRegistry).where(
+                    and_(
+                        SignalSourceRegistry.tenant_id == tenant_id,
+                        SignalSourceRegistry.project_id == project_id,
+                    )
+                ).order_by(SignalSourceRegistry.updated_at.desc()).limit(max(1, min(limit, 100)))
+            )
+        ).scalars().all()
+        data: list[dict[str, Any]] = []
+        for row in rows:
+            executions = (
+                await self.db.execute(
+                    select(AIWorkerExecution).where(
+                        and_(
+                            AIWorkerExecution.tenant_id == tenant_id,
+                            AIWorkerExecution.signal_source_registry_id == int(row.id),
+                        )
+                    ).order_by(AIWorkerExecution.created_at.desc()).limit(6)
+                )
+            ).scalars().all()
+            data.append(
+                {
+                    "id": int(row.id),
+                    "project_id": int(row.project_id),
+                    "signal_id": int(row.signal_id) if row.signal_id else None,
+                    "source_kind": str(row.source_kind or ""),
+                    "source_name": str(row.source_name or ""),
+                    "source_identifier": str(row.source_identifier or ""),
+                    "requested_via": str(row.requested_via or ""),
+                    "status": str(row.status or ""),
+                    "onboarding_status": str(row.onboarding_status or ""),
+                    "updated_at": row.updated_at.isoformat() if row.updated_at else None,
+                    "worker_executions": [
+                        {
+                            "id": int(item.id),
+                            "worker_role_key": str(item.worker_role_key or ""),
+                            "worker_role_name": str(item.worker_role_name or ""),
+                            "status": str(item.status or ""),
+                            "summary": str(item.summary or ""),
+                            "completed_at": item.completed_at.isoformat() if item.completed_at else None,
+                        }
+                        for item in executions
+                    ],
+                }
+            )
+        return data
+
+    async def _plan_onboarding(
+        self,
+        *,
+        tenant_id: int,
+        project: Project,
+        signal: Signal,
+        initiating_text: str,
+        requested_via: str,
+        response_mode: str,
+    ) -> tuple[dict[str, Any], str, dict[str, Any]]:
+        request_payload = {
+            "requested_via": requested_via,
+            "initiating_text": initiating_text[:600],
+        }
+        project_payload = {
+            "project_id": int(project.id),
+            "title": str(project.title or project.name or "").strip(),
+            "description": str(project.description or "").strip()[:400],
+        }
+        signal_payload = {
+            "signal_id": int(signal.id) if getattr(signal, "id", None) is not None else None,
+            "title": str(signal.title or "").strip(),
+            "summary": str(signal.summary or "").strip()[:400],
+            "signal_type": str(signal.signal_type or "").strip(),
+            "recommended_action": str(signal.recommended_action or "").strip(),
+        }
+        prompt = PROMPT_SIGNAL_SOURCE_ONBOARDING.format(
+            request_json=json.dumps(request_payload, ensure_ascii=False),
+            project_json=json.dumps(project_payload, ensure_ascii=False),
+            signal_json=json.dumps(signal_payload, ensure_ascii=False),
+        )
+        try:
+            extraction = await llm_service.generate_structured_object(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="signal_source_onboarding_plan",
+                use_prompt_profile=False,
+                max_tokens=420,
+                response_mode=response_mode,
+            )
+            diagnostics = self._build_planner_diagnostics(extraction)
+            if extraction.payload is None:
+                return (
+                    self._heuristic_onboarding_plan(
+                        initiating_text=initiating_text,
+                        requested_via=requested_via,
+                        error=str(extraction.error_code or "structured_unknown"),
+                    ),
+                    "fallback",
+                    diagnostics,
+                )
+            plan = extraction.payload
+            normalized = self._normalize_onboarding_plan(
+                plan=plan,
+                initiating_text=initiating_text,
+                requested_via=requested_via,
+            )
+            return normalized, "planned", diagnostics
+        except Exception as exc:
+            return (
+                self._heuristic_onboarding_plan(
+                    initiating_text=initiating_text,
+                    requested_via=requested_via,
+                    error=str(exc),
+                ),
+                "fallback",
+                {
+                    "used_fallback": True,
+                    "error_code": "structured_empty",
+                    "source_field": None,
+                    "provider": "",
+                    "model": "",
+                },
+            )
+
+    def _normalize_onboarding_plan(
+        self,
+        *,
+        plan: dict[str, Any],
+        initiating_text: str,
+        requested_via: str,
+    ) -> dict[str, Any]:
+        should_onboard = bool(plan.get("should_onboard"))
+        summary = str(plan.get("summary") or "").strip()[:300]
+        source = plan.get("source") if isinstance(plan.get("source"), dict) else {}
+        worker_roles = self._normalize_worker_roles(plan.get("worker_roles"))
+        if not should_onboard:
+            return {
+                "should_onboard": False,
+                "summary": summary or "未识别为新增信号源请求。",
+                "source": None,
+                "worker_roles": [],
+                "openclaw_instruction": "",
+            }
+        normalized_source = self._normalize_source_payload(
+            source_payload=source,
+            fallback_text=initiating_text,
+            requested_via=requested_via,
+        )
+        return {
+            "should_onboard": True,
+            "summary": summary or f"准备接入信号源 {normalized_source['source_name']}。",
+            "source": normalized_source,
+            "worker_roles": worker_roles,
+            "openclaw_instruction": (
+                str(plan.get("openclaw_instruction") or "").strip()
+                or self._build_openclaw_instruction(normalized_source, initiating_text)
+            )[:600],
+        }
+
+    def _heuristic_onboarding_plan(
+        self,
+        *,
+        initiating_text: str,
+        requested_via: str,
+        error: str,
+    ) -> dict[str, Any]:
+        source = self._normalize_source_payload(
+            source_payload={},
+            fallback_text=initiating_text,
+            requested_via=requested_via,
+        )
+        return {
+            "should_onboard": True,
+            "summary": "系统已按规则兜底为信号源接入请求，并生成最小接入闭环。",
+            "source": source,
+            "worker_roles": list(DEFAULT_WORKER_ROLE_KEYS),
+            "openclaw_instruction": self._build_openclaw_instruction(source, initiating_text),
+        }
+
+    def _normalize_source_payload(
+        self,
+        *,
+        source_payload: dict[str, Any],
+        fallback_text: str,
+        requested_via: str,
+    ) -> dict[str, Any]:
+        source_kind = self._normalize_source_kind(source_payload.get("source_kind"), text=fallback_text, requested_via=requested_via)
+        source_name = self._clean_source_name_candidate(str(source_payload.get("source_name") or "").strip())
+        if not source_name:
+            source_name = self._extract_source_name(fallback_text, default=f"{source_kind}信号源")
+        source_name = source_name[:MAX_SOURCE_NAME_LEN] or f"{source_kind}信号源"
+        source_identifier = str(source_payload.get("source_identifier") or "").strip()
+        if source_identifier:
+            existing_kind = source_identifier.split(":", 1)[0].strip().lower()
+            if existing_kind and existing_kind != source_kind:
+                source_identifier = ""
+        if not source_identifier:
+            source_identifier = self._build_source_identifier(
+                source_kind=source_kind,
+                source_name=source_name,
+                text=fallback_text,
+                source_payload=source_payload,
+            )
+        description = str(source_payload.get("description") or "").strip()[:1000] or fallback_text[:300]
+        config = source_payload.get("config") if isinstance(source_payload.get("config"), dict) else {}
+        return {
+            "source_kind": source_kind,
+            "source_name": source_name,
+            "source_identifier": source_identifier[:MAX_SOURCE_IDENTIFIER_LEN],
+            "description": description,
+            "config": config,
+        }
+
+    @staticmethod
+    def _normalize_source_kind(raw_kind: Any, *, text: str, requested_via: str) -> str:
+        cleaned = str(raw_kind or "").strip().lower()
+        for kind, hints in SOURCE_KIND_HINTS:
+            if cleaned == kind:
+                return kind
+            if any(token in cleaned for token in hints):
+                return kind
+        text_lower = str(text or "").lower()
+        for kind, hints in SOURCE_KIND_HINTS:
+            if any(token in text_lower for token in hints):
+                return kind
+        via = str(requested_via or "").strip().lower()
+        if via in {"feishu", "pmflow", "github", "browser"}:
+            return via
+        return "custom"
+
+    @classmethod
+    def _extract_source_name(cls, text: str, *, default: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(text or "")).strip()
+        if not cleaned:
+            return default
+        repo_match = re.search(r"\b([A-Za-z0-9_.-]+/[A-Za-z0-9_.-]+)\b", cleaned)
+        if repo_match:
+            return repo_match.group(1)[:MAX_SOURCE_NAME_LEN]
+        quoted_match = re.search(r"[\"'“](.+?)[\"'”]", cleaned)
+        if quoted_match:
+            candidate = cls._clean_source_name_candidate(quoted_match.group(1))
+            if candidate:
+                return candidate[:MAX_SOURCE_NAME_LEN]
+        for pattern in (
+            r"(?:新增|增加|添加|接入|配置|注册|建立)(?:一个|1个|一条|一组|新的|一个新的)?\s*(.+?)(?:信号源|来源|监控源)",
+            r"(?:信号源|来源|监控源)(?:为|是|叫|命名为)?\s*(.+)",
+        ):
+            match = re.search(pattern, cleaned, re.IGNORECASE)
+            if match:
+                value = cls._clean_source_name_candidate(match.group(1))
+                if value:
+                    return value[:MAX_SOURCE_NAME_LEN]
+        return cls._clean_source_name_candidate(cleaned)[:MAX_SOURCE_NAME_LEN] or default
+
+    @classmethod
+    def _build_source_identifier(
+        cls,
+        *,
+        source_kind: str,
+        source_name: str,
+        text: str,
+        source_payload: dict[str, Any],
+    ) -> str:
+        normalized_text = re.sub(r"\s+", " ", str(text or "")).strip()
+        explicit_locator = cls._extract_explicit_source_locator(source_payload=source_payload, text=text)
+        if explicit_locator:
+            basis = re.sub(r"[^a-zA-Z0-9:/._-]+", "-", explicit_locator).strip("-").lower()
+            return f"{source_kind}:{basis[:120]}"
+        if cls._source_name_is_generic(source_name, source_kind=source_kind):
+            digest = hashlib.sha1(f"{source_kind}:{normalized_text}".encode("utf-8")).hexdigest()[:16]
+            return f"{source_kind}:req-{digest}"
+        basis = re.sub(r"[^a-zA-Z0-9\u4e00-\u9fff]+", "-", source_name).strip("-").lower()
+        if basis:
+            return f"{source_kind}:{basis[:120]}"
+        digest = hashlib.sha1(f"{source_kind}:{normalized_text}".encode("utf-8")).hexdigest()[:16]
+        return f"{source_kind}:{digest}"
+
+    @staticmethod
+    def _clean_source_name_candidate(value: str) -> str:
+        cleaned = re.sub(r"\s+", " ", str(value or "")).strip()
+        cleaned = re.sub(r"^(?:给)?项目id[:：]?\s*\d+\s*", "", cleaned, flags=re.IGNORECASE)
+        cleaned = re.sub(r"[，。,.;；:：].*$", "", cleaned)
+        cleaned = cleaned.strip(" []【】()（）:：\"'“”")
+        cleaned = re.sub(r"^(?:请|帮我|帮忙|把|将|需要|想要|用于|用来)\s*", "", cleaned, flags=re.IGNORECASE)
+        cleaned = re.sub(
+            r"^(?:一个|1个|一条|一组|一个新的|新的|新增的)\s+",
+            "",
+            cleaned,
+            flags=re.IGNORECASE,
+        )
+        cleaned = re.sub(r"(?:信号源|来源|监控源|source|channel)$", "", cleaned, flags=re.IGNORECASE).strip()
+        if cleaned.lower() == "github":
+            return "GitHub"
+        if cleaned.lower() == "feishu":
+            return "Feishu"
+        normalized = re.sub(r"[^a-zA-Z0-9\u4e00-\u9fff]+", "", cleaned).lower()
+        if normalized in {"", "项目id", "给项目id", "复合", "复合信号源", "信号源", "来源", "监控源"}:
+            return ""
+        return cleaned
+
+    @classmethod
+    def _extract_explicit_source_locator(cls, *, source_payload: dict[str, Any], text: str) -> str:
+        config = source_payload.get("config") if isinstance(source_payload.get("config"), dict) else {}
+        direct_candidates = [
+            source_payload.get("repo"),
+            source_payload.get("repository"),
+            source_payload.get("url"),
+            source_payload.get("webhook_url"),
+            source_payload.get("channel_id"),
+            source_payload.get("chat_id"),
+            config.get("repo"),
+            config.get("repository"),
+            config.get("url"),
+            config.get("webhook_url"),
+            config.get("channel_id"),
+            config.get("chat_id"),
+        ]
+        for candidate in direct_candidates:
+            normalized = cls._normalize_locator_fragment(candidate)
+            if normalized:
+                return normalized
+
+        url_match = re.search(r"https?://[^\s，。,；;）)】\\]]+", str(text or ""), re.IGNORECASE)
+        if url_match:
+            return cls._normalize_locator_fragment(url_match.group(0))
+
+        repo_match = re.search(r"\b([A-Za-z0-9_.-]+/[A-Za-z0-9_.-]+)\b", str(text or ""))
+        if repo_match:
+            return f"repo:{repo_match.group(1).lower()}"
+
+        chat_match = re.search(
+            r"\b(?:chat_id|channel_id|群聊id|频道id)[:：=\s]+([A-Za-z0-9_-]{6,})\b",
+            str(text or ""),
+            re.IGNORECASE,
+        )
+        if chat_match:
+            return f"chat:{chat_match.group(1).lower()}"
+        return ""
+
+    @staticmethod
+    def _normalize_locator_fragment(value: Any) -> str:
+        text = str(value or "").strip().strip("\"'“”")
+        if not text:
+            return ""
+        text = re.sub(r"\?.*$", "", text).rstrip("/")
+        return text[:140].lower()
+
+    @classmethod
+    def _source_name_is_generic(cls, source_name: str, *, source_kind: str) -> bool:
+        normalized = re.sub(r"[^a-zA-Z0-9\u4e00-\u9fff]+", "", str(source_name or "")).lower()
+        if not normalized:
+            return True
+        generic_tokens = {
+            re.sub(r"[^a-zA-Z0-9\u4e00-\u9fff]+", "", source_kind).lower(),
+            re.sub(r"[^a-zA-Z0-9\u4e00-\u9fff]+", "", f"{source_kind}信号源").lower(),
+            "项目id",
+            "给项目id",
+            "复合",
+            "复合信号源",
+        }
+        for kind, hints in SOURCE_KIND_HINTS:
+            if kind != source_kind:
+                continue
+            generic_tokens.update(
+                re.sub(r"[^a-zA-Z0-9\u4e00-\u9fff]+", "", str(token)).lower()
+                for token in hints
+            )
+            break
+        return normalized in generic_tokens
+
+    @staticmethod
+    def _normalize_worker_roles(raw_roles: Any) -> list[str]:
+        roles: list[str] = []
+        if isinstance(raw_roles, list):
+            for item in raw_roles:
+                key = str(item or "").strip()
+                if key in DEFAULT_WORKER_ROLE_KEYS and key not in roles:
+                    roles.append(key)
+        if roles:
+            return roles[: len(DEFAULT_WORKER_ROLE_KEYS)]
+        return list(DEFAULT_WORKER_ROLE_KEYS)
+
+    @staticmethod
+    def _build_openclaw_instruction(source: dict[str, Any], initiating_text: str) -> str:
+        return (
+            f"围绕新增信号源“{source.get('source_name') or '未命名信号源'}”完成接入准备。"
+            "请创建必要任务、分配合适负责人，并通知相关成员。"
+            f"补充上下文：{str(initiating_text or '').strip()[:240]}"
+        )
+
+    async def _create_or_reuse_registry(
+        self,
+        *,
+        tenant_id: int,
+        project_id: int,
+        signal_id: Optional[int],
+        operator_user_id: Optional[int],
+        requested_via: str,
+        source_payload: dict[str, Any],
+        initiating_text: str,
+        dry_run: bool,
+    ) -> dict[str, Any]:
+        normalized_source = self._normalize_source_payload(
+            source_payload=source_payload,
+            fallback_text=initiating_text,
+            requested_via=requested_via,
+        )
+        if dry_run:
+            return {
+                "status": "planned",
+                "registry_id": None,
+                "source_kind": normalized_source.get("source_kind"),
+                "source_name": normalized_source.get("source_name"),
+                "source_identifier": normalized_source.get("source_identifier"),
+                "description": str(normalized_source.get("description") or "").strip(),
+            }
+
+        row = SignalSourceRegistry(
+            tenant_id=tenant_id,
+            project_id=project_id,
+            signal_id=signal_id,
+            created_by_user_id=operator_user_id,
+            source_kind=str(normalized_source["source_kind"]),
+            source_name=str(normalized_source["source_name"]),
+            source_identifier=str(normalized_source["source_identifier"]),
+            description=str(normalized_source.get("description") or "").strip() or None,
+            requested_via=str(requested_via or "system").strip()[:40] or "system",
+            status="pending",
+            onboarding_status="pending",
+            config=normalized_source.get("config") if isinstance(normalized_source.get("config"), dict) else {},
+            raw_context={
+                "initiating_text": str(initiating_text or "")[:2000],
+                "requested_via": requested_via,
+            },
+            last_signal_at=_utc_now_naive(),
+        )
+        self.db.add(row)
+        try:
+            await self.db.flush()
+            await self.db.commit()
+            await self.db.refresh(row)
+            return {
+                "status": "created",
+                "registry_id": int(row.id),
+                "source_kind": str(row.source_kind or ""),
+                "source_name": str(row.source_name or ""),
+                "source_identifier": str(row.source_identifier or ""),
+                "description": str(row.description or ""),
+            }
+        except IntegrityError:
+            await self.db.rollback()
+            existing = (
+                await self.db.execute(
+                    select(SignalSourceRegistry).where(
+                        and_(
+                            SignalSourceRegistry.tenant_id == tenant_id,
+                            SignalSourceRegistry.project_id == project_id,
+                            SignalSourceRegistry.source_kind == str(normalized_source["source_kind"]),
+                            SignalSourceRegistry.source_identifier == str(normalized_source["source_identifier"]),
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if existing is None:
+                raise
+            existing.signal_id = signal_id
+            if str(normalized_source.get("source_name") or "").strip():
+                existing.source_name = str(normalized_source.get("source_name") or "").strip()[:MAX_SOURCE_NAME_LEN]
+            existing.description = str(normalized_source.get("description") or existing.description or "").strip() or None
+            existing.config = normalized_source.get("config") if isinstance(normalized_source.get("config"), dict) else existing.config
+            existing.raw_context = {
+                "initiating_text": str(initiating_text or "")[:2000],
+                "requested_via": requested_via,
+            }
+            if not isinstance(existing.config, dict) or not existing.config:
+                existing.status = "pending"
+            existing.onboarding_status = "pending"
+            existing.last_signal_at = _utc_now_naive()
+            await self.db.commit()
+            return {
+                "status": "reused",
+                "registry_id": int(existing.id),
+                "source_kind": str(existing.source_kind or ""),
+                "source_name": str(existing.source_name or ""),
+                "source_identifier": str(existing.source_identifier or ""),
+                "description": str(existing.description or ""),
+            }
+
+    async def _dispatch_openclaw(
+        self,
+        *,
+        tenant_id: int,
+        operator_user: Optional[User],
+        project: Project,
+        instruction: str,
+        dry_run: bool,
+        response_mode: str,
+    ) -> dict[str, Any]:
+        if not instruction:
+            return {
+                "status": "skipped",
+                "reason": "empty_instruction",
+            }
+        if operator_user is None:
+            return {
+                "status": "skipped",
+                "reason": "no_manager_operator",
+            }
+        try:
+            from .openclaw_commander_service import OpenClawCommanderService
+
+            result = await OpenClawCommanderService(self.db).dispatch_command(
+                tenant_id=tenant_id,
+                operator_user=operator_user,
+                project_id=int(project.id),
+                instruction=instruction,
+                dry_run=dry_run,
+                max_operations=6,
+                response_mode=response_mode,
+                persist_run=False,
+            )
+            return {
+                "status": "planned" if dry_run else "completed",
+                "result": result,
+            }
+        except Exception as exc:
+            return {
+                "status": "failed",
+                "reason": str(exc),
+            }
+
+    async def _execute_worker_roles(
+        self,
+        *,
+        tenant_id: int,
+        project: Project,
+        signal: Signal,
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_role_keys: list[str],
+        trace_id: str,
+        operator_user: Optional[User],
+        dry_run: bool,
+        response_mode: str,
+    ) -> list[dict[str, Any]]:
+        bundle = await prompt_profile_service.get_profiles(self.db, tenant_id=tenant_id)
+        roles = bundle.get("roles") if isinstance(bundle.get("roles"), list) else []
+        role_map = {
+            str(item.get("role_key") or "").strip(): item
+            for item in roles
+            if isinstance(item, dict) and str(item.get("role_key") or "").strip()
+        }
+
+        if dry_run:
+            return self._build_planned_worker_results(worker_role_keys, role_map=role_map)
+
+        semaphore = asyncio.Semaphore(WORKER_MAX_CONCURRENCY)
+
+        async def _run_role(role_key: str) -> dict[str, Any]:
+            role = role_map.get(role_key)
+            if not isinstance(role, dict):
+                return {
+                    "worker_role_key": role_key,
+                    "worker_role_name": role_key,
+                    "status": "failed",
+                    "reason": "role_not_configured",
+                }
+            try:
+                async with semaphore:
+                    return await type(self)._execute_single_worker_role_isolated(
+                        tenant_id=tenant_id,
+                        project_payload=self._serialize_snapshot(project),
+                        signal_payload=self._serialize_snapshot(signal),
+                        registry_result=dict(registry_result),
+                        openclaw_result=dict(openclaw_result),
+                        role=dict(role),
+                        trace_id=trace_id,
+                        operator_user_payload=self._serialize_snapshot(operator_user),
+                        response_mode=response_mode,
+                    )
+            except Exception as exc:
+                return {
+                    "worker_role_key": role_key,
+                    "worker_role_name": str(role.get("role_name") or role_key).strip()[:64] or role_key,
+                    "status": "failed",
+                    "reason": str(exc)[:300],
+                }
+
+        tasks = [asyncio.create_task(_run_role(role_key)) for role_key in worker_role_keys]
+        return list(await asyncio.gather(*tasks, return_exceptions=False))
+
+    async def _execute_single_worker_role(
+        self,
+        *,
+        tenant_id: int,
+        project: Project,
+        signal: Signal,
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        role: dict[str, Any],
+        trace_id: str,
+        operator_user: Optional[User],
+        dry_run: bool,
+        response_mode: str,
+    ) -> dict[str, Any]:
+        role_key = str(role.get("role_key") or "").strip()
+        role_name = str(role.get("role_name") or role_key).strip()[:64] or role_key
+        registry_id = registry_result.get("registry_id")
+        source_payload = {
+            "registry_id": int(registry_id) if registry_id is not None else None,
+            "source_kind": str(registry_result.get("source_kind") or ""),
+            "source_name": str(registry_result.get("source_name") or ""),
+            "source_identifier": str(registry_result.get("source_identifier") or ""),
+            "description": str(registry_result.get("description") or ""),
+        }
+
+        execution_row: Optional[AIWorkerExecution] = None
+        if not dry_run:
+            execution_row = AIWorkerExecution(
+                tenant_id=tenant_id,
+                project_id=int(project.id),
+                signal_id=int(signal.id) if getattr(signal, "id", None) is not None else None,
+                signal_source_registry_id=int(registry_id) if registry_id is not None else None,
+                operator_user_id=int(operator_user.id) if operator_user is not None else None,
+                trace_id=trace_id,
+                operation_trace_id=(
+                    str((openclaw_result.get("result") or {}).get("trace_id") or "").strip()
+                    if isinstance(openclaw_result.get("result"), dict)
+                    else None
+                ),
+                execution_type="signal_source_onboarding",
+                worker_role_key=role_key,
+                worker_role_name=role_name,
+                status="running",
+                input_payload=source_payload,
+                started_at=_utc_now_naive(),
+            )
+            self.db.add(execution_row)
+            await self.db.flush()
+            # 先提交 running 状态，避免 SQLite 事务跨 LLM 调用长期持有写锁。
+            await self.db.commit()
+
+        prompt = self._build_worker_prompt(
+            role=role,
+            project=project,
+            signal=signal,
+            source_payload=source_payload,
+            openclaw_result=openclaw_result,
+        )
+        if dry_run:
+            return {
+                "worker_role_key": role_key,
+                "worker_role_name": role_name,
+                "status": "planned",
+            }
+        try:
+            extraction = await llm_service.generate_structured_object(
+                prompt,
+                db=self.db,
+                tenant_id=tenant_id,
+                function_name="signal_source_ai_worker_execution",
+                use_prompt_profile=False,
+                max_tokens=420,
+                response_mode="complete" if response_mode == "fast" else response_mode,
+            )
+            if extraction.payload is None:
+                raise ValueError(str(extraction.error_code or "structured_empty"))
+            parsed = self._normalize_worker_output(
+                extraction.payload,
+                role_name=role_name,
+                source_name=source_payload.get("source_name"),
+            )
+            if execution_row is not None:
+                execution_row.status = "completed"
+                execution_row.summary = str(parsed.get("summary") or "").strip()[:1000]
+                execution_row.output_payload = parsed
+                execution_row.completed_at = _utc_now_naive()
+                await self.db.commit()
+            return {
+                "worker_role_key": role_key,
+                "worker_role_name": role_name,
+                "status": "completed",
+                "summary": str(parsed.get("summary") or "").strip(),
+                "execution_id": int(execution_row.id) if execution_row is not None else None,
+                "used_fallback": False,
+            }
+        except Exception as exc:
+            fallback = self._fallback_worker_output(
+                role_name=role_name,
+                source_name=str(source_payload.get("source_name") or "信号源"),
+            )
+            if execution_row is not None:
+                execution_row.status = "completed"
+                execution_row.summary = str(fallback.get("summary") or "").strip()[:1000]
+                execution_row.output_payload = fallback
+                execution_row.error_message = str(exc)[:1000]
+                execution_row.completed_at = _utc_now_naive()
+                await self.db.commit()
+            return {
+                "worker_role_key": role_key,
+                "worker_role_name": role_name,
+                "status": "completed",
+                "reason": str(exc),
+                "summary": str(fallback.get("summary") or "").strip(),
+                "used_fallback": True,
+                "execution_id": int(execution_row.id) if execution_row is not None else None,
+            }
+
+    @classmethod
+    async def _execute_single_worker_role_isolated(
+        cls,
+        *,
+        tenant_id: int,
+        project_payload: dict[str, Any],
+        signal_payload: dict[str, Any],
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        role: dict[str, Any],
+        trace_id: str,
+        operator_user_payload: Optional[dict[str, Any]],
+        response_mode: str,
+    ) -> dict[str, Any]:
+        session_factory = getattr(cls, "session_factory", AsyncSessionLocal)
+        async with session_factory() as db:
+            service = cls(db)
+            return await service._execute_single_worker_role(
+                tenant_id=tenant_id,
+                project=SimpleNamespace(**project_payload),
+                signal=SimpleNamespace(**signal_payload),
+                registry_result=registry_result,
+                openclaw_result=openclaw_result,
+                role=role,
+                trace_id=trace_id,
+                operator_user=SimpleNamespace(**operator_user_payload) if isinstance(operator_user_payload, dict) else None,
+                dry_run=False,
+                response_mode=response_mode,
+            )
+
+    @classmethod
+    async def _background_finalize_with_guard(
+        cls,
+        *,
+        tenant_id: int,
+        trace_id: str,
+        project_payload: dict[str, Any],
+        signal_payload: dict[str, Any],
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_role_keys: list[str],
+        operator_user_payload: Optional[dict[str, Any]],
+        planner_status: str,
+        planner_diagnostics: Optional[dict[str, Any]],
+        summary: str,
+        signal_id: Optional[int],
+    ) -> dict[str, Any]:
+        try:
+            return await cls._execute_workers_and_finalize(
+                tenant_id=tenant_id,
+                trace_id=trace_id,
+                project_payload=project_payload,
+                signal_payload=signal_payload,
+                registry_result=registry_result,
+                openclaw_result=openclaw_result,
+                worker_role_keys=worker_role_keys,
+                operator_user_payload=operator_user_payload,
+                planner_status=planner_status,
+                planner_diagnostics=planner_diagnostics,
+                summary=summary,
+                signal_id=signal_id,
+            )
+        except Exception as exc:
+            session_factory = getattr(cls, "session_factory", AsyncSessionLocal)
+            async with session_factory() as db:
+                service = cls(db)
+                run_service = AIOrchestrationRunService(db)
+                row = await run_service.get_run(tenant_id=tenant_id, trace_id=trace_id)
+                response_mode = str(row.response_mode or "fast") if row is not None else "fast"
+                existing_payload = dict(row.result_payload or {}) if row is not None and isinstance(row.result_payload, dict) else {}
+                source_payload = existing_payload.get("signal_source") if isinstance(existing_payload.get("signal_source"), dict) else {}
+                worker_results = (
+                    existing_payload.get("worker_executions")
+                    if isinstance(existing_payload.get("worker_executions"), list)
+                    else []
+                )
+                completed_keys = {
+                    str(item.get("worker_role_key") or "").strip()
+                    for item in worker_results
+                    if isinstance(item, dict) and str(item.get("worker_role_key") or "").strip()
+                }
+                pending_workers = [key for key in worker_role_keys if key not in completed_keys] or list(worker_role_keys)
+                failed_result = service._build_orchestration_result(
+                    trace_id=trace_id,
+                    response_mode=response_mode,
+                    planner_status=planner_status,
+                    planner_diagnostics=planner_diagnostics,
+                    summary=summary,
+                    registry_result={
+                        "registry_id": source_payload.get("registry_id"),
+                        "status": source_payload.get("status"),
+                    },
+                    source_payload={
+                        "source_kind": source_payload.get("source_kind"),
+                        "source_name": source_payload.get("source_name"),
+                        "source_identifier": source_payload.get("source_identifier"),
+                    },
+                    openclaw_result=openclaw_result,
+                    worker_results=worker_results,
+                    pending_workers=pending_workers,
+                    status="failed",
+                    run_status="failed",
+                )
+                failed_result["background_error"] = str(exc)[:500]
+                if row is not None:
+                    await run_service.update_run(
+                        tenant_id=tenant_id,
+                        trace_id=trace_id,
+                        status="failed",
+                        planner_status=planner_status,
+                        openclaw_status=str(openclaw_result.get("status") or row.openclaw_status or ""),
+                        worker_total=max(len(worker_role_keys), int(row.worker_total or 0)),
+                        worker_completed=int(row.worker_completed or 0),
+                        worker_failed=int(row.worker_failed or 0),
+                        summary=summary,
+                        planner_diagnostics=planner_diagnostics or {},
+                        result_payload=failed_result,
+                        completed_at=_utc_now_naive(),
+                    )
+                    await db.commit()
+                return failed_result
+
+    @classmethod
+    async def _execute_workers_and_finalize(
+        cls,
+        *,
+        tenant_id: int,
+        trace_id: str,
+        project_payload: dict[str, Any],
+        signal_payload: dict[str, Any],
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_role_keys: list[str],
+        operator_user_payload: Optional[dict[str, Any]],
+        planner_status: str,
+        planner_diagnostics: Optional[dict[str, Any]],
+        summary: str,
+        signal_id: Optional[int],
+    ) -> dict[str, Any]:
+        session_factory = getattr(cls, "session_factory", AsyncSessionLocal)
+        async with session_factory() as db:
+            service = cls(db)
+            bundle = await prompt_profile_service.get_profiles(db, tenant_id=tenant_id)
+            roles = bundle.get("roles") if isinstance(bundle.get("roles"), list) else []
+            role_map = {
+                str(item.get("role_key") or "").strip(): item
+                for item in roles
+                if isinstance(item, dict) and str(item.get("role_key") or "").strip()
+            }
+            semaphore = asyncio.Semaphore(WORKER_MAX_CONCURRENCY)
+
+            async def _run_role(role_key: str) -> dict[str, Any]:
+                role = role_map.get(role_key)
+                if not isinstance(role, dict):
+                    return {
+                        "worker_role_key": role_key,
+                        "worker_role_name": role_key,
+                        "status": "failed",
+                        "reason": "role_not_configured",
+                    }
+                try:
+                    async with semaphore:
+                        return await cls._execute_single_worker_role_isolated(
+                            tenant_id=tenant_id,
+                            project_payload=project_payload,
+                            signal_payload=signal_payload,
+                            registry_result=registry_result,
+                            openclaw_result=openclaw_result,
+                            role=dict(role),
+                            trace_id=trace_id,
+                            operator_user_payload=operator_user_payload,
+                            response_mode="complete",
+                        )
+                except Exception as exc:
+                    return {
+                        "worker_role_key": role_key,
+                        "worker_role_name": str(role.get("role_name") or role_key).strip()[:64] or role_key,
+                        "status": "failed",
+                        "reason": str(exc)[:300],
+                    }
+
+            tasks = [asyncio.create_task(_run_role(role_key)) for role_key in worker_role_keys]
+            worker_results: list[dict[str, Any]] = []
+            for task in asyncio.as_completed(tasks):
+                result = await task
+                worker_results.append(result)
+                await service._update_run_after_worker(
+                    tenant_id=tenant_id,
+                    trace_id=trace_id,
+                    summary=summary,
+                    planner_status=planner_status,
+                    planner_diagnostics=planner_diagnostics,
+                    registry_result=registry_result,
+                    openclaw_result=openclaw_result,
+                    worker_role_keys=worker_role_keys,
+                    worker_results=worker_results,
+                )
+
+            final_result = await service._finalize_orchestration_run(
+                tenant_id=tenant_id,
+                trace_id=trace_id,
+                signal_id=signal_id,
+                summary=summary,
+                planner_status=planner_status,
+                planner_diagnostics=planner_diagnostics,
+                registry_result=registry_result,
+                openclaw_result=openclaw_result,
+                worker_role_keys=worker_role_keys,
+                worker_results=worker_results,
+            )
+            return final_result
+
+    @staticmethod
+    def _serialize_snapshot(value: Any) -> Optional[dict[str, Any]]:
+        if value is None:
+            return None
+        if isinstance(value, dict):
+            return dict(value)
+        data = getattr(value, "__dict__", None)
+        if isinstance(data, dict):
+            return {str(key): data[key] for key in data}
+        return None
+
+    @staticmethod
+    def _build_planner_diagnostics(extraction: Any) -> dict[str, Any]:
+        return {
+            "used_fallback": extraction.error_code is not None,
+            "error_code": extraction.error_code,
+            "source_field": extraction.source_field or None,
+            "provider": str(extraction.provider or ""),
+            "model": str(extraction.model or ""),
+        }
+
+    def _build_planned_worker_results(
+        self,
+        worker_role_keys: list[str],
+        *,
+        role_map: Optional[dict[str, dict[str, Any]]] = None,
+    ) -> list[dict[str, Any]]:
+        results: list[dict[str, Any]] = []
+        for role_key in worker_role_keys:
+            role = role_map.get(role_key) if isinstance(role_map, dict) else None
+            role_name = str((role or {}).get("role_name") or role_key).strip()[:64] or role_key
+            results.append(
+                {
+                    "worker_role_key": role_key,
+                    "worker_role_name": role_name,
+                    "status": "planned",
+                }
+            )
+        return results
+
+    @staticmethod
+    def _build_signal_source_snapshot(registry_result: dict[str, Any], source_payload: dict[str, Any]) -> dict[str, Any]:
+        return {
+            "registry_id": registry_result.get("registry_id"),
+            "status": registry_result.get("status"),
+            "onboarding_status": registry_result.get("onboarding_status"),
+            "source_kind": source_payload.get("source_kind"),
+            "source_name": source_payload.get("source_name"),
+            "source_identifier": source_payload.get("source_identifier"),
+            "description": source_payload.get("description"),
+            "config": source_payload.get("config") if isinstance(source_payload.get("config"), dict) else {},
+        }
+
+    def _build_orchestration_result(
+        self,
+        *,
+        trace_id: str,
+        response_mode: str,
+        planner_status: str,
+        planner_diagnostics: Optional[dict[str, Any]],
+        summary: str,
+        registry_result: dict[str, Any],
+        source_payload: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_results: list[dict[str, Any]],
+        pending_workers: list[str],
+        status: str,
+        run_status: str,
+    ) -> dict[str, Any]:
+        return {
+            "trace_id": trace_id,
+            "status": status,
+            "run_status": run_status,
+            "response_mode": response_mode,
+            "planner_status": planner_status,
+            "planner_diagnostics": planner_diagnostics,
+            "summary": summary,
+            "pending_workers": list(pending_workers),
+            "signal_source": self._build_signal_source_snapshot(registry_result, source_payload),
+            "openclaw": openclaw_result,
+            "worker_executions": list(worker_results),
+        }
+
+    @classmethod
+    def _has_actionable_source_config(cls, *, source_payload: dict[str, Any]) -> bool:
+        if not isinstance(source_payload, dict):
+            return False
+        config = source_payload.get("config") if isinstance(source_payload.get("config"), dict) else {}
+        if config:
+            return True
+        locator = cls._extract_explicit_source_locator(
+            source_payload=source_payload,
+            text=str(source_payload.get("description") or ""),
+        )
+        if locator:
+            return True
+        source_kind = str(source_payload.get("source_kind") or "").strip().lower()
+        source_identifier = str(source_payload.get("source_identifier") or "").strip().lower()
+        if source_kind in {"github", "feishu", "webhook", "url"} and source_identifier and ":req-" not in source_identifier:
+            return True
+        if source_kind in {"pmflow", "custom", "browser"} and (
+            not source_identifier or source_identifier.startswith(f"{source_kind}:req-")
+        ):
+            return False
+        return False
+
+    @staticmethod
+    def _has_non_fallback_worker_result(worker_results: list[dict[str, Any]]) -> bool:
+        for item in worker_results:
+            if not isinstance(item, dict):
+                continue
+            status = str(item.get("status") or "").strip().lower()
+            if status != "completed":
+                continue
+            if bool(item.get("used_fallback")):
+                continue
+            return True
+        return False
+
+    @staticmethod
+    def _extract_openclaw_execution_rows(openclaw_result: dict[str, Any]) -> list[dict[str, Any]]:
+        if not isinstance(openclaw_result, dict):
+            return []
+        result = openclaw_result.get("result") if isinstance(openclaw_result.get("result"), dict) else openclaw_result
+        execution = result.get("execution") if isinstance(result, dict) and isinstance(result.get("execution"), dict) else {}
+        rows = execution.get("results") if isinstance(execution.get("results"), list) else []
+        return [row for row in rows if isinstance(row, dict)]
+
+    @classmethod
+    def _has_material_openclaw_execution(cls, openclaw_result: dict[str, Any]) -> bool:
+        for row in cls._extract_openclaw_execution_rows(openclaw_result):
+            status = str(row.get("status") or "").strip().lower()
+            if status not in {"applied", "skipped"}:
+                continue
+            operation_code = str(row.get("operation_code") or "").strip().lower()
+            if operation_code == "notification.create":
+                continue
+            return True
+        return False
+
+    @classmethod
+    def _resolve_registry_activation(
+        cls,
+        *,
+        source_payload: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_results: list[dict[str, Any]],
+    ) -> tuple[str, str]:
+        actionable_config = cls._has_actionable_source_config(source_payload=source_payload)
+        has_real_worker = cls._has_non_fallback_worker_result(worker_results)
+        has_material_openclaw = cls._has_material_openclaw_execution(openclaw_result)
+        if actionable_config and has_real_worker and has_material_openclaw:
+            return "active", "completed"
+        return "pending", "pending_manual"
+
+    @staticmethod
+    def _resolve_orchestration_status(*, openclaw_status: str, worker_results: list[dict[str, Any]]) -> str:
+        normalized_openclaw = str(openclaw_status or "").strip().lower()
+        worker_statuses = [str(item.get("status") or "").strip().lower() for item in worker_results]
+        has_worker_completed = any(status == "completed" for status in worker_statuses)
+        has_worker_failed = any(status == "failed" for status in worker_statuses)
+        if normalized_openclaw == "failed" and not has_worker_completed:
+            return "failed"
+        if has_worker_failed or normalized_openclaw == "failed":
+            return "partial"
+        if has_worker_completed or normalized_openclaw == "completed":
+            return "completed"
+        return "failed"
+
+    @staticmethod
+    def _worker_counts(worker_results: list[dict[str, Any]]) -> tuple[int, int]:
+        completed = 0
+        failed = 0
+        for item in worker_results:
+            status = str(item.get("status") or "").strip().lower()
+            if status == "failed":
+                failed += 1
+                continue
+            if status in {"completed", "planned", "skipped"}:
+                completed += 1
+        return completed, failed
+
+    async def _update_run_after_worker(
+        self,
+        *,
+        tenant_id: int,
+        trace_id: str,
+        summary: str,
+        planner_status: str,
+        planner_diagnostics: Optional[dict[str, Any]],
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_role_keys: list[str],
+        worker_results: list[dict[str, Any]],
+    ) -> None:
+        run_service = AIOrchestrationRunService(self.db)
+        row = await run_service.get_run(tenant_id=tenant_id, trace_id=trace_id)
+        if row is None:
+            return
+        source_payload = row.result_payload.get("signal_source") if isinstance(row.result_payload, dict) else {}
+        completed_count, failed_count = self._worker_counts(worker_results)
+        completed_keys = {
+            str(item.get("worker_role_key") or "").strip()
+            for item in worker_results
+            if str(item.get("worker_role_key") or "").strip()
+        }
+        pending_workers = [key for key in worker_role_keys if key not in completed_keys]
+        status = "partial" if pending_workers else self._resolve_orchestration_status(
+            openclaw_status=str(openclaw_result.get("status") or ""),
+            worker_results=worker_results,
+        )
+        result_payload = self._build_orchestration_result(
+            trace_id=trace_id,
+            response_mode=str(row.response_mode or "fast"),
+            planner_status=planner_status,
+            planner_diagnostics=planner_diagnostics,
+            summary=summary,
+            registry_result={
+                "registry_id": (source_payload or {}).get("registry_id"),
+                "status": (source_payload or {}).get("status"),
+            },
+            source_payload={
+                "source_kind": (source_payload or {}).get("source_kind"),
+                "source_name": (source_payload or {}).get("source_name"),
+                "source_identifier": (source_payload or {}).get("source_identifier"),
+            },
+            openclaw_result=openclaw_result,
+            worker_results=worker_results,
+            pending_workers=pending_workers,
+            status=status,
+            run_status="running" if pending_workers else status,
+        )
+        await run_service.update_run(
+            tenant_id=tenant_id,
+            trace_id=trace_id,
+            status="running" if pending_workers else status,
+            planner_status=planner_status,
+            openclaw_status=str(openclaw_result.get("status") or ""),
+            worker_total=len(worker_role_keys),
+            worker_completed=completed_count,
+            worker_failed=failed_count,
+            summary=summary,
+            planner_diagnostics=planner_diagnostics or {},
+            result_payload=result_payload,
+            completed_at=None if pending_workers else _utc_now_naive(),
+        )
+        await self.db.commit()
+
+    async def _finalize_orchestration_run(
+        self,
+        *,
+        tenant_id: int,
+        trace_id: str,
+        signal_id: Optional[int],
+        summary: str,
+        planner_status: str,
+        planner_diagnostics: Optional[dict[str, Any]],
+        registry_result: dict[str, Any],
+        openclaw_result: dict[str, Any],
+        worker_role_keys: list[str],
+        worker_results: list[dict[str, Any]],
+    ) -> dict[str, Any]:
+        run_service = AIOrchestrationRunService(self.db)
+        row = await run_service.get_run(tenant_id=tenant_id, trace_id=trace_id)
+        response_mode = str(row.response_mode or "fast") if row is not None else "fast"
+        source_payload = row.result_payload.get("signal_source") if row is not None and isinstance(row.result_payload, dict) else {}
+        final_status = self._resolve_orchestration_status(
+            openclaw_status=str(openclaw_result.get("status") or ""),
+            worker_results=worker_results,
+        )
+        registry_status, onboarding_status = self._resolve_registry_activation(
+            source_payload=source_payload if isinstance(source_payload, dict) else {},
+            openclaw_result=openclaw_result,
+            worker_results=worker_results,
+        )
+        if final_status == "completed" and onboarding_status != "completed":
+            final_status = "pending_manual"
+        completed_count, failed_count = self._worker_counts(worker_results)
+        final_result = self._build_orchestration_result(
+            trace_id=trace_id,
+            response_mode=response_mode,
+            planner_status=planner_status,
+            planner_diagnostics=planner_diagnostics,
+            summary=summary,
+            registry_result={
+                "registry_id": (source_payload or {}).get("registry_id"),
+                "status": registry_status,
+                "onboarding_status": onboarding_status,
+            },
+            source_payload={
+                "source_kind": (source_payload or {}).get("source_kind"),
+                "source_name": (source_payload or {}).get("source_name"),
+                "source_identifier": (source_payload or {}).get("source_identifier"),
+            },
+            openclaw_result=openclaw_result,
+            worker_results=worker_results,
+            pending_workers=[],
+            status=final_status,
+            run_status=final_status,
+        )
+        await run_service.update_run(
+            tenant_id=tenant_id,
+            trace_id=trace_id,
+            status=final_status,
+            planner_status=planner_status,
+            openclaw_status=str(openclaw_result.get("status") or ""),
+            worker_total=len(worker_role_keys),
+            worker_completed=completed_count,
+            worker_failed=failed_count,
+            summary=summary,
+            planner_diagnostics=planner_diagnostics or {},
+            result_payload=final_result,
+            completed_at=_utc_now_naive(),
+        )
+        if registry_result.get("registry_id") is not None:
+            registry_row = await self._load_signal_source_registry(
+                tenant_id=tenant_id,
+                registry_id=int(registry_result["registry_id"]),
+            )
+            if registry_row is not None:
+                registry_row.last_signal_at = _utc_now_naive()
+                registry_row.last_worker_run_at = _utc_now_naive()
+                registry_row.status = registry_status
+                registry_row.onboarding_status = onboarding_status
+        if signal_id:
+            persisted_signal = await self._load_signal(tenant_id=tenant_id, signal_id=int(signal_id))
+            if persisted_signal is not None:
+                await self._persist_signal_orchestration_result(signal=persisted_signal, orchestration_result=final_result)
+        await self.db.commit()
+        return final_result
+
+    @staticmethod
+    def _build_worker_prompt(
+        *,
+        role: dict[str, Any],
+        project: Project,
+        signal: Signal,
+        source_payload: dict[str, Any],
+        openclaw_result: dict[str, Any],
+    ) -> str:
+        role_payload = {
+            "role_key": str(role.get("role_key") or "").strip(),
+            "role_name": str(role.get("role_name") or "").strip(),
+            "role_description": str(role.get("role_description") or "").strip(),
+            "system_prompt": str(role.get("system_prompt") or "").strip(),
+        }
+        project_payload = {
+            "project_id": int(project.id),
+            "title": str(project.title or project.name or "").strip(),
+            "description": str(project.description or "").strip()[:1000],
+        }
+        signal_payload = {
+            "signal_id": int(signal.id) if getattr(signal, "id", None) is not None else None,
+            "title": str(signal.title or "").strip(),
+            "summary": str(signal.summary or signal.content or "").strip()[:1200],
+        }
+        commander_payload = openclaw_result.get("result") if isinstance(openclaw_result.get("result"), dict) else openclaw_result
+        embedded_prompt = PROMPT_AI_WORKER_SIGNAL_SOURCE_EXECUTION.format(
+            role_json=json.dumps(role_payload, ensure_ascii=False),
+            project_json=json.dumps(project_payload, ensure_ascii=False),
+            source_json=json.dumps(source_payload, ensure_ascii=False),
+            signal_json=json.dumps(signal_payload, ensure_ascii=False),
+            commander_json=json.dumps(commander_payload or {}, ensure_ascii=False),
+        )
+        system_prompt = str(role.get("system_prompt") or "").strip()
+        if system_prompt:
+            return f"[Role System Prompt]\n{system_prompt}\n\n{embedded_prompt}"
+        return embedded_prompt
+
+    @staticmethod
+    def _normalize_worker_output(parsed: dict[str, Any], *, role_name: str, source_name: Any) -> dict[str, Any]:
+        deliverables = parsed.get("deliverables") if isinstance(parsed.get("deliverables"), list) else []
+        risks = parsed.get("risks") if isinstance(parsed.get("risks"), list) else []
+        next_actions = parsed.get("next_actions") if isinstance(parsed.get("next_actions"), list) else []
+        return {
+            "summary": str(parsed.get("summary") or f"{role_name} 已完成 {source_name} 的接入分析。").strip()[:1000],
+            "deliverables": [str(item).strip()[:200] for item in deliverables if str(item).strip()][:4],
+            "risks": [str(item).strip()[:200] for item in risks if str(item).strip()][:4],
+            "next_actions": [str(item).strip()[:200] for item in next_actions if str(item).strip()][:4],
+            "used_fallback": False,
+        }
+
+    @staticmethod
+    def _fallback_worker_output(*, role_name: str, source_name: str) -> dict[str, Any]:
+        return {
+            "summary": f"{role_name} 已为 {source_name} 生成兜底执行摘要。",
+            "deliverables": [f"梳理 {source_name} 的接入目标与边界"],
+            "risks": [f"{source_name} 的标识与接入细节仍需补充确认"],
+            "next_actions": ["等待 OpenClaw 生成任务并推进执行"],
+            "used_fallback": True,
+        }
+
+    async def _persist_signal_orchestration_result(
+        self,
+        *,
+        signal: Signal,
+        orchestration_result: dict[str, Any],
+    ) -> None:
+        strategic_payload: dict[str, Any] = {}
+        raw_strategic = signal.strategic_analysis
+        if isinstance(raw_strategic, dict):
+            strategic_payload = dict(raw_strategic)
+        elif isinstance(raw_strategic, str) and raw_strategic.strip():
+            try:
+                parsed = json.loads(raw_strategic)
+            except Exception:
+                parsed = {}
+            if isinstance(parsed, dict):
+                strategic_payload = parsed
+        strategic_payload["signal_source_orchestration"] = orchestration_result
+        signal.strategic_analysis = json.dumps(strategic_payload, ensure_ascii=False)
+        orchestration_status = str(orchestration_result.get("status") or "").lower()
+        signal.processed = orchestration_status in {"completed", "partial"}
+        signal.status = self._map_orchestration_status_to_signal_status(orchestration_status)
+        signal.tasks_generated = json.dumps(
+            self._extract_task_operations(orchestration_result.get("openclaw")),
+            ensure_ascii=False,
+        )
+        await self.db.commit()
+
+    @staticmethod
+    def _extract_task_operations(openclaw_payload: Any) -> list[dict[str, Any]]:
+        if not isinstance(openclaw_payload, dict):
+            return []
+        result = openclaw_payload.get("result") if isinstance(openclaw_payload.get("result"), dict) else openclaw_payload
+        execution = result.get("execution") if isinstance(result, dict) and isinstance(result.get("execution"), dict) else {}
+        rows = execution.get("results") if isinstance(execution.get("results"), list) else []
+        tasks: list[dict[str, Any]] = []
+        for row in rows:
+            if not isinstance(row, dict):
+                continue
+            operation_code = str(row.get("operation_code") or "").strip()
+            payload = row.get("payload") if isinstance(row.get("payload"), dict) else {}
+            if operation_code not in {"task.create", "task.assign", "task.prioritize", "task.status_transition"}:
+                continue
+            tasks.append(
+                {
+                    "operation_code": operation_code,
+                    "status": str(row.get("status") or "").strip(),
+                    "entity_id": row.get("entity_id"),
+                    "task_id": payload.get("task_id"),
+                    "title": payload.get("title"),
+                    "summary": str(row.get("message") or "").strip(),
+                }
+            )
+        return tasks
+
+    @staticmethod
+    def _map_orchestration_status_to_signal_status(status: str) -> str:
+        normalized = str(status or "").strip().lower()
+        if normalized == "completed":
+            return "已编排"
+        if normalized == "pending_manual":
+            return "待确认接入"
+        if normalized == "partial":
+            return "部分编排"
+        if normalized == "failed":
+            return "编排失败"
+        if normalized == "planned":
+            return "已规划"
+        if normalized == "skipped":
+            return "已跳过"
+        return "待处理"
+
+    async def _resolve_operator_user(
+        self,
+        *,
+        tenant_id: int,
+        project: Project,
+        operator_user_id: Optional[int],
+    ) -> Optional[User]:
+        preferred_ids: list[int] = []
+        if operator_user_id is not None:
+            preferred_ids.append(int(operator_user_id))
+        if project.owner_id:
+            preferred_ids.append(int(project.owner_id))
+
+        for user_id in preferred_ids:
+            row = (
+                await self.db.execute(
+                    select(User).where(
+                        and_(
+                            User.tenant_id == tenant_id,
+                            User.id == user_id,
+                            User.is_active.is_(True),
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+            if row is not None and str(row.role or "").strip().upper() in {"BOSS", "ADMIN"}:
+                return row
+
+        return (
+            await self.db.execute(
+                select(User).where(
+                    and_(
+                        User.tenant_id == tenant_id,
+                        User.is_active.is_(True),
+                        User.role.in_(["BOSS", "ADMIN"]),
+                    )
+                ).order_by(User.id.asc()).limit(1)
+                )
+            ).scalar_one_or_none()
+
+    @staticmethod
+    def _build_user_snapshot(user: Optional[User]) -> Optional[SimpleNamespace]:
+        if user is None:
+            return None
+        user_id = getattr(user, "id", None)
+        if user_id is None:
+            return None
+        return SimpleNamespace(
+            id=int(user_id),
+            name=str(getattr(user, "name", "") or "").strip(),
+            username=str(getattr(user, "username", "") or "").strip(),
+            role=str(getattr(user, "role", "") or "").strip(),
+        )
+
+    async def _load_project(self, *, tenant_id: int, project_id: int) -> Optional[Project]:
+        return (
+            await self.db.execute(
+                select(Project).where(
+                    and_(
+                        Project.tenant_id == tenant_id,
+                        Project.id == project_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _load_signal(self, *, tenant_id: int, signal_id: int) -> Optional[Signal]:
+        return (
+            await self.db.execute(
+                select(Signal).where(
+                    and_(
+                        Signal.tenant_id == tenant_id,
+                        Signal.id == signal_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def _load_signal_source_registry(
+        self,
+        *,
+        tenant_id: int,
+        registry_id: int,
+    ) -> Optional[SignalSourceRegistry]:
+        return (
+            await self.db.execute(
+                select(SignalSourceRegistry).where(
+                    and_(
+                        SignalSourceRegistry.tenant_id == tenant_id,
+                        SignalSourceRegistry.id == registry_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+
+    @staticmethod
+    def _build_ephemeral_signal(*, initiating_text: str) -> Signal:
+        signal = Signal(
+            tenant_id=1,
+            source="PMFlow",
+            content=initiating_text[:2000],
+            impact_score=0.7,
+            priority="medium",
+            processed=False,
+            title=initiating_text[:80] or "新增信号源",
+            signal_type="新需求",
+            summary=initiating_text[:300],
+            recommended_action="enter_only",
+            status="待处理",
+        )
+        signal.id = 0  # type: ignore[assignment]
+        return signal
+
+    @staticmethod
+    def _parse_json_object(raw: str) -> dict[str, Any]:
+        text = str(raw or "").strip()
+        if "```" in text:
+            match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.S)
+            if match:
+                text = match.group(1)
+        if not text.startswith("{"):
+            match = re.search(r"(\{.*\})", text, re.S)
+            if match:
+                text = match.group(1)
+        parsed = json.loads(text)
+        if not isinstance(parsed, dict):
+            raise ValueError("json_object_required")
+        return parsed
diff --git a/backend/app/services/skill_score_service.py b/backend/app/services/skill_score_service.py
new file mode 100644
index 00000000..d2a9fe06
--- /dev/null
+++ b/backend/app/services/skill_score_service.py
@@ -0,0 +1,221 @@
+"""任务完成后的技能分更新服务。"""
+
+from __future__ import annotations
+
+import json
+from datetime import datetime
+from typing import Optional
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import PROMPT_SKILL_SCORE_UPDATE
+from ..database import AsyncSessionLocal
+from ..models import AICallLog, Task, UserSkillBinding
+from .llm import llm_service
+
+PRIORITY_WEIGHT = {
+    "critical": 1.5,
+    "high": 1.2,
+    "medium": 1.0,
+    "low": 0.8,
+}
+
+
+def _clamp(value: float, lower: float, upper: float) -> float:
+    return max(lower, min(upper, value))
+
+
+def _normalize_status(status: Optional[str]) -> str:
+    return (status or "").strip().lower()
+
+
+def _parse_json(text: str) -> dict:
+    body = (text or "").strip()
+    if body.startswith("```json"):
+        body = body[7:].strip()
+    if body.startswith("```"):
+        body = body[3:].strip()
+    if body.endswith("```"):
+        body = body[:-3].strip()
+    return json.loads(body)
+
+
+def _heuristic_deltas(
+    task: Task,
+    required_skill_ids: list[int],
+) -> list[dict]:
+    quality_score = float(task.quality_score or 75.0)
+    estimated = max(float(task.estimated_hours or 0.0), 0.0)
+    actual = max(float(task.actual_hours or estimated or 1.0), 1.0)
+    efficiency_ratio = estimated / actual if estimated > 0 else 1.0
+
+    quality_delta = _clamp((quality_score - 75.0) / 10.0, -3.0, 3.0)
+    efficiency_delta = _clamp((efficiency_ratio - 1.0) * 4.0, -3.0, 3.0)
+    return [
+        {
+            "skill_tag_id": skill_id,
+            "quality_delta": quality_delta,
+            "efficiency_delta": efficiency_delta,
+        }
+        for skill_id in required_skill_ids
+    ]
+
+
+async def _apply_skill_deltas(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    user_id: int,
+    priority: str,
+    deltas: list[dict],
+) -> int:
+    binding_result = await db.execute(
+        select(UserSkillBinding).where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.user_id == user_id,
+            )
+        )
+    )
+    bindings = list(binding_result.scalars().all())
+    binding_map = {int(row.skill_tag_id): row for row in bindings}
+    weight = PRIORITY_WEIGHT.get(priority, 1.0)
+
+    touched = 0
+    for item in deltas:
+        try:
+            skill_tag_id = int(item.get("skill_tag_id"))
+        except (TypeError, ValueError):
+            continue
+        row = binding_map.get(skill_tag_id)
+        if not row:
+            row = UserSkillBinding(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                skill_tag_id=skill_tag_id,
+                proficiency_level="beginner",
+                level="beginner",
+                proficiency=0,
+                quality_score=0.0,
+                efficiency_score=0.0,
+            )
+            db.add(row)
+            binding_map[skill_tag_id] = row
+
+        quality_delta = float(item.get("quality_delta", 0.0)) * weight
+        efficiency_delta = float(item.get("efficiency_delta", 0.0)) * weight
+        row.quality_score = round(
+            _clamp(float(row.quality_score or 0.0) + quality_delta, 0.0, 100.0),
+            1,
+        )
+        row.efficiency_score = round(
+            _clamp(float(row.efficiency_score or 0.0) + efficiency_delta, 0.0, 100.0),
+            1,
+        )
+        combined = (float(row.quality_score or 0.0) * 0.5) + (
+            float(row.efficiency_score or 0.0) * 0.5
+        )
+        row.proficiency = int(_clamp(round(combined), 0, 100))
+        row.last_used_at = datetime.utcnow()
+        row.updated_at = datetime.utcnow()
+        touched += 1
+
+    return touched
+
+
+async def process_task_completion(
+    *,
+    task_id: int,
+    tenant_id: int,
+    operator_user_id: Optional[int] = None,
+) -> dict:
+    """任务完成后更新 assignee 的技能质量/效率分。"""
+    async with AsyncSessionLocal() as db:
+        task = (
+            await db.execute(
+                select(Task).where(
+                    and_(
+                        Task.tenant_id == tenant_id,
+                        Task.id == task_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not task:
+            return {"updated_count": 0, "reason": "task_not_found"}
+        if _normalize_status(task.status) not in {"done", "completed", "closed"}:
+            return {"updated_count": 0, "reason": "task_not_done"}
+        if not task.assignee_id:
+            return {"updated_count": 0, "reason": "task_no_assignee"}
+
+        required_skill_ids = [
+            int(v) for v in (task.required_skill_tag_ids or []) if v is not None
+        ]
+        if not required_skill_ids:
+            return {"updated_count": 0, "reason": "task_no_required_skills"}
+
+        prompt = PROMPT_SKILL_SCORE_UPDATE.format(
+            task_id=int(task.id),
+            priority=str(task.priority or "medium"),
+            quality_score=float(task.quality_score or 75.0),
+            estimated_hours=float(task.estimated_hours or 0.0),
+            actual_hours=float(task.actual_hours or float(task.estimated_hours or 1.0) or 1.0),
+            required_skill_tag_ids=required_skill_ids,
+        )
+
+        response_text = ""
+        status = "success"
+        deltas: list[dict]
+        try:
+            response_text = await llm_service.generate(
+                prompt,
+                function_name="skill_score_update",
+                tenant_id=tenant_id,
+            )
+            payload = _parse_json(response_text)
+            parsed = payload.get("deltas", []) if isinstance(payload, dict) else []
+            deltas = []
+            for row in parsed:
+                if not isinstance(row, dict):
+                    continue
+                try:
+                    deltas.append(
+                        {
+                            "skill_tag_id": int(row.get("skill_tag_id")),
+                            "quality_delta": float(row.get("quality_delta", 0.0)),
+                            "efficiency_delta": float(row.get("efficiency_delta", 0.0)),
+                        }
+                    )
+                except (TypeError, ValueError):
+                    continue
+            if not deltas:
+                deltas = _heuristic_deltas(task, required_skill_ids)
+                status = "fallback"
+        except Exception as exc:
+            response_text = str(exc)
+            deltas = _heuristic_deltas(task, required_skill_ids)
+            status = "fallback"
+
+        updated = await _apply_skill_deltas(
+            db,
+            tenant_id=tenant_id,
+            user_id=int(task.assignee_id),
+            priority=str(task.priority or "medium").lower(),
+            deltas=deltas,
+        )
+        db.add(
+            AICallLog(
+                tenant_id=tenant_id,
+                user_id=operator_user_id or int(task.assignee_id),
+                function_name="skill_score_update",
+                prompt=prompt,
+                response=(response_text or json.dumps({"deltas": deltas}, ensure_ascii=False))[:8000],
+                tokens_used=max(len(prompt) // 4, 1) + max(len(response_text) // 4, 0),
+                cost=((max(len(prompt) // 4, 1) * 0.00015 + max(len(response_text) // 4, 0) * 0.0006) / 1000),
+                status=status,
+            )
+        )
+        await db.commit()
+        return {"updated_count": updated, "status": status}
+
diff --git a/backend/app/services/skill_service.py b/backend/app/services/skill_service.py
new file mode 100644
index 00000000..c6d0eed7
--- /dev/null
+++ b/backend/app/services/skill_service.py
@@ -0,0 +1,623 @@
+"""
+PMFlow 人员能力图谱 - 技能标签与用户技能绑定服务
+
+- 系统预设 AI 开发技能标签初始化
+- SkillTag CRUD（系统标签禁止删改）
+- UserSkillBinding 批量更新与查询
+- UserWorkload 自动更新
+"""
+
+import uuid as uuid_module
+from datetime import datetime, timedelta
+from typing import List, Optional
+
+from sqlalchemy import and_, func, select, text
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import selectinload
+
+from ..models import SkillTag, UserSkillBinding, UserWorkload, User, UserStatus, Task
+from ..database import _is_sqlite
+
+
+# 系统预设 AI 开发专属技能标签
+SYSTEM_SKILL_TAGS = [
+    {"name": "Python", "slug": "python", "category": "language", "sort_order": 1},
+    {"name": "TypeScript", "slug": "typescript", "category": "language", "sort_order": 2},
+    {"name": "FastAPI", "slug": "fastapi", "category": "framework", "sort_order": 3},
+    {"name": "React", "slug": "react", "category": "framework", "sort_order": 4},
+    {"name": "SQLAlchemy", "slug": "sqlalchemy", "category": "framework", "sort_order": 5},
+    {"name": "LangChain", "slug": "langchain", "category": "framework", "sort_order": 6},
+    {"name": "LLM 调优", "slug": "llm-tuning", "category": "skill", "sort_order": 7},
+    {"name": "RAG 构建", "slug": "rag", "category": "skill", "sort_order": 8},
+    {"name": "向量数据库", "slug": "vector-db", "category": "tool", "sort_order": 9},
+    {"name": "Ollama", "slug": "ollama", "category": "tool", "sort_order": 10},
+    {"name": "Docker", "slug": "docker", "category": "tool", "sort_order": 11},
+    {"name": "Git", "slug": "git", "category": "tool", "sort_order": 12},
+]
+
+
+async def init_system_skill_tags(db: AsyncSession, tenant_id: int = 1) -> int:
+    """
+    初始化系统预设 AI 开发技能标签。
+    项目启动时自动执行，已存在的标签不会重复创建。
+    SQLite 下用 raw INSERT 显式指定 id，避免 NOT NULL constraint failed: skill_tags.id。
+    """
+    result = await db.execute(
+        select(SkillTag).where(
+            and_(SkillTag.tenant_id == tenant_id, SkillTag.is_system == True)
+        )
+    )
+    existing = result.scalars().all()
+    existing_slugs = {t.slug for t in existing}
+
+    to_create = [item for item in SYSTEM_SKILL_TAGS if item["slug"] not in existing_slugs]
+    if not to_create:
+        return 0
+
+    if _is_sqlite():
+        # SQLite 下部分表 id 无自增或为 UUID，用 raw SQL 显式指定 id
+        r = await db.execute(text("SELECT MAX(id) FROM skill_tags"))
+        max_id = r.scalar()
+        try:
+            next_id = int(max_id or 0) + 1
+            use_int_id = True
+        except (TypeError, ValueError):
+            use_int_id = False
+        for item in to_create:
+            if use_int_id:
+                tag_id = next_id
+                next_id += 1
+            else:
+                tag_id = str(uuid_module.uuid4()).replace("-", "")[:32]
+            await db.execute(
+                text("""
+                    INSERT INTO skill_tags (id, name, slug, category, description, is_system, sort_order, tenant_id)
+                    VALUES (:id, :name, :slug, :category, :description, :is_system, :sort_order, :tenant_id)
+                """),
+                {
+                    "id": tag_id,
+                    "name": item["name"],
+                    "slug": item["slug"],
+                    "category": item.get("category") or "tool",
+                    "description": None,
+                    "is_system": 1,
+                    "sort_order": item.get("sort_order", 0),
+                    "tenant_id": tenant_id,
+                },
+            )
+        await db.commit()
+        return len(to_create)
+
+    created = 0
+    for item in to_create:
+        tag = SkillTag(
+            tenant_id=tenant_id,
+            name=item["name"],
+            slug=item["slug"],
+            category=item.get("category"),
+            is_system=True,
+            sort_order=item.get("sort_order", 0),
+        )
+        db.add(tag)
+        created += 1
+    await db.commit()
+    return created
+
+
+# --------------------------------------------------------------------------
+# SkillTag CRUD
+# --------------------------------------------------------------------------
+
+
+async def create_skill_tag(
+    db: AsyncSession,
+    tenant_id: int,
+    name: str,
+    slug: str,
+    category: Optional[str] = None,
+    description: Optional[str] = None,
+    sort_order: int = 0,
+) -> SkillTag:
+    """创建技能标签（仅自定义标签）"""
+    existing = await db.execute(
+        select(SkillTag).where(
+            and_(SkillTag.tenant_id == tenant_id, SkillTag.slug == slug)
+        )
+    )
+    if existing.scalar_one_or_none():
+        raise ValueError(f"技能标签 slug 已存在: {slug}")
+
+    tag = SkillTag(
+        tenant_id=tenant_id,
+        name=name,
+        slug=slug,
+        category=category,
+        description=description,
+        is_system=False,
+        sort_order=sort_order,
+    )
+    db.add(tag)
+    await db.commit()
+    await db.refresh(tag)
+    return tag
+
+
+async def list_skill_tags(
+    db: AsyncSession,
+    tenant_id: int,
+    category: Optional[str] = None,
+    include_system: bool = True,
+) -> List[SkillTag]:
+    """列出技能标签"""
+    query = select(SkillTag).where(SkillTag.tenant_id == tenant_id)
+    if category:
+        query = query.where(SkillTag.category == category)
+    if not include_system:
+        query = query.where(SkillTag.is_system == False)
+    query = query.order_by(SkillTag.sort_order, SkillTag.name)
+    result = await db.execute(query)
+    return list(result.scalars().all())
+
+
+async def get_skill_tag(
+    db: AsyncSession, tag_id: int, tenant_id: int
+) -> Optional[SkillTag]:
+    """获取单个技能标签"""
+    result = await db.execute(
+        select(SkillTag).where(
+            and_(SkillTag.id == tag_id, SkillTag.tenant_id == tenant_id)
+        )
+    )
+    return result.scalar_one_or_none()
+
+
+async def update_skill_tag(
+    db: AsyncSession,
+    tag_id: int,
+    tenant_id: int,
+    name: Optional[str] = None,
+    category: Optional[str] = None,
+    description: Optional[str] = None,
+    sort_order: Optional[int] = None,
+) -> Optional[SkillTag]:
+    """更新技能标签（系统标签禁止修改）"""
+    tag = await get_skill_tag(db, tag_id, tenant_id)
+    if not tag:
+        return None
+    if tag.is_system:
+        raise ValueError("系统预设技能标签不可修改")
+
+    if name is not None:
+        tag.name = name
+    if category is not None:
+        tag.category = category
+    if description is not None:
+        tag.description = description
+    if sort_order is not None:
+        tag.sort_order = sort_order
+    tag.updated_at = datetime.utcnow()
+    await db.commit()
+    await db.refresh(tag)
+    return tag
+
+
+async def delete_skill_tag(
+    db: AsyncSession, tag_id: int, tenant_id: int
+) -> bool:
+    """删除技能标签（系统标签禁止删除）"""
+    tag = await get_skill_tag(db, tag_id, tenant_id)
+    if not tag:
+        return False
+    if tag.is_system:
+        raise ValueError("系统预设技能标签不可删除")
+
+    await db.delete(tag)
+    await db.commit()
+    return True
+
+
+# --------------------------------------------------------------------------
+# UserSkillBinding
+# --------------------------------------------------------------------------
+
+
+async def init_user_skill_bindings(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    *,
+    include_system_only: bool = True,
+    commit: bool = False,
+) -> int:
+    """
+    初始化用户技能绑定（空基线）。
+
+    为用户补齐未绑定的标签，默认 `beginner + proficiency=0`。
+    """
+    tags_query = select(SkillTag.id).where(SkillTag.tenant_id == tenant_id)
+    if include_system_only:
+        tags_query = tags_query.where(SkillTag.is_system == True)
+    tag_rows = await db.execute(tags_query)
+    tag_ids = [int(v) for v in tag_rows.scalars().all()]
+    if not tag_ids:
+        return 0
+
+    existing_rows = await db.execute(
+        select(UserSkillBinding.skill_tag_id).where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.user_id == user_id,
+                UserSkillBinding.skill_tag_id.in_(tag_ids),
+            )
+        )
+    )
+    existing_ids = {int(v) for v in existing_rows.scalars().all()}
+
+    created = 0
+    for skill_tag_id in tag_ids:
+        if skill_tag_id in existing_ids:
+            continue
+        db.add(
+            UserSkillBinding(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                skill_tag_id=skill_tag_id,
+                proficiency_level="beginner",
+                level="beginner",
+                proficiency=0,
+                quality_score=0.0,
+                efficiency_score=0.0,
+            )
+        )
+        created += 1
+
+    if created:
+        await db.flush()
+        if commit:
+            await db.commit()
+    return created
+
+
+async def batch_update_user_skill_bindings(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    bindings: List[dict],
+) -> List[UserSkillBinding]:
+    """
+    批量更新用户技能绑定。
+    若 skill_tag_id 已存在则更新，否则创建。
+    """
+    result_bindings = []
+
+    for item in bindings:
+        skill_tag_id = item.get("skill_tag_id")
+        if skill_tag_id is None:
+            continue
+        if isinstance(skill_tag_id, str):
+            try:
+                skill_tag_id = int(skill_tag_id)
+            except ValueError:
+                continue
+
+        proficiency = item.get("proficiency", 50)
+        level = item.get("level", "intermediate")
+        years = item.get("years_experience")
+
+        existing = await db.execute(
+            select(UserSkillBinding).where(
+                and_(
+                    UserSkillBinding.tenant_id == tenant_id,
+                    UserSkillBinding.user_id == user_id,
+                    UserSkillBinding.skill_tag_id == skill_tag_id,
+                )
+            )
+        )
+        binding = existing.scalar_one_or_none()
+
+        if binding:
+            binding.proficiency = proficiency
+            binding.level = level
+            binding.years_experience = years
+            binding.proficiency_level = level
+        else:
+            binding = UserSkillBinding(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                skill_tag_id=skill_tag_id,
+                proficiency_level=level,
+                proficiency=proficiency,
+                level=level,
+                years_experience=years,
+            )
+            db.add(binding)
+
+        result_bindings.append(binding)
+
+    await db.commit()
+    for b in result_bindings:
+        await db.refresh(b)
+    return result_bindings
+
+
+async def get_user_skill_bindings(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+) -> List[UserSkillBinding]:
+    """按用户查询技能绑定"""
+    result = await db.execute(
+        select(UserSkillBinding)
+        .options(selectinload(UserSkillBinding.skill_tag))
+        .where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.user_id == user_id,
+            )
+        )
+        .order_by(UserSkillBinding.proficiency.desc())
+    )
+    return list(result.scalars().all())
+
+
+async def get_skill_tag_user_bindings(
+    db: AsyncSession,
+    tenant_id: int,
+    skill_tag_id: int,
+) -> List[UserSkillBinding]:
+    """按技能标签查询用户绑定"""
+    result = await db.execute(
+        select(UserSkillBinding)
+        .options(
+            selectinload(UserSkillBinding.user),
+            selectinload(UserSkillBinding.skill_tag),
+        )
+        .where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.skill_tag_id == skill_tag_id,
+            )
+        )
+        .order_by(UserSkillBinding.proficiency.desc())
+    )
+    return list(result.scalars().all())
+
+
+async def get_user_skill_task_history(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    skill_tag_id: int,
+    limit: int = 50,
+) -> List[dict]:
+    """
+    查询用户某技能下的历史任务。
+
+    口径：
+    - assignee_id == user_id
+    - required_skill_tag_ids 包含 skill_tag_id
+    """
+    if _is_sqlite():
+        # SQLite 下 ARRAY 以 JSON 文本存储，先拉取后在 Python 侧做包含判断。
+        result = await db.execute(
+            select(Task)
+            .where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.assignee_id == user_id,
+                )
+            )
+            .order_by(Task.updated_at.desc())
+            .limit(max(limit * 5, 100))
+        )
+        tasks = result.scalars().all()
+        filtered = [
+            t
+            for t in tasks
+            if isinstance(t.required_skill_tag_ids, list)
+            and skill_tag_id in t.required_skill_tag_ids
+        ][:limit]
+    else:
+        result = await db.execute(
+            select(Task)
+            .where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.assignee_id == user_id,
+                    Task.required_skill_tag_ids.contains([skill_tag_id]),
+                )
+            )
+            .order_by(Task.updated_at.desc())
+            .limit(limit)
+        )
+        filtered = list(result.scalars().all())
+
+    return [
+        {
+            "id": str(t.id),
+            "title": t.title,
+            "description": t.description,
+            "status": t.status,
+            "priority": t.priority,
+            "estimated_hours": float(t.estimated_hours) if t.estimated_hours is not None else None,
+            "project_id": str(t.project_id) if t.project_id is not None else None,
+            "truth_node_id": str(t.truth_node_id) if t.truth_node_id is not None else None,
+            "updated_at": t.updated_at.isoformat() if t.updated_at else None,
+        }
+        for t in filtered
+    ]
+
+
+# --------------------------------------------------------------------------
+# UserWorkload 自动更新
+# --------------------------------------------------------------------------
+
+
+async def update_user_workload(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    total_tasks_delta: int = 0,
+    completed_tasks_delta: int = 0,
+    in_progress_delta: int = 0,
+    estimated_hours_delta: float = 0,
+    completed_hours_delta: float = 0,
+) -> Optional[UserWorkload]:
+    """
+    更新用户工作量。
+    新 schema: current_assigned_hours, workload_ratio
+    """
+    result = await db.execute(
+        select(UserWorkload).where(
+            and_(
+                UserWorkload.tenant_id == tenant_id,
+                UserWorkload.user_id == user_id,
+            )
+        )
+    )
+    workload = result.scalar_one_or_none()
+
+    if not workload:
+        workload = UserWorkload(
+            tenant_id=tenant_id,
+            user_id=user_id,
+            current_assigned_hours=0.0,
+            max_available_hours_per_week=40.0,
+            workload_ratio=0.0,
+        )
+        db.add(workload)
+        await db.flush()
+
+    delta = estimated_hours_delta + completed_hours_delta
+    workload.current_assigned_hours = max(
+        0, float(workload.current_assigned_hours or 0) + delta
+    )
+    max_hours = float(workload.max_available_hours_per_week or 40)
+    workload.workload_ratio = min(
+        100.0, (workload.current_assigned_hours / max_hours) * 100
+    )
+
+    await db.commit()
+    await db.refresh(workload)
+
+    user_result = await db.execute(select(User).where(User.id == user_id))
+    user = user_result.scalar_one_or_none()
+    if user:
+        user.workload = int(workload.workload_ratio)
+        await db.commit()
+
+    return workload
+
+
+# --------------------------------------------------------------------------
+# 人员推荐
+# --------------------------------------------------------------------------
+
+
+async def recommend_personnel(
+    db: AsyncSession,
+    tenant_id: int,
+    required_skills: List[str],
+    task_description: Optional[str] = None,
+    task_id: Optional[str] = None,
+    limit: int = 5,
+) -> dict:
+    """
+    根据所需技能推荐匹配人员。
+    返回匹配度评分、推荐理由等。
+    """
+    if not required_skills:
+        return {
+            "task_id": task_id,
+            "task_description": task_description,
+            "required_skills": [],
+            "recommendations": [],
+            "total_count": 0,
+        }
+
+    # 通过 slug 或 name 匹配技能标签
+    skill_tags_result = await db.execute(
+        select(SkillTag).where(
+            and_(
+                SkillTag.tenant_id == tenant_id,
+                (SkillTag.slug.in_(required_skills) | SkillTag.name.in_(required_skills)),
+            )
+        )
+    )
+    tags = skill_tags_result.scalars().all()
+    tag_ids = [t.id for t in tags]
+
+    if not tag_ids:
+        return {
+            "task_id": task_id,
+            "task_description": task_description,
+            "required_skills": required_skills,
+            "recommendations": [],
+            "total_count": 0,
+        }
+
+    # 查询拥有这些技能的用户及熟练度
+    bindings_result = await db.execute(
+        select(UserSkillBinding, SkillTag, User)
+        .join(SkillTag, UserSkillBinding.skill_tag_id == SkillTag.id)
+        .join(User, UserSkillBinding.user_id == User.id)
+        .where(
+            and_(
+                UserSkillBinding.tenant_id == tenant_id,
+                UserSkillBinding.skill_tag_id.in_(tag_ids),
+                User.status == UserStatus.ACTIVE,
+            )
+        )
+    )
+    rows = bindings_result.all()
+
+    # 按用户聚合
+    user_scores: dict = {}
+    for b, st, u in rows:
+        uid = str(u.id)
+        if uid not in user_scores:
+            user_scores[uid] = {
+                "user": u,
+                "skills": [],
+                "proficiencies": [],
+                "match_count": 0,
+            }
+        user_scores[uid]["skills"].append(st.name)
+        user_scores[uid]["proficiencies"].append(b.proficiency)
+        user_scores[uid]["match_count"] = len(user_scores[uid]["skills"])
+
+    # 计算匹配度并排序
+    recommendations = []
+    for uid, data in user_scores.items():
+        u = data["user"]
+        avg_prof = (
+            sum(data["proficiencies"]) / len(data["proficiencies"])
+            if data["proficiencies"]
+            else 0
+        )
+        match_ratio = len(data["skills"]) / len(required_skills) if required_skills else 0
+        match_score = min(100, (match_ratio * 60 + avg_prof * 0.4))
+        reason = f"匹配技能: {', '.join(data['skills'])}"
+        recommendations.append(
+            {
+                "user_id": uid,
+                "user_name": u.name,
+                "match_score": round(match_score, 1),
+                "reason": reason,
+                "skill_match_count": data["match_count"],
+                "avg_proficiency": round(avg_prof, 1),
+                "workload_rate": float(u.workload) if u.workload else 0,
+                "recommended_role": u.role.value if hasattr(u.role, "value") else str(u.role),
+            }
+        )
+
+    recommendations.sort(key=lambda x: (-x["match_score"], -x["skill_match_count"]))
+    recommendations = recommendations[:limit]
+
+    return {
+        "task_id": task_id,
+        "task_description": task_description,
+        "required_skills": required_skills,
+        "recommendations": recommendations,
+        "total_count": len(recommendations),
+    }
diff --git a/backend/app/services/sstc_service.py b/backend/app/services/sstc_service.py
new file mode 100644
index 00000000..cacab6df
--- /dev/null
+++ b/backend/app/services/sstc_service.py
@@ -0,0 +1,594 @@
+"""SSTC 看板聚合服务"""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from typing import Any, Iterable
+
+from sqlalchemy import and_, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import (
+    AlignmentRecord,
+    CognitiveConflict,
+    Project,
+    ProjectSignalLink,
+    Signal,
+    Task,
+    TruthNode,
+    User,
+    UserProject,
+    UserWorkload,
+)
+from ..utils.project_members import ensure_project_default_members
+
+DONE_STATUSES = {"done", "completed"}
+ACTIVE_STATUSES = {"in_progress", "review"}
+BLOCKED_STATUSES = {"blocked"}
+ALIGNED_STATUSES = {"aligned", "confirmed", "已对齐"}
+ALIGNING_STATUSES = {"aligning", "alignment_in_progress", "对齐中"}
+ARCHIVED_STATUSES = {"archived", "已归档"}
+RELEASE_VERIFIED_STATUSES = {"verified", "已验证"}
+RELEASE_LIVE_STATUSES = {"online", "deployed", "已上线"}
+RELEASE_RELEASING_STATUSES = {"releasing", "发布中"}
+
+
+def _to_json(value: Any, fallback: dict | None = None) -> dict:
+    if fallback is None:
+        fallback = {}
+    if isinstance(value, dict):
+        return value
+    if isinstance(value, str):
+        import json
+        try:
+            parsed = json.loads(value)
+            return parsed if isinstance(parsed, dict) else {"text": value}
+        except (TypeError, ValueError):
+            return {"text": value}
+    return fallback
+
+
+def _avg(values: Iterable[float]) -> float:
+    values = list(values)
+    if not values:
+        return 0.0
+    return float(sum(values) / len(values))
+
+
+class SSTCService:
+    def __init__(self, db: AsyncSession, tenant_id: int):
+        self.db = db
+        self.tenant_id = tenant_id
+
+    async def get_project(self, project_id: int) -> Project | None:
+        result = await self.db.execute(
+            select(Project).where(
+                and_(Project.id == project_id, Project.tenant_id == self.tenant_id)
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def get_signals(self, project_id: int) -> list[Signal]:
+        linked_result = await self.db.execute(
+            select(Signal)
+            .join(ProjectSignalLink, ProjectSignalLink.signal_id == Signal.id)
+            .where(
+                and_(
+                    ProjectSignalLink.project_id == project_id,
+                    Signal.tenant_id == self.tenant_id,
+                )
+            )
+            .order_by(Signal.created_at.desc())
+        )
+        linked_rows = list(linked_result.scalars().all())
+        merged: dict[int, Signal] = {
+            int(row.id): row
+            for row in linked_rows
+            if getattr(row, "id", None) is not None
+        }
+
+        fallback_result = await self.db.execute(
+            select(Signal)
+            .where(
+                and_(
+                    Signal.matched_project_id == project_id,
+                    Signal.tenant_id == self.tenant_id,
+                )
+            )
+            .order_by(Signal.created_at.desc())
+        )
+        for row in fallback_result.scalars().all():
+            if getattr(row, "id", None) is None:
+                continue
+            merged.setdefault(int(row.id), row)
+
+        return sorted(
+            merged.values(),
+            key=lambda row: row.created_at or datetime.min,
+            reverse=True,
+        )
+
+    async def get_charters(self, project_id: int) -> list[TruthNode]:
+        result = await self.db.execute(
+            select(TruthNode)
+            .where(
+                and_(
+                    TruthNode.project_id == project_id,
+                    TruthNode.tenant_id == self.tenant_id,
+                    TruthNode.node_type == "charter",
+                )
+            )
+            .order_by(TruthNode.created_at.desc())
+        )
+        return list(result.scalars().all())
+
+    async def get_tasks(self, project_id: int) -> list[Task]:
+        result = await self.db.execute(
+            select(Task).where(
+                and_(Task.project_id == project_id, Task.tenant_id == self.tenant_id)
+            )
+        )
+        return list(result.scalars().all())
+
+    async def get_releases(self, project_id: int) -> list[TruthNode]:
+        result = await self.db.execute(
+            select(TruthNode)
+            .where(
+                and_(
+                    TruthNode.project_id == project_id,
+                    TruthNode.tenant_id == self.tenant_id,
+                    TruthNode.node_type == "release",
+                )
+            )
+            .order_by(TruthNode.created_at.desc())
+        )
+        return list(result.scalars().all())
+
+    async def get_reviews(self, project_id: int) -> list[TruthNode]:
+        result = await self.db.execute(
+            select(TruthNode)
+            .where(
+                and_(
+                    TruthNode.project_id == project_id,
+                    TruthNode.tenant_id == self.tenant_id,
+                    TruthNode.node_type == "review",
+                )
+            )
+            .order_by(TruthNode.created_at.desc())
+        )
+        return list(result.scalars().all())
+
+    async def get_alignment_stats(self, charter_ids: list[int]) -> dict[str, Any]:
+        if not charter_ids:
+            return {"score": 0.0, "pending_nodes": 0, "unresolved_conflicts": 0}
+
+        ar_result = await self.db.execute(
+            select(AlignmentRecord).where(
+                and_(
+                    AlignmentRecord.truth_node_id.in_(charter_ids),
+                    AlignmentRecord.tenant_id == self.tenant_id,
+                )
+            )
+        )
+        records = list(ar_result.scalars().all())
+        score = _avg([float(r.alignment_score or 0) for r in records])
+
+        node_confirmed: dict[int, set[int]] = {}
+        for r in records:
+            if r.is_confirmed:
+                node_confirmed.setdefault(int(r.truth_node_id), set()).add(int(r.user_id))
+        pending_nodes = len([nid for nid in charter_ids if not node_confirmed.get(nid)])
+
+        conflict_result = await self.db.execute(
+            select(func.count(CognitiveConflict.id)).where(
+                and_(
+                    CognitiveConflict.truth_node_id.in_(charter_ids),
+                    CognitiveConflict.tenant_id == self.tenant_id,
+                    CognitiveConflict.status != "resolved",
+                )
+            )
+        )
+        unresolved_conflicts = int(conflict_result.scalar() or 0)
+        return {
+            "score": round(score, 2),
+            "pending_nodes": pending_nodes,
+            "unresolved_conflicts": unresolved_conflicts,
+        }
+
+    async def get_workload_health(self, assignee_ids: list[int]) -> dict[str, Any]:
+        if not assignee_ids:
+            return {
+                "avg_ratio": 0.0,
+                "overloaded_count": 0,
+                "idle_count": 0,
+                "member_count": 0,
+            }
+
+        result = await self.db.execute(
+            select(UserWorkload).where(
+                and_(
+                    UserWorkload.tenant_id == self.tenant_id,
+                    UserWorkload.user_id.in_(assignee_ids),
+                )
+            )
+        )
+        workloads = list(result.scalars().all())
+        ratios = []
+        for row in workloads:
+            ratio = float(row.workload_ratio or 0)
+            if ratio > 1.5:
+                ratio = ratio / 100.0
+            ratios.append(ratio)
+        avg_ratio = _avg(ratios)
+        overloaded_count = len([x for x in ratios if x > 1.0])
+        idle_count = len([x for x in ratios if x < 0.3])
+        return {
+            "avg_ratio": round(avg_ratio, 2),
+            "overloaded_count": overloaded_count,
+            "idle_count": idle_count,
+            "member_count": len(assignee_ids),
+        }
+
+    @staticmethod
+    def _task_metrics(tasks: list[Task]) -> dict[str, Any]:
+        total = len(tasks)
+        done = len([t for t in tasks if (t.status or "").lower() in DONE_STATUSES])
+        blocked = len([t for t in tasks if (t.status or "").lower() in BLOCKED_STATUSES])
+        in_progress = len([t for t in tasks if (t.status or "").lower() in ACTIVE_STATUSES])
+        today_done = len(
+            [
+                t
+                for t in tasks
+                if (t.status or "").lower() in DONE_STATUSES
+                and t.completed_at
+                and t.completed_at.date() == datetime.now().date()
+            ]
+        )
+        overdue = len(
+            [
+                t
+                for t in tasks
+                if t.due_date and t.due_date < datetime.now(timezone.utc).replace(tzinfo=None) and (t.status or "").lower() not in DONE_STATUSES
+            ]
+        )
+        return {
+            "total": total,
+            "done": done,
+            "blocked": blocked,
+            "in_progress": in_progress,
+            "today_done": today_done,
+            "overdue": overdue,
+            "progress": round((done / total) * 100, 2) if total else 0,
+        }
+
+    async def build_overview(self, project_id: int) -> dict[str, Any]:
+        project = await self.get_project(project_id)
+        if not project:
+            return {}
+        signals = await self.get_signals(project_id)
+        charters = await self.get_charters(project_id)
+        tasks = await self.get_tasks(project_id)
+        charters_ids = [int(c.id) for c in charters]
+        alignment = await self.get_alignment_stats(charters_ids)
+        metrics = self._task_metrics(tasks)
+        member_ids = await self.get_project_member_ids(project_id)
+        workload = await self.get_workload_health(member_ids)
+        chain_id = str(charters[0].id) if charters else None
+        return {
+            "project": {
+                "id": str(project.id),
+                "title": project.title or project.name,
+                "description": project.description,
+                "status": project.status,
+                "owner_id": str(project.owner_id) if project.owner_id else None,
+                "created_at": project.created_at.isoformat() if project.created_at else None,
+                "chain_id": chain_id,
+                "signal_count": len(signals),
+                "charter_count": len(charters),
+                "task_count": len(tasks),
+            },
+            "chain_id": chain_id,
+            "alignment_score": alignment["score"],
+            "pending_alignment_nodes": alignment["pending_nodes"],
+            "unresolved_conflicts": alignment["unresolved_conflicts"],
+            "task_progress": metrics,
+            "workload_health": workload,
+        }
+
+    async def build_stages(self, project_id: int) -> list[dict[str, Any]]:
+        signals = await self.get_signals(project_id)
+        charters = await self.get_charters(project_id)
+        tasks = await self.get_tasks(project_id)
+        releases = await self.get_releases(project_id)
+        reviews = await self.get_reviews(project_id)
+
+        has_p0 = any((s.priority or "").lower() == "p0" for s in signals)
+        has_pending_signal = any(not bool(s.processed) for s in signals)
+        signal_status = "已同步"
+        if has_p0 or has_pending_signal:
+            signal_status = "待处理"
+
+        latest_charter = charters[0] if charters else None
+        charter_status_raw = (latest_charter.status if latest_charter else "").lower()
+        if not latest_charter:
+            charter_status = "草稿"
+        elif charter_status_raw in ALIGNING_STATUSES:
+            charter_status = "对齐中"
+        elif charter_status_raw in ARCHIVED_STATUSES:
+            charter_status = "已归档"
+        elif charter_status_raw in ALIGNED_STATUSES:
+            charter_status = "已对齐"
+        else:
+            charter_status = "草稿"
+        charter_aligned = charter_status == "已对齐"
+
+        if not charter_aligned:
+            task_status = "未解锁"
+        elif not tasks:
+            task_status = "待确认"
+        elif all((t.status or "").lower() in DONE_STATUSES for t in tasks):
+            task_status = "已确认"
+        else:
+            task_status = "待确认"
+
+        if not tasks:
+            execution_status = "未开始"
+        elif all((t.status or "").lower() in DONE_STATUSES for t in tasks):
+            execution_status = "已完成"
+        elif any((t.status or "").lower() in (ACTIVE_STATUSES | BLOCKED_STATUSES) for t in tasks):
+            execution_status = "进行中"
+        else:
+            execution_status = "未开始"
+        metrics = self._task_metrics(tasks)
+        execution_done = execution_status == "已完成"
+        release_pre_unlocked = charter_aligned and metrics["done"] > 0
+
+        latest_release = releases[0] if releases else None
+        release_raw = (latest_release.status if latest_release else "").lower()
+        if not release_pre_unlocked:
+            release_status = "未解锁"
+        elif not execution_done and not latest_release:
+            release_status = "可预发布"
+        elif not latest_release:
+            release_status = "待发布"
+        elif release_raw in RELEASE_VERIFIED_STATUSES:
+            release_status = "已验证"
+        elif release_raw in RELEASE_RELEASING_STATUSES:
+            release_status = "发布中"
+        elif release_raw in RELEASE_LIVE_STATUSES:
+            release_status = "已上线"
+        else:
+            release_status = "可预发布" if not execution_done else "待发布"
+        release_verified = release_status == "已验证"
+        review_pre_unlocked = release_status in {"已上线", "已验证", "发布中"}
+
+        latest_review = reviews[0] if reviews else None
+        review_raw = (latest_review.status if latest_review else "").lower()
+        if not review_pre_unlocked:
+            review_status = "未解锁"
+        elif not release_verified and not latest_review:
+            review_status = "可阶段复盘"
+        elif not latest_review:
+            review_status = "草稿"
+        elif review_raw in ARCHIVED_STATUSES:
+            review_status = "已归档"
+        else:
+            review_status = "草稿"
+
+        return [
+            {
+                "key": "signal",
+                "label": "信号源",
+                "status": signal_status,
+                "summary": f"{len(signals)} 条有效信号",
+                "locked": False,
+                "editable": True,
+                "risk_count": int(has_p0),
+            },
+            {
+                "key": "charter",
+                "label": "需求宪章",
+                "status": charter_status,
+                "summary": f"{len(charters)} 份宪章",
+                "locked": False,
+                "editable": True,
+                "risk_count": 1 if charter_status == "对齐中" else 0,
+            },
+            {
+                "key": "task",
+                "label": "任务拆解",
+                "status": task_status,
+                "summary": f"{len(tasks)} 个任务",
+                "locked": not charter_aligned,
+                "editable": charter_aligned,
+                "risk_count": metrics["blocked"],
+            },
+            {
+                "key": "execution",
+                "label": "执行落地",
+                "status": execution_status,
+                "summary": f"完成率 {metrics['progress']}%",
+                "locked": not charter_aligned,
+                "editable": charter_aligned,
+                "risk_count": metrics["blocked"] + metrics["overdue"],
+            },
+            {
+                "key": "release",
+                "label": "发布验证",
+                "status": release_status,
+                "summary": f"{len(releases)} 次发布",
+                "locked": not release_pre_unlocked,
+                "editable": release_pre_unlocked,
+                "risk_count": metrics["overdue"],
+            },
+            {
+                "key": "review",
+                "label": "复盘沉淀",
+                "status": review_status,
+                "summary": f"{len(reviews)} 份复盘",
+                "locked": not review_pre_unlocked,
+                "editable": review_pre_unlocked,
+                "risk_count": 0,
+            },
+        ]
+
+    async def build_risks(self, project_id: int) -> dict[str, Any]:
+        charters = await self.get_charters(project_id)
+        tasks = await self.get_tasks(project_id)
+        charter_ids = [int(c.id) for c in charters]
+        conflict_rows = []
+        if charter_ids:
+            conflict_result = await self.db.execute(
+                select(CognitiveConflict).where(
+                    and_(
+                        CognitiveConflict.truth_node_id.in_(charter_ids),
+                        CognitiveConflict.tenant_id == self.tenant_id,
+                        CognitiveConflict.status != "resolved",
+                    )
+                )
+            )
+            conflict_rows = list(conflict_result.scalars().all())
+
+        blocked_tasks = [t for t in tasks if (t.status or "").lower() in BLOCKED_STATUSES]
+        overdue_tasks = [
+            t
+            for t in tasks
+            if t.due_date and t.due_date < datetime.now(timezone.utc).replace(tzinfo=None) and (t.status or "").lower() not in DONE_STATUSES
+        ]
+        risk_items = [
+            {
+                "id": f"task-blocked-{t.id}",
+                "type": "task_blocked",
+                "priority": "high",
+                "title": f"任务阻塞：{t.title}",
+                "entity_id": str(t.id),
+            }
+            for t in blocked_tasks
+        ] + [
+            {
+                "id": f"task-overdue-{t.id}",
+                "type": "task_overdue",
+                "priority": "medium",
+                "title": f"任务逾期：{t.title}",
+                "entity_id": str(t.id),
+            }
+            for t in overdue_tasks
+        ]
+
+        conflicts = [
+            {
+                "id": str(c.id),
+                "type": c.conflict_type,
+                "priority": "high",
+                "title": c.description[:80],
+                "entity_id": str(c.truth_node_id),
+            }
+            for c in conflict_rows
+        ]
+        return {
+            "conflicts": conflicts,
+            "risks": risk_items,
+            "summary": {
+                "conflict_count": len(conflicts),
+                "risk_count": len(risk_items),
+                "blocked_task_count": len(blocked_tasks),
+                "overdue_task_count": len(overdue_tasks),
+            },
+        }
+
+    async def build_chain(self, project_id: int) -> dict[str, Any]:
+        project = await self.get_project(project_id)
+        if not project:
+            return {}
+
+        signals = await self.get_signals(project_id)
+        charters = await self.get_charters(project_id)
+        tasks = await self.get_tasks(project_id)
+
+        tasks_by_charter: dict[str, list[dict[str, Any]]] = {}
+        for c in charters:
+            tasks_by_charter[str(c.id)] = [
+                {
+                    "id": str(t.id),
+                    "title": t.title,
+                    "status": t.status,
+                    "priority": t.priority,
+                    "charter_id": str(c.id),
+                    "description": t.description,
+                    "estimated_hours": float(t.estimated_hours) if t.estimated_hours else None,
+                    "assignee_id": str(t.assignee_id) if t.assignee_id else None,
+                }
+                for t in tasks
+                if int(t.truth_node_id) == int(c.id)
+            ]
+
+        return {
+            "project": {
+                "id": str(project.id),
+                "title": project.title or project.name,
+                "description": project.description,
+            },
+            "signals": [
+                {
+                    "id": str(s.id),
+                    "source": s.source,
+                    "content": (s.content or "")[:240],
+                    "impact_score": s.impact_score,
+                    "priority": s.priority,
+                    "processed": s.processed,
+                }
+                for s in signals
+            ],
+            "charters": [
+                {
+                    "id": str(c.id),
+                    "title": c.title,
+                    "content": _to_json(c.content),
+                    "signal_ids": _to_json(c.content).get("signal_ids", []),
+                    "status": c.status,
+                    "version": c.version,
+                    "parent_id": str(c.parent_id) if c.parent_id else None,
+                }
+                for c in charters
+            ],
+            "tasks_by_charter": tasks_by_charter,
+        }
+
+    async def has_member_confirmed(self, charter_id: int, user_id: int) -> bool:
+        result = await self.db.execute(
+            select(func.count(AlignmentRecord.id)).where(
+                and_(
+                    AlignmentRecord.tenant_id == self.tenant_id,
+                    AlignmentRecord.truth_node_id == charter_id,
+                    AlignmentRecord.user_id == user_id,
+                    AlignmentRecord.is_confirmed.is_(True),
+                )
+            )
+        )
+        return int(result.scalar() or 0) > 0
+
+    async def get_project_member_ids(self, project_id: int) -> list[int]:
+        member_result = await self.db.execute(
+            select(UserProject.user_id).where(
+                and_(
+                    UserProject.project_id == project_id,
+                    UserProject.tenant_id == self.tenant_id,
+                )
+            )
+        )
+        member_ids = {int(x) for x in member_result.scalars().all() if x is not None}
+        project = await self.get_project(project_id)
+        if not member_ids and project:
+            member_ids = set(
+                await ensure_project_default_members(
+                    self.db,
+                    tenant_id=self.tenant_id,
+                    project_id=int(project_id),
+                    owner_id=int(project.owner_id) if project.owner_id else None,
+                    assigned_role=project.assigned_role,
+                )
+            )
+            await self.db.commit()
+        if project and project.owner_id:
+            member_ids.add(int(project.owner_id))
+        return sorted(member_ids)
diff --git a/backend/app/services/symbol_index_service.py b/backend/app/services/symbol_index_service.py
new file mode 100644
index 00000000..a8a9a1d6
--- /dev/null
+++ b/backend/app/services/symbol_index_service.py
@@ -0,0 +1,776 @@
+"""Symbol 索引服务。
+
+目标：
+1. 对代码库执行增量符号扫描（def/ref）
+2. 提供符号查询、详情、概览
+3. 为 GraphRAG / Agent 检索提供结构化工具层
+"""
+
+from __future__ import annotations
+
+import ast
+import hashlib
+import re
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Optional
+
+from sqlalchemy import and_, delete, func, or_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import SymbolDefinition, SymbolReference
+
+
+TS_SYMBOL_RE = re.compile(
+    r"^\s*(?:export\s+)?(?:default\s+)?(?:async\s+)?(?:function|class|interface|type|const)\s+([A-Za-z_][A-Za-z0-9_]*)",
+    re.MULTILINE,
+)
+TS_IMPORT_RE = re.compile(r"^\s*import\s+.*?\s+from\s+['\"]([^'\"]+)['\"]", re.MULTILINE)
+TS_CALL_RE = re.compile(r"\b([A-Za-z_][A-Za-z0-9_]*)\s*\(")
+TS_KEYWORDS = {"if", "for", "while", "switch", "catch", "return", "new", "typeof", "await"}
+
+
+@dataclass(slots=True)
+class ParsedSymbol:
+    symbol_type: str
+    symbol_name: str
+    qualified_name: str
+    start_line: int
+    end_line: int
+    signature: str
+    docstring: str
+    payload: dict[str, Any]
+
+
+@dataclass(slots=True)
+class ParsedReference:
+    relation_type: str
+    target_symbol_name: str
+    target_qualified_name: str
+    source_qualified_name: str
+    start_line: int
+    end_line: int
+    payload: dict[str, Any]
+
+
+class SymbolIndexService:
+    def __init__(self, db: AsyncSession, *, tenant_id: int) -> None:
+        self.db = db
+        self.tenant_id = int(tenant_id)
+
+    @staticmethod
+    def _default_scan_roots(*, backend_root: Path, workspace_root: Path) -> list[Path]:
+        roots: list[Path] = []
+        for candidate in [
+            (backend_root / "app").resolve(),
+            (workspace_root / "frontend" / "src").resolve(),
+        ]:
+            if candidate.exists() and candidate.is_dir():
+                roots.append(candidate)
+        return roots
+
+    @staticmethod
+    def _to_rel_path(*, file_path: Path, workspace_root: Path, backend_root: Path) -> str:
+        try:
+            return str(file_path.resolve().relative_to(workspace_root))
+        except Exception:
+            try:
+                return str(file_path.resolve().relative_to(backend_root))
+            except Exception:
+                return str(file_path.resolve())
+
+    async def sync_codebase_symbols(
+        self,
+        *,
+        base_dir: Optional[str] = None,
+        max_files: int = 500,
+        force_reindex: bool = False,
+    ) -> dict[str, Any]:
+        backend_root = Path(__file__).resolve().parents[2]
+        workspace_root = backend_root.parent
+        if base_dir:
+            root = Path(base_dir).resolve()
+            if not root.exists() or not root.is_dir():
+                raise ValueError(f"symbol scan path not found: {root}")
+            scan_roots = [root]
+        else:
+            scan_roots = self._default_scan_roots(backend_root=backend_root, workspace_root=workspace_root)
+        if not scan_roots:
+            raise ValueError("symbol scan path not found: backend/app and frontend/src")
+
+        allowed = {".py", ".ts", ".tsx", ".js", ".jsx"}
+        files: list[Path] = []
+        for root in scan_roots:
+            files.extend([p for p in root.rglob("*") if p.is_file() and p.suffix.lower() in allowed])
+        files.sort(key=lambda x: str(x))
+        selected = files[: max(1, int(max_files))]
+
+        stats = {
+            "root": str(scan_roots[0]),
+            "roots": [str(x) for x in scan_roots],
+            "candidate_files": len(files),
+            "selected_files": len(selected),
+            "changed_files": 0,
+            "skipped_unchanged_files": 0,
+            "indexed_symbols": 0,
+            "indexed_references": 0,
+            "failed_files": 0,
+            "errors": [],
+        }
+
+        for file_path in selected:
+            language = self._detect_language(file_path)
+            if not language:
+                continue
+            try:
+                text = file_path.read_text(encoding="utf-8", errors="ignore")
+                rel_path = self._to_rel_path(file_path=file_path, workspace_root=workspace_root, backend_root=backend_root)
+                file_hash = hashlib.sha1(text.encode("utf-8")).hexdigest()
+                mtime = datetime.fromtimestamp(file_path.stat().st_mtime, tz=timezone.utc).replace(tzinfo=None)
+                if not force_reindex:
+                    latest_hash = await self._get_latest_file_hash(rel_path=rel_path)
+                    if latest_hash and latest_hash == file_hash:
+                        stats["skipped_unchanged_files"] += 1
+                        continue
+
+                symbols, refs = self._parse_file(
+                    file_path=rel_path,
+                    language=language,
+                    content=text,
+                )
+                s_count, r_count = await self._replace_file_index(
+                    file_path=rel_path,
+                    language=language,
+                    file_hash=file_hash,
+                    source_mtime=mtime,
+                    symbols=symbols,
+                    refs=refs,
+                )
+                stats["changed_files"] += 1
+                stats["indexed_symbols"] += s_count
+                stats["indexed_references"] += r_count
+            except Exception as exc:
+                stats["failed_files"] += 1
+                stats["errors"].append(f"{str(file_path)}: {str(exc)[:180]}")
+
+        await self.db.commit()
+        stats["errors"] = stats["errors"][:20]
+        stats["finished_at"] = datetime.now(timezone.utc).isoformat()
+        return stats
+
+    async def search_symbols(
+        self,
+        *,
+        query: str,
+        language: Optional[str] = None,
+        symbol_type: Optional[str] = None,
+        limit: int = 30,
+    ) -> dict[str, Any]:
+        text = str(query or "").strip()
+        if not text:
+            return {"count": 0, "items": []}
+
+        like = f"%{text}%"
+        stmt = select(SymbolDefinition).where(
+            and_(
+                SymbolDefinition.tenant_id == self.tenant_id,
+                SymbolDefinition.status == "active",
+                or_(
+                    SymbolDefinition.symbol_name.ilike(like),
+                    SymbolDefinition.qualified_name.ilike(like),
+                    SymbolDefinition.file_path.ilike(like),
+                ),
+            )
+        )
+        if language:
+            stmt = stmt.where(SymbolDefinition.language == str(language).strip().lower()[:20])
+        if symbol_type:
+            stmt = stmt.where(SymbolDefinition.symbol_type == str(symbol_type).strip().lower()[:32])
+        rows = (
+            await self.db.execute(
+                stmt.order_by(SymbolDefinition.updated_at.desc(), SymbolDefinition.id.desc()).limit(max(1, min(int(limit), 200)))
+            )
+        ).scalars().all()
+        return {
+            "count": len(rows),
+            "items": [self._serialize_symbol(row) for row in rows],
+        }
+
+    async def get_symbol_detail(self, *, symbol_id: int) -> dict[str, Any]:
+        symbol = (
+            await self.db.execute(
+                select(SymbolDefinition).where(
+                    and_(
+                        SymbolDefinition.tenant_id == self.tenant_id,
+                        SymbolDefinition.id == int(symbol_id),
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if symbol is None:
+            raise ValueError("symbol_not_found")
+
+        refs = (
+            await self.db.execute(
+                select(SymbolReference).where(
+                    and_(
+                        SymbolReference.tenant_id == self.tenant_id,
+                        SymbolReference.source_symbol_id == int(symbol.id),
+                    )
+                )
+                .order_by(SymbolReference.id.desc())
+                .limit(120)
+            )
+        ).scalars().all()
+
+        inbound_refs = (
+            await self.db.execute(
+                select(SymbolReference)
+                .where(
+                    and_(
+                        SymbolReference.tenant_id == self.tenant_id,
+                        or_(
+                            SymbolReference.target_symbol_name == str(symbol.symbol_name),
+                            SymbolReference.target_qualified_name == str(symbol.qualified_name or ""),
+                        ),
+                    )
+                )
+                .order_by(SymbolReference.id.desc())
+                .limit(120)
+            )
+        ).scalars().all()
+
+        return {
+            "symbol": self._serialize_symbol(symbol),
+            "outbound_refs": [self._serialize_ref(row) for row in refs],
+            "inbound_refs": [self._serialize_ref(row) for row in inbound_refs],
+        }
+
+    async def get_overview(
+        self,
+        *,
+        top_n_files: int = 10,
+        auto_sync_if_empty: bool = False,
+        base_dir: Optional[str] = None,
+    ) -> dict[str, Any]:
+        total = int(
+            (
+                await self.db.execute(
+                    select(func.count(SymbolDefinition.id)).where(SymbolDefinition.tenant_id == self.tenant_id)
+                )
+            ).scalar_one()
+            or 0
+        )
+        if total == 0 and auto_sync_if_empty:
+            await self.sync_codebase_symbols(base_dir=base_dir, max_files=300, force_reindex=False)
+            total = int(
+                (
+                    await self.db.execute(
+                        select(func.count(SymbolDefinition.id)).where(SymbolDefinition.tenant_id == self.tenant_id)
+                    )
+                ).scalar_one()
+                or 0
+            )
+
+        language_groups = (
+            await self.db.execute(
+                select(SymbolDefinition.language, func.count(SymbolDefinition.id))
+                .where(SymbolDefinition.tenant_id == self.tenant_id)
+                .group_by(SymbolDefinition.language)
+            )
+        ).all()
+        type_groups = (
+            await self.db.execute(
+                select(SymbolDefinition.symbol_type, func.count(SymbolDefinition.id))
+                .where(SymbolDefinition.tenant_id == self.tenant_id)
+                .group_by(SymbolDefinition.symbol_type)
+            )
+        ).all()
+        file_groups = (
+            await self.db.execute(
+                select(SymbolDefinition.file_path, func.count(SymbolDefinition.id))
+                .where(SymbolDefinition.tenant_id == self.tenant_id)
+                .group_by(SymbolDefinition.file_path)
+                .order_by(func.count(SymbolDefinition.id).desc())
+                .limit(max(1, min(int(top_n_files), 50)))
+            )
+        ).all()
+        ref_total = int(
+            (
+                await self.db.execute(
+                    select(func.count(SymbolReference.id)).where(SymbolReference.tenant_id == self.tenant_id)
+                )
+            ).scalar_one()
+            or 0
+        )
+        indexed_files_total = int(
+            (
+                await self.db.execute(
+                    select(func.count(func.distinct(SymbolDefinition.file_path))).where(
+                        SymbolDefinition.tenant_id == self.tenant_id
+                    )
+                )
+            ).scalar_one()
+            or 0
+        )
+        return {
+            "tenant_id": self.tenant_id,
+            "indexed_files": indexed_files_total,
+            "symbol_total": total,
+            "reference_total": ref_total,
+            "python_symbols": int(dict(language_groups).get("python", 0) or 0),
+            "typescript_symbols": int(dict(language_groups).get("typescript", 0) or 0),
+            "symbols_by_language": {str(k): int(v or 0) for k, v in language_groups},
+            "symbols_by_type": {str(k): int(v or 0) for k, v in type_groups},
+            "top_modules": [
+                {"module": str(path), "symbols": int(cnt or 0)}
+                for path, cnt in file_groups
+            ],
+            "snapshot_at": datetime.now(timezone.utc).isoformat(),
+        }
+
+    async def _get_latest_file_hash(self, *, rel_path: str) -> Optional[str]:
+        row = (
+            await self.db.execute(
+                select(SymbolDefinition.file_hash).where(
+                    and_(
+                        SymbolDefinition.tenant_id == self.tenant_id,
+                        SymbolDefinition.file_path == rel_path,
+                    )
+                )
+                .order_by(SymbolDefinition.updated_at.desc(), SymbolDefinition.id.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+        return str(row).strip() if row else None
+
+    async def _replace_file_index(
+        self,
+        *,
+        file_path: str,
+        language: str,
+        file_hash: str,
+        source_mtime: datetime,
+        symbols: list[ParsedSymbol],
+        refs: list[ParsedReference],
+    ) -> tuple[int, int]:
+        await self.db.execute(
+            delete(SymbolReference).where(
+                and_(
+                    SymbolReference.tenant_id == self.tenant_id,
+                    SymbolReference.file_path == file_path,
+                )
+            )
+        )
+        await self.db.execute(
+            delete(SymbolDefinition).where(
+                and_(
+                    SymbolDefinition.tenant_id == self.tenant_id,
+                    SymbolDefinition.file_path == file_path,
+                )
+            )
+        )
+
+        created_symbols: list[SymbolDefinition] = []
+        qn_to_id: dict[str, int] = {}
+        name_to_id: dict[str, int] = {}
+        for item in symbols:
+            row = SymbolDefinition(
+                tenant_id=self.tenant_id,
+                file_path=file_path,
+                language=language,
+                symbol_type=item.symbol_type[:32],
+                symbol_name=item.symbol_name[:255],
+                qualified_name=item.qualified_name[:700] if item.qualified_name else None,
+                start_line=int(item.start_line) if item.start_line > 0 else None,
+                end_line=int(item.end_line) if item.end_line > 0 else None,
+                signature=item.signature[:1000] if item.signature else None,
+                docstring=item.docstring[:2000] if item.docstring else None,
+                file_hash=file_hash,
+                source_mtime=source_mtime,
+                status="active",
+                payload=item.payload,
+                last_indexed_at=datetime.now(timezone.utc).replace(tzinfo=None),
+            )
+            self.db.add(row)
+            created_symbols.append(row)
+        await self.db.flush()
+        for row in created_symbols:
+            if row.qualified_name:
+                qn_to_id[str(row.qualified_name)] = int(row.id)
+            name_to_id[str(row.symbol_name)] = int(row.id)
+
+        ref_count = 0
+        # DB 层有 (tenant_id,file_path,relation_type,target_symbol_name,start_line,source_symbol_id) 唯一约束，
+        # 这里先去重，避免同一行重复调用（例如链式调用）触发批量写入冲突。
+        ref_seen_keys: set[tuple[str, str, Optional[int], Optional[int]]] = set()
+        for ref in refs:
+            source_id = None
+            source_qn = str(ref.source_qualified_name or "").strip()
+            if source_qn:
+                source_id = qn_to_id.get(source_qn)
+                if source_id is None:
+                    source_tail = source_qn.split(".")[-1]
+                    source_id = name_to_id.get(source_tail)
+            relation_type = ref.relation_type[:32]
+            target_symbol_name = ref.target_symbol_name[:255]
+            start_line = int(ref.start_line) if ref.start_line > 0 else None
+            dedupe_key = (
+                relation_type,
+                target_symbol_name,
+                start_line,
+                int(source_id) if source_id is not None else None,
+            )
+            if dedupe_key in ref_seen_keys:
+                continue
+            ref_seen_keys.add(dedupe_key)
+            row = SymbolReference(
+                tenant_id=self.tenant_id,
+                source_symbol_id=int(source_id) if source_id is not None else None,
+                file_path=file_path,
+                relation_type=relation_type,
+                target_symbol_name=target_symbol_name,
+                target_qualified_name=ref.target_qualified_name[:700] if ref.target_qualified_name else None,
+                start_line=start_line,
+                end_line=int(ref.end_line) if ref.end_line > 0 else None,
+                payload=ref.payload,
+            )
+            self.db.add(row)
+            ref_count += 1
+        await self.db.flush()
+        return (len(created_symbols), ref_count)
+
+    def _parse_file(
+        self,
+        *,
+        file_path: str,
+        language: str,
+        content: str,
+    ) -> tuple[list[ParsedSymbol], list[ParsedReference]]:
+        if language == "python":
+            return self._parse_python_file(file_path=file_path, content=content)
+        return self._parse_typescript_file(file_path=file_path, content=content)
+
+    def _parse_python_file(self, *, file_path: str, content: str) -> tuple[list[ParsedSymbol], list[ParsedReference]]:
+        text = str(content or "")
+        if not text.strip():
+            return ([], [])
+        try:
+            tree = ast.parse(text)
+        except Exception:
+            return ([], [])
+        symbols: list[ParsedSymbol] = []
+        refs: list[ParsedReference] = []
+
+        class Visitor(ast.NodeVisitor):
+            def __init__(self):
+                self.class_stack: list[str] = []
+                self.func_stack: list[str] = []
+
+            def _current_source(self) -> str:
+                if self.func_stack:
+                    return self.func_stack[-1]
+                if self.class_stack:
+                    return self.class_stack[-1]
+                return ""
+
+            def _qualify(self, name: str) -> str:
+                chain = [*self.class_stack]
+                if self.func_stack:
+                    chain.extend(self.func_stack[-1].split(".")[len(self.class_stack):])
+                chain.append(name)
+                uniq: list[str] = []
+                for item in chain:
+                    if item and (not uniq or uniq[-1] != item):
+                        uniq.append(item)
+                return ".".join(uniq) if uniq else name
+
+            def visit_ClassDef(self, node: ast.ClassDef) -> Any:
+                qualified = ".".join([*self.class_stack, node.name]) if self.class_stack else node.name
+                symbols.append(
+                    ParsedSymbol(
+                        symbol_type="class",
+                        symbol_name=node.name,
+                        qualified_name=qualified,
+                        start_line=int(getattr(node, "lineno", 0) or 0),
+                        end_line=int(getattr(node, "end_lineno", getattr(node, "lineno", 0)) or 0),
+                        signature=f"class {node.name}",
+                        docstring=str(ast.get_docstring(node) or ""),
+                        payload={"file_path": file_path},
+                    )
+                )
+                self.class_stack.append(node.name)
+                self.generic_visit(node)
+                self.class_stack.pop()
+
+            def visit_FunctionDef(self, node: ast.FunctionDef) -> Any:
+                self._visit_function_node(node=node, is_async=False)
+
+            def visit_AsyncFunctionDef(self, node: ast.AsyncFunctionDef) -> Any:
+                self._visit_function_node(node=node, is_async=True)
+
+            def _visit_function_node(self, *, node: Any, is_async: bool) -> None:
+                if self.class_stack:
+                    qualified = ".".join([*self.class_stack, node.name])
+                    symbol_type = "method"
+                else:
+                    qualified = node.name
+                    symbol_type = "function"
+                arg_names = []
+                for arg in list(node.args.posonlyargs) + list(node.args.args):
+                    arg_names.append(str(arg.arg))
+                if node.args.vararg is not None:
+                    arg_names.append(f"*{node.args.vararg.arg}")
+                for arg in node.args.kwonlyargs:
+                    arg_names.append(str(arg.arg))
+                if node.args.kwarg is not None:
+                    arg_names.append(f"**{node.args.kwarg.arg}")
+                prefix = "async def" if is_async else "def"
+                signature = f"{prefix} {node.name}({', '.join(arg_names)})"
+                symbols.append(
+                    ParsedSymbol(
+                        symbol_type=symbol_type,
+                        symbol_name=node.name,
+                        qualified_name=qualified,
+                        start_line=int(getattr(node, "lineno", 0) or 0),
+                        end_line=int(getattr(node, "end_lineno", getattr(node, "lineno", 0)) or 0),
+                        signature=signature[:1000],
+                        docstring=str(ast.get_docstring(node) or ""),
+                        payload={"file_path": file_path},
+                    )
+                )
+                self.func_stack.append(qualified)
+                self.generic_visit(node)
+                self.func_stack.pop()
+
+            def visit_Import(self, node: ast.Import) -> Any:
+                source = self._current_source()
+                for alias in node.names:
+                    target = str(alias.name or "").strip()
+                    if not target:
+                        continue
+                    refs.append(
+                        ParsedReference(
+                            relation_type="imports",
+                            target_symbol_name=target.split(".")[-1][:255],
+                            target_qualified_name=target[:700],
+                            source_qualified_name=source[:700],
+                            start_line=int(getattr(node, "lineno", 0) or 0),
+                            end_line=int(getattr(node, "end_lineno", getattr(node, "lineno", 0)) or 0),
+                            payload={"alias": str(alias.asname or "")[:255] if alias.asname else None},
+                        )
+                    )
+                self.generic_visit(node)
+
+            def visit_ImportFrom(self, node: ast.ImportFrom) -> Any:
+                source = self._current_source()
+                module = str(node.module or "").strip()
+                for alias in node.names:
+                    name = str(alias.name or "").strip()
+                    if not name:
+                        continue
+                    target_qn = f"{module}.{name}" if module else name
+                    refs.append(
+                        ParsedReference(
+                            relation_type="imports",
+                            target_symbol_name=name[:255],
+                            target_qualified_name=target_qn[:700],
+                            source_qualified_name=source[:700],
+                            start_line=int(getattr(node, "lineno", 0) or 0),
+                            end_line=int(getattr(node, "end_lineno", getattr(node, "lineno", 0)) or 0),
+                            payload={"module": module[:255] if module else None},
+                        )
+                    )
+                self.generic_visit(node)
+
+            def visit_Call(self, node: ast.Call) -> Any:
+                source = self._current_source()
+                target = ""
+                if isinstance(node.func, ast.Name):
+                    target = str(node.func.id or "").strip()
+                elif isinstance(node.func, ast.Attribute):
+                    target = str(node.func.attr or "").strip()
+                if target:
+                    refs.append(
+                        ParsedReference(
+                            relation_type="calls",
+                            target_symbol_name=target[:255],
+                            target_qualified_name=target[:700],
+                            source_qualified_name=source[:700],
+                            start_line=int(getattr(node, "lineno", 0) or 0),
+                            end_line=int(getattr(node, "end_lineno", getattr(node, "lineno", 0)) or 0),
+                            payload={},
+                        )
+                    )
+                self.generic_visit(node)
+
+        Visitor().visit(tree)
+        return (symbols[:5000], refs[:12000])
+
+    def _parse_typescript_file(self, *, file_path: str, content: str) -> tuple[list[ParsedSymbol], list[ParsedReference]]:
+        text = str(content or "")
+        if not text.strip():
+            return ([], [])
+        symbols: list[ParsedSymbol] = []
+        refs: list[ParsedReference] = []
+        lines = text.splitlines()
+        line_symbol_map: dict[int, str] = {}
+
+        for idx, line in enumerate(lines, start=1):
+            match = re.match(
+                r"^\s*(?:export\s+)?(?:default\s+)?(?:async\s+)?(function|class|interface|type|const)\s+([A-Za-z_][A-Za-z0-9_]*)",
+                line,
+            )
+            if not match:
+                continue
+            symbol_type = str(match.group(1) or "").strip().lower()
+            name = str(match.group(2) or "").strip()
+            if not name:
+                continue
+            signature = line.strip()
+            symbols.append(
+                ParsedSymbol(
+                    symbol_type=symbol_type,
+                    symbol_name=name,
+                    qualified_name=name,
+                    start_line=idx,
+                    end_line=idx,
+                    signature=signature[:1000],
+                    docstring="",
+                    payload={"file_path": file_path},
+                )
+            )
+            line_symbol_map[idx] = name
+
+        if "frontend/src/router/" in str(file_path).replace("\\", "/"):
+            for idx, line in enumerate(lines, start=1):
+                m = re.search(
+                    r'path:\s*"([^"]+)"\s*,\s*element:\s*<([A-Za-z_][A-Za-z0-9_]*)',
+                    line,
+                )
+                if not m:
+                    continue
+                route_path = str(m.group(1) or "").strip()
+                component = str(m.group(2) or "").strip()
+                if not route_path:
+                    continue
+                clean_route = route_path if route_path.startswith("/") else f"/{route_path}"
+                symbol_name = f"route:{clean_route}"
+                symbols.append(
+                    ParsedSymbol(
+                        symbol_type="route",
+                        symbol_name=symbol_name[:255],
+                        qualified_name=symbol_name[:700],
+                        start_line=idx,
+                        end_line=idx,
+                        signature=line.strip()[:1000],
+                        docstring=f"component:{component}",
+                        payload={"file_path": file_path, "route_path": clean_route, "component": component},
+                    )
+                )
+
+        if str(file_path).replace("\\", "/").endswith("components/layout/Sidebar.tsx"):
+            for idx, line in enumerate(lines, start=1):
+                m = re.search(r'to:\s*"([^"]+)"\s*,\s*label:\s*"([^"]+)"', line)
+                if not m:
+                    continue
+                route_path = str(m.group(1) or "").strip()
+                label = str(m.group(2) or "").strip()
+                if not route_path or not label:
+                    continue
+                clean_route = route_path if route_path.startswith("/") else f"/{route_path}"
+                symbol_name = f"feature:{label}"
+                symbols.append(
+                    ParsedSymbol(
+                        symbol_type="feature",
+                        symbol_name=symbol_name[:255],
+                        qualified_name=f"{label}@{clean_route}"[:700],
+                        start_line=idx,
+                        end_line=idx,
+                        signature=line.strip()[:1000],
+                        docstring=f"menu:{clean_route}",
+                        payload={"file_path": file_path, "feature_label": label, "route_path": clean_route},
+                    )
+                )
+
+        imports = TS_IMPORT_RE.finditer(text)
+        for match in imports:
+            module = str(match.group(1) or "").strip()
+            if not module:
+                continue
+            start = text[: match.start()].count("\n") + 1
+            refs.append(
+                ParsedReference(
+                    relation_type="imports",
+                    target_symbol_name=module.split("/")[-1][:255],
+                    target_qualified_name=module[:700],
+                    source_qualified_name="",
+                    start_line=start,
+                    end_line=start,
+                    payload={},
+                )
+            )
+
+        for idx, line in enumerate(lines, start=1):
+            source_symbol = ""
+            candidate_lines = [x for x in line_symbol_map.keys() if x <= idx]
+            if candidate_lines:
+                source_symbol = line_symbol_map[max(candidate_lines)]
+            for call in TS_CALL_RE.finditer(line):
+                name = str(call.group(1) or "").strip()
+                if not name or name.lower() in TS_KEYWORDS:
+                    continue
+                refs.append(
+                    ParsedReference(
+                        relation_type="calls",
+                        target_symbol_name=name[:255],
+                        target_qualified_name=name[:700],
+                        source_qualified_name=source_symbol[:700],
+                        start_line=idx,
+                        end_line=idx,
+                        payload={},
+                    )
+                )
+        return (symbols[:5000], refs[:12000])
+
+    @staticmethod
+    def _detect_language(path: Path) -> str:
+        ext = path.suffix.lower()
+        if ext == ".py":
+            return "python"
+        if ext in {".ts", ".tsx", ".js", ".jsx"}:
+            return "typescript"
+        return ""
+
+    @staticmethod
+    def _serialize_symbol(row: SymbolDefinition) -> dict[str, Any]:
+        return {
+            "id": int(row.id),
+            "file_path": str(row.file_path or ""),
+            "language": str(row.language or ""),
+            "symbol_type": str(row.symbol_type or ""),
+            "symbol_name": str(row.symbol_name or ""),
+            "qualified_name": str(row.qualified_name or ""),
+            "start_line": int(row.start_line) if row.start_line is not None else None,
+            "end_line": int(row.end_line) if row.end_line is not None else None,
+            "signature": str(row.signature or ""),
+            "docstring": str(row.docstring or ""),
+            "status": str(row.status or ""),
+            "updated_at": row.updated_at.isoformat() if row.updated_at else None,
+            "last_indexed_at": row.last_indexed_at.isoformat() if row.last_indexed_at else None,
+        }
+
+    @staticmethod
+    def _serialize_ref(row: SymbolReference) -> dict[str, Any]:
+        return {
+            "id": int(row.id),
+            "source_symbol_id": int(row.source_symbol_id) if row.source_symbol_id is not None else None,
+            "file_path": str(row.file_path or ""),
+            "relation_type": str(row.relation_type or ""),
+            "target_symbol_name": str(row.target_symbol_name or ""),
+            "target_qualified_name": str(row.target_qualified_name or ""),
+            "start_line": int(row.start_line) if row.start_line is not None else None,
+            "end_line": int(row.end_line) if row.end_line is not None else None,
+            "updated_at": row.updated_at.isoformat() if row.updated_at else None,
+        }
diff --git a/backend/app/services/task_persistence.py b/backend/app/services/task_persistence.py
new file mode 100644
index 00000000..a4e0ca3b
--- /dev/null
+++ b/backend/app/services/task_persistence.py
@@ -0,0 +1,324 @@
+"""共享任务持久化规范。"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import SkillTag, Task
+from ..utils.title import compact_title
+from . import workload_service
+from .operation_log_service import log_operation
+
+TASK_PRIORITY_MAP = {
+    "p0": "high",
+    "critical": "high",
+    "high": "high",
+    "p1": "medium",
+    "medium": "medium",
+    "normal": "medium",
+    "default": "medium",
+    "p2": "low",
+    "p3": "low",
+    "low": "low",
+}
+TASK_STATUS_MAP = {
+    "pending": "todo",
+    "todo": "todo",
+    "open": "todo",
+    "new": "todo",
+    "in_progress": "in_progress",
+    "doing": "in_progress",
+    "processing": "in_progress",
+    "review": "review",
+    "qa": "review",
+    "testing": "review",
+    "done": "done",
+    "completed": "done",
+    "blocked": "blocked",
+    "closed": "closed",
+}
+SUPPORTED_TASK_STATUSES = {"todo", "in_progress", "review", "done", "blocked", "closed"}
+
+
+@dataclass(slots=True)
+class SkillTagResolution:
+    input_names: list[str]
+    matched_ids: list[int]
+    matched_items: list[dict[str, Any]]
+    unmatched_names: list[str]
+
+
+@dataclass(slots=True)
+class TaskCreateDraft:
+    title: str
+    description: Optional[str] = None
+    assignee_id: Optional[int] = None
+    required_skill_tag_ids: list[int] = field(default_factory=list)
+    priority: str = "medium"
+    status: str = "todo"
+    estimated_hours: Optional[float] = None
+    due_date: Optional[datetime] = None
+    actual_hours: Optional[float] = None
+    quality_score: Optional[float] = None
+    ai_recommended_assignees: Optional[dict[str, Any]] = None
+
+
+def normalize_task_priority(value: Any, *, default: str = "medium") -> str:
+    key = str(value or default).strip().lower()
+    return TASK_PRIORITY_MAP.get(key, TASK_PRIORITY_MAP.get(default, "medium"))
+
+
+def normalize_task_status(value: Any, *, default: str = "todo") -> str:
+    key = str(value or default).strip().lower()
+    normalized = TASK_STATUS_MAP.get(key, TASK_STATUS_MAP.get(default, "todo"))
+    if normalized not in SUPPORTED_TASK_STATUSES:
+        return default
+    return normalized
+
+
+def normalize_estimated_hours(value: Any) -> Optional[float]:
+    if value is None or value == "":
+        return None
+    try:
+        hours = round(float(value), 2)
+    except (TypeError, ValueError) as exc:
+        raise ValueError("invalid estimated_hours") from exc
+    if hours < 0 or hours > 999.99:
+        raise ValueError("invalid estimated_hours")
+    return hours
+
+
+def normalize_due_date(value: Any) -> Optional[datetime]:
+    if value is None or value == "":
+        return None
+    if isinstance(value, datetime):
+        if value.tzinfo is not None:
+            return value.astimezone(timezone.utc).replace(tzinfo=None)
+        return value
+    text = str(value).strip()
+    if not text:
+        return None
+    try:
+        return datetime.fromisoformat(text.replace("Z", "+00:00")).replace(tzinfo=None)
+    except ValueError as exc:
+        raise ValueError("invalid due_date") from exc
+
+
+def normalize_int_list(values: Any) -> list[int]:
+    if values is None:
+        return []
+    if not isinstance(values, (list, tuple, set)):
+        values = [values]
+    normalized: list[int] = []
+    seen: set[int] = set()
+    for raw in values:
+        if isinstance(raw, bool):
+            continue
+        try:
+            value = int(str(raw).strip())
+        except (TypeError, ValueError):
+            continue
+        if value <= 0 or value in seen:
+            continue
+        seen.add(value)
+        normalized.append(value)
+    return normalized
+
+
+def normalize_skill_names(values: Any) -> list[str]:
+    if values is None:
+        return []
+    if not isinstance(values, (list, tuple, set)):
+        values = [values]
+    normalized: list[str] = []
+    seen: set[str] = set()
+    for raw in values:
+        text = str(raw or "").strip()
+        if not text:
+            continue
+        key = text.lower()
+        if key in seen:
+            continue
+        seen.add(key)
+        normalized.append(text[:120])
+    return normalized
+
+
+def build_task_create_draft(
+    *,
+    title: str,
+    description: Optional[str] = None,
+    assignee_id: Optional[int] = None,
+    required_skill_tag_ids: Any = None,
+    priority: Any = "medium",
+    status: Any = "todo",
+    estimated_hours: Any = None,
+    due_date: Any = None,
+    actual_hours: Any = None,
+    quality_score: Any = None,
+    ai_recommended_assignees: Optional[dict[str, Any]] = None,
+) -> TaskCreateDraft:
+    normalized_title = compact_title(str(title or "").strip(), max_len=32)
+    if not normalized_title:
+        raise ValueError("task title is empty")
+    return TaskCreateDraft(
+        title=normalized_title,
+        description=str(description or "").strip() or None,
+        assignee_id=int(assignee_id) if assignee_id is not None else None,
+        required_skill_tag_ids=normalize_int_list(required_skill_tag_ids),
+        priority=normalize_task_priority(priority),
+        status=normalize_task_status(status),
+        estimated_hours=normalize_estimated_hours(estimated_hours),
+        due_date=normalize_due_date(due_date),
+        actual_hours=normalize_estimated_hours(actual_hours),
+        quality_score=None if quality_score in {None, ""} else float(quality_score),
+        ai_recommended_assignees=ai_recommended_assignees,
+    )
+
+
+async def resolve_skill_tag_ids_by_names(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    skill_names: Any,
+) -> SkillTagResolution:
+    input_names = normalize_skill_names(skill_names)
+    if not input_names:
+        return SkillTagResolution(
+            input_names=[],
+            matched_ids=[],
+            matched_items=[],
+            unmatched_names=[],
+        )
+
+    rows = (
+        await db.execute(
+            select(SkillTag).where(SkillTag.tenant_id == int(tenant_id))
+        )
+    ).scalars().all()
+    if not rows:
+        return SkillTagResolution(
+            input_names=input_names,
+            matched_ids=[],
+            matched_items=[],
+            unmatched_names=input_names,
+        )
+
+    exact: dict[str, SkillTag] = {}
+    fuzzy: list[tuple[str, SkillTag]] = []
+    for row in rows:
+        name = str(row.name or "").strip().lower()
+        slug = str(row.slug or "").strip().lower()
+        if name:
+            exact[name] = row
+            fuzzy.append((name, row))
+        if slug:
+            exact[slug] = row
+            fuzzy.append((slug, row))
+
+    matched_ids: list[int] = []
+    matched_items: list[dict[str, Any]] = []
+    unmatched_names: list[str] = []
+    seen_ids: set[int] = set()
+    for name in input_names:
+        key = name.lower()
+        matched = exact.get(key)
+        if matched is None:
+            for candidate, row in fuzzy:
+                if key in candidate or candidate in key:
+                    matched = row
+                    break
+        if matched is None:
+            unmatched_names.append(name)
+            continue
+        skill_id = int(matched.id)
+        if skill_id not in seen_ids:
+            seen_ids.add(skill_id)
+            matched_ids.append(skill_id)
+        matched_items.append(
+            {
+                "input_name": name,
+                "skill_tag_id": skill_id,
+                "skill_tag_name": str(matched.name or "").strip(),
+                "skill_tag_slug": str(matched.slug or "").strip(),
+            }
+        )
+
+    return SkillTagResolution(
+        input_names=input_names,
+        matched_ids=matched_ids,
+        matched_items=matched_items,
+        unmatched_names=unmatched_names,
+    )
+
+
+async def persist_task(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    truth_node_id: int,
+    project_id: int,
+    draft: TaskCreateDraft,
+    actor_user_id: Optional[int] = None,
+    action: str = "task.create",
+    remark: Optional[str] = None,
+    commit: bool = False,
+    refresh_workload: bool = False,
+) -> Task:
+    task = Task(
+        tenant_id=int(tenant_id),
+        truth_node_id=int(truth_node_id),
+        project_id=int(project_id),
+        assignee_id=draft.assignee_id,
+        title=draft.title,
+        description=draft.description,
+        required_skill_tag_ids=draft.required_skill_tag_ids,
+        status=draft.status,
+        priority=draft.priority,
+        estimated_hours=draft.estimated_hours,
+        actual_hours=draft.actual_hours,
+        quality_score=draft.quality_score,
+        due_date=draft.due_date,
+        ai_recommended_assignees=draft.ai_recommended_assignees,
+    )
+    db.add(task)
+    await db.flush()
+
+    if actor_user_id is not None:
+        await log_operation(
+            db,
+            tenant_id=int(tenant_id),
+            user_id=int(actor_user_id),
+            project_id=int(project_id),
+            entity_type="task",
+            entity_id=str(task.id),
+            action=action,
+            before_data=None,
+            after_data={
+                "title": task.title,
+                "truth_node_id": int(truth_node_id),
+                "priority": task.priority,
+                "status": task.status,
+                "assignee_id": int(task.assignee_id) if task.assignee_id is not None else None,
+                "required_skill_tag_ids": list(task.required_skill_tag_ids or []),
+                "estimated_hours": float(task.estimated_hours) if task.estimated_hours is not None else None,
+                "due_date": task.due_date.isoformat() if task.due_date else None,
+            },
+            remark=remark,
+        )
+
+    if refresh_workload and task.assignee_id is not None:
+        await workload_service.refresh_user_assigned_hours_from_tasks(
+            db,
+            tenant_id=int(tenant_id),
+            user_id=int(task.assignee_id),
+        )
+    elif commit:
+        await db.commit()
+
+    return task
diff --git a/backend/app/services/wecom_integration.py b/backend/app/services/wecom_integration.py
new file mode 100644
index 00000000..510e0dc6
--- /dev/null
+++ b/backend/app/services/wecom_integration.py
@@ -0,0 +1,293 @@
+"""
+WeChat Work (企业微信) Integration Service
+
+功能：
+1. 接收企业微信群聊消息事件
+2. AI 判断消息是否为有价值的信号
+3. 自动创建信号并关联到项目
+4. 推送通知给相关角色
+"""
+import asyncio
+import json
+import hashlib
+import hmac
+import base64
+import struct
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional
+import httpx
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from sqlalchemy import text
+
+from ..models import (
+    Signal, Project, Integration
+)
+from .signal_capture import ProactiveSignalCaptureService
+
+
+class WeChatWorkIntegrationService:
+    """企业微信集成服务"""
+
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+        # 从数据库获取企业微信配置
+        self.corp_id = None
+        self.agent_id = None
+        self.app_secret = None
+        self.token = None
+        self.encoding_aes_key = None
+        self.tenant_id = 1
+
+    async def load_config(self, tenant_id: int = 1):
+        """从数据库加载企业微信配置"""
+        self.tenant_id = tenant_id
+
+        result = await self.db.execute(
+            select(Integration).where(
+                Integration.tenant_id == tenant_id,
+                Integration.type == 'wecom',
+                Integration.is_active.is_(True),
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration:
+            if isinstance(integration.config, str):
+                config = json.loads(integration.config) if integration.config else {}
+            else:
+                config = integration.config or {}
+            self.corp_id = config.get('corp_id')
+            self.agent_id = config.get('agent_id')
+            self.app_secret = config.get('app_secret')
+            self.token = config.get('token')
+            self.encoding_aes_key = config.get('encoding_aes_key')
+        else:
+            self.corp_id = None
+            self.agent_id = None
+            self.app_secret = None
+            self.token = None
+            self.encoding_aes_key = None
+
+    def is_configured(self) -> bool:
+        """检查是否已配置"""
+        return all([
+            self.corp_id,
+            self.agent_id,
+            self.app_secret,
+            self.token,
+            self.encoding_aes_key
+        ])
+
+    async def save_config(
+        self,
+        corp_id: str,
+        agent_id: str,
+        app_secret: str,
+        token: str,
+        encoding_aes_key: str,
+        tenant_id: int = 1
+    ):
+        """保存企业微信配置到数据库"""
+        config = {
+            'corp_id': corp_id,
+            'agent_id': agent_id,
+            'app_secret': app_secret,
+            'token': token,
+            'encoding_aes_key': encoding_aes_key
+        }
+
+        # 查找或创建集成记录
+        result = await self.db.execute(
+            select(Integration).where(
+                Integration.tenant_id == tenant_id,
+                Integration.type == 'wecom'
+            )
+        )
+        integration = result.scalar_one_or_none()
+
+        if integration:
+            integration.config = config
+            integration.is_active = True
+        else:
+            integration = Integration(
+                type='wecom',
+                config=config,
+                is_active=True,
+                tenant_id=tenant_id
+            )
+            self.db.add(integration)
+
+        await self.db.commit()
+
+        # 更新内存配置
+        self.corp_id = corp_id
+        self.agent_id = agent_id
+        self.app_secret = app_secret
+        self.token = token
+        self.encoding_aes_key = encoding_aes_key
+        self.tenant_id = tenant_id
+
+    def verify_url(self, msg_signature: str, timestamp: str, nonce: str, echostr: str) -> str:
+        """
+        验证企业微信事件订阅 URL
+
+        企业微信会发送一个 GET 请求来验证 URL 有效性
+        """
+        if not self.is_configured():
+            raise ValueError("企业微信未配置")
+
+        # 计算签名
+        sign_str = f"{self.token}{timestamp}{nonce}"
+        signature = hashlib.sha256(sign_str.encode()).hexdigest()
+
+        if signature != msg_signature:
+            raise ValueError("签名验证失败")
+
+        # 解密 echostr
+        decrypted = self._decrypt(echostr)
+        return decrypted
+
+    def verify_callback(self, msg_signature: str, timestamp: str, nonce: str, data: str) -> bool:
+        """
+        验证企业微信回调签名
+        """
+        if not self.token:
+            return False
+
+        # 计算签名
+        sign_str = f"{self.token}{timestamp}{nonce}{data}"
+        computed_signature = hashlib.sha256(sign_str.encode()).hexdigest()
+
+        return computed_signature == msg_signature
+
+    async def handle_message_event(
+        self,
+        event_data: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """
+        处理企业微信消息事件
+
+        支持的事件类型：
+        - add_half_user: 添加企业成员
+        - update_user: 更新成员信息
+        - send_msg_success: 消息发送成功
+        """
+        try:
+            event_type = event_data.get('Event', '')
+
+            # 企业微信群聊消息通过其他方式处理，这里主要是事件处理
+            return {
+                'code': 0,
+                'msg': f'Event {event_type} received'
+            }
+
+        except Exception as e:
+            print(f"[WeCom] Handle event error: {e}")
+            return {
+                'code': 1,
+                'msg': str(e)
+            }
+
+    def _decrypt(self, encoded_text: str) -> str:
+        """
+        解密企业微信加密消息
+
+        AES 加密的消息格式: base64编码的16字节随机字符串 + 4字节消息长度 + 消息内容
+        """
+        try:
+            # Base64 解码
+            cipher_text = base64.b64decode(encoded_text)
+
+            # AES 密钥处理
+            aes_key = base64.b64decode(self.encoding_aes_key + "=")
+
+            # 解密
+            from Crypto.Cipher import AES
+            cipher = AES.new(aes_key, AES.MODE_CBC, aes_key[:16])
+
+            decrypted = cipher.decrypt(cipher_text)
+
+            # 去除补位字符
+            pad = decrypted[-1]
+            if isinstance(pad, str):
+                pad = ord(pad)
+            decrypted = decrypted[:-pad]
+
+            # 提取消息内容（前16字节是随机字符串，接着4字节是消息长度）
+            content = decrypted[16:]
+            msg_len = struct.unpack(">I", content[:4])[0]
+            message = content[4:4 + msg_len].decode('utf-8')
+
+            return message
+
+        except Exception as e:
+            print(f"[WeCom] Decrypt error: {e}")
+            raise ValueError("消息解密失败")
+
+    async def send_message_to_group(
+        self,
+        chat_id: str,
+        message: str
+    ) -> Dict[str, Any]:
+        """
+        发送消息到企业微信群聊
+
+        用于推送通知、简报等
+        """
+        if not self.is_configured():
+            raise ValueError("企业微信未配置")
+
+        # 获取 access_token
+        access_token = await self._get_access_token()
+
+        # 发送消息 API
+        url = "https://qyapi.weixin.qq.com/cgi-bin/message/send"
+        params = {"access_token": access_token}
+
+        message_data = {
+            "touser": "",
+            "msgtype": "text",
+            "agentid": int(self.agent_id),
+            "text": {
+                "content": message
+            },
+            "safe": 0
+        }
+
+        # 如果是群聊，使用 chatid
+        if chat_id:
+            message_data["chatid"] = chat_id
+            message_data["msgtype"] = "text"
+
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False) as client:
+            resp = await client.post(url, params=params, json=message_data)
+            resp.raise_for_status()
+            return resp.json()
+
+    async def _get_access_token(self) -> str:
+        """获取企业微信 access_token"""
+        if not self.corp_id or not self.app_secret:
+            raise ValueError("企业微信未配置")
+
+        url = "https://qyapi.weixin.qq.com/cgi-bin/gettoken"
+        params = {
+            "corpid": self.corp_id,
+            "corpsecret": self.app_secret
+        }
+
+        async with httpx.AsyncClient(timeout=20.0, trust_env=False) as client:
+            resp = await client.get(url, params=params)
+            resp.raise_for_status()
+            result = resp.json()
+
+            if result.get('errcode') == 0:
+                return result.get('access_token')
+            raise Exception(f"获取 token 失败: {result}")
+
+
+# 全局服务实例
+def get_wecom_service(db: AsyncSession) -> WeChatWorkIntegrationService:
+    """获取企业微信服务实例"""
+    return WeChatWorkIntegrationService(db)
diff --git a/backend/app/services/workload_alert_service.py b/backend/app/services/workload_alert_service.py
new file mode 100644
index 00000000..d451af3e
--- /dev/null
+++ b/backend/app/services/workload_alert_service.py
@@ -0,0 +1,252 @@
+"""团队负载告警服务。"""
+
+from __future__ import annotations
+
+from datetime import datetime, timedelta
+from typing import Optional
+
+from sqlalchemy import and_, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import OperationLog, Task, User, UserProject, UserWorkload
+from .lark import lark_service
+
+DONE_STATUSES = {"done", "completed", "closed", "resolved", "archived"}
+
+
+async def _list_alert_logs(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    user_id: int,
+) -> list[OperationLog]:
+    rows = await db.execute(
+        select(OperationLog).where(
+            and_(
+                OperationLog.tenant_id == tenant_id,
+                OperationLog.entity_type == "workload_alert",
+                OperationLog.entity_id == str(user_id),
+            )
+        ).order_by(OperationLog.created_at.desc())
+    )
+    return list(rows.scalars().all())
+
+
+def _latest_by_action(logs: list[OperationLog], action: str) -> Optional[OperationLog]:
+    for row in logs:
+        if row.action == action:
+            return row
+    return None
+
+
+def _is_unresolved(logs: list[OperationLog]) -> bool:
+    latest_detected = _latest_by_action(logs, "overload_detected")
+    if not latest_detected:
+        return False
+    latest_resolved = _latest_by_action(logs, "overload_resolved")
+    if not latest_resolved:
+        return True
+    return bool(latest_detected.created_at and latest_resolved.created_at and latest_detected.created_at > latest_resolved.created_at)
+
+
+async def on_user_workload_changed(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    user_id: int,
+    workload_ratio: float,
+) -> dict:
+    """用户负载变化时触发告警/恢复。"""
+    logs = await _list_alert_logs(db, tenant_id=tenant_id, user_id=user_id)
+    unresolved = _is_unresolved(logs)
+
+    if workload_ratio > 100 and not unresolved:
+        text = (
+            f"成员 {user_id} 负载过载：{workload_ratio:.1f}% (>100%)，"
+            f"请尽快处理任务转派。"
+        )
+        notify_result = await lark_service.send_card_message(
+            chat_id=f"tenant-{tenant_id}-workload-alert",
+            content=text,
+            tasks=[],
+        )
+        db.add(
+            OperationLog(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                project_id=None,
+                entity_type="workload_alert",
+                entity_id=str(user_id),
+                action="overload_detected",
+                before_data={"unresolved": unresolved},
+                after_data={
+                    "workload_ratio": round(workload_ratio, 2),
+                    "notify_channel": "feishu",
+                    "notify_status": notify_result.get("status"),
+                },
+                remark=text,
+            )
+        )
+        return {"alerted": True, "resolved": False}
+
+    if workload_ratio <= 100 and unresolved:
+        db.add(
+            OperationLog(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                project_id=None,
+                entity_type="workload_alert",
+                entity_id=str(user_id),
+                action="overload_resolved",
+                before_data={"unresolved": True},
+                after_data={"workload_ratio": round(workload_ratio, 2)},
+                remark="workload_back_to_normal",
+            )
+        )
+        return {"alerted": False, "resolved": True}
+
+    return {"alerted": False, "resolved": False}
+
+
+async def _build_transfer_suggestions(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+    user_id: int,
+    max_items: int = 3,
+) -> list[dict]:
+    tasks_result = await db.execute(
+        select(Task).where(
+            and_(
+                Task.tenant_id == tenant_id,
+                Task.assignee_id == user_id,
+                func.lower(Task.status).notin_(DONE_STATUSES),
+            )
+        ).order_by(Task.updated_at.desc())
+    )
+    tasks = list(tasks_result.scalars().all())
+    if not tasks:
+        return []
+
+    suggestions: list[dict] = []
+    for task in tasks:
+        member_result = await db.execute(
+            select(UserProject.user_id).where(
+                and_(
+                    UserProject.tenant_id == tenant_id,
+                    UserProject.project_id == int(task.project_id),
+                )
+            )
+        )
+        member_ids = [int(v) for v in member_result.scalars().all() if int(v) != user_id]
+        if not member_ids:
+            continue
+
+        candidate_result = await db.execute(
+            select(UserWorkload.user_id, UserWorkload.workload_ratio)
+            .join(User, User.id == UserWorkload.user_id)
+            .where(
+                and_(
+                    UserWorkload.tenant_id == tenant_id,
+                    UserWorkload.user_id.in_(member_ids),
+                    User.tenant_id == tenant_id,
+                    User.status == "active",
+                    User.is_active.is_(True),
+                )
+            )
+            .order_by(UserWorkload.workload_ratio.asc())
+        )
+        candidates = candidate_result.all()
+        if not candidates:
+            continue
+
+        target_user_id, target_ratio = candidates[0]
+        suggestions.append(
+            {
+                "task_id": int(task.id),
+                "project_id": int(task.project_id),
+                "old_assignee_id": user_id,
+                "new_assignee_id": int(target_user_id),
+                "new_assignee_ratio": round(float(target_ratio or 0.0), 2),
+                "reason": "auto_transfer_after_24h_overload",
+            }
+        )
+        if len(suggestions) >= max_items:
+            break
+    return suggestions
+
+
+async def process_tenant_unresolved_overloads(
+    db: AsyncSession,
+    *,
+    tenant_id: int,
+) -> dict:
+    """扫描租户过载并为持续 24h 未恢复的成员生成转派建议。"""
+    rows = await db.execute(
+        select(UserWorkload.user_id, UserWorkload.workload_ratio).join(
+            User,
+            and_(
+                User.id == UserWorkload.user_id,
+                User.tenant_id == tenant_id,
+            ),
+        ).where(
+            and_(
+                UserWorkload.tenant_id == tenant_id,
+                User.status == "active",
+                User.is_active.is_(True),
+            )
+        )
+    )
+    user_rows = rows.all()
+
+    created = 0
+    now = datetime.utcnow()
+    for user_id_raw, ratio_raw in user_rows:
+        user_id = int(user_id_raw)
+        ratio = float(ratio_raw or 0.0)
+        await on_user_workload_changed(
+            db,
+            tenant_id=tenant_id,
+            user_id=user_id,
+            workload_ratio=ratio,
+        )
+        if ratio <= 100:
+            continue
+
+        logs = await _list_alert_logs(db, tenant_id=tenant_id, user_id=user_id)
+        latest_detected = _latest_by_action(logs, "overload_detected")
+        latest_resolved = _latest_by_action(logs, "overload_resolved")
+        latest_suggestion = _latest_by_action(logs, "transfer_suggestion")
+        if not latest_detected or not latest_detected.created_at:
+            continue
+        if latest_resolved and latest_resolved.created_at and latest_resolved.created_at > latest_detected.created_at:
+            continue
+        if now - latest_detected.created_at < timedelta(hours=24):
+            continue
+        if latest_suggestion and latest_suggestion.created_at and latest_suggestion.created_at > latest_detected.created_at:
+            continue
+
+        suggestions = await _build_transfer_suggestions(
+            db,
+            tenant_id=tenant_id,
+            user_id=user_id,
+        )
+        if not suggestions:
+            continue
+        db.add(
+            OperationLog(
+                tenant_id=tenant_id,
+                user_id=user_id,
+                project_id=None,
+                entity_type="workload_alert",
+                entity_id=str(user_id),
+                action="transfer_suggestion",
+                before_data={"workload_ratio": ratio},
+                after_data={"suggestions": suggestions},
+                remark="overload_unresolved_24h",
+            )
+        )
+        created += 1
+
+    await db.commit()
+    return {"suggestion_count": created, "user_count": len(user_rows)}
diff --git a/backend/app/services/workload_service.py b/backend/app/services/workload_service.py
new file mode 100644
index 00000000..6f91ae3e
--- /dev/null
+++ b/backend/app/services/workload_service.py
@@ -0,0 +1,739 @@
+"""PMFlow 团队负载服务。"""
+
+from __future__ import annotations
+
+import json
+from datetime import date, datetime, timedelta
+from typing import Optional
+
+from sqlalchemy import and_, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config.ai_prompts import PROMPT_WORKLOAD_OPTIMIZATION
+from ..models import AICallLog, Task, User, UserProject, UserSkillBinding, UserWorkload
+from .llm import llm_service
+
+LEVEL_HOURS_MULTIPLIER = {
+    "expert": 0.8,
+    "advanced": 0.9,
+    "intermediate": 1.0,
+    "beginner": 1.1,
+}
+
+
+def remaining_workdays_this_week(reference_date: Optional[date] = None) -> int:
+    """计算本周剩余工作日（含当天）。"""
+    today = reference_date or date.today()
+    weekday = today.weekday()  # Monday = 0
+    days_to_friday = max(0, 4 - weekday)
+    count = 0
+    for offset in range(days_to_friday + 1):
+        d = today + timedelta(days=offset)
+        if d.weekday() < 5:
+            count += 1
+    return max(count, 1)
+
+
+def calculate_workload_ratio(
+    assigned_hours: float,
+    max_weekly_hours: float,
+    reference_date: Optional[date] = None,
+) -> float:
+    """
+    负载率公式（百分比）：
+    ratio = assigned_hours / ((max_weekly_hours / 5) * remaining_workdays_this_week)
+    """
+    remaining_days = remaining_workdays_this_week(reference_date)
+    denominator = (max_weekly_hours / 5.0) * remaining_days
+    if denominator <= 0:
+        return 0.0
+    return round((assigned_hours / denominator) * 100.0, 2)
+
+
+async def ensure_user_workload(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+) -> UserWorkload:
+    result = await db.execute(
+        select(UserWorkload).where(
+            and_(
+                UserWorkload.tenant_id == tenant_id,
+                UserWorkload.user_id == user_id,
+            )
+        )
+    )
+    workload = result.scalar_one_or_none()
+    if workload:
+        return workload
+
+    workload = UserWorkload(
+        tenant_id=tenant_id,
+        user_id=user_id,
+        current_assigned_hours=0.0,
+        max_available_hours_per_week=40.0,
+        workload_ratio=0.0,
+    )
+    db.add(workload)
+    await db.flush()
+    return workload
+
+
+async def recalculate_user_workload_ratio(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    *,
+    commit: bool = True,
+) -> Optional[UserWorkload]:
+    """重算单用户负载率并同步 users.workload。"""
+    user_result = await db.execute(
+        select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+    )
+    user = user_result.scalar_one_or_none()
+    if not user:
+        return None
+
+    workload = await ensure_user_workload(db, tenant_id, user_id)
+    ratio = calculate_workload_ratio(
+        assigned_hours=float(workload.current_assigned_hours or 0),
+        max_weekly_hours=float(workload.max_available_hours_per_week or 40),
+    )
+    workload.workload_ratio = ratio
+    user.workload = int(max(0, min(200, ratio)))
+    user.updated_at = datetime.utcnow()
+    try:
+        from . import workload_alert_service
+
+        await workload_alert_service.on_user_workload_changed(
+            db,
+            tenant_id=tenant_id,
+            user_id=user_id,
+            workload_ratio=ratio,
+        )
+    except Exception:
+        pass
+
+    if commit:
+        await db.commit()
+        await db.refresh(workload)
+    return workload
+
+
+async def recalculate_all_users_workload_ratio(db: AsyncSession, tenant_id: int) -> int:
+    """批量重算租户内所有用户负载率。"""
+    result = await db.execute(select(User.id).where(User.tenant_id == tenant_id))
+    user_ids = [int(uid) for uid in result.scalars().all()]
+    for uid in user_ids:
+        await recalculate_user_workload_ratio(db, tenant_id, uid, commit=False)
+    await db.commit()
+    return len(user_ids)
+
+
+async def list_tenant_ids(db: AsyncSession) -> list[int]:
+    """获取当前系统存在用户数据的租户 ID 列表。"""
+    result = await db.execute(select(User.tenant_id).distinct())
+    return [int(tid) for tid in result.scalars().all() if tid is not None]
+
+
+async def get_team_workload_board(
+    db: AsyncSession,
+    tenant_id: int = 1,
+    project_id: Optional[int] = None,
+) -> list[dict]:
+    """团队负载看板数据。"""
+    query = (
+        select(UserWorkload, User)
+        .join(User, UserWorkload.user_id == User.id)
+        .where(
+            and_(
+                UserWorkload.tenant_id == tenant_id,
+                User.tenant_id == tenant_id,
+            )
+        )
+    )
+    if project_id is not None:
+        query = query.join(
+            UserProject,
+            and_(
+                UserProject.tenant_id == tenant_id,
+                UserProject.user_id == User.id,
+            ),
+        ).where(UserProject.project_id == project_id)
+
+    query = query.order_by(UserWorkload.workload_ratio.desc(), User.id.asc())
+    rows = (await db.execute(query)).all()
+
+    return [
+        {
+            "user_id": str(w.user_id),
+            "user_name": u.name,
+            "username": u.username,
+            "role": u.role,
+            "is_active": u.is_active,
+            "current_assigned_hours": float(w.current_assigned_hours or 0),
+            "max_weekly_hours": float(w.max_available_hours_per_week or 40),
+            "workload_ratio": float(w.workload_ratio or 0),
+            "remaining_hours": round(
+                max(
+                    0.0,
+                    float(w.max_available_hours_per_week or 40)
+                    - float(w.current_assigned_hours or 0),
+                ),
+                2,
+            ),
+            "updated_at": w.updated_at.isoformat() if w.updated_at else None,
+        }
+        for w, u in rows
+    ]
+
+
+async def get_user_personal_workload(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+) -> Optional[dict]:
+    """个人负载详情。"""
+    row = (
+        await db.execute(
+            select(UserWorkload, User)
+            .join(User, UserWorkload.user_id == User.id)
+            .where(
+                and_(
+                    UserWorkload.tenant_id == tenant_id,
+                    UserWorkload.user_id == user_id,
+                    User.tenant_id == tenant_id,
+                )
+            )
+        )
+    ).first()
+    if row:
+        w, u = row
+        return {
+            "user_id": str(u.id),
+            "user_name": u.name,
+            "username": u.username,
+            "is_active": u.is_active,
+            "current_assigned_hours": float(w.current_assigned_hours or 0),
+            "max_weekly_hours": float(w.max_available_hours_per_week or 40),
+            "workload_ratio": float(w.workload_ratio or 0),
+            "remaining_hours": round(
+                max(
+                    0.0,
+                    float(w.max_available_hours_per_week or 40)
+                    - float(w.current_assigned_hours or 0),
+                ),
+                2,
+            ),
+            "updated_at": w.updated_at.isoformat() if w.updated_at else None,
+        }
+
+    user = (
+        await db.execute(
+            select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+        )
+    ).scalar_one_or_none()
+    if not user:
+        return None
+
+    workload = await ensure_user_workload(db, tenant_id, user_id)
+    await db.commit()
+    return {
+        "user_id": str(user.id),
+        "user_name": user.name,
+        "username": user.username,
+        "is_active": user.is_active,
+        "current_assigned_hours": float(workload.current_assigned_hours or 0),
+        "max_weekly_hours": float(workload.max_available_hours_per_week or 40),
+        "workload_ratio": float(workload.workload_ratio or 0),
+        "remaining_hours": round(
+            max(
+                0.0,
+                float(workload.max_available_hours_per_week or 40)
+                - float(workload.current_assigned_hours or 0),
+            ),
+            2,
+        ),
+        "updated_at": workload.updated_at.isoformat() if workload.updated_at else None,
+    }
+
+
+async def update_user_max_weekly_hours(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    max_weekly_hours: float,
+) -> Optional[dict]:
+    """更新每周可用工时并重算负载率。"""
+    user = (
+        await db.execute(
+            select(User).where(and_(User.tenant_id == tenant_id, User.id == user_id))
+        )
+    ).scalar_one_or_none()
+    if not user:
+        return None
+
+    workload = await ensure_user_workload(db, tenant_id, user_id)
+    workload.max_available_hours_per_week = max(1.0, float(max_weekly_hours))
+    await recalculate_user_workload_ratio(db, tenant_id, user_id, commit=False)
+    await db.commit()
+    await db.refresh(workload)
+
+    return {
+        "user_id": str(user.id),
+        "max_weekly_hours": float(workload.max_available_hours_per_week),
+        "workload_ratio": float(workload.workload_ratio or 0),
+    }
+
+
+async def list_user_assigned_tasks(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    *,
+    status: Optional[str] = None,
+    priority: Optional[str] = None,
+) -> list[dict]:
+    """查询用户任务列表。"""
+    query = select(Task).where(
+        and_(
+            Task.tenant_id == tenant_id,
+            Task.assignee_id == user_id,
+        )
+    )
+    if status:
+        query = query.where(Task.status == status)
+    if priority:
+        query = query.where(Task.priority == priority)
+
+    rows = (await db.execute(query.order_by(Task.updated_at.desc()))).scalars().all()
+    return [
+        {
+            "id": str(t.id),
+            "title": t.title,
+            "project_id": str(t.project_id),
+            "priority": t.priority,
+            "status": t.status,
+            "estimated_hours": float(t.estimated_hours or 0),
+            "due_date": t.due_date.isoformat() if t.due_date else None,
+        }
+        for t in rows
+    ]
+
+
+async def get_user_workload_trend(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+    days: int = 7,
+) -> list[dict]:
+    """
+    用户负载趋势。
+
+    当前无历史快照表时，返回最近 N 天的平滑序列用于前端趋势图展示。
+    """
+    workload = (
+        await db.execute(
+            select(UserWorkload).where(
+                and_(
+                    UserWorkload.tenant_id == tenant_id,
+                    UserWorkload.user_id == user_id,
+                )
+            )
+        )
+    ).scalar_one_or_none()
+    ratio_now = float(workload.workload_ratio or 0) if workload else 0.0
+
+    days = max(1, min(days, 30))
+    start = date.today() - timedelta(days=days - 1)
+    series: list[dict] = []
+    for idx in range(days):
+        current_day = start + timedelta(days=idx)
+        # 缺少历史表时做线性平滑，保留可视化可用性。
+        ratio = round(max(0.0, ratio_now * (0.92 + 0.08 * (idx + 1) / days)), 2)
+        series.append(
+            {
+                "date": current_day.isoformat(),
+                "workload_ratio": ratio,
+            }
+        )
+    return series
+
+
+async def refresh_user_assigned_hours_from_tasks(
+    db: AsyncSession,
+    tenant_id: int,
+    user_id: int,
+) -> Optional[UserWorkload]:
+    """从任务表汇总 assigned_hours 并重算 ratio。"""
+    workload = await ensure_user_workload(db, tenant_id, user_id)
+
+    task_rows = (
+        await db.execute(
+            select(Task).where(
+                and_(
+                    Task.tenant_id == tenant_id,
+                    Task.assignee_id == user_id,
+                    func.lower(Task.status).notin_(["done", "completed", "closed"]),
+                )
+            )
+        )
+    ).scalars().all()
+
+    skill_rows = (
+        await db.execute(
+            select(
+                UserSkillBinding.skill_tag_id,
+                UserSkillBinding.level,
+                UserSkillBinding.proficiency_level,
+                UserSkillBinding.proficiency,
+            ).where(
+                and_(
+                    UserSkillBinding.tenant_id == tenant_id,
+                    UserSkillBinding.user_id == user_id,
+                )
+            )
+        )
+    ).all()
+
+    def normalize_level(
+        level: Optional[str],
+        proficiency_level: Optional[str],
+        proficiency: Optional[int],
+    ) -> str:
+        raw = (level or proficiency_level or "").strip().lower()
+        if raw in LEVEL_HOURS_MULTIPLIER:
+            return raw
+        score = int(proficiency or 0)
+        if score >= 85:
+            return "expert"
+        if score >= 70:
+            return "advanced"
+        if score >= 50:
+            return "intermediate"
+        return "beginner"
+
+    level_map: dict[int, str] = {
+        int(skill_tag_id): normalize_level(level, proficiency_level, proficiency)
+        for skill_tag_id, level, proficiency_level, proficiency in skill_rows
+    }
+
+    def effective_estimated_hours(task: Task) -> float:
+        base = float(task.estimated_hours or 0.0)
+        if base <= 0:
+            return 0.0
+        required_ids = _safe_int_list(task.required_skill_tag_ids)
+        if not required_ids:
+            return base
+        multipliers = [
+            LEVEL_HOURS_MULTIPLIER.get(level_map[sid], 1.0)
+            for sid in required_ids
+            if sid in level_map
+        ]
+        if not multipliers:
+            return base
+        factor = sum(multipliers) / len(multipliers)
+        return round(base * factor, 2)
+
+    workload.current_assigned_hours = round(
+        sum(effective_estimated_hours(task) for task in task_rows),
+        2,
+    )
+
+    await recalculate_user_workload_ratio(db, tenant_id, user_id, commit=False)
+    await db.commit()
+    await db.refresh(workload)
+    return workload
+
+
+def _skill_match_score(task_skill_ids: list[int], user_skill_ids: set[int]) -> float:
+    if not task_skill_ids:
+        return 0.0
+    matched = sum(1 for sid in task_skill_ids if sid in user_skill_ids)
+    return matched / len(task_skill_ids)
+
+
+def _safe_int_list(values: object) -> list[int]:
+    """兼容历史脏数据：数组中可能混入技能名称等非数字值。"""
+    source = values
+    if source is None:
+        return []
+    if isinstance(source, str):
+        text = source.strip()
+        if not text:
+            return []
+        try:
+            source = json.loads(text)
+        except (TypeError, ValueError, json.JSONDecodeError):
+            return []
+    if not isinstance(source, list):
+        return []
+
+    parsed: list[int] = []
+    for item in source:
+        if item is None or isinstance(item, bool):
+            continue
+        if isinstance(item, int):
+            parsed.append(item)
+            continue
+        if isinstance(item, float):
+            if item.is_integer():
+                parsed.append(int(item))
+            continue
+        if isinstance(item, str):
+            text = item.strip()
+            if text and text.lstrip("-").isdigit():
+                parsed.append(int(text))
+    return parsed
+
+
+def _parse_json_response(text: str) -> dict:
+    body = (text or "").strip()
+    if body.startswith("```json"):
+        body = body[7:].strip()
+    if body.startswith("```"):
+        body = body[3:].strip()
+    if body.endswith("```"):
+        body = body[:-3].strip()
+    return json.loads(body)
+
+
+async def build_ai_optimize_preview(
+    db: AsyncSession,
+    tenant_id: int,
+    *,
+    project_id: Optional[int] = None,
+    max_moves: int = 20,
+    unassigned_only: bool = False,
+) -> list[dict]:
+    """生成任务重分配建议（预览，不落库）。"""
+    task_query = select(Task).where(
+        and_(
+            Task.tenant_id == tenant_id,
+            func.lower(Task.status).notin_(["done", "completed", "closed"]),
+        )
+    )
+    if unassigned_only:
+        task_query = task_query.where(Task.assignee_id.is_(None))
+    if project_id is not None:
+        task_query = task_query.where(Task.project_id == project_id)
+
+    task_rows = (await db.execute(task_query.order_by(Task.updated_at.desc()))).scalars().all()
+    if not task_rows:
+        return []
+
+    active_users_query = select(User).where(
+        and_(
+            User.tenant_id == tenant_id,
+            User.status == "active",
+            User.is_active.is_(True),
+        )
+    )
+    users = (await db.execute(active_users_query)).scalars().all()
+    if not users:
+        return []
+
+    workload_rows = (
+        await db.execute(
+            select(UserWorkload).where(
+                and_(
+                    UserWorkload.tenant_id == tenant_id,
+                    UserWorkload.user_id.in_([int(u.id) for u in users]),
+                )
+            )
+        )
+    ).scalars().all()
+    workload_map = {int(w.user_id): float(w.workload_ratio or 0.0) for w in workload_rows}
+
+    skill_rows = (
+        await db.execute(
+            select(UserSkillBinding.user_id, UserSkillBinding.skill_tag_id).where(
+                UserSkillBinding.tenant_id == tenant_id
+            )
+        )
+    ).all()
+    user_skill_map: dict[int, set[int]] = {}
+    for uid, sid in skill_rows:
+        user_skill_map.setdefault(int(uid), set()).add(int(sid))
+
+    users_by_project: dict[int, list[User]] = {}
+    for task in task_rows:
+        if int(task.project_id) in users_by_project:
+            continue
+        links = (
+            await db.execute(
+                select(UserProject.user_id).where(
+                    and_(
+                        UserProject.tenant_id == tenant_id,
+                        UserProject.project_id == int(task.project_id),
+                    )
+                )
+            )
+        ).scalars().all()
+        allowed_ids = {int(uid) for uid in links}
+        users_by_project[int(task.project_id)] = [u for u in users if int(u.id) in allowed_ids]
+
+    llm_suggestion_map: dict[int, dict] = {}
+    members_payload = []
+    for user in users:
+        uid = int(user.id)
+        members_payload.append(
+            {
+                "user_id": uid,
+                "user_name": user.name,
+                "workload_ratio": round(workload_map.get(uid, float(user.workload or 0)), 2),
+                "skill_tag_ids": sorted(list(user_skill_map.get(uid, set()))),
+            }
+        )
+    tasks_payload = [
+        {
+            "task_id": int(task.id),
+            "project_id": int(task.project_id),
+            "title": task.title,
+            "assignee_id": int(task.assignee_id) if task.assignee_id else None,
+            "required_skill_tag_ids": _safe_int_list(task.required_skill_tag_ids),
+            "estimated_hours": float(task.estimated_hours or 0),
+            "status": task.status,
+        }
+        for task in task_rows
+    ]
+    llm_prompt = PROMPT_WORKLOAD_OPTIMIZATION.format(
+        members_json=json.dumps(members_payload, ensure_ascii=False),
+        tasks_json=json.dumps(tasks_payload, ensure_ascii=False),
+    )
+    llm_text = ""
+    llm_status = "success"
+    try:
+        llm_text = await llm_service.generate(
+            llm_prompt,
+            function_name="workload_optimize_prompt",
+            tenant_id=tenant_id,
+        )
+        parsed = _parse_json_response(llm_text)
+        for item in parsed.get("proposals", []) if isinstance(parsed, dict) else []:
+            try:
+                tid = int(item.get("task_id"))
+                llm_suggestion_map[tid] = {
+                    "new_assignee_id": int(item.get("new_assignee_id")),
+                    "reason": str(item.get("reason") or "llm_suggested"),
+                }
+            except (TypeError, ValueError):
+                continue
+    except Exception as exc:
+        llm_status = "failed"
+        llm_text = str(exc)
+
+    db.add(
+        AICallLog(
+            tenant_id=tenant_id,
+            user_id=None,
+            function_name="workload_optimize_prompt",
+            prompt=llm_prompt,
+            response=(llm_text or "")[:8000],
+            tokens_used=max(len(llm_prompt) // 4, 1) + max(len(llm_text) // 4, 0),
+            cost=((max(len(llm_prompt) // 4, 1) * 0.00015 + max(len(llm_text) // 4, 0) * 0.0006) / 1000),
+            status=llm_status,
+        )
+    )
+    await db.commit()
+
+    proposals: list[dict] = []
+    for task in task_rows:
+        current_assignee = int(task.assignee_id or 0)
+        is_unassigned = current_assignee <= 0
+        candidates = users_by_project.get(int(task.project_id), [])
+        if not candidates:
+            continue
+
+        required_skill_ids = _safe_int_list(task.required_skill_tag_ids)
+        ranked = []
+        llm_target_id = llm_suggestion_map.get(int(task.id), {}).get("new_assignee_id")
+        llm_reason = llm_suggestion_map.get(int(task.id), {}).get("reason")
+        for candidate in candidates:
+            uid = int(candidate.id)
+            if not is_unassigned and uid == current_assignee:
+                continue
+            ratio = workload_map.get(uid, float(candidate.workload or 0))
+            skill_score = _skill_match_score(required_skill_ids, user_skill_map.get(uid, set()))
+            # workload 越低越优，skill_score 越高越优
+            total_score = (skill_score * 0.7) + ((max(0.0, 120.0 - ratio) / 120.0) * 0.3)
+            if llm_target_id and uid == int(llm_target_id):
+                total_score += 0.2
+            ranked.append((total_score, ratio, candidate))
+
+        ranked.sort(key=lambda x: (-x[0], x[1]))
+        if not ranked:
+            continue
+
+        best_score, _best_ratio, best_user = ranked[0]
+        current_ratio = workload_map.get(current_assignee, 0.0) if not is_unassigned else 0.0
+        if is_unassigned:
+            if best_score < 0.2:
+                continue
+            reason = f"initial_assignment_score={best_score:.2f}"
+        else:
+            if current_ratio <= 100 and best_score < 0.5:
+                continue
+            reason = f"skill_match+workload_score={best_score:.2f}"
+        if llm_target_id and int(best_user.id) == int(llm_target_id):
+            reason = f"{reason};llm={llm_reason or 'suggested'}"
+
+        proposals.append(
+            {
+                "task_id": int(task.id),
+                "task_title": task.title,
+                "project_id": int(task.project_id),
+                "old_assignee_id": current_assignee if not is_unassigned else None,
+                "new_assignee_id": int(best_user.id),
+                "old_ratio": round(current_ratio, 2) if not is_unassigned else None,
+                "new_ratio": round(workload_map.get(int(best_user.id), 0.0), 2),
+                "reason": reason,
+            }
+        )
+        if len(proposals) >= max_moves:
+            break
+
+    return proposals
+
+
+async def apply_ai_optimize_plan(
+    db: AsyncSession,
+    tenant_id: int,
+    proposals: list[dict],
+) -> dict:
+    """应用优化方案并重算涉及成员负载。"""
+    touched_user_ids: set[int] = set()
+    applied = 0
+
+    for item in proposals:
+        task_id = int(item.get("task_id", 0))
+        new_assignee_id = int(item.get("new_assignee_id", 0))
+        if task_id <= 0 or new_assignee_id <= 0:
+            continue
+
+        task = (
+            await db.execute(
+                select(Task).where(
+                    and_(
+                        Task.tenant_id == tenant_id,
+                        Task.id == task_id,
+                    )
+                )
+            )
+        ).scalar_one_or_none()
+        if not task:
+            continue
+
+        if task.assignee_id:
+            touched_user_ids.add(int(task.assignee_id))
+        task.assignee_id = new_assignee_id
+        touched_user_ids.add(new_assignee_id)
+        applied += 1
+
+    if touched_user_ids:
+        for uid in touched_user_ids:
+            await refresh_user_assigned_hours_from_tasks(db, tenant_id, uid)
+    else:
+        await db.commit()
+
+    return {"applied_count": applied, "affected_users": len(touched_user_ids)}
diff --git a/backend/app/state.py b/backend/app/state.py
new file mode 100644
index 00000000..6870f7d5
--- /dev/null
+++ b/backend/app/state.py
@@ -0,0 +1,18 @@
+"""
+应用级全局状态
+
+event_bus、ws_manager 等由 main.py startup 设置，各 router 通过 `from app import state` 使用。
+"""
+
+event_bus = None
+ws_manager = None
+
+
+def set_event_bus(eb):
+    global event_bus
+    event_bus = eb
+
+
+def set_ws_manager(wm):
+    global ws_manager
+    ws_manager = wm
diff --git a/backend/app/websocket.py b/backend/app/websocket.py
new file mode 100644
index 00000000..f97e5e90
--- /dev/null
+++ b/backend/app/websocket.py
@@ -0,0 +1,263 @@
+"""
+Neural Sync - WebSocket Manager
+
+实现实时事件广播和 WebSocket 连接管理。
+"""
+
+from typing import Dict, List, Set, Optional, Any
+from datetime import datetime
+import json
+import logging
+from fastapi import WebSocket, WebSocketDisconnect
+from enum import Enum
+
+logger = logging.getLogger(__name__)
+
+
+class ConnectionEventType(str, Enum):
+    """连接事件类型"""
+    CONNECTED = "connected"
+    DISCONNECTED = "disconnected"
+    SUBSCRIBED = "subscribed"
+    UNSUBSCRIBED = "unsubscribed"
+
+
+class WebSocketConnection:
+    """WebSocket 连接包装器"""
+
+    def __init__(self, websocket: WebSocket, client_id: str):
+        self.websocket = websocket
+        self.client_id = client_id
+        self.subscriptions: Set[str] = set()
+        self.connected_at = datetime.utcnow()
+        self.last_ping: Optional[datetime] = None
+
+    async def send(self, data: Dict[str, Any]) -> bool:
+        """发送数据到客户端"""
+        try:
+            await self.websocket.send_json(data)
+            return True
+        except Exception as e:
+            logger.error(f"Error sending to client {self.client_id}: {e}")
+            return False
+
+    async def ping(self) -> bool:
+        """发送 ping"""
+        try:
+            await self.websocket.send_json({"type": "ping", "timestamp": datetime.utcnow().timestamp()})
+            self.last_ping = datetime.utcnow()
+            return True
+        except Exception as e:
+            logger.error(f"Error pinging client {self.client_id}: {e}")
+            return False
+
+    def subscribe(self, event_type: str):
+        """订阅事件类型"""
+        self.subscriptions.add(event_type)
+
+    def unsubscribe(self, event_type: str):
+        """取消订阅事件类型"""
+        self.subscriptions.discard(event_type)
+
+    def is_subscribed(self, event_type: str) -> bool:
+        """检查是否订阅了事件类型"""
+        return event_type in self.subscriptions or "*" in self.subscriptions
+
+
+class WebSocketManager:
+    """
+    WebSocket 连接管理器
+
+    特性:
+    - 连接管理
+    - 事件订阅过滤
+    - 心跳检测
+    - 广播和单播
+    """
+
+    def __init__(self):
+        self._connections: Dict[str, WebSocketConnection] = {}
+        self._client_counter = 0
+        self._heartbeat_interval = 30  # seconds
+        self._heartbeat_task: Optional[Any] = None
+
+    @property
+    def connections(self) -> Dict[str, WebSocketConnection]:
+        """获取所有连接"""
+        return self._connections.copy()
+
+    @property
+    def connection_count(self) -> int:
+        """获取连接数"""
+        return len(self._connections)
+
+    async def connect(self, websocket: WebSocket) -> str:
+        """
+        接受新的 WebSocket 连接
+
+        Args:
+            websocket: WebSocket 连接
+
+        Returns:
+            客户端 ID
+        """
+        await websocket.accept()
+
+        self._client_counter += 1
+        client_id = f"client_{self._client_counter}"
+
+        connection = WebSocketConnection(websocket, client_id)
+        self._connections[client_id] = connection
+
+        # 发送连接确认
+        await connection.send({
+            "type": ConnectionEventType.CONNECTED,
+            "client_id": client_id,
+            "timestamp": datetime.utcnow().timestamp(),
+        })
+
+        logger.info(f"WebSocket client connected: {client_id}")
+        return client_id
+
+    async def disconnect(self, client_id: str):
+        """断开连接"""
+        if client_id in self._connections:
+            del self._connections[client_id]
+            logger.info(f"WebSocket client disconnected: {client_id}")
+
+    async def subscribe(self, client_id: str, event_types: List[str]):
+        """订阅事件类型"""
+        if client_id in self._connections:
+            connection = self._connections[client_id]
+            for event_type in event_types:
+                connection.subscribe(event_type)
+
+            await connection.send({
+                "type": ConnectionEventType.SUBSCRIBED,
+                "event_types": list(connection.subscriptions),
+                "timestamp": datetime.utcnow().timestamp(),
+            })
+
+    async def unsubscribe(self, client_id: str, event_types: List[str]):
+        """取消订阅事件类型"""
+        if client_id in self._connections:
+            connection = self._connections[client_id]
+            for event_type in event_types:
+                connection.unsubscribe(event_type)
+
+            await connection.send({
+                "type": ConnectionEventType.UNSUBSCRIBED,
+                "event_types": list(connection.subscriptions),
+                "timestamp": datetime.utcnow().timestamp(),
+            })
+
+    async def broadcast(self, event_type: str, data: Dict[str, Any]) -> int:
+        """
+        广播事件到所有订阅的客户端
+
+        Args:
+            event_type: 事件类型
+            data: 事件数据
+
+        Returns:
+            发送成功的客户端数量
+        """
+        message = {
+            "type": "event",
+            "event_type": event_type,
+            "data": data,
+            "timestamp": datetime.utcnow().timestamp(),
+        }
+
+        success_count = 0
+        failed_clients = []
+
+        for client_id, connection in self._connections.items():
+            if connection.is_subscribed(event_type):
+                if await connection.send(message):
+                    success_count += 1
+                else:
+                    failed_clients.append(client_id)
+
+        # 清理失败的连接
+        for client_id in failed_clients:
+            await self.disconnect(client_id)
+
+        if success_count > 0:
+            logger.debug(f"Broadcast {event_type} to {success_count} clients")
+
+        return success_count
+
+    async def send_to_client(self, client_id: str, data: Dict[str, Any]) -> bool:
+        """发送数据到指定客户端"""
+        if client_id in self._connections:
+            return await self._connections[client_id].send(data)
+        return False
+
+    async def start_heartbeat(self):
+        """启动心跳检测"""
+        import asyncio
+
+        async def heartbeat_loop():
+            while True:
+                await asyncio.sleep(self._heartbeat_interval)
+                await self._send_heartbeat()
+
+        self._heartbeat_task = asyncio.create_task(heartbeat_loop())
+        logger.info("WebSocket heartbeat started")
+
+    async def stop_heartbeat(self):
+        """停止心跳检测"""
+        if self._heartbeat_task:
+            self._heartbeat_task.cancel()
+            self._heartbeat_task = None
+            logger.info("WebSocket heartbeat stopped")
+
+    async def _send_heartbeat(self):
+        """发送心跳到所有连接"""
+        failed_clients = []
+
+        for client_id, connection in self._connections.items():
+            if not await connection.ping():
+                failed_clients.append(client_id)
+
+        # 清理失败的连接
+        for client_id in failed_clients:
+            await self.disconnect(client_id)
+
+    async def close_all(self):
+        """关闭所有连接"""
+        for client_id in list(self._connections.keys()):
+            await self.disconnect(client_id)
+
+        await self.stop_heartbeat()
+
+
+# 全局 WebSocket 管理器实例
+_global_manager: Optional[WebSocketManager] = None
+
+
+def get_websocket_manager() -> WebSocketManager:
+    """获取全局 WebSocket 管理器实例"""
+    global _global_manager
+    if _global_manager is None:
+        _global_manager = WebSocketManager()
+    return _global_manager
+
+
+# 事件总线桥接 - 自动将事件广播到 WebSocket 客户端
+async def setup_event_bridge(event_bus):
+    """设置事件总线与 WebSocket 的桥接"""
+    from .events import EventType
+
+    async def broadcast_event(event):
+        """广播事件到 WebSocket 客户端"""
+        manager = get_websocket_manager()
+        await manager.broadcast(
+            event_type=event.type.value,
+            data=event.data,
+        )
+
+    # 订阅所有事件
+    event_bus.subscribe_wildcard(broadcast_event)
+    logger.info("Event bus to WebSocket bridge established")
diff --git a/backend/check_users.py b/backend/check_users.py
new file mode 100644
index 00000000..f5b57c38
--- /dev/null
+++ b/backend/check_users.py
@@ -0,0 +1,27 @@
+
+import asyncio
+import sys
+from pathlib import Path
+
+# Add backend directory to sys.path
+sys.path.append(str(Path(__file__).parent.absolute()))
+
+from app.database import AsyncSessionLocal
+from app.models.user import User
+from sqlalchemy import select, func
+
+async def check_users():
+    async with AsyncSessionLocal() as session:
+        result = await session.execute(select(func.count(User.id)))
+        count = result.scalar()
+        print(f"User count: {count}")
+        
+        if count > 0:
+            result = await session.execute(select(User.username, User.role).limit(5))
+            users = result.fetchall()
+            print("Existing users:")
+            for u in users:
+                print(f"- {u.username} ({u.role})")
+
+if __name__ == "__main__":
+    asyncio.run(check_users())
diff --git a/backend/init_db.py b/backend/init_db.py
new file mode 100644
index 00000000..1a0306a8
--- /dev/null
+++ b/backend/init_db.py
@@ -0,0 +1,41 @@
+#!/usr/bin/env python3
+"""
+Initialize/fix SQLite database by creating all missing tables.
+Run this when the database schema is out of sync.
+"""
+import asyncio
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from sqlalchemy.ext.asyncio import create_async_engine, AsyncSession
+from sqlalchemy.orm import sessionmaker
+from backend.app.models import Base
+from backend.app.database import DATABASE_URL
+
+from sqlalchemy import text
+
+async def init_database():
+    """Create all tables that don't exist yet."""
+    print(f"Using database: {DATABASE_URL}")
+
+    engine = create_async_engine(DATABASE_URL, echo=True)
+    async_session = sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
+
+    async with engine.begin() as conn:
+        # Create all tables that don't exist
+        await conn.run_sync(Base.metadata.create_all)
+        print("✅ All tables created successfully")
+
+    # Verify tables were created
+    async with engine.begin() as conn:
+        result = await conn.run_sync(lambda sync_conn: sync_conn.execute(text("SELECT name FROM sqlite_master WHERE type='table'")).fetchall())
+        tables = [row[0] for row in result]
+        print(f"📊 Tables in database: {', '.join(sorted(tables))}")
+
+    await engine.dispose()
+
+if __name__ == "__main__":
+    asyncio.run(init_database())
diff --git a/backend/scripts/migrate_sqlite_to_postgres.py b/backend/scripts/migrate_sqlite_to_postgres.py
new file mode 100644
index 00000000..95a9c112
--- /dev/null
+++ b/backend/scripts/migrate_sqlite_to_postgres.py
@@ -0,0 +1,290 @@
+#!/usr/bin/env python3
+"""将本地 SQLite 数据迁移到当前 PostgreSQL（Docker）。
+
+默认源库: /app/backend/neural_sync.db
+默认目标库: 环境变量 DATABASE_URL
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import json
+import os
+import sqlite3
+from datetime import datetime
+from typing import Any
+
+import asyncpg
+
+TABLE_ORDER = [
+    "skill_tags",
+    "users",
+    "projects",
+    "feishu_items",
+    "signals",
+    "truth_nodes",
+    "tasks",
+    "project_signal_links",
+    "user_skill_bindings",
+    "user_workloads",
+    "qa_sessions",
+    "qa_records",
+    "qa_feedback",
+]
+
+
+def _parse_json(value: Any, default: Any) -> Any:
+    if value is None:
+        return default
+    if isinstance(value, (dict, list)):
+        return value
+    text = str(value).strip()
+    if not text:
+        return default
+    try:
+        return json.loads(text)
+    except Exception:
+        return default
+
+
+def _parse_ts(value: Any) -> Any:
+    if value is None:
+        return None
+    if isinstance(value, datetime):
+        return value
+    text = str(value).strip()
+    if not text:
+        return None
+    text = text.replace("T", " ").replace("Z", "")
+    try:
+        return datetime.fromisoformat(text)
+    except Exception:
+        return text
+
+
+def _parse_int_array(value: Any) -> list[int]:
+    if value is None:
+        return []
+    if isinstance(value, list):
+        raw = value
+    else:
+        raw = _parse_json(value, default=[])
+        if not isinstance(raw, list):
+            raw = []
+    result: list[int] = []
+    for item in raw:
+        try:
+            num = int(item)
+            result.append(num)
+        except Exception:
+            continue
+    return result
+
+
+def _transform(table: str, row: dict[str, Any], pg_types: dict[str, str]) -> dict[str, Any]:
+    data = dict(row)
+
+    if table == "users":
+        data["assigned_domains"] = _parse_json(data.get("assigned_domains"), default=[])
+
+    if table == "truth_nodes":
+        content = data.get("content")
+        parsed = _parse_json(content, default=None)
+        if parsed is None:
+            parsed = {"raw": str(content or "")}
+        data["content"] = parsed
+
+    if table == "tasks":
+        data["required_skill_tag_ids"] = _parse_int_array(data.get("required_skill_tag_ids"))
+        data["ai_recommended_assignees"] = _parse_json(data.get("ai_recommended_assignees"), default=[])
+
+    if table == "qa_records":
+        data["sources"] = _parse_json(data.get("sources"), default=[])
+        data["domains_used"] = _parse_json(data.get("domains_used"), default=[])
+
+    if table == "qa_feedback":
+        data["error_types"] = _parse_json(data.get("error_types"), default=[])
+
+    # 通用类型修正
+    for col, typ in pg_types.items():
+        if col not in data:
+            continue
+        if typ in {"json", "jsonb"}:
+            value = data[col]
+            if value is None:
+                continue
+            if isinstance(value, str):
+                data[col] = value
+            else:
+                data[col] = json.dumps(value, ensure_ascii=False)
+            continue
+        if typ == "boolean":
+            value = data[col]
+            if isinstance(value, bool):
+                continue
+            if value is None:
+                data[col] = False
+            else:
+                data[col] = str(value).strip().lower() in {"1", "true", "t", "yes", "y"}
+        elif typ == "timestamp without time zone":
+            data[col] = _parse_ts(data.get(col))
+
+    return data
+
+
+async def _pg_table_columns(conn: asyncpg.Connection, table: str) -> list[str]:
+    rows = await conn.fetch(
+        """
+        SELECT column_name
+        FROM information_schema.columns
+        WHERE table_schema='public' AND table_name=$1
+        ORDER BY ordinal_position
+        """,
+        table,
+    )
+    return [str(r["column_name"]) for r in rows]
+
+
+async def _pg_table_types(conn: asyncpg.Connection, table: str) -> dict[str, str]:
+    rows = await conn.fetch(
+        """
+        SELECT column_name, data_type
+        FROM information_schema.columns
+        WHERE table_schema='public' AND table_name=$1
+        """,
+        table,
+    )
+    return {str(r["column_name"]): str(r["data_type"]) for r in rows}
+
+
+async def _pg_existing_tables(conn: asyncpg.Connection) -> set[str]:
+    rows = await conn.fetch(
+        """
+        SELECT tablename
+        FROM pg_tables
+        WHERE schemaname='public'
+        """
+    )
+    return {str(r["tablename"]) for r in rows}
+
+
+def _sqlite_columns(sqlite_conn: sqlite3.Connection, table: str) -> list[str]:
+    rows = sqlite_conn.execute(f"PRAGMA table_info({table})").fetchall()
+    return [str(r[1]) for r in rows]
+
+
+def _sqlite_rows(sqlite_conn: sqlite3.Connection, table: str) -> list[dict[str, Any]]:
+    sqlite_conn.row_factory = sqlite3.Row
+    rows = sqlite_conn.execute(f"SELECT * FROM {table} ORDER BY id ASC").fetchall()
+    return [dict(r) for r in rows]
+
+
+async def migrate(sqlite_path: str, postgres_url: str, *, mode: str = "merge") -> None:
+    sqlite_conn = sqlite3.connect(sqlite_path)
+    sqlite_conn.row_factory = sqlite3.Row
+
+    pg_url = postgres_url.replace("postgresql+asyncpg://", "postgresql://")
+    pg = await asyncpg.connect(pg_url)
+
+    try:
+        pg_tables = await _pg_existing_tables(pg)
+        sqlite_tables = {
+            r[0]
+            for r in sqlite_conn.execute("SELECT name FROM sqlite_master WHERE type='table'").fetchall()
+        }
+
+        tables = [t for t in TABLE_ORDER if t in pg_tables and t in sqlite_tables]
+        if not tables:
+            raise RuntimeError("无可迁移表，请检查库路径与目标库")
+
+        print(f"[INFO] migrate tables: {tables}")
+        print(f"[INFO] mode: {mode}")
+
+        if mode == "replace":
+            # 清空目标核心表（含级联）
+            await pg.execute("TRUNCATE TABLE " + ", ".join(tables) + " RESTART IDENTITY CASCADE")
+
+        for table in tables:
+            pg_cols = await _pg_table_columns(pg, table)
+            pg_types = await _pg_table_types(pg, table)
+            sqlite_cols = _sqlite_columns(sqlite_conn, table)
+            common_cols = [c for c in pg_cols if c in sqlite_cols]
+            if not common_cols:
+                print(f"[WARN] skip {table}: no common cols")
+                continue
+
+            rows = _sqlite_rows(sqlite_conn, table)
+            inserted = 0
+            for row in rows:
+                row = _transform(table, row, pg_types)
+                values = [row.get(c) for c in common_cols]
+                placeholders = ", ".join(f"${i+1}" for i in range(len(common_cols)))
+                cols_sql = ", ".join(common_cols)
+                if mode == "merge" and "id" in common_cols and "id" in pg_cols:
+                    sql = (
+                        f"INSERT INTO {table} ({cols_sql}) VALUES ({placeholders}) "
+                        "ON CONFLICT (id) DO NOTHING"
+                    )
+                else:
+                    sql = f"INSERT INTO {table} ({cols_sql}) VALUES ({placeholders})"
+                try:
+                    await pg.execute(sql, *values)
+                    inserted += 1
+                except Exception as exc:
+                    print(f"[WARN] {table} row id={row.get('id')} failed: {exc}")
+
+            if "id" in common_cols:
+                await pg.execute(
+                    f"SELECT setval(pg_get_serial_sequence('{table}','id'), COALESCE((SELECT MAX(id) FROM {table}), 1), true)"
+                )
+
+            print(f"[OK] {table}: {inserted}/{len(rows)}")
+
+        # 摘要
+        summary_tables = [
+            "users",
+            "projects",
+            "signals",
+            "truth_nodes",
+            "tasks",
+            "qa_sessions",
+            "qa_records",
+            "qa_feedback",
+        ]
+        print("[SUMMARY]")
+        for table in summary_tables:
+            if table not in tables:
+                continue
+            count = await pg.fetchval(f"SELECT count(*) FROM {table}")
+            print(f"  - {table}: {int(count or 0)}")
+    finally:
+        await pg.close()
+        sqlite_conn.close()
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--sqlite", default="/app/backend/neural_sync.db")
+    parser.add_argument("--postgres", default=os.getenv("DATABASE_URL", ""))
+    parser.add_argument("--mode", choices=["merge", "replace"], default="merge")
+    args = parser.parse_args()
+
+    if not args.postgres:
+        print("[ERROR] DATABASE_URL 为空，请使用 --postgres 指定")
+        return 2
+
+    if not os.path.exists(args.sqlite):
+        print(f"[ERROR] SQLite 文件不存在: {args.sqlite}")
+        return 2
+
+    try:
+        asyncio.run(migrate(args.sqlite, args.postgres, mode=args.mode))
+        return 0
+    except Exception as exc:
+        print(f"[ERROR] migration failed: {exc}")
+        return 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/backend/seed_all.py b/backend/seed_all.py
new file mode 100644
index 00000000..8195cb37
--- /dev/null
+++ b/backend/seed_all.py
@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+"""
+Comprehensive seed script for PMFlow/Neural Sync.
+Creates users, domains, and initial data.
+"""
+import asyncio
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.future import select
+from app.database import AsyncSessionLocal
+from app.models import User, ProjectDomain, Project
+from app.services.auth import AuthService
+
+
+async def seed_users():
+    """Create default users with known passwords."""
+    # Create initial users (with simple password hashes for demo)
+    # In production, these should be created with proper password hashing
+    # admin_hash = AuthService.hash_password("admin123")
+    # boss_hash = AuthService.hash_password("boss123")
+    # ai_hash = AuthService.hash_password("ai123")
+    # be_hash = AuthService.hash_password("be123")
+    # fe_hash = AuthService.hash_password("fe123")
+
+    async with AsyncSessionLocal() as session:
+        # Check if users already exist
+        result = await session.execute(select(User).where(User.username == "admin"))
+        admin = result.scalar_one_or_none()
+
+        if not admin:
+            # Create admin user
+            admin = User(
+                username="admin",
+                password_hash=AuthService.hash_password("admin123"),
+                name="系统管理员",
+                email="admin@pmflow.ai",
+                role="BOSS",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+                tenant_id=1,
+            )
+            session.add(admin)
+            print("✅ Created admin user (username: admin, password: admin123)")
+
+        # Create boss user
+        result = await session.execute(select(User).where(User.username == "boss"))
+        boss = result.scalar_one_or_none()
+        if not boss:
+            boss = User(
+                username="boss",
+                password_hash=AuthService.hash_password("boss123"),
+                name="老板",
+                email="boss@pmflow.ai",
+                role="BOSS",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+                tenant_id=1,
+            )
+            session.add(boss)
+            print("✅ Created boss user (username: boss, password: boss123)")
+
+        # Create AI engineer user
+        result = await session.execute(select(User).where(User.username == "ai_engineer"))
+        ai_user = result.scalar_one_or_none()
+        if not ai_user:
+            ai_user = User(
+                username="ai_engineer",
+                password_hash=AuthService.hash_password("ai123"),
+                name="AI工程师",
+                email="ai@pmflow.ai",
+                role="AI",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+                tenant_id=1,
+            )
+            session.add(ai_user)
+            print("✅ Created ai_engineer user (username: ai_engineer, password: ai123)")
+
+        # Create backend user
+        result = await session.execute(select(User).where(User.username == "backend"))
+        backend_user = result.scalar_one_or_none()
+        if not backend_user:
+            backend_user = User(
+                username="backend",
+                password_hash=AuthService.hash_password("be123"),
+                name="后端开发",
+                email="backend@pmflow.ai",
+                role="BE",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+                tenant_id=1,
+            )
+            session.add(backend_user)
+            print("✅ Created backend user (username: backend, password: be123)")
+
+        # Create frontend user
+        result = await session.execute(select(User).where(User.username == "frontend"))
+        frontend_user = result.scalar_one_or_none()
+        if not frontend_user:
+            frontend_user = User(
+                username="frontend",
+                password_hash=AuthService.hash_password("fe123"),
+                name="前端开发",
+                email="frontend@pmflow.ai",
+                role="FE",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+                tenant_id=1,
+            )
+            session.add(frontend_user)
+            print("✅ Created frontend user (username: frontend, password: fe123)")
+
+        await session.commit()
+
+
+async def seed_domains():
+    """Create default project domains."""
+    async with AsyncSessionLocal() as session:
+        # Check if domains already exist
+        result = await session.execute(select(ProjectDomain))
+        domains = result.scalars().all()
+
+        if not domains:
+            default_domains = [
+                ProjectDomain(
+                    name="AI模型工程",
+                    slug="ai-engineering",
+                    description="AI模型训练、评估和部署相关项目",
+                    is_default=True,
+                    signal_count=0,
+                    context_node_count=0,
+                    experiment_count=0,
+                    tenant_id=1,
+                ),
+                ProjectDomain(
+                    name="后端服务",
+                    slug="backend-services",
+                    description="后端API、数据库、微服务等项目",
+                    is_default=False,
+                    signal_count=0,
+                    context_node_count=0,
+                    experiment_count=0,
+                    tenant_id=1,
+                ),
+                ProjectDomain(
+                    name="前端应用",
+                    slug="frontend-apps",
+                    description="Vue/React前端应用、UI组件等",
+                    is_default=False,
+                    signal_count=0,
+                    context_node_count=0,
+                    experiment_count=0,
+                    tenant_id=1,
+                ),
+            ]
+
+            for domain in default_domains:
+                session.add(domain)
+
+            await session.commit()
+            print("✅ Created 3 default project domains: ai-engineering, backend-services, frontend-apps")
+        else:
+            print(f"ℹ️  Project domains already exist ({len(domains)} found)")
+
+
+async def main():
+    print("🌱 Seeding PMFlow database...")
+    print("=" * 50)
+
+    await seed_users()
+    await seed_domains()
+
+    print("=" * 50)
+    print("✅ Seeding completed!")
+    print()
+    print("📝 Default users created:")
+    print("  - admin / admin123 (BOSS)")
+    print("  - boss / boss123 (BOSS)")
+    print("  - ai_engineer / ai123 (AI)")
+    print("  - backend / be123 (BE)")
+    print("  - frontend / fe123 (FE)")
+    print()
+    print("🔐 Use these credentials to login at: http://127.0.0.1:5174/login")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/backend/tests/test_adaptive_rag_intent_corrections.py b/backend/tests/test_adaptive_rag_intent_corrections.py
new file mode 100644
index 00000000..2401989d
--- /dev/null
+++ b/backend/tests/test_adaptive_rag_intent_corrections.py
@@ -0,0 +1,36 @@
+import unittest
+
+from app.services.adaptive_rag_service import AdaptiveRAGService
+
+
+class AdaptiveRAGIntentCorrectionTest(unittest.TestCase):
+    def test_build_intent_correction_record_is_structured_and_stable(self) -> None:
+        first = AdaptiveRAGService._build_intent_correction_record(
+            user_utterance="不对，是新增进 PMFlow 的项目列表里面",
+            wrong_interpretation="signal_intake",
+            correct_intent="create_project",
+            canonical_action="project.create",
+            clarification_template="请补充项目名称、项目描述、项目范围。",
+            risk_level="high",
+            project_id=None,
+        )
+        second = AdaptiveRAGService._build_intent_correction_record(
+            user_utterance="不对，是新增进 PMFlow 的项目列表里面",
+            wrong_interpretation="signal_intake",
+            correct_intent="create_project",
+            canonical_action="project.create",
+            clarification_template="请补充项目名称、项目描述、项目范围。",
+            risk_level="high",
+            project_id=None,
+        )
+
+        self.assertEqual(first[0], second[0])
+        self.assertEqual(first[1], "意图纠错/create_project")
+        self.assertIn("用户原话:", first[2])
+        self.assertEqual(first[3]["record_type"], "intent_correction")
+        self.assertEqual(first[3]["correct_intent"], "create_project")
+        self.assertEqual(first[3]["canonical_action"], "project.create")
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/backend/tests/test_ai_logs_api.py b/backend/tests/test_ai_logs_api.py
new file mode 100644
index 00000000..1ae3d2d0
--- /dev/null
+++ b/backend/tests/test_ai_logs_api.py
@@ -0,0 +1,184 @@
+import os
+import unittest
+from tempfile import NamedTemporaryFile
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
+
+from app.api.ai_logs import _apply_project_filter
+from app.models import AIProcessingLog, Base, IngestEvent, Project, User, WorkItem
+
+
+class AILogsApiTest(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self) -> None:
+        tmp = NamedTemporaryFile(prefix="pmflow-ai-logs-test-", suffix=".db", delete=False)
+        tmp.close()
+        self.db_path = tmp.name
+        self.engine = create_async_engine(f"sqlite+aiosqlite:///{self.db_path}", future=True)
+        self.session_factory = async_sessionmaker(self.engine, expire_on_commit=False, class_=AsyncSession)
+        async with self.engine.begin() as conn:
+            await conn.run_sync(Base.metadata.create_all)
+
+        async with self.session_factory() as db:
+            user = User(
+                tenant_id=1,
+                username="boss",
+                email="boss@example.com",
+                password_hash="hash",
+                name="老板",
+                role="BOSS",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+            )
+            db.add(user)
+            await db.flush()
+
+            project_a = Project(
+                tenant_id=1,
+                name="项目A",
+                title="项目A",
+                description="A",
+                owner_id=int(user.id),
+                status="active",
+                domain_id=None,
+                assigned_role="pm",
+                signal_count=0,
+            )
+            project_b = Project(
+                tenant_id=1,
+                name="项目B",
+                title="项目B",
+                description="B",
+                owner_id=int(user.id),
+                status="active",
+                domain_id=None,
+                assigned_role="pm",
+                signal_count=0,
+            )
+            db.add_all([project_a, project_b])
+            await db.flush()
+
+            ingest_a = IngestEvent(
+                tenant_id=1,
+                channel="feishu_message",
+                event_key="om-project-a",
+                status="done",
+                payload={},
+                retry_count=0,
+                error_message=None,
+                processed_at=None,
+            )
+            ingest_b = IngestEvent(
+                tenant_id=1,
+                channel="feishu_message",
+                event_key="om-project-b",
+                status="done",
+                payload={},
+                retry_count=0,
+                error_message=None,
+                processed_at=None,
+            )
+            db.add_all([ingest_a, ingest_b])
+            await db.flush()
+
+            db.add_all(
+                [
+                    WorkItem(
+                        tenant_id=1,
+                        ingest_event_id=int(ingest_a.id),
+                        source_system="feishu",
+                        source_message_id="om-project-a",
+                        state="succeeded",
+                        intent="project.create",
+                        decision_level="project",
+                        project_id=int(project_a.id),
+                    ),
+                    WorkItem(
+                        tenant_id=1,
+                        ingest_event_id=int(ingest_b.id),
+                        source_system="feishu",
+                        source_message_id="om-project-b",
+                        state="succeeded",
+                        intent="signal.create",
+                        decision_level="project",
+                        project_id=int(project_b.id),
+                    ),
+                ]
+            )
+            db.add_all(
+                [
+                    AIProcessingLog(
+                        tenant_id=1,
+                        feishu_message_id="om-project-a",
+                        feishu_item_id=None,
+                        raw_message="创建项目A",
+                        ai_raw_output="{}",
+                        project_score=None,
+                        signal_confidence=None,
+                        level="ai_pmo",
+                        normalized_text="创建项目A",
+                        route="ai_pmo.project.create",
+                        intent="project.create",
+                        executor="ai_pmo_kernel",
+                        operation_code="project.create",
+                        operation_result={"project_id": int(project_a.id)},
+                        reply_source="action_succeeded",
+                        routing_decision="ai_pmo.project.create",
+                        routing_reason="AI PMO kernel route=ai_pmo.project.create",
+                        reasoning_bullets=["work_item_id=1"],
+                        processing_status="success",
+                        error_message=None,
+                        duration_ms=100,
+                    ),
+                    AIProcessingLog(
+                        tenant_id=1,
+                        feishu_message_id="om-project-b",
+                        feishu_item_id=None,
+                        raw_message="创建信号B",
+                        ai_raw_output="{}",
+                        project_score=None,
+                        signal_confidence=None,
+                        level="ai_pmo",
+                        normalized_text="创建信号B",
+                        route="ai_pmo.signal.create",
+                        intent="signal.create",
+                        executor="ai_pmo_kernel",
+                        operation_code="signal.enter",
+                        operation_result={"project_id": int(project_b.id)},
+                        reply_source="action_succeeded",
+                        routing_decision="ai_pmo.signal.create",
+                        routing_reason="AI PMO kernel route=ai_pmo.signal.create",
+                        reasoning_bullets=["work_item_id=2"],
+                        processing_status="success",
+                        error_message=None,
+                        duration_ms=100,
+                    ),
+                ]
+            )
+            await db.commit()
+
+    async def asyncTearDown(self) -> None:
+        await self.engine.dispose()
+        if getattr(self, "db_path", None):
+            try:
+                os.remove(self.db_path)
+            except FileNotFoundError:
+                pass
+
+    async def test_apply_project_filter_includes_ai_pmo_work_item_message_ids(self) -> None:
+        async with self.session_factory() as db:
+            stmt = select(AIProcessingLog).where(AIProcessingLog.tenant_id == 1).order_by(AIProcessingLog.id.asc())
+            rows = (await db.execute(_apply_project_filter(stmt, tenant_id=1, project_id=1))).scalars().all()
+
+        self.assertEqual(len(rows), 1)
+        self.assertEqual(rows[0].feishu_message_id, "om-project-a")
+
+    async def test_apply_project_filter_keeps_other_project_logs_out(self) -> None:
+        async with self.session_factory() as db:
+            stmt = select(AIProcessingLog).where(AIProcessingLog.tenant_id == 1).order_by(AIProcessingLog.id.asc())
+            rows = (await db.execute(_apply_project_filter(stmt, tenant_id=1, project_id=2))).scalars().all()
+
+        self.assertEqual(len(rows), 1)
+        self.assertEqual(rows[0].feishu_message_id, "om-project-b")
diff --git a/backend/tests/test_ai_pmo_full_scenario.py b/backend/tests/test_ai_pmo_full_scenario.py
new file mode 100644
index 00000000..ae513df2
--- /dev/null
+++ b/backend/tests/test_ai_pmo_full_scenario.py
@@ -0,0 +1,640 @@
+import json
+import os
+import unittest
+from datetime import datetime
+from tempfile import NamedTemporaryFile
+from unittest.mock import AsyncMock, patch
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
+
+from app.api.users import _create_user_with_defaults
+from app.models import (
+    Base,
+    Budget,
+    Project,
+    ProjectDomain,
+    ProjectSignalLink,
+    Signal,
+    SkillTag,
+    Task,
+    TruthNode,
+    User,
+    UserProject,
+    UserSkillBinding,
+    UserWorkload,
+    WorkItem,
+)
+from app.schemas import UserCreate
+from app.services.ai_pmo_cadence_service import AIPMOCadenceService
+from app.services.ai_pmo_kernel_service import AIPMOKernelService
+from app.services.signal_capture import ProactiveSignalCaptureService
+from app.services.skill_service import recommend_personnel
+from app.services import workload_service
+
+
+class AIPMOFullScenarioTest(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self) -> None:
+        tmp = NamedTemporaryFile(prefix="pmflow-ai-pmo-scenario-", suffix=".db", delete=False)
+        tmp.close()
+        self.db_path = tmp.name
+        self.engine = create_async_engine(f"sqlite+aiosqlite:///{self.db_path}", future=True)
+        self.session_factory = async_sessionmaker(self.engine, expire_on_commit=False, class_=AsyncSession)
+        async with self.engine.begin() as conn:
+            await conn.run_sync(Base.metadata.create_all)
+
+        async with self.session_factory() as db:
+            self.policy_skill_id = await self._create_skill_tag(db, name="政策分析", slug="policy-analysis")
+            self.research_skill_id = await self._create_skill_tag(db, name="行业研究", slug="market-research")
+            self.web_skill_id = await self._create_skill_tag(db, name="爬虫采集", slug="web-crawling")
+            self.graph_skill_id = await self._create_skill_tag(db, name="知识图谱", slug="knowledge-graph")
+            self.pipeline_skill_id = await self._create_skill_tag(db, name="数据管道", slug="data-pipeline")
+
+            boss = User(
+                tenant_id=1,
+                username="boss",
+                email="boss@example.com",
+                password_hash="hash",
+                name="老板",
+                role="BOSS",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+            )
+            db.add(boss)
+            await db.flush()
+            self.boss_id = int(boss.id)
+
+            domain = ProjectDomain(
+                tenant_id=1,
+                name="医药",
+                slug="pharma",
+                description="医药产业情报域",
+                config={},
+                is_active=True,
+                is_default=False,
+                signal_count=0,
+                context_node_count=0,
+                experiment_count=0,
+                owner_id=str(boss.id),
+            )
+            db.add(domain)
+            await db.flush()
+            self.domain_id = int(domain.id)
+
+            project = Project(
+                tenant_id=1,
+                name="医药行业十五五规划布局研究",
+                title="医药行业十五五规划布局研究",
+                description="跟踪医保支付改革、器械招采与创新药方向。",
+                owner_id=int(boss.id),
+                status="active",
+                domain_id=int(domain.id),
+                assigned_role="pm",
+                signal_count=0,
+            )
+            db.add(project)
+            await db.flush()
+            self.base_project_id = int(project.id)
+
+            charter = TruthNode(
+                tenant_id=1,
+                node_type="charter",
+                project_id=int(project.id),
+                signal_id=None,
+                created_by=int(boss.id),
+                title="医药研究项目宪章",
+                content={"success_criteria": ["两周内形成华东医保支付改革观察报告"]},
+                version=1,
+                status="active",
+            )
+            db.add(charter)
+            await db.flush()
+            self.base_charter_id = int(charter.id)
+
+            db.add(
+                Budget(
+                    tenant_id=1,
+                    budget_type="project",
+                    project_id=int(project.id),
+                    period="monthly",
+                    budget_amount_usd=1500.0,
+                    alert_threshold_percent=80.0,
+                    hard_limit_enabled=True,
+                    current_usage_usd=620.0,
+                    current_period_start=datetime(2026, 3, 1),
+                    current_period_end=datetime(2026, 3, 31),
+                    status="active",
+                    override_history={},
+                )
+            )
+            db.add(UserProject(tenant_id=1, user_id=int(boss.id), project_id=int(project.id)))
+            db.add(
+                UserWorkload(
+                    tenant_id=1,
+                    user_id=int(boss.id),
+                    current_assigned_hours=18.0,
+                    max_available_hours_per_week=40.0,
+                    workload_ratio=0.0,
+                )
+            )
+            await self._bind_skill(db, user_id=int(boss.id), skill_tag_id=self.policy_skill_id, level="advanced", proficiency=86)
+            await self._bind_skill(db, user_id=int(boss.id), skill_tag_id=self.research_skill_id, level="expert", proficiency=92)
+            db.add(
+                Task(
+                    tenant_id=1,
+                    truth_node_id=int(charter.id),
+                    project_id=int(project.id),
+                    assignee_id=int(boss.id),
+                    ai_recommended_assignees=None,
+                    title="拆解华东医保支付改革样本",
+                    description="需要比对江苏/浙江 DRG-DIP 改革口径和器械招采变化。",
+                    required_skill_tag_ids=[self.policy_skill_id, self.web_skill_id],
+                    status="blocked",
+                    priority="high",
+                    estimated_hours=12.0,
+                    actual_hours=0.0,
+                    quality_score=None,
+                    due_date=datetime(2026, 3, 12),
+                    completed_at=None,
+                )
+            )
+            await workload_service.recalculate_user_workload_ratio(db, tenant_id=1, user_id=int(boss.id), commit=False)
+            await db.commit()
+
+    async def asyncTearDown(self) -> None:
+        await self.engine.dispose()
+        if getattr(self, "db_path", None):
+            try:
+                os.remove(self.db_path)
+            except FileNotFoundError:
+                pass
+
+    async def _create_skill_tag(self, db: AsyncSession, *, name: str, slug: str) -> int:
+        row = SkillTag(
+            tenant_id=1,
+            name=name,
+            slug=slug,
+            category="skill",
+            description=f"{name}能力",
+            is_system=False,
+            sort_order=0,
+        )
+        db.add(row)
+        await db.flush()
+        return int(row.id)
+
+    async def _bind_skill(
+        self,
+        db: AsyncSession,
+        *,
+        user_id: int,
+        skill_tag_id: int,
+        level: str,
+        proficiency: int,
+    ) -> None:
+        db.add(
+            UserSkillBinding(
+                tenant_id=1,
+                user_id=int(user_id),
+                skill_tag_id=int(skill_tag_id),
+                proficiency_level=level,
+                proficiency=int(proficiency),
+                level=level,
+                years_experience=5.0,
+                quality_score=8.5,
+                efficiency_score=8.0,
+            )
+        )
+        await db.flush()
+
+    async def _set_workload_hours(self, db: AsyncSession, *, user_id: int, assigned_hours: float) -> None:
+        workload = (
+            await db.execute(
+                select(UserWorkload).where(
+                    and_(
+                        UserWorkload.tenant_id == 1,
+                        UserWorkload.user_id == int(user_id),
+                    )
+                )
+            )
+        ).scalar_one()
+        workload.current_assigned_hours = float(assigned_hours)
+        await workload_service.recalculate_user_workload_ratio(db, tenant_id=1, user_id=int(user_id), commit=False)
+
+    async def test_full_difficult_scenario_covers_feishu_context_project_signal_workflow_and_workload(self) -> None:
+        async with self.session_factory() as db:
+            strategy_user = await _create_user_with_defaults(
+                db,
+                tenant_id=1,
+                user_data=UserCreate(
+                    username="hanqing",
+                    password="password-1",
+                    name="韩青",
+                    email="hanqing@example.com",
+                    role="ADMIN",
+                    assigned_domains=["pharma"],
+                    project_ids=[self.base_project_id],
+                ),
+                actor_id=self.boss_id,
+                commit=False,
+            )
+            crawler_user = await _create_user_with_defaults(
+                db,
+                tenant_id=1,
+                user_data=UserCreate(
+                    username="zhouzhan",
+                    password="password-2",
+                    name="周湛",
+                    email="zhouzhan@example.com",
+                    role="BE",
+                    assigned_domains=["pharma"],
+                    project_ids=[],
+                ),
+                actor_id=self.boss_id,
+                commit=False,
+            )
+            graph_user = await _create_user_with_defaults(
+                db,
+                tenant_id=1,
+                user_data=UserCreate(
+                    username="guyan",
+                    password="password-3",
+                    name="顾言",
+                    email="guyan@example.com",
+                    role="AI",
+                    assigned_domains=["pharma"],
+                    project_ids=[],
+                ),
+                actor_id=self.boss_id,
+                commit=False,
+            )
+            await self._bind_skill(db, user_id=int(strategy_user.id), skill_tag_id=self.policy_skill_id, level="expert", proficiency=93)
+            await self._bind_skill(db, user_id=int(strategy_user.id), skill_tag_id=self.research_skill_id, level="expert", proficiency=90)
+            await self._bind_skill(db, user_id=int(crawler_user.id), skill_tag_id=self.web_skill_id, level="expert", proficiency=95)
+            await self._bind_skill(db, user_id=int(crawler_user.id), skill_tag_id=self.policy_skill_id, level="advanced", proficiency=84)
+            await self._bind_skill(db, user_id=int(crawler_user.id), skill_tag_id=self.pipeline_skill_id, level="advanced", proficiency=88)
+            await self._bind_skill(db, user_id=int(graph_user.id), skill_tag_id=self.graph_skill_id, level="expert", proficiency=91)
+            await self._bind_skill(db, user_id=int(graph_user.id), skill_tag_id=self.pipeline_skill_id, level="advanced", proficiency=82)
+            await self._set_workload_hours(db, user_id=int(strategy_user.id), assigned_hours=28.0)
+            await self._set_workload_hours(db, user_id=int(crawler_user.id), assigned_hours=8.0)
+            await self._set_workload_hours(db, user_id=int(graph_user.id), assigned_hours=10.0)
+            await db.commit()
+
+            kernel = AIPMOKernelService(db, tenant_id=1)
+            capture_service = ProactiveSignalCaptureService(db)
+
+            hard_question = (
+                "项目ID: 1 如果把样本窗口收紧到 2025Q4-2026Q1，只看江苏/浙江的 DRG-DIP 支付、"
+                "三甲院器械招采与设备更新预算，当前状态下最可能拖慢结论闭环的到底是预算、"
+                "blocked task 还是技能缺口？为什么？别给套话，按证据优先级直接判断。"
+            )
+            with (
+                patch(
+                    "app.services.ai_pmo_kernel_service.AdaptiveRAGService.answer_question",
+                    AsyncMock(
+                        return_value={
+                            "answer": "优先级 1 是 blocked task，其次是技能缺口，预算暂未触顶。",
+                            "sources": [],
+                        }
+                    ),
+                ),
+                patch(
+                    "app.services.signal_capture.context_cache_service.get_session_messages",
+                    AsyncMock(return_value=[]),
+                ),
+                patch(
+                    "app.services.signal_capture.context_cache_service.append_session_messages",
+                    AsyncMock(),
+                ) as cache_append_mock,
+                patch(
+                    "app.services.signal_capture.AdaptiveRAGService.ingest_feishu_message_async",
+                    AsyncMock(),
+                ),
+            ):
+                qa_result = await capture_service.capture_from_feishu(
+                    group_chat_id="oc_complex_case",
+                    message_content=hard_question,
+                    sender="老板",
+                    tenant_id=1,
+                    proposer_open_id="ou_boss",
+                    captured_at=datetime(2026, 3, 9, 9, 0, 0),
+                    feishu_message_id="om-qa-hard-1",
+                    chat_type="group",
+                )
+
+            async def create_project_side_effect(*, input_text: str, creator_id: int, tenant_id: int) -> dict:
+                project = Project(
+                    tenant_id=int(tenant_id),
+                    name="华东医保控费与器械国产替代观察台",
+                    title="华东医保控费与器械国产替代观察台",
+                    description="聚焦华东 DRG-DIP、器械带量采购和国产替代节奏。",
+                    owner_id=int(creator_id),
+                    status="active",
+                    domain_id=self.domain_id,
+                    assigned_role="ai_engineer",
+                    signal_count=0,
+                )
+                db.add(project)
+                await db.flush()
+
+                charter = TruthNode(
+                    tenant_id=1,
+                    node_type="charter",
+                    project_id=int(project.id),
+                    signal_id=None,
+                    created_by=int(creator_id),
+                    title="华东医保观察项目宪章",
+                    content={"success_criteria": ["10 天内交付策略 memo、政策雷达和竞品跟踪框架"]},
+                    version=1,
+                    status="active",
+                )
+                db.add(charter)
+                await db.flush()
+
+                bootstrap_signal = Signal(
+                    tenant_id=1,
+                    source="PMFlow",
+                    title="医保控费与器械国产替代立项信号",
+                    content=input_text,
+                    summary="立项信号",
+                    signal_type="新需求",
+                    priority="high",
+                    impact_score=0.85,
+                    matched_project_id=int(project.id),
+                    reasoning=json.dumps(["来自高强度情景立项指令"], ensure_ascii=False),
+                    confidence=0.92,
+                    recommended_action="enter_only",
+                    processed=True,
+                    tasks_generated=json.dumps([], ensure_ascii=False),
+                    status="已处理",
+                )
+                db.add(bootstrap_signal)
+                await db.flush()
+                db.add(
+                    ProjectSignalLink(
+                        tenant_id=1,
+                        project_id=int(project.id),
+                        signal_id=int(bootstrap_signal.id),
+                        link_type="related",
+                        relevance_score=1.0,
+                    )
+                )
+                db.add(
+                    Budget(
+                        tenant_id=1,
+                        budget_type="project",
+                        project_id=int(project.id),
+                        period="monthly",
+                        budget_amount_usd=3200.0,
+                        alert_threshold_percent=80.0,
+                        hard_limit_enabled=True,
+                        current_usage_usd=240.0,
+                        current_period_start=datetime(2026, 3, 1),
+                        current_period_end=datetime(2026, 3, 31),
+                        status="active",
+                        override_history={},
+                    )
+                )
+                db.add(UserProject(tenant_id=1, user_id=int(creator_id), project_id=int(project.id)))
+                task_one = Task(
+                    tenant_id=1,
+                    truth_node_id=int(charter.id),
+                    project_id=int(project.id),
+                    assignee_id=None,
+                    ai_recommended_assignees=None,
+                    title="搭建政策抓取与公告解析管道",
+                    description="每天跟踪江苏/浙江/安徽 医保支付改革、招采公告与价格体系拐点。",
+                    required_skill_tag_ids=[self.policy_skill_id, self.web_skill_id],
+                    status="todo",
+                    priority="high",
+                    estimated_hours=18.0,
+                    actual_hours=0.0,
+                    quality_score=None,
+                    due_date=datetime(2026, 3, 19),
+                    completed_at=None,
+                )
+                task_two = Task(
+                    tenant_id=1,
+                    truth_node_id=int(charter.id),
+                    project_id=int(project.id),
+                    assignee_id=None,
+                    ai_recommended_assignees=None,
+                    title="沉淀器械中标结构知识图谱",
+                    description="把区域政策、招采公告和竞品动作串成结构化图谱。",
+                    required_skill_tag_ids=[self.graph_skill_id],
+                    status="todo",
+                    priority="medium",
+                    estimated_hours=12.0,
+                    actual_hours=0.0,
+                    quality_score=None,
+                    due_date=datetime(2026, 3, 20),
+                    completed_at=None,
+                )
+                db.add_all([task_one, task_two])
+                await db.flush()
+                return {
+                    "project": {"id": int(project.id), "title": project.title, "name": project.name},
+                    "signal": {"id": int(bootstrap_signal.id), "title": bootstrap_signal.title},
+                    "tasks": [
+                        {"id": int(task_one.id), "title": task_one.title},
+                        {"id": int(task_two.id), "title": task_two.title},
+                    ],
+                }
+
+            with patch(
+                "app.services.ai_pmo_kernel_service.AIStewardService.create_project_from_natural_language",
+                AsyncMock(side_effect=create_project_side_effect),
+            ):
+                project_result = await kernel.process_ingress(
+                    source_system="feishu",
+                    source_message_id="om-project-hard-1",
+                    message_text=(
+                        "不要复用现在这个宏观研究盘子，单独立一个项目，名字就叫"
+                        "“华东医保控费与器械国产替代观察台”，目标不是写周报，"
+                        "而是 10 天内拿出策略 memo、政策雷达和竞品跟踪框架。"
+                    ),
+                    sender="老板",
+                    channel_ref="oc_complex_case",
+                    source_sender_open_id="ou_boss",
+                )
+
+            new_project_id = int(project_result["created_project_id"])
+
+            signal_result = await kernel.process_ingress(
+                source_system="feishu",
+                source_message_id="om-signal-hard-1",
+                message_text=(
+                    f"给项目ID: {new_project_id} 新增一条信号：持续盯江苏、浙江、安徽 2026 年器械带量采购、"
+                    "DRG-DIP 支付改革和国产替代中标结构，只要价格体系或准入口径出现拐点就预警，"
+                    "不要只抓新闻标题。"
+                ),
+                sender="老板",
+                channel_ref="oc_complex_case",
+                source_sender_open_id="ou_boss",
+                project_hint_id=new_project_id,
+            )
+
+            source_result = await kernel.process_ingress(
+                source_system="feishu",
+                source_message_id="om-source-hard-1",
+                message_text=(
+                    f"给项目ID: {new_project_id} 接一个复合信号源，不是简单 RSS：每天自动抓医保支付改革、"
+                    "器械招采公告、上市公司电话会和 GitHub 开源竞品动态，优先走自动接入。"
+                ),
+                sender="老板",
+                channel_ref="oc_complex_case",
+                source_sender_open_id="ou_boss",
+                project_hint_id=new_project_id,
+            )
+
+            db.add_all(
+                [
+                    UserProject(tenant_id=1, user_id=int(strategy_user.id), project_id=new_project_id),
+                    UserProject(tenant_id=1, user_id=int(crawler_user.id), project_id=new_project_id),
+                    UserProject(tenant_id=1, user_id=int(graph_user.id), project_id=new_project_id),
+                ]
+            )
+            await db.flush()
+
+            with patch(
+                "app.services.ai_pmo_kernel_service.SignalSourceDomainWorkflow.run",
+                AsyncMock(
+                    return_value={
+                        "status": "completed",
+                        "registry_id": 9,
+                        "registry_status": "active",
+                        "onboarding_status": "completed",
+                        "openclaw_status": "completed",
+                        "typed_actions": [{"operation_code": "task.create", "status": "applied"}],
+                        "material_action_count": 1,
+                        "worker_executions": [{"used_fallback": False}],
+                        "pending_workers": [],
+                        "next_step": "继续自动执行后续任务",
+                        "fallback_used": False,
+                        "raw_result": {},
+                    }
+                ),
+            ) as onboarding_run_mock:
+                resume_result = await kernel.process_ingress(
+                    source_system="ai_pmo_ui",
+                    source_message_id="resume-work-item-hard-1",
+                    message_text=f"继续处理 work_item #{int(source_result['work_item_id'])}",
+                    sender="老板",
+                    channel_ref="ai_pmo_ui",
+                    project_hint_id=new_project_id,
+                )
+
+            target_work_item = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.id == int(source_result["work_item_id"]),
+                        )
+                    )
+                )
+            ).scalar_one()
+
+            new_project_tasks = (
+                await db.execute(
+                    select(Task)
+                    .where(and_(Task.tenant_id == 1, Task.project_id == new_project_id))
+                    .order_by(Task.id.asc())
+                )
+            ).scalars().all()
+            task_policy_stream = next(task for task in new_project_tasks if "抓取" in task.title)
+            task_graph = next(task for task in new_project_tasks if "知识图谱" in task.title)
+
+            recommendation_result = await recommend_personnel(
+                db=db,
+                tenant_id=1,
+                required_skills=["policy-analysis", "web-crawling"],
+                task_description=task_policy_stream.description,
+                task_id=str(task_policy_stream.id),
+                limit=5,
+            )
+
+            with patch(
+                "app.services.workload_service.llm_service.generate",
+                AsyncMock(
+                    return_value=json.dumps(
+                        {
+                            "proposals": [
+                                {
+                                    "task_id": int(task_policy_stream.id),
+                                    "new_assignee_id": int(crawler_user.id),
+                                    "reason": "crawler engineer has exact skill mix and lower load",
+                                },
+                                {
+                                    "task_id": int(task_graph.id),
+                                    "new_assignee_id": int(graph_user.id),
+                                    "reason": "graph engineer owns ontology and extraction pipeline",
+                                },
+                            ]
+                        },
+                        ensure_ascii=False,
+                    )
+                ),
+            ):
+                optimize_preview = await workload_service.build_ai_optimize_preview(
+                    db,
+                    tenant_id=1,
+                    project_id=new_project_id,
+                    max_moves=5,
+                    unassigned_only=True,
+                )
+
+            optimize_result = await workload_service.apply_ai_optimize_plan(
+                db,
+                tenant_id=1,
+                proposals=optimize_preview,
+            )
+            board_rows = await workload_service.get_team_workload_board(db, tenant_id=1, project_id=new_project_id)
+            cadence_service = AIPMOCadenceService(db, tenant_id=1)
+            program_view = await cadence_service.build_program_view()
+
+        self.assertEqual(qa_result["status"], "answered")
+        self.assertIn("blocked task", qa_result["direct_reply_text"])
+        cache_append_mock.assert_awaited_once()
+        self.assertEqual(cache_append_mock.await_args.kwargs["question"], hard_question)
+
+        self.assertEqual(project_result["status"], "captured")
+        self.assertIsNotNone(project_result["created_project_id"])
+        self.assertIn("已立项", project_result["direct_reply_text"])
+
+        self.assertEqual(signal_result["status"], "captured")
+        self.assertEqual(signal_result["linked_project_ids"], [new_project_id])
+        self.assertIsNotNone(signal_result["signal_id"])
+
+        self.assertEqual(source_result["status"], "captured")
+        self.assertIn("技能缺口", source_result["direct_reply_text"])
+        self.assertFalse(bool(source_result["review_summary"]["verified"]))
+
+        self.assertEqual(onboarding_run_mock.await_count, 1)
+        self.assertEqual(resume_result["status"], "answered")
+        self.assertIn("已继续处理", resume_result["direct_reply_text"])
+        self.assertEqual(target_work_item.state, "succeeded")
+        self.assertEqual(target_work_item.execution_result["signal_source_workflow"]["status"], "completed")
+
+        self.assertGreaterEqual(recommendation_result["total_count"], 1)
+        self.assertEqual(recommendation_result["recommendations"][0]["user_name"], "周湛")
+
+        self.assertGreaterEqual(len(optimize_preview), 2)
+        self.assertEqual(optimize_result["applied_count"], 2)
+
+        task_assignee_map = {int(task.id): int(task.assignee_id or 0) for task in new_project_tasks}
+        self.assertEqual(task_assignee_map[int(task_policy_stream.id)], int(crawler_user.id))
+        self.assertEqual(task_assignee_map[int(task_graph.id)], int(graph_user.id))
+
+        board_by_name = {str(row["user_name"]): row for row in board_rows}
+        self.assertIn("周湛", board_by_name)
+        self.assertIn("顾言", board_by_name)
+        self.assertGreater(float(board_by_name["周湛"]["current_assigned_hours"]), 8.0)
+        self.assertGreaterEqual(float(board_by_name["顾言"]["current_assigned_hours"]), 9.5)
+
+        self.assertGreaterEqual(program_view["summary"]["program_count"], 1)
+        self.assertTrue(
+            any(new_project_id in program["project_ids"] for program in program_view["programs"])
+        )
diff --git a/backend/tests/test_ai_pmo_kernel_service.py b/backend/tests/test_ai_pmo_kernel_service.py
new file mode 100644
index 00000000..60363902
--- /dev/null
+++ b/backend/tests/test_ai_pmo_kernel_service.py
@@ -0,0 +1,590 @@
+import os
+import unittest
+from datetime import datetime
+from tempfile import NamedTemporaryFile
+from unittest.mock import AsyncMock, patch
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
+
+from app.models import Base, Budget, PortfolioSnapshot, Project, ProjectDomain, Task, TruthNode, User, UserProject, UserSkillBinding, UserWorkload, WorkItem
+from app.services.ai_pmo_cadence_service import AIPMOCadenceService
+from app.services.ai_pmo_kernel_service import (
+    AIPMOKernelService,
+    INTENT_SIGNAL_SOURCE_ONBOARD,
+    IntentResolver,
+)
+from app.services.ai_pmo_domain_workflow import SignalSourceDomainWorkflow
+from app.services.signal_source_orchestration_service import SignalSourceOrchestrationService
+
+
+class AIPMOKernelServiceTest(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self) -> None:
+        tmp = NamedTemporaryFile(prefix="pmflow-ai-pmo-test-", suffix=".db", delete=False)
+        tmp.close()
+        self.db_path = tmp.name
+        self.engine = create_async_engine(f"sqlite+aiosqlite:///{self.db_path}", future=True)
+        self.session_factory = async_sessionmaker(self.engine, expire_on_commit=False, class_=AsyncSession)
+        async with self.engine.begin() as conn:
+            await conn.run_sync(Base.metadata.create_all)
+
+        async with self.session_factory() as db:
+            user = User(
+                tenant_id=1,
+                username="boss",
+                email="boss@example.com",
+                password_hash="hash",
+                name="老板",
+                role="BOSS",
+                assigned_domains=[],
+                workload=0,
+                status="active",
+                is_active=True,
+            )
+            db.add(user)
+            await db.flush()
+
+            domain = ProjectDomain(
+                tenant_id=1,
+                name="医药",
+                slug="pharma",
+                description="医药研究域",
+                config={},
+                is_active=True,
+                is_default=False,
+                signal_count=0,
+                context_node_count=0,
+                experiment_count=0,
+                owner_id=str(user.id),
+            )
+            db.add(domain)
+            await db.flush()
+
+            project = Project(
+                tenant_id=1,
+                name="医药行业十五五规划布局研究",
+                title="医药行业十五五规划布局研究",
+                description="研究医药行业方向和机会。",
+                owner_id=int(user.id),
+                status="active",
+                domain_id=int(domain.id),
+                assigned_role="pm",
+                signal_count=0,
+            )
+            db.add(project)
+            await db.flush()
+
+            charter = TruthNode(
+                tenant_id=1,
+                node_type="charter",
+                project_id=int(project.id),
+                signal_id=None,
+                created_by=int(user.id),
+                title="项目宪章",
+                content={"success_criteria": ["两周内形成研究报告"]},
+                version=1,
+                status="active",
+            )
+            db.add(charter)
+            await db.flush()
+            db.add(
+                Budget(
+                    tenant_id=1,
+                    budget_type="project",
+                    project_id=int(project.id),
+                    period="monthly",
+                    budget_amount_usd=1000.0,
+                    alert_threshold_percent=80.0,
+                    hard_limit_enabled=True,
+                    current_usage_usd=500.0,
+                    current_period_start=datetime(2026, 3, 1),
+                    current_period_end=datetime(2026, 3, 31),
+                    status="active",
+                    override_history={},
+                )
+            )
+            db.add(UserProject(tenant_id=1, user_id=int(user.id), project_id=int(project.id)))
+            db.add(
+                UserWorkload(
+                    tenant_id=1,
+                    user_id=int(user.id),
+                    current_assigned_hours=20.0,
+                    max_available_hours_per_week=40.0,
+                    workload_ratio=0.5,
+                )
+            )
+            db.add(
+                UserSkillBinding(
+                    tenant_id=1,
+                    user_id=int(user.id),
+                    skill_tag_id=1,
+                    proficiency_level="expert",
+                    proficiency=5,
+                    level="expert",
+                    years_experience=8.0,
+                    quality_score=9.0,
+                    efficiency_score=8.5,
+                )
+            )
+            db.add(
+                Task(
+                    tenant_id=1,
+                    truth_node_id=int(charter.id),
+                    project_id=int(project.id),
+                    assignee_id=int(user.id),
+                    ai_recommended_assignees=None,
+                    title="梳理外部依赖",
+                    description="当前依赖较多，任务处于 blocked。",
+                    required_skill_tag_ids=[2],
+                    status="blocked",
+                    priority="high",
+                    estimated_hours=8.0,
+                    actual_hours=0.0,
+                    quality_score=None,
+                    due_date=datetime(2026, 3, 1),
+                    completed_at=None,
+                )
+            )
+            await db.commit()
+
+    async def asyncTearDown(self) -> None:
+        await self.engine.dispose()
+        if getattr(self, "db_path", None):
+            try:
+                os.remove(self.db_path)
+            except FileNotFoundError:
+                pass
+
+    def test_intent_resolver_detects_signal_source_request(self) -> None:
+        intent = IntentResolver.resolve(
+            text="给项目ID: 1 新增一个 GitHub 信号源，监控 openai/codex 仓库。",
+            project_id=1,
+        )
+        self.assertEqual(intent.name, INTENT_SIGNAL_SOURCE_ONBOARD)
+
+    async def test_build_world_model_includes_budget_capacity_and_skill_graph(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            world_model = await service._build_world_model(
+                message_text="给项目ID: 1 新增一个医药信号。",
+                sender="老板",
+                project_hint_id=1,
+            )
+
+        self.assertEqual(world_model["budget"]["budget_id"], 1)
+        self.assertEqual(len(world_model["capacity"]), 1)
+        self.assertEqual(len(world_model["skill_graph"]), 1)
+        self.assertEqual(world_model["resolved_project_id"], 1)
+        self.assertEqual(world_model["project_domain"]["slug"], "pharma")
+        self.assertEqual(world_model["dependencies"]["pressure"], "high")
+        self.assertEqual(world_model["task_summary"]["uncovered_skill_tag_ids"], [2])
+
+    async def test_process_query_message_persists_work_item_and_reply_receipt(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            with patch(
+                "app.services.ai_pmo_kernel_service.AdaptiveRAGService.answer_question",
+                AsyncMock(return_value={"answer": "当前项目状态正常。", "sources": []}),
+            ):
+                result = await service.process_ingress(
+                    source_system="pmflow_manual",
+                    source_message_id="msg-query-1",
+                    message_text="项目ID: 1 现在状态如何？",
+                    sender="老板",
+                    channel_ref="pmflow",
+                    project_hint_id=1,
+                )
+
+            work_item = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.source_message_id == "msg-query-1",
+                        )
+                    )
+                )
+            ).scalar_one()
+            review_cycle = (
+                await db.execute(
+                    select(PortfolioSnapshot).where(
+                        and_(
+                            PortfolioSnapshot.tenant_id == 1,
+                            PortfolioSnapshot.scope_type == "review_cycle",
+                            PortfolioSnapshot.scope_ref == f"work_item:{int(work_item.id)}",
+                        )
+                    )
+                )
+            ).scalar_one()
+
+        self.assertTrue(result["handled"])
+        self.assertEqual(result["status"], "answered")
+        self.assertTrue(result["direct_reply_required"])
+        self.assertEqual(result["direct_reply_text"], "当前项目状态正常。")
+        self.assertEqual(result["decision_level"], "project")
+        self.assertIn("agent_quality_score", work_item.review_summary)
+        self.assertTrue(bool(work_item.review_summary.get("verified")))
+        self.assertEqual(review_cycle.summary["work_item_id"], int(work_item.id))
+
+    async def test_duplicate_message_is_filtered(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            with patch(
+                "app.services.ai_pmo_kernel_service.AdaptiveRAGService.answer_question",
+                AsyncMock(return_value={"answer": "OK", "sources": []}),
+            ):
+                first = await service.process_ingress(
+                    source_system="pmflow_manual",
+                    source_message_id="dup-1",
+                    message_text="项目ID: 1 现在状态如何？",
+                    sender="老板",
+                    channel_ref="pmflow",
+                    project_hint_id=1,
+                )
+                second = await service.process_ingress(
+                    source_system="pmflow_manual",
+                    source_message_id="dup-1",
+                    message_text="项目ID: 1 现在状态如何？",
+                    sender="老板",
+                    channel_ref="pmflow",
+                    project_hint_id=1,
+                )
+
+        self.assertEqual(first["status"], "answered")
+        self.assertEqual(second["status"], "ignored")
+        self.assertFalse(second["direct_reply_required"])
+
+    async def test_policy_gate_blocks_archive_like_request(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            result = await service.process_ingress(
+                source_system="pmflow_manual",
+                source_message_id="gate-1",
+                message_text="请把项目ID: 1 直接归档。",
+                sender="老板",
+                channel_ref="pmflow",
+                project_hint_id=1,
+            )
+
+        self.assertTrue(result["handled"])
+        self.assertEqual(result["status"], "captured")
+        self.assertIn("人工确认", result["direct_reply_text"])
+
+    async def test_process_project_create_uses_new_kernel(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            with patch(
+                "app.services.ai_pmo_kernel_service.AIStewardService.create_project_from_natural_language",
+                AsyncMock(
+                    return_value={
+                        "project": {"id": 9, "title": "医药新项目", "name": "医药新项目"},
+                        "signal": {"id": 11},
+                        "tasks": [{"id": 21, "title": "任务A"}],
+                    }
+                ),
+            ):
+                result = await service.process_ingress(
+                    source_system="pmflow_manual",
+                    source_message_id="project-1",
+                    message_text="请创建项目：医药新项目，做医药趋势研究。",
+                    sender="老板",
+                    channel_ref="pmflow",
+                )
+            work_item = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.source_message_id == "project-1",
+                        )
+                    )
+                )
+            ).scalar_one()
+
+        self.assertEqual(result["status"], "captured")
+        self.assertEqual(result["created_project_id"], 9)
+        self.assertEqual(result["linked_project_ids"], [9])
+        self.assertTrue(result["direct_reply_required"])
+        self.assertEqual(work_item.project_id, 9)
+
+    async def test_signal_source_request_hits_simulation_policy_gate_when_skill_gap_exists(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            result = await service.process_ingress(
+                source_system="pmflow_manual",
+                source_message_id="source-1",
+                message_text="给项目ID: 1 新增一个 GitHub 信号源，监控 openai/codex。",
+                sender="老板",
+                channel_ref="pmflow",
+                project_hint_id=1,
+            )
+
+        self.assertTrue(result["handled"])
+        self.assertEqual(result["status"], "captured")
+        self.assertIn("技能缺口", result["direct_reply_text"])
+
+    async def test_signal_create_stays_auto_executable_under_dependency_pressure(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            result = await service.process_ingress(
+                source_system="pmflow_manual",
+                source_message_id="signal-auto-1",
+                message_text="给项目ID: 1 新增一条信号：跟踪江苏医保支付改革和器械招采拐点。",
+                sender="老板",
+                channel_ref="pmflow",
+                project_hint_id=1,
+            )
+
+        self.assertTrue(result["handled"])
+        self.assertEqual(result["status"], "captured")
+        self.assertIsNotNone(result["signal_id"])
+        self.assertEqual(result["linked_project_ids"], [1])
+        self.assertIn("已入库", result["direct_reply_text"])
+
+    async def test_final_phase_generates_risk_cadence_and_program_view(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            result = await service.process_ingress(
+                source_system="pmflow_manual",
+                source_message_id="source-risk-1",
+                message_text="给项目ID: 1 新增一个 GitHub 信号源，监控 openai/codex。",
+                sender="老板",
+                channel_ref="pmflow",
+                project_hint_id=1,
+            )
+            cadence_service = AIPMOCadenceService(db, tenant_id=1)
+            latest_risk = await cadence_service.get_latest_snapshot(
+                cadence_type="risk_review",
+                project_id=1,
+            )
+            program_view = await cadence_service.build_program_view()
+
+        self.assertTrue(result["handled"])
+        self.assertEqual(result["status"], "captured")
+        self.assertIsNotNone(latest_risk)
+        self.assertEqual(latest_risk["scope_type"], "risk_review")
+        self.assertGreaterEqual(int(latest_risk["summary"]["blocked_task_count"]), 1)
+        self.assertEqual(program_view["summary"]["program_count"], 1)
+        self.assertGreaterEqual(int(program_view["summary"]["pending_manual_total"]), 1)
+
+    async def test_followup_continue_resumes_single_waiting_manual_work_item(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            first = await service.process_ingress(
+                source_system="feishu",
+                source_message_id="feishu-source-1",
+                message_text="给项目ID: 1 新增一个 GitHub 信号源，监控 openai/codex。",
+                sender="老板",
+                channel_ref="oc_phase3",
+                source_sender_open_id="ou_boss",
+                project_hint_id=1,
+            )
+            with patch(
+                "app.services.ai_pmo_kernel_service.SignalSourceDomainWorkflow.run",
+                AsyncMock(
+                    return_value={
+                        "status": "completed",
+                        "registry_id": 7,
+                        "typed_actions": [{"operation_code": "task.create", "status": "applied"}],
+                        "material_action_count": 1,
+                        "fallback_used": False,
+                        "raw_result": {},
+                    }
+                ),
+            ):
+                followup = await service.process_ingress(
+                    source_system="feishu",
+                    source_message_id="feishu-followup-1",
+                    message_text="继续处理",
+                    sender="老板",
+                    channel_ref="oc_phase3",
+                    source_sender_open_id="ou_boss",
+                    project_hint_id=1,
+                )
+            original = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.source_message_id == "feishu-source-1",
+                        )
+                    )
+                )
+            ).scalar_one()
+
+        self.assertEqual(first["status"], "captured")
+        self.assertEqual(followup["status"], "answered")
+        self.assertIn("已继续处理", followup["direct_reply_text"])
+        self.assertEqual(original.state, "succeeded")
+
+    async def test_followup_continue_requires_explicit_target_when_multiple_waiting_manual_items(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            await service.process_ingress(
+                source_system="feishu",
+                source_message_id="feishu-source-2a",
+                message_text="给项目ID: 1 新增一个 GitHub 信号源，监控 openai/codex。",
+                sender="老板",
+                channel_ref="oc_phase3_multi",
+                source_sender_open_id="ou_boss",
+                project_hint_id=1,
+            )
+            await service.process_ingress(
+                source_system="feishu",
+                source_message_id="feishu-source-2b",
+                message_text="给项目ID: 1 新增一个 Web 搜索信号源，监控医保动态。",
+                sender="老板",
+                channel_ref="oc_phase3_multi",
+                source_sender_open_id="ou_boss",
+                project_hint_id=1,
+            )
+            followup = await service.process_ingress(
+                source_system="feishu",
+                source_message_id="feishu-followup-2",
+                message_text="继续处理",
+                sender="老板",
+                channel_ref="oc_phase3_multi",
+                source_sender_open_id="ou_boss",
+                project_hint_id=1,
+            )
+            followup_item = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.source_message_id == "feishu-followup-2",
+                        )
+                    )
+                )
+            ).scalar_one()
+
+        self.assertEqual(followup["status"], "captured")
+        self.assertIn("多个待确认工作项", followup["direct_reply_text"])
+        self.assertEqual(followup_item.state, "waiting_manual")
+
+    async def test_followup_explicit_work_item_id_can_resume_across_source_system(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            first = await service.process_ingress(
+                source_system="feishu",
+                source_message_id="feishu-source-3",
+                message_text="给项目ID: 1 新增一个 GitHub 信号源，监控 openai/codex。",
+                sender="老板",
+                channel_ref="oc_phase3_explicit",
+                source_sender_open_id="ou_boss",
+                project_hint_id=1,
+            )
+            target_id = int(first["work_item_id"])
+            with patch(
+                "app.services.ai_pmo_kernel_service.SignalSourceDomainWorkflow.run",
+                AsyncMock(
+                    return_value={
+                        "status": "completed",
+                        "registry_id": 9,
+                        "typed_actions": [{"operation_code": "task.create", "status": "applied"}],
+                        "material_action_count": 1,
+                        "fallback_used": False,
+                        "raw_result": {},
+                    }
+                ),
+            ):
+                followup = await service.process_ingress(
+                    source_system="ai_pmo_ui",
+                    source_message_id="resume-ui-1",
+                    message_text=f"继续处理 work_item #{target_id}",
+                    sender="老板",
+                    channel_ref="ai_pmo_ui",
+                    project_hint_id=1,
+                )
+
+        self.assertEqual(followup["status"], "answered")
+        self.assertIn("已继续处理", followup["direct_reply_text"])
+
+    def test_signal_source_domain_workflow_keeps_notification_only_pending_manual(self) -> None:
+        normalized = SignalSourceDomainWorkflow.normalize(
+            {
+                "status": "partial",
+                "signal_source": {
+                    "registry_id": 7,
+                    "status": "pending",
+                    "onboarding_status": "pending_manual",
+                },
+                "openclaw": {
+                    "status": "completed",
+                    "result": {
+                        "execution": {
+                            "results": [
+                                {
+                                    "operation_code": "notification.create",
+                                    "status": "applied",
+                                    "payload": {"title": "待确认"},
+                                }
+                            ]
+                        }
+                    },
+                },
+                "worker_executions": [
+                    {
+                        "worker_role_key": "pm_dispatcher",
+                        "status": "completed",
+                        "used_fallback": True,
+                    }
+                ],
+                "pending_workers": [],
+            }
+        )
+
+        self.assertEqual(normalized["status"], "pending_manual")
+        self.assertEqual(normalized["material_action_count"], 0)
+
+    async def test_transport_mode_statement_is_answered_without_creating_work_item(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            result = await service.process_ingress(
+                source_system="pmflow_manual",
+                source_message_id="transport-mode-1",
+                message_text="现在都是以长连接的形式，并不是webhook。",
+                sender="老板",
+                channel_ref="pmflow",
+                project_hint_id=1,
+            )
+            persisted = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.source_message_id == "transport-mode-1",
+                        )
+                    )
+                )
+            ).scalar_one_or_none()
+
+        self.assertTrue(result["handled"])
+        self.assertEqual(result["status"], "answered")
+        self.assertTrue(result["direct_reply_required"])
+        self.assertIn("长连接模式", result["direct_reply_text"])
+        self.assertIsNone(persisted)
+
+    async def test_signal_source_registry_write_re_normalizes_invalid_kind_and_generic_name(self) -> None:
+        async with self.session_factory() as db:
+            service = SignalSourceOrchestrationService(db)
+            result = await service._create_or_reuse_registry(
+                tenant_id=1,
+                project_id=1,
+                signal_id=None,
+                operator_user_id=None,
+                requested_via="feishu",
+                source_payload={
+                    "source_kind": "long",
+                    "source_name": "给项目ID",
+                    "source_identifier": "long:给项目id",
+                    "description": "每日搜寻最新的医药行业变革动态",
+                    "config": {"keywords": ["医药行业", "变革"]},
+                },
+                initiating_text="给项目ID: 1 接一个复合信号源：每日搜寻最新的医药行业变革动态。",
+                dry_run=False,
+            )
+
+        self.assertEqual(result["source_kind"], "feishu")
+        self.assertNotEqual(result["source_name"], "给项目ID")
+        self.assertTrue(str(result["source_identifier"]).startswith("feishu:"))
diff --git a/backend/tests/test_ai_steward_fail_fast.py b/backend/tests/test_ai_steward_fail_fast.py
new file mode 100644
index 00000000..b0852ccd
--- /dev/null
+++ b/backend/tests/test_ai_steward_fail_fast.py
@@ -0,0 +1,107 @@
+import asyncio
+import unittest
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+from app.services.ai_steward import AIStewardService
+from app.models import CognitiveConflict
+
+
+class AIStewardFailFastTest(unittest.IsolatedAsyncioTestCase):
+    async def test_analyze_project_input_times_out_to_structured_fallback(self) -> None:
+        service = AIStewardService(db=None)  # type: ignore[arg-type]
+
+        with patch.object(service.llm, "generate", AsyncMock(side_effect=asyncio.TimeoutError())):
+            result = await service._analyze_project_input(
+                "创建一个 3 周上线的 AI 客服项目，先出可运行 MVP。",
+                tenant_id=1,
+            )
+
+        self.assertEqual(result["primary_role"], "ai_engineer")
+        self.assertEqual(result["timeline"], "3周")
+
+    async def test_generate_task_chain_draft_times_out_to_default_tasks(self) -> None:
+        service = AIStewardService(db=None)  # type: ignore[arg-type]
+
+        with patch.object(service.llm, "generate", AsyncMock(side_effect=asyncio.TimeoutError())):
+            tasks = await service._generate_task_chain_draft(
+                analysis={
+                    "project_title": "AI 客服项目",
+                    "tech_stack": ["FastAPI", "React"],
+                    "timeline": "3周",
+                    "primary_role": "ai_engineer",
+                },
+                tenant_id=1,
+            )
+
+        self.assertGreaterEqual(len(tasks), 2)
+        self.assertEqual(tasks[0]["title"], "技术方案设计")
+
+    async def test_persist_project_creation_draft_writes_conflict_user_ids_as_list(self) -> None:
+        added_rows = []
+
+        class FakeDB:
+            def add(self, row):
+                added_rows.append(row)
+
+            async def flush(self):
+                current = len(added_rows)
+                for index, row in enumerate(added_rows, start=1):
+                    if getattr(row, "id", None) is None:
+                        row.id = index
+
+        service = AIStewardService(db=FakeDB())  # type: ignore[arg-type]
+
+        with patch(
+            "app.services.ai_steward.persist_task",
+            AsyncMock(
+                return_value=SimpleNamespace(
+                    id=101,
+                    title="任务A",
+                    description="desc",
+                    estimated_hours=8.0,
+                    priority="high",
+                    status="todo",
+                )
+            ),
+        ):
+            payload = await service._persist_project_creation_draft(
+                normalized={
+                    "project": {
+                        "name": "测试项目",
+                        "title": "测试项目",
+                        "description": "desc",
+                        "assigned_role": "ai_engineer",
+                        "domain_id": 1,
+                    },
+                    "signal": {
+                        "content": "signal",
+                        "impact_score": 0.8,
+                        "priority": "high",
+                        "strategic_analysis": {},
+                    },
+                    "charter": {
+                        "title": "项目宪章",
+                        "content": {"success_criteria": ["ok"]},
+                    },
+                    "tasks": [SimpleNamespace(title="任务A", description="desc")],
+                    "task_audit": [],
+                    "risks": [
+                        {
+                            "type": "schedule_delay",
+                            "description": "任务超时预警",
+                            "threshold_hours": 48.0,
+                            "severity": "medium",
+                        }
+                    ],
+                    "analysis": {},
+                },
+                input_text="创建测试项目",
+                creator_id=2,
+                tenant_id=1,
+            )
+
+        risk_rows = [row for row in added_rows if isinstance(row, CognitiveConflict)]
+        self.assertEqual(len(risk_rows), 1)
+        self.assertEqual(risk_rows[0].involved_user_ids, [])
+        self.assertEqual(payload["risks"][0]["description"], "任务超时预警")
diff --git a/backend/tests/test_api_router_order.py b/backend/tests/test_api_router_order.py
new file mode 100644
index 00000000..575c228e
--- /dev/null
+++ b/backend/tests/test_api_router_order.py
@@ -0,0 +1,22 @@
+from app.api.router import api_router
+
+
+def test_project_signal_route_uses_single_sstc_canonical_endpoint() -> None:
+    canonical = []
+    legacy = []
+    for route in api_router.routes:
+        path = getattr(route, "path", "")
+        methods = set(getattr(route, "methods", set()) or set())
+        if path == "/projects/{project_id}/signals" and "GET" in methods:
+            endpoint = getattr(route, "endpoint", None)
+            canonical.append(
+                (
+                    getattr(endpoint, "__module__", ""),
+                    getattr(endpoint, "__name__", ""),
+                )
+            )
+        if path == "/projects/{project_id}/linked-signals" and "GET" in methods:
+            legacy.append(path)
+
+    assert canonical == [("app.api.sstc", "list_signals")]
+    assert legacy == []
diff --git a/backend/tests/test_auth_password_validation.py b/backend/tests/test_auth_password_validation.py
new file mode 100644
index 00000000..2b7747cc
--- /dev/null
+++ b/backend/tests/test_auth_password_validation.py
@@ -0,0 +1,40 @@
+from app.services.auth import AuthService
+from app.services import auth as auth_module
+
+
+def test_validate_password_rejects_over_72_utf8_bytes_with_friendly_message() -> None:
+    password = "a" * 73
+
+    try:
+        AuthService.hash_password(password)
+        raise AssertionError("expected ValueError for overlong password")
+    except ValueError as exc:
+        message = str(exc)
+
+    assert "72 个 UTF-8 字节" in message
+    assert "72 bytes" not in message.lower()
+
+
+def test_validate_password_counts_utf8_bytes_not_only_character_length() -> None:
+    password = "密" * 25
+
+    try:
+        AuthService.validate_password(password)
+        raise AssertionError("expected ValueError for multibyte password overflow")
+    except ValueError as exc:
+        message = str(exc)
+
+    assert "72 个 UTF-8 字节" in message
+
+
+def test_hash_password_falls_back_when_bcrypt_backend_misreports_72_bytes(monkeypatch) -> None:
+    class _BrokenPwdContext:
+        @staticmethod
+        def hash(_password: str) -> str:
+            raise ValueError("password cannot be longer than 72 bytes, truncate manually if necessary")
+
+    monkeypatch.setattr(auth_module, "HAS_PASSLIB", True)
+    monkeypatch.setattr(auth_module, "pwd_context", _BrokenPwdContext(), raising=False)
+    hashed = AuthService.hash_password("Passw0rd!")
+
+    assert hashed == AuthService._simple_hash("Passw0rd!")
diff --git a/backend/tests/test_feishu_long_connection.py b/backend/tests/test_feishu_long_connection.py
new file mode 100644
index 00000000..1f80f2ec
--- /dev/null
+++ b/backend/tests/test_feishu_long_connection.py
@@ -0,0 +1,131 @@
+import unittest
+from unittest.mock import patch
+
+from app.services.feishu_long_connection import FeishuLongConnectionManager
+
+
+class _AliveThread:
+    def __init__(self) -> None:
+        self._alive = True
+
+    def is_alive(self) -> bool:
+        return self._alive
+
+    def join(self, timeout: float | None = None) -> None:
+        _ = timeout
+        self._alive = False
+
+
+class _StubFeishuLongConnectionManager(FeishuLongConnectionManager):
+    def __init__(self, config: dict | None, preflight_exc: Exception | None = None) -> None:
+        super().__init__(session_factory=None)  # type: ignore[arg-type]
+        self._sdk_available = True
+        self._stub_config = config
+        self._preflight_exc = preflight_exc
+        self.start_calls: list[tuple[str, str]] = []
+
+    async def _load_active_config(self) -> dict | None:
+        if self._stub_config is None:
+            return None
+        return dict(self._stub_config)
+
+    async def _preflight_auth(self, *, tenant_id: int) -> None:
+        _ = tenant_id
+        if self._preflight_exc is not None:
+            raise self._preflight_exc
+
+    def _start_ws_thread(self, app_id: str, app_secret: str) -> None:
+        self.start_calls.append((app_id, app_secret))
+        self._thread = _AliveThread()
+
+
+class FeishuLongConnectionManagerTest(unittest.IsolatedAsyncioTestCase):
+    async def asyncTearDown(self) -> None:
+        manager = getattr(self, "manager", None)
+        if manager is not None and manager.status().get("running"):
+            await manager.stop()
+
+    async def test_auth_failure_is_classified_without_starting_thread(self) -> None:
+        self.manager = _StubFeishuLongConnectionManager(
+            config={
+                "tenant_id": 9,
+                "event_mode": "long_connection",
+                "app_id": "cli_xxx",
+                "app_secret": "secret_xxx",
+            },
+            preflight_exc=ValueError("获取 tenant_access_token 失败(http=401): {'code': 99991663}"),
+        )
+
+        await self.manager.start()
+
+        status = self.manager.status()
+        self.assertEqual(status["tenant_id"], 9)
+        self.assertEqual(status["last_error_code"], "auth_failed")
+        self.assertFalse(status["last_error_retryable"])
+        self.assertFalse(status["thread_alive"])
+        self.assertEqual(self.manager.start_calls, [])
+
+    async def test_network_failure_is_marked_retryable(self) -> None:
+        self.manager = _StubFeishuLongConnectionManager(
+            config={
+                "tenant_id": 3,
+                "event_mode": "long_connection",
+                "app_id": "cli_xxx",
+                "app_secret": "secret_xxx",
+            },
+            preflight_exc=ValueError("访问飞书 auth 接口失败: timed out"),
+        )
+
+        await self.manager.start()
+
+        status = self.manager.status()
+        self.assertEqual(status["last_error_code"], "network_error")
+        self.assertTrue(status["last_error_retryable"])
+        self.assertEqual(status["restart_attempts"], 1)
+        self.assertIsNotNone(status["next_retry_at"])
+        self.assertEqual(self.manager.start_calls, [])
+
+    async def test_successful_start_records_tenant_and_auth_timestamp(self) -> None:
+        self.manager = _StubFeishuLongConnectionManager(
+            config={
+                "tenant_id": 12,
+                "event_mode": "long_connection",
+                "app_id": "cli_xxx",
+                "app_secret": "secret_xxx",
+            }
+        )
+
+        await self.manager.start()
+
+        status = self.manager.status()
+        self.assertEqual(status["tenant_id"], 12)
+        self.assertTrue(status["thread_alive"])
+        self.assertIsNotNone(status["last_auth_at"])
+        self.assertEqual(self.manager.start_calls, [("cli_xxx", "secret_xxx")])
+
+    async def test_retry_strategy_is_1_2_4_and_stops_after_three_attempts(self) -> None:
+        self.manager = _StubFeishuLongConnectionManager(
+            config={
+                "tenant_id": 7,
+                "event_mode": "long_connection",
+                "app_id": "cli_xxx",
+                "app_secret": "secret_xxx",
+            }
+        )
+        self.manager._set_error("network", code="network_error", retryable=True)
+
+        with patch("app.services.feishu_long_connection.time.time", return_value=100.0):
+            self.manager._schedule_retry()
+            self.assertEqual(self.manager._restart_attempts, 1)
+            self.assertEqual(self.manager._next_retry_at, 101.0)
+            self.manager._schedule_retry()
+            self.assertEqual(self.manager._restart_attempts, 2)
+            self.assertEqual(self.manager._next_retry_at, 102.0)
+            self.manager._schedule_retry()
+            self.assertEqual(self.manager._restart_attempts, 3)
+            self.assertEqual(self.manager._next_retry_at, 104.0)
+            # 第 4 次触发应熔断，不再继续重试。
+            self.manager._schedule_retry()
+            self.assertEqual(self.manager._restart_attempts, 3)
+            self.assertEqual(self.manager._next_retry_at, 0.0)
+            self.assertFalse(self.manager._last_error_retryable)
diff --git a/backend/tests/test_feishu_routing_receipts.py b/backend/tests/test_feishu_routing_receipts.py
new file mode 100644
index 00000000..9babf0b6
--- /dev/null
+++ b/backend/tests/test_feishu_routing_receipts.py
@@ -0,0 +1,177 @@
+import unittest
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+from app.services.feishu_integration import FeishuIntegrationService
+from app.services.feishu_signal_pipeline import FeishuSignalPipelineService
+from app.services.signal_capture import ProactiveSignalCaptureService
+from app.services.context_cache import context_cache_service
+
+
+class FeishuRoutingReceiptsTest(unittest.TestCase):
+    def test_normalize_incoming_text_strips_bot_mentions_and_quote_noise(self) -> None:
+        text = FeishuSignalPipelineService._normalize_incoming_text(
+            "@PMFlow\n回复：\n> 上一条消息\n[附带图片]\n麻烦看下统一通知入口"
+        )
+
+        self.assertEqual(text, "麻烦看下统一通知入口")
+
+    def test_pending_receipt_prefers_clarification_message(self) -> None:
+        reply = FeishuIntegrationService._build_execution_receipt_reply(
+            message_content="@PMFlow 继续处理",
+            capture_result={
+                "status": "captured",
+                "results": [{"route_status": "queued_for_decision"}],
+                "clarification_message": "当前群最近同时涉及项目#12、#18，请明确归属项目。",
+            },
+        )
+
+        self.assertEqual(reply, "当前群最近同时涉及项目#12、#18，请明确归属项目。")
+
+    def test_failed_receipt_uses_controlled_executor_message(self) -> None:
+        reply = FeishuIntegrationService._build_execution_receipt_reply(
+            message_content="@PMFlow 创建任务",
+            capture_result={
+                "status": "captured",
+                "results": [{"route_status": "failed"}],
+            },
+        )
+
+        self.assertEqual(reply, "未执行：受控执行器拒绝了这次动作，请补充缺失字段或检查项目上下文。")
+
+    def test_followup_confirmation_reply_uses_real_db_ids(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+
+        reply = service._build_followup_confirmation_reply(
+            target_item=SimpleNamespace(id=23, level="signal"),
+            confirm_result={"signal_id": "91", "project_id": "12"},
+        )
+
+        self.assertEqual(reply, "已继续处理：信号#91 已确认并关联项目#12。")
+
+    def test_provider_error_guard_blocks_aborted_error_text(self) -> None:
+        self.assertTrue(FeishuIntegrationService._looks_like_provider_error("Request was aborted."))
+
+    def test_normalize_outgoing_reply_text_replaces_provider_error(self) -> None:
+        self.assertEqual(
+            FeishuIntegrationService._normalize_outgoing_reply_text("Request was aborted."),
+            "系统处理中出现异常，本次未执行。请稍后重试，或补充更明确的目标、对象和动作。",
+        )
+
+    def test_pipeline_decorate_capture_result_sanitizes_provider_error_reply(self) -> None:
+        result = FeishuSignalPipelineService._decorate_capture_result(
+            {
+                "status": "answered",
+                "classification": "question",
+                "direct_reply_text": "Request was aborted.",
+            },
+            chat_type="group",
+            normalized_text="@PMFlow 现在状态如何",
+            route="group.qa_direct",
+            db_mutations=[],
+        )
+
+        self.assertEqual(
+            result["direct_reply_text"],
+            "系统处理中出现异常，本次未执行。请稍后重试，或补充更明确的目标、对象和动作。",
+        )
+
+    def test_handle_message_receive_skips_reply_for_duplicate_event(self) -> None:
+        service = FeishuIntegrationService(db=None)  # type: ignore[arg-type]
+        event = {
+            "message": {
+                "message_id": "om_duplicate",
+                "message_type": "text",
+                "chat_id": "oc_test",
+                "chat_type": "group",
+                "content": '{"text":"@PMFlow 帮我继续处理"}',
+            },
+            "chat": {
+                "chat_id": "oc_test",
+                "chat_type": "group",
+            },
+            "sender": {
+                "sender_type": "user",
+                "sender_id": {"open_id": "ou_test"},
+                "sender_name": "林宇航",
+            },
+            "type": "im.message.receive_v1",
+        }
+
+        with (
+            patch.object(service, "_record_ingress_trace", AsyncMock()),
+            patch.object(service, "_record_message_trace", AsyncMock()),
+            patch.object(service, "_resolve_sender_display_name", AsyncMock(return_value="林宇航")),
+            patch.object(service, "_safe_reply_text", AsyncMock()) as reply_mock,
+            patch(
+                "app.services.feishu_integration.ProactiveSignalCaptureService.capture_from_feishu",
+                AsyncMock(
+                    return_value={
+                        "status": "ignored",
+                        "reason": "duplicate_filtered",
+                        "classification": "duplicate",
+                        "direct_reply_required": True,
+                        "direct_reply_text": "Request was aborted.",
+                    }
+                ),
+            ),
+        ):
+            asyncio.run(service._handle_message_receive(event))
+
+        reply_mock.assert_not_awaited()
+
+    def test_sync_feishu_recent_cache_backfills_missing_rag_turn(self) -> None:
+        service = ProactiveSignalCaptureService(db=None)  # type: ignore[arg-type]
+
+        with (
+            patch.object(context_cache_service, "get_session_messages", AsyncMock(return_value=[])),
+            patch.object(context_cache_service, "append_session_messages", AsyncMock()) as append_mock,
+        ):
+            asyncio.run(
+                service._sync_feishu_recent_cache(
+                    tenant_id=1,
+                    group_chat_id="oc_test",
+                    sender="林宇航",
+                    sender_open_id="ou_test",
+                    message="目前医药方向值得关注什么？",
+                    pipeline_result={
+                        "rag_result": {"answer": "可关注创新药、器械和公共卫生。"},
+                        "direct_reply_text": "可关注创新药、器械和公共卫生。",
+                    },
+                )
+            )
+
+        append_mock.assert_awaited_once()
+
+    def test_sync_feishu_recent_cache_skips_duplicate_turn(self) -> None:
+        service = ProactiveSignalCaptureService(db=None)  # type: ignore[arg-type]
+
+        with (
+            patch.object(
+                context_cache_service,
+                "get_session_messages",
+                AsyncMock(
+                    return_value=[
+                        {"role": "user", "content": "目前医药方向值得关注什么？"},
+                        {"role": "assistant", "content": "可关注创新药、器械和公共卫生。"},
+                    ]
+                ),
+            ),
+            patch.object(context_cache_service, "append_session_messages", AsyncMock()) as append_mock,
+        ):
+            asyncio.run(
+                service._sync_feishu_recent_cache(
+                    tenant_id=1,
+                    group_chat_id="oc_test",
+                    sender="林宇航",
+                    sender_open_id="ou_test",
+                    message="目前医药方向值得关注什么？",
+                    pipeline_result={
+                        "rag_result": {"answer": "可关注创新药、器械和公共卫生。"},
+                        "direct_reply_text": "可关注创新药、器械和公共卫生。",
+                    },
+                )
+            )
+
+        append_mock.assert_not_awaited()
diff --git a/backend/tests/test_inbox_queue_view.py b/backend/tests/test_inbox_queue_view.py
new file mode 100644
index 00000000..e78bc9cb
--- /dev/null
+++ b/backend/tests/test_inbox_queue_view.py
@@ -0,0 +1,369 @@
+import unittest
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+from app.models.feishu_item import FeishuItem
+from app.services.feishu_signal_pipeline import FOLLOWUP_STATUS_RE, FeishuSignalPipelineService
+
+
+class InboxQueueViewTest(unittest.TestCase):
+    def test_has_processed_message_ignores_ingress_trace_only(self) -> None:
+        service = FeishuSignalPipelineService(db=SimpleNamespace(), tenant_id=1)  # type: ignore[arg-type]
+
+        async def fake_execute(stmt):
+            sql = str(stmt)
+            if "FROM feishu_items" in sql:
+                return SimpleNamespace(scalar_one_or_none=lambda: None)
+            self.assertIn("ai_processing_logs.level IS NULL OR ai_processing_logs.level !=", sql)
+            self.assertIn(
+                "ai_processing_logs.routing_decision IS NULL OR ai_processing_logs.routing_decision !=",
+                sql,
+            )
+            return SimpleNamespace(scalar_one_or_none=lambda: None)
+
+        service.db.execute = AsyncMock(side_effect=fake_execute)  # type: ignore[attr-defined]
+
+        exists = asyncio.run(service._has_processed_message("om_ingress_only"))
+
+        self.assertFalse(exists)
+
+    def test_dedupe_queue_items_aggregates_occurrence_count(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+        first = FeishuItem(
+            tenant_id=1,
+            raw_message="接入 GitHub 信号源",
+            level="signal",
+            card_type="signal",
+            route_status="pending",
+            dedupe_key="signal:3:group:external_integration",
+            occurrence_count=2,
+        )
+        second = FeishuItem(
+            tenant_id=1,
+            raw_message="接入 GitHub 信号源",
+            level="signal",
+            card_type="signal",
+            route_status="pending",
+            dedupe_key="signal:3:group:external_integration",
+            occurrence_count=1,
+        )
+
+        rows = service._dedupe_queue_items([first, second])
+
+        self.assertEqual(len(rows), 1)
+        self.assertEqual(int(rows[0].occurrence_count), 3)
+
+    def test_resolve_message_class_falls_back_from_route_status(self) -> None:
+        pending_item = FeishuItem(
+            tenant_id=1,
+            raw_message="同步 GitHub issue",
+            level="signal",
+            card_type="signal",
+            route_status="pending",
+        )
+        auto_done_item = FeishuItem(
+            tenant_id=1,
+            raw_message="同步 GitHub issue",
+            level="signal",
+            card_type="signal",
+            route_status="auto_processed",
+        )
+
+        self.assertEqual(FeishuSignalPipelineService._resolve_message_class(pending_item), "action_required")
+        self.assertEqual(FeishuSignalPipelineService._resolve_message_class(auto_done_item), "auto_done")
+
+    def test_resolve_message_class_supports_new_status_machine(self) -> None:
+        queued_item = FeishuItem(
+            tenant_id=1,
+            raw_message="新增一个项目",
+            level="project",
+            card_type="new_project_candidate",
+            route_status="queued_for_decision",
+        )
+        undone_item = FeishuItem(
+            tenant_id=1,
+            raw_message="新增一个项目",
+            level="command",
+            card_type="project_create",
+            route_status="undone",
+        )
+
+        self.assertEqual(FeishuSignalPipelineService._resolve_message_class(queued_item), "action_required")
+        self.assertEqual(FeishuSignalPipelineService._resolve_message_class(undone_item), "auto_done")
+
+    def test_extract_project_intent_fields_requires_title_description_and_scope(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+
+        fields = service._extract_project_intent_fields("新增一个项目，名称叫飞书协同中台，用于统一多群消息分发，范围是 PMFlow 主数据入库。")
+
+        self.assertTrue(fields["ready"])
+        self.assertEqual(fields["title"], "飞书协同中台")
+        self.assertIn("统一多群消息分发", fields["description"])
+        self.assertIn("PMFlow 主数据入库", fields["project_scope"])
+
+    def test_resolve_intent_gate_recognizes_explicit_project_create(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+
+        gate = asyncio.run(
+            service._resolve_intent_gate(
+                message="不对，是新增进 PMFlow 的项目列表里面，名称叫飞书协同中台，用于统一多群消息分发，范围是 PMFlow 主数据入库。",
+                inferred_project_id=None,
+            )
+        )
+
+        self.assertEqual(gate["intent"], "create_project")
+        self.assertEqual(gate["canonical_action"], "project.create")
+
+    def test_resolve_intent_gate_uses_context_prefill_for_project_create(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+        service._maybe_prefill_project_fields_from_context = AsyncMock(  # type: ignore[method-assign]
+            return_value={
+                "title": "医药十五五方向研究",
+                "description": "基于最近一轮医药方向问答整理投资与立项建议。",
+                "project_scope": "围绕创新药、医疗器械、公共卫生与老龄化方向做机会梳理。",
+                "missing_fields": [],
+                "ready": True,
+            }
+        )
+
+        gate = asyncio.run(
+            service._resolve_intent_gate(
+                message="汇总以上内容给我创建个项目在项目列表",
+                inferred_project_id=None,
+                source_chat_id="oc_test",
+                source_sender_open_id="ou_test",
+                sender="林宇航",
+            )
+        )
+
+        self.assertEqual(gate["intent"], "create_project")
+        self.assertEqual(gate["fields"]["title"], "医药十五五方向研究")
+
+    def test_resolve_intent_gate_uses_embedded_message_context_for_project_create(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+
+        gate = asyncio.run(
+            service._resolve_intent_gate(
+                message=(
+                    '目前"十五五"规划（2026-2030）还在制定中，但从已有信号来看，医药行业这几个方向值得关注：'
+                    " 创新药会是核心 中医药传承创新 应对老龄化 供应链安全 出海国际化"
+                    " 根据这个汇总一下，给我创建一个项目在项目列表"
+                ),
+                inferred_project_id=None,
+            )
+        )
+
+        self.assertEqual(gate["intent"], "create_project")
+        self.assertEqual(gate["fields"]["title"], "十五五医药方向研究")
+        self.assertIn("创新药会是核心", gate["fields"]["description"])
+
+    def test_resolve_intent_gate_blocks_project_create_when_recent_context_is_failure_text(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+        service._load_recent_chat_turns = AsyncMock(  # type: ignore[method-assign]
+            return_value=[
+                {
+                    "question": "目前十五五规划里，医药行业值得关注哪些方向？",
+                    "answer": "我暂时无法生成可靠回答（模型服务异常）。为避免误导，本次不直接给结论；请稍后重试。",
+                }
+            ]
+        )
+        service._load_recent_chat_reply_fallback_text = AsyncMock(return_value="")  # type: ignore[method-assign]
+
+        gate = asyncio.run(
+            service._resolve_intent_gate(
+                message="根据上面汇总一下，给我创建一个项目在项目列表",
+                inferred_project_id=None,
+                source_chat_id="oc_test",
+                source_sender_open_id="ou_test",
+                sender="林宇航",
+            )
+        )
+
+        self.assertEqual(gate["intent"], "clarification_required")
+        self.assertIn("title", gate["fields"]["missing_fields"])
+
+    def test_followup_continue_accepts_project_reference_message(self) -> None:
+        self.assertTrue(
+            FeishuSignalPipelineService._looks_like_continue_followup("你应该继续创建我刚才要求的项目")
+        )
+        self.assertFalse(
+            FeishuSignalPipelineService._looks_like_continue_followup("继续创建一个项目，名称叫飞书协同中台")
+        )
+
+    def test_followup_status_regex_matches_creation_question(self) -> None:
+        self.assertTrue(bool(FOLLOWUP_STATUS_RE.search("是否已经创建")))
+        self.assertTrue(bool(FOLLOWUP_STATUS_RE.search("创建了吗")))
+
+    def test_select_followup_status_target_prefers_project_create(self) -> None:
+        source_item = FeishuItem(
+            tenant_id=1,
+            raw_message="新增一个Web搜索信号源",
+            level="command",
+            card_type="command",
+            route_status="auto_done",
+            result_payload={"operation_code": "source.connect"},
+        )
+        project_item = FeishuItem(
+            tenant_id=1,
+            raw_message="给我创建一个项目在项目列表",
+            level="command",
+            card_type="project_create",
+            route_status="auto_done",
+            result_payload={"operation_code": "project.create"},
+        )
+
+        target = FeishuSignalPipelineService._select_followup_status_target(
+            message="是否已经创建",
+            recent_items=[source_item, project_item],
+        )
+
+        self.assertIs(target, project_item)
+
+    def test_resolve_project_candidate_fields_blocks_placeholder_project(self) -> None:
+        service = FeishuSignalPipelineService(db=None, tenant_id=1)  # type: ignore[arg-type]
+        item = FeishuItem(
+            tenant_id=1,
+            raw_message="给我创建一个项目在项目列表",
+            level="project",
+            card_type="new_project_candidate",
+            route_status="queued_for_decision",
+            result_payload={
+                "operation_code": "project.create",
+                "operation_payload": {
+                    "title": "",
+                    "description": "",
+                },
+                "project_scope": "",
+            },
+            blocking_reason="未执行：当前建项目请求还缺少 项目名称。",
+        )
+
+        fields = service._resolve_project_candidate_fields(item=item, suggestion=None)
+
+        self.assertFalse(fields["ready"])
+        self.assertIn("title", fields["missing_fields"])
+
+    def test_create_project_gate_missing_fields_only_returns_clarification(self) -> None:
+        service = FeishuSignalPipelineService(db=SimpleNamespace(), tenant_id=1)  # type: ignore[arg-type]
+        service._resolve_command_operator_user = AsyncMock(return_value={"id": 1})  # type: ignore[method-assign]
+        service._can_execute_high_value_action = lambda _operator: True  # type: ignore[method-assign]
+        service._queue_project_intent_item = AsyncMock()  # type: ignore[method-assign]
+        service._write_processing_log = AsyncMock()  # type: ignore[method-assign]
+        service._execute_controlled_operation = AsyncMock(  # type: ignore[method-assign]
+            side_effect=AssertionError("executor should not run")
+        )
+
+        result = asyncio.run(
+            service._handle_create_project_gate(
+                message="汇总以上内容给我创建个项目在项目列表",
+                sender="林宇航",
+                source_chat_id="oc_test",
+                source_sender_open_id="ou_test",
+                feishu_message_id="om_test",
+                captured_at=None,
+                intent_gate={
+                    "intent": "clarification_required",
+                    "canonical_action": "project.create",
+                    "fields": {
+                        "title": "",
+                        "description": "汇总以上内容给我创建个项目在项目列表",
+                        "project_scope": "",
+                        "missing_fields": ["title"],
+                        "ready": False,
+                    },
+                    "clarification_template": "未执行：当前建项目请求还缺少 项目名称。",
+                },
+                process_started_at=0.0,
+            )
+        )
+
+        self.assertEqual(result["intent"], "clarification_required")
+        self.assertEqual(result["direct_reply_text"], "未执行：当前建项目请求还缺少 项目名称。")
+        self.assertEqual(result["results"], [])
+        service._queue_project_intent_item.assert_not_awaited()
+        service._execute_controlled_operation.assert_not_awaited()
+
+    def test_create_task_gate_missing_fields_only_returns_clarification(self) -> None:
+        service = FeishuSignalPipelineService(db=SimpleNamespace(), tenant_id=1)  # type: ignore[arg-type]
+        service._resolve_command_operator_user = AsyncMock(return_value={"id": 1})  # type: ignore[method-assign]
+        service._can_execute_high_value_action = lambda _operator: True  # type: ignore[method-assign]
+        service._queue_command_intent_item = AsyncMock()  # type: ignore[method-assign]
+        service._write_processing_log = AsyncMock()  # type: ignore[method-assign]
+        service._execute_controlled_operation = AsyncMock(  # type: ignore[method-assign]
+            side_effect=AssertionError("executor should not run")
+        )
+
+        result = asyncio.run(
+            service._handle_create_task_gate(
+                message="帮我建个任务",
+                sender="林宇航",
+                source_chat_id="oc_test",
+                source_sender_open_id="ou_test",
+                feishu_message_id="om_test",
+                captured_at=None,
+                intent_gate={
+                    "intent": "clarification_required",
+                    "canonical_action": "task.create",
+                    "fields": {
+                        "project_id": None,
+                        "title": "",
+                        "description": "",
+                        "missing_fields": ["project_id", "title"],
+                        "ready": False,
+                    },
+                    "clarification_template": "未执行：当前建任务请求还缺少 归属项目、任务标题。",
+                },
+                process_started_at=0.0,
+            )
+        )
+
+        self.assertEqual(result["intent"], "clarification_required")
+        self.assertEqual(result["direct_reply_text"], "未执行：当前建任务请求还缺少 归属项目、任务标题。")
+        self.assertEqual(result["results"], [])
+        service._queue_command_intent_item.assert_not_awaited()
+        service._execute_controlled_operation.assert_not_awaited()
+
+    def test_should_hide_pending_item_from_inbox_for_query_and_missing_fields(self) -> None:
+        lookup_item = FeishuItem(
+            tenant_id=1,
+            raw_message="现在有哪些项目",
+            level="signal",
+            card_type="signal",
+            route_status="pending",
+        )
+        clarification_item = FeishuItem(
+            tenant_id=1,
+            raw_message="汇总以上内容给我创建个项目在项目列表",
+            level="project",
+            card_type="new_project_candidate",
+            route_status="queued_for_decision",
+            decision_reason="missing_required_fields",
+        )
+        status_item = FeishuItem(
+            tenant_id=1,
+            raw_message="你好 长连接时都已通",
+            level="signal",
+            card_type="signal",
+            route_status="pending",
+        )
+
+        self.assertTrue(FeishuSignalPipelineService._should_hide_pending_item_from_inbox(lookup_item))
+        self.assertTrue(FeishuSignalPipelineService._should_hide_pending_item_from_inbox(clarification_item))
+        self.assertTrue(FeishuSignalPipelineService._should_hide_pending_item_from_inbox(status_item))
+
+    def test_build_project_preview_prefers_structured_title_and_detail(self) -> None:
+        title = FeishuSignalPipelineService._build_project_preview_title(
+            title="",
+            description="基于最近一轮医药方向问答整理投资与立项建议。",
+            project_scope="围绕创新药、医疗器械、公共卫生与老龄化方向做机会梳理。",
+        )
+        summary = FeishuSignalPipelineService._build_project_preview_summary(
+            description="基于最近一轮医药方向问答整理投资与立项建议。",
+            project_scope="围绕创新药、医疗器械、公共卫生与老龄化方向做机会梳理。",
+            blocking_reason="未执行：当前建项目请求还缺少 项目名称。",
+            fallback_text="汇总以上内容给我创建个项目在项目列表",
+        )
+
+        self.assertEqual(title, "围绕创新药")
+        self.assertIn("投资与立项建议", summary)
diff --git a/backend/tests/test_llm_structured_extraction.py b/backend/tests/test_llm_structured_extraction.py
new file mode 100644
index 00000000..d1e7cf25
--- /dev/null
+++ b/backend/tests/test_llm_structured_extraction.py
@@ -0,0 +1,207 @@
+import unittest
+from unittest.mock import AsyncMock
+
+from app.services.llm import ChatCompletionResult, LLMService, StructuredExtractionResult
+
+
+class LLMStructuredExtractionTest(unittest.TestCase):
+    def setUp(self) -> None:
+        self.service = LLMService()
+
+    def _extract(self, response: dict, *, function_name: str = "openclaw_command_dispatch") -> StructuredExtractionResult:
+        return self.service._extract_structured_object_from_response(
+            function_name=function_name,
+            response=response,
+            fallback_text="",
+            provider="glm",
+            model="GLM-4.7",
+        )
+
+    def test_extracts_object_from_plain_content(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "content": '{"summary":"ok","confidence":0.9,"operations":[]}',
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.validation, "ok")
+        self.assertEqual(result.payload["summary"], "ok")
+
+    def test_extracts_object_from_typed_chunk_content(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "content": [
+                                {"type": "output_text", "text": '{"summary":"chunk","confidence":0.8,"operations":[]}'}
+                            ]
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.validation, "ok")
+        self.assertEqual(result.payload["summary"], "chunk")
+
+    def test_extracts_object_from_tool_call_arguments(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "tool_calls": [
+                                {
+                                    "function": {
+                                        "arguments": '{"summary":"args","confidence":0.7,"operations":[]}'
+                                    }
+                                }
+                            ],
+                            "content": "",
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.source_field, "tool_calls.function.arguments")
+        self.assertEqual(result.payload["summary"], "args")
+
+    def test_extracts_json_from_reasoning_content_with_prefix_text(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "content": "",
+                            "reasoning_content": "先分析上下文，再给结果：\n{\"summary\":\"from_reasoning\",\"confidence\":0.66,\"operations\":[]}",
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.source_field, "message.reasoning_content")
+        self.assertEqual(result.payload["summary"], "from_reasoning")
+
+    def test_returns_json_not_found_when_reasoning_has_only_natural_language(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "content": "",
+                            "reasoning_content": "我认为还需要更多上下文。",
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.error_code, "structured_json_not_found")
+        self.assertEqual(result.source_field, "message.reasoning_content")
+
+    def test_unwraps_nested_json_field_once(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "content": '{"json":"{\\"summary\\":\\"nested\\",\\"confidence\\":0.55,\\"operations\\":[]}"}',
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.validation, "ok")
+        self.assertEqual(result.payload["summary"], "nested")
+
+    def test_rejects_top_level_array_for_object_only_function(self) -> None:
+        result = self._extract(
+            {
+                "choices": [
+                    {
+                        "message": {
+                            "content": '[{"summary":"bad"}]',
+                        }
+                    }
+                ]
+            }
+        )
+
+        self.assertEqual(result.error_code, "structured_schema_invalid")
+
+
+class LLMStructuredCacheTest(unittest.IsolatedAsyncioTestCase):
+    async def test_generate_structured_object_uses_cache_after_first_hit(self) -> None:
+        service = LLMService()
+        calls: list[str] = []
+        cache: dict[str, str] = {}
+
+        class _FakeRedis:
+            async def get(self, key: str):
+                return cache.get(key)
+
+            async def setex(self, key: str, _ttl: int, value: str):
+                cache[key] = value
+
+        service._get_redis = AsyncMock(return_value=_FakeRedis())
+        service._resolve_engine_for_call = AsyncMock(
+            return_value={
+                "provider": "glm",
+                "model": "GLM-4.7",
+                "api_key": "fake",
+                "base_url": "http://unit.test",
+            }
+        )
+        service._resolve_active_prompt_profile = AsyncMock(return_value={"role_name": "", "system_prompt": ""})
+        service._log_ai_call = AsyncMock()
+
+        async def _fake_chat_completion_raw(**_: object) -> ChatCompletionResult:
+            calls.append("hit")
+            return ChatCompletionResult(
+                text='{"summary":"cached","confidence":0.9,"operations":[]}',
+                raw_response={
+                    "choices": [
+                        {
+                            "message": {
+                                "content": '{"summary":"cached","confidence":0.9,"operations":[]}',
+                            }
+                        }
+                    ]
+                },
+            )
+
+        service._chat_completion_raw = AsyncMock(side_effect=_fake_chat_completion_raw)
+
+        first = await service.generate_structured_object(
+            "planner prompt",
+            db=object(),
+            tenant_id=1,
+            function_name="openclaw_command_dispatch",
+            response_mode="fast",
+        )
+        second = await service.generate_structured_object(
+            "planner prompt",
+            db=object(),
+            tenant_id=1,
+            function_name="openclaw_command_dispatch",
+            response_mode="fast",
+        )
+
+        self.assertEqual(first.payload["summary"], "cached")
+        self.assertEqual(second.payload["summary"], "cached")
+        self.assertTrue(second.from_cache)
+        self.assertEqual(len(calls), 1)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/backend/tests/test_openclaw_commander_service.py b/backend/tests/test_openclaw_commander_service.py
new file mode 100644
index 00000000..15de8509
--- /dev/null
+++ b/backend/tests/test_openclaw_commander_service.py
@@ -0,0 +1,445 @@
+import unittest
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+from app.models import Project
+from app.services.openclaw_commander_service import (
+    OpenClawCommanderService,
+    OperationExecutorService,
+)
+from app.services.llm import StructuredExtractionResult
+
+
+class _DuplicatePersistentExecutor(OperationExecutorService):
+    def __init__(self) -> None:
+        super().__init__(db=None)  # type: ignore[arg-type]
+
+    async def _claim_persistent_idempotency(self, **_: object):
+        return None, {
+            "status": "skipped",
+            "message": "duplicate_persistent",
+            "entity_type": "notification",
+            "entity_id": "21",
+            "trace_id": "openclaw-dup-1",
+        }
+
+    async def _execute_single_operation(self, **_: object):
+        raise AssertionError("duplicate notification should not execute")
+
+
+class _SignalSourceFallbackCommander(OpenClawCommanderService):
+    def __init__(self, *, charter_ready: bool) -> None:
+        super().__init__(db=None)  # type: ignore[arg-type]
+        self.charter_ready = charter_ready
+
+    async def _load_project(self, **_: object):
+        return SimpleNamespace(id=3, owner_id=7, title="项目A", status="active")
+
+    async def _load_latest_aligned_charter(self, **_: object):
+        if not self.charter_ready:
+            return None
+        return SimpleNamespace(id=2, status="aligned")
+
+    async def _resolve_signal_source_notification_recipient(self, **_: object):
+        return 7
+
+
+class _ProjectFallbackCommander(OpenClawCommanderService):
+    def __init__(self) -> None:
+        super().__init__(db=None)  # type: ignore[arg-type]
+
+    async def _load_project_directory(self, **_: object):
+        return [{"project_id": 8, "title": "财务冲账协同", "status": "active"}]
+
+
+class _MemoryDB:
+    def __init__(self) -> None:
+        self.added: list[object] = []
+
+    def add(self, obj: object) -> None:
+        self.added.append(obj)
+
+    async def flush(self) -> None:
+        for obj in self.added:
+            if isinstance(obj, Project) and getattr(obj, "id", None) in {None, 0}:
+                obj.id = 8
+
+    async def commit(self) -> None:
+        return None
+
+    async def rollback(self) -> None:
+        return None
+
+
+class _ProjectExecutor(OperationExecutorService):
+    def __init__(self, *, existing_project: object | None = None) -> None:
+        self.memory_db = _MemoryDB()
+        super().__init__(db=self.memory_db)  # type: ignore[arg-type]
+        self.existing_project = existing_project
+
+    async def _load_project_by_title(self, **_: object):
+        return self.existing_project
+
+    async def _load_active_user(self, *, tenant_id: int, user_id: int):
+        _ = tenant_id
+        return SimpleNamespace(id=user_id, status="active")
+
+
+class OpenClawCommanderServiceTest(unittest.TestCase):
+    def setUp(self) -> None:
+        self.executor = OperationExecutorService(db=None)  # type: ignore[arg-type]
+
+    def test_plan_operations_accepts_tool_call_json(self) -> None:
+        service = OpenClawCommanderService(db=None)  # type: ignore[arg-type]
+        service._commander_operation_catalog = lambda: [  # type: ignore[method-assign]
+            {
+                "operation_code": "task.assign",
+                "description": "assign",
+                "input_schema": {"properties": {"task_id": {}, "assignee_id": {}}},
+            }
+        ]
+
+        with patch(
+            "app.services.openclaw_commander_service.llm_service.generate_structured_object",
+            AsyncMock(
+                return_value=StructuredExtractionResult(
+                    payload={
+                        "summary": "先补负责人，再发通知",
+                        "confidence": 0.82,
+                        "operations": [
+                            {
+                                "operation_code": "task.assign",
+                                "payload": {"task_id": 12, "assignee_id": 5},
+                                "reason": "技能更匹配",
+                            }
+                        ],
+                    },
+                    source_field="tool_calls.function.arguments",
+                    provider="glm",
+                    model="GLM-4.7",
+                    validation="ok",
+                )
+            ),
+        ):
+            plan, planner_status, diagnostics = self.async_run(
+                service._plan_operations(
+                    tenant_id=1,
+                    operator_user=SimpleNamespace(id=9, name="Boss", role="BOSS"),
+                    project=SimpleNamespace(id=3, title="项目A", name="项目A", description="desc", status="active", assigned_role=""),
+                    tasks=[],
+                    task_summary={},
+                    members=[],
+                    instruction="把任务 12 分给张三",
+                    dry_run=False,
+                    max_operations=4,
+                    response_mode="fast",
+                )
+            )
+
+        self.assertEqual(planner_status, "planned")
+        self.assertEqual(diagnostics["source_field"], "tool_calls.function.arguments")
+        self.assertEqual(plan["operations"][0]["payload"]["task_id"], 12)
+
+    def test_plan_operations_returns_schema_invalid_diagnostics_on_bad_payload(self) -> None:
+        service = OpenClawCommanderService(db=None)  # type: ignore[arg-type]
+        service._commander_operation_catalog = lambda: []  # type: ignore[method-assign]
+        service._build_fallback_plan = AsyncMock(return_value={"summary": "等待补充上下文。", "confidence": 0.0, "operations": []})  # type: ignore[method-assign]
+
+        with patch(
+            "app.services.openclaw_commander_service.llm_service.generate_structured_object",
+            AsyncMock(
+                return_value=StructuredExtractionResult(
+                    payload=None,
+                    source_field="message.content",
+                    provider="glm",
+                    model="GLM-4.7",
+                    error_code="structured_schema_invalid",
+                    validation="invalid",
+                )
+            ),
+        ):
+            plan, planner_status, diagnostics = self.async_run(
+                service._plan_operations(
+                    tenant_id=1,
+                    operator_user=SimpleNamespace(id=9, name="Boss", role="BOSS"),
+                    project=SimpleNamespace(id=3, title="项目A", name="项目A", description="desc", status="active", assigned_role=""),
+                    tasks=[],
+                    task_summary={},
+                    members=[],
+                    instruction="把任务 12 分给张三",
+                    dry_run=False,
+                    max_operations=4,
+                    response_mode="fast",
+                )
+            )
+
+        self.assertEqual(planner_status, "fallback")
+        self.assertEqual(diagnostics["error_code"], "structured_schema_invalid")
+        self.assertEqual(plan["operations"], [])
+
+    def test_validate_payload_coerces_integer_and_rejects_extra_fields(self) -> None:
+        normalized = self.executor._validate_and_normalize_payload(
+            operation_code="task.assign",
+            payload={"task_id": "12", "assignee_id": "8"},
+        )
+
+        self.assertEqual(normalized, {"task_id": 12, "assignee_id": 8})
+
+        with self.assertRaisesRegex(ValueError, "unsupported fields"):
+            self.executor._validate_and_normalize_payload(
+                operation_code="task.assign",
+                payload={"task_id": 12, "assignee_id": 8, "foo": "bar"},
+            )
+
+    def test_validate_task_create_payload_supports_array_and_numeric_fields(self) -> None:
+        normalized = self.executor._validate_and_normalize_payload(
+            operation_code="task.create",
+            payload={
+                "project_id": "7",
+                "title": "补充告警兜底",
+                "required_skill_tag_ids": ["1", 2, "3"],
+                "estimated_hours": "4.5",
+                "priority": "P0",
+            },
+        )
+
+        self.assertEqual(normalized["project_id"], 7)
+        self.assertEqual(normalized["required_skill_tag_ids"], [1, 2, 3])
+        self.assertEqual(normalized["estimated_hours"], 4.5)
+        self.assertEqual(normalized["priority"], "P0")
+
+    def test_validate_project_create_payload_supports_optional_owner_and_description(self) -> None:
+        normalized = self.executor._validate_and_normalize_payload(
+            operation_code="project.create",
+            payload={
+                "title": "财务冲账协同",
+                "description": "用于协调财务冲账任务。",
+                "owner_id": "7",
+            },
+        )
+
+        self.assertEqual(
+            normalized,
+            {
+                "title": "财务冲账协同",
+                "description": "用于协调财务冲账任务。",
+                "owner_id": 7,
+            },
+        )
+
+    def test_workload_update_payload_requires_max_weekly_hours(self) -> None:
+        normalized = self.executor._validate_and_normalize_payload(
+            operation_code="personnel.workload_update",
+            payload={"user_id": "5", "max_weekly_hours": "32"},
+        )
+
+        self.assertEqual(normalized, {"user_id": 5, "max_weekly_hours": 32.0})
+
+        with self.assertRaisesRegex(ValueError, "unsupported fields"):
+            self.executor._validate_and_normalize_payload(
+                operation_code="personnel.workload_update",
+                payload={"user_id": 5, "workload": 70},
+            )
+
+    def test_notification_idempotency_key_uses_trimmed_content_hash(self) -> None:
+        base_payload = {
+            "user_id": 3,
+            "title": "请处理阻塞",
+            "content": "  任务 A 已重新分配给你  ",
+        }
+        same_after_trim = {
+            "user_id": 3,
+            "title": "请处理阻塞",
+            "content": "任务 A 已重新分配给你",
+        }
+        changed_content = {
+            "user_id": 3,
+            "title": "请处理阻塞",
+            "content": "任务 B 已重新分配给你",
+        }
+
+        key_a = self.executor._build_idempotency_key(
+            operation_code="notification.create",
+            payload=base_payload,
+            tenant_id=1,
+            project_id=9,
+        )
+        key_b = self.executor._build_idempotency_key(
+            operation_code="notification.create",
+            payload=same_after_trim,
+            tenant_id=1,
+            project_id=9,
+        )
+        key_c = self.executor._build_idempotency_key(
+            operation_code="notification.create",
+            payload=changed_content,
+            tenant_id=1,
+            project_id=9,
+        )
+        key_other_project = self.executor._build_idempotency_key(
+            operation_code="notification.create",
+            payload=same_after_trim,
+            tenant_id=1,
+            project_id=10,
+        )
+
+        self.assertEqual(key_a, key_b)
+        self.assertNotEqual(key_a, key_c)
+        self.assertNotEqual(key_a, key_other_project)
+
+    def test_execute_operations_skips_persistent_duplicate_notification(self) -> None:
+        executor = _DuplicatePersistentExecutor()
+
+        result = self.async_run(
+            executor.execute_operations(
+                tenant_id=1,
+                operator_user=SimpleNamespace(id=9),
+                project=SimpleNamespace(id=4),
+                operations=[
+                    {
+                        "operation_code": "notification.create",
+                        "payload": {
+                            "user_id": 3,
+                            "title": "请处理阻塞",
+                            "content": "任务 A 已重新分配给你",
+                        },
+                        "reason": "重复回放",
+                    }
+                ],
+                dry_run=False,
+                trace_id="openclaw-1",
+            )
+        )
+
+        self.assertEqual(result["requested_count"], 1)
+        self.assertEqual(result["skipped_count"], 1)
+        self.assertEqual(result["results"][0]["message"], "duplicate_persistent")
+        self.assertEqual(result["results"][0]["entity_id"], "21")
+
+    def test_normalize_task_priority_maps_p_levels(self) -> None:
+        self.assertEqual(self.executor._normalize_task_priority("P0"), "high")
+        self.assertEqual(self.executor._normalize_task_priority("P1"), "medium")
+        self.assertEqual(self.executor._normalize_task_priority("P3"), "low")
+
+    def test_signal_source_fallback_without_charter_creates_notification_only(self) -> None:
+        commander = _SignalSourceFallbackCommander(charter_ready=False)
+
+        plan = self.async_run(
+            commander._build_fallback_plan(
+                tenant_id=1,
+                project_id=3,
+                instruction="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                max_operations=4,
+                error="llm offline",
+            )
+        )
+
+        self.assertEqual(plan["summary"], "OpenClaw 规划失败，已回退为信号源接入最小闭环。")
+        self.assertEqual(len(plan["operations"]), 1)
+        self.assertEqual(plan["operations"][0]["operation_code"], "notification.create")
+        self.assertEqual(plan["operations"][0]["payload"]["title"], "GitHub信号源接入待确认")
+
+    def test_signal_source_fallback_with_charter_creates_task_and_notification(self) -> None:
+        commander = _SignalSourceFallbackCommander(charter_ready=True)
+
+        plan = self.async_run(
+            commander._build_fallback_plan(
+                tenant_id=1,
+                project_id=3,
+                instruction="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                max_operations=4,
+                error="llm offline",
+            )
+        )
+
+        self.assertEqual([item["operation_code"] for item in plan["operations"]], ["task.create", "notification.create"])
+        self.assertEqual(plan["operations"][0]["payload"]["project_id"], 3)
+        self.assertEqual(plan["operations"][0]["payload"]["title"], "接入GitHub信号源")
+
+    def test_project_create_fallback_builds_project_create_operation(self) -> None:
+        commander = _ProjectFallbackCommander()
+
+        plan = self.async_run(
+            commander._build_fallback_plan(
+                tenant_id=1,
+                project_id=None,
+                instruction="帮我创建一个项目，财务冲账协同",
+                max_operations=4,
+                error="structured_http_error",
+                project_directory=None,
+            )
+        )
+
+        self.assertEqual(plan["operations"][0]["operation_code"], "project.create")
+        self.assertEqual(plan["operations"][0]["payload"]["title"], "财务冲账协同")
+
+    def test_project_status_fallback_resolves_project_id_from_directory(self) -> None:
+        commander = _ProjectFallbackCommander()
+
+        plan = self.async_run(
+            commander._build_fallback_plan(
+                tenant_id=1,
+                project_id=None,
+                instruction="把财务冲账协同项目归档",
+                max_operations=4,
+                error="structured_http_error",
+                project_directory=None,
+            )
+        )
+
+        self.assertEqual(plan["operations"][0]["operation_code"], "project.archive")
+        self.assertEqual(plan["operations"][0]["payload"]["project_id"], 8)
+
+    def test_execute_project_create_creates_active_project(self) -> None:
+        executor = _ProjectExecutor()
+
+        with patch(
+            "app.services.openclaw_commander_service.log_operation",
+            AsyncMock(),
+        ), patch(
+            "app.services.openclaw_commander_service.context_cache_service.invalidate_system_context",
+            AsyncMock(),
+        ):
+            result = self.async_run(
+                executor._execute_project_create(
+                    tenant_id=1,
+                    operator_user=SimpleNamespace(id=9, role="BOSS"),
+                    payload={"title": "财务冲账协同", "description": "用于财务冲账。"},
+                    reason="explicit_project_create",
+                    trace_id="openclaw-1",
+                )
+            )
+
+        created_project = next(obj for obj in executor.memory_db.added if isinstance(obj, Project))
+        self.assertEqual(result["status"], "applied")
+        self.assertEqual(created_project.title, "财务冲账协同")
+        self.assertEqual(created_project.status, "active")
+        self.assertEqual(result["result"]["project_id"], 8)
+
+    def test_execute_project_create_skips_existing_project(self) -> None:
+        executor = _ProjectExecutor(
+            existing_project=SimpleNamespace(id=5, title="财务冲账协同", name="财务冲账协同", status="active")
+        )
+
+        result = self.async_run(
+            executor._execute_project_create(
+                tenant_id=1,
+                operator_user=SimpleNamespace(id=9, role="BOSS"),
+                payload={"title": "财务冲账协同"},
+                reason="explicit_project_create",
+                trace_id="openclaw-1",
+            )
+        )
+
+        self.assertEqual(result["status"], "skipped")
+        self.assertEqual(result["entity_id"], "5")
+        self.assertEqual(result["result"]["status"], "active")
+
+    def async_run(self, awaitable):
+        import asyncio
+
+        return asyncio.run(awaitable)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/backend/tests/test_signal_source_orchestration_service.py b/backend/tests/test_signal_source_orchestration_service.py
new file mode 100644
index 00000000..266118f3
--- /dev/null
+++ b/backend/tests/test_signal_source_orchestration_service.py
@@ -0,0 +1,422 @@
+import unittest
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+from app.models.signal_source_registry import SignalSourceRegistry
+from app.services.llm import StructuredExtractionResult
+from app.services.signal_source_orchestration_service import SignalSourceOrchestrationService
+
+
+class _StubSignalSourceOrchestrationService(SignalSourceOrchestrationService):
+    def __init__(self) -> None:
+        super().__init__(db=None)  # type: ignore[arg-type]
+
+    async def _load_project(self, **_: object):
+        return SimpleNamespace(id=3, title="项目A", name="项目A", description="desc", owner_id=1)
+
+    async def _load_signal(self, **_: object):
+        return SimpleNamespace(
+            id=9,
+            title="新增 GitHub 信号源",
+            summary="监控关键仓库 issue/PR",
+            signal_type="新需求",
+            recommended_action="enter_only",
+            content="新增 GitHub 信号源并自动拆任务",
+        )
+
+    async def _plan_onboarding(self, **_: object):
+        return (
+            {
+                "should_onboard": True,
+                "summary": "接入 GitHub 仓库信号源并下发 AI worker。",
+                "source": {
+                    "source_kind": "github",
+                    "source_name": "核心仓库",
+                    "source_identifier": "github:core-repo",
+                    "description": "跟踪 issue、pr、release",
+                },
+                "worker_roles": ["pm_dispatcher", "tech_architect"],
+                "openclaw_instruction": "为 GitHub 信号源创建接入任务并通知负责人。",
+            },
+            "planned",
+            {
+                "used_fallback": False,
+                "error_code": None,
+                "source_field": "message.content",
+                "provider": "glm",
+                "model": "GLM-4.7",
+            },
+        )
+
+    async def _create_or_reuse_registry(self, **_: object):
+        row = SignalSourceRegistry(
+            tenant_id=1,
+            project_id=3,
+            signal_id=9,
+            source_kind="github",
+            source_name="核心仓库",
+            source_identifier="github:core-repo",
+            description="跟踪 issue、pr、release",
+            requested_via="pmflow",
+            status="active",
+            onboarding_status="pending",
+        )
+        row.id = 11  # type: ignore[assignment]
+        return {
+            "status": "created",
+            "registry_id": 11,
+            "row": row,
+        }
+
+    async def _resolve_operator_user(self, **_: object):
+        return SimpleNamespace(id=1, role="BOSS")
+
+    async def _dispatch_openclaw(self, **_: object):
+        return {
+            "status": "completed",
+            "result": {
+                "trace_id": "openclaw-123",
+                "execution": {
+                    "results": [
+                        {
+                            "operation_code": "task.create",
+                            "status": "applied",
+                            "entity_id": "21",
+                            "payload": {"title": "接入 GitHub 监听"},
+                        }
+                    ]
+                },
+            },
+        }
+
+    async def _execute_worker_roles(self, **_: object):
+        return [
+            {
+                "worker_role_key": "pm_dispatcher",
+                "worker_role_name": "产品经理（任务分派）",
+                "status": "completed",
+                "summary": "已输出接入拆解。",
+            },
+            {
+                "worker_role_key": "tech_architect",
+                "worker_role_name": "技术架构师",
+                "status": "completed",
+                "summary": "已输出接入约束。",
+            },
+        ]
+
+
+class SignalSourceOrchestrationServiceTest(unittest.TestCase):
+    def async_run(self, awaitable):
+        import asyncio
+
+        return asyncio.run(awaitable)
+
+    def test_detect_signal_source_request(self) -> None:
+        self.assertTrue(
+            SignalSourceOrchestrationService.looks_like_signal_source_request(
+                "请新增一个 GitHub 信号源，监控核心仓库 issue。"
+            )
+        )
+        self.assertFalse(
+            SignalSourceOrchestrationService.looks_like_signal_source_request(
+                "把任务 12 分给张三。"
+            )
+        )
+
+    def test_orchestrate_from_signal_returns_completed_payload(self) -> None:
+        service = _StubSignalSourceOrchestrationService()
+
+        result = self.async_run(
+            service.orchestrate_from_signal(
+                tenant_id=1,
+                project_id=3,
+                signal_id=9,
+                initiating_text="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                requested_via="pmflow",
+                operator_user_id=1,
+                dry_run=True,
+            )
+        )
+
+        self.assertEqual(result["status"], "planned")
+        self.assertEqual(result["signal_source"]["registry_id"], 11)
+        self.assertEqual(result["signal_source"]["source_kind"], "github")
+        self.assertEqual(len(result["worker_executions"]), 2)
+        self.assertEqual(result["openclaw"]["status"], "completed")
+
+    def test_orchestrate_from_signal_skips_non_source_request(self) -> None:
+        service = _StubSignalSourceOrchestrationService()
+
+        result = self.async_run(
+            service.orchestrate_from_signal(
+                tenant_id=1,
+                project_id=3,
+                signal_id=9,
+                initiating_text="把任务 12 调成 P0。",
+                requested_via="pmflow",
+                operator_user_id=1,
+                dry_run=True,
+            )
+        )
+
+        self.assertEqual(result["status"], "skipped")
+        self.assertEqual(result["reason"], "not_signal_source_request")
+
+    def test_normalize_source_payload_uses_request_hash_for_generic_source_name(self) -> None:
+        service = _StubSignalSourceOrchestrationService()
+
+        normalized = service._normalize_source_payload(
+            source_payload={},
+            fallback_text="请新增一个 GitHub 信号源，监控核心仓库 issue 和 PR。",
+            requested_via="pmflow",
+        )
+
+        self.assertEqual(normalized["source_kind"], "github")
+        self.assertEqual(normalized["source_name"], "GitHub")
+        self.assertTrue(normalized["source_identifier"].startswith("github:req-"))
+
+    def test_build_source_identifier_prefers_explicit_repo_locator(self) -> None:
+        identifier = SignalSourceOrchestrationService._build_source_identifier(
+            source_kind="github",
+            source_name="GitHub",
+            text="新增 GitHub 信号源，仓库是 openai/codex。",
+            source_payload={"repo": "OpenAI/Codex"},
+        )
+
+        self.assertEqual(identifier, "github:openai/codex")
+
+    def test_map_orchestration_status_updates_signal_status_label(self) -> None:
+        self.assertEqual(
+            SignalSourceOrchestrationService._map_orchestration_status_to_signal_status("completed"),
+            "已编排",
+        )
+        self.assertEqual(
+            SignalSourceOrchestrationService._map_orchestration_status_to_signal_status("pending_manual"),
+            "待确认接入",
+        )
+        self.assertEqual(
+            SignalSourceOrchestrationService._map_orchestration_status_to_signal_status("partial"),
+            "部分编排",
+        )
+        self.assertEqual(
+            SignalSourceOrchestrationService._map_orchestration_status_to_signal_status("failed"),
+            "编排失败",
+        )
+
+    def test_resolve_registry_activation_requires_real_config_and_non_fallback_result(self) -> None:
+        registry_status, onboarding_status = SignalSourceOrchestrationService._resolve_registry_activation(
+            source_payload={
+                "source_kind": "pmflow",
+                "source_name": "Web搜索",
+                "source_identifier": "pmflow:req-abc123",
+                "description": "新增一个Web搜索信号源，关键词：医药行业。",
+                "config": {},
+            },
+            openclaw_result={
+                "status": "completed",
+                "result": {
+                    "execution": {
+                        "results": [
+                            {
+                                "operation_code": "task.create",
+                                "status": "applied",
+                                "payload": {"title": "接入 GitHub 监听"},
+                            }
+                        ]
+                    }
+                },
+            },
+            worker_results=[
+                {
+                    "worker_role_key": "pm_dispatcher",
+                    "status": "completed",
+                    "used_fallback": True,
+                }
+            ],
+        )
+
+        self.assertEqual(registry_status, "pending")
+        self.assertEqual(onboarding_status, "pending_manual")
+
+    def test_resolve_registry_activation_allows_explicit_repo_with_real_worker_output(self) -> None:
+        registry_status, onboarding_status = SignalSourceOrchestrationService._resolve_registry_activation(
+            source_payload={
+                "source_kind": "github",
+                "source_name": "核心仓库",
+                "source_identifier": "github:openai/codex",
+                "description": "监控 openai/codex issue 和 PR",
+                "config": {},
+                "repo": "openai/codex",
+            },
+            openclaw_result={
+                "status": "completed",
+                "result": {
+                    "execution": {
+                        "results": [
+                            {
+                                "operation_code": "task.create",
+                                "status": "applied",
+                                "payload": {"title": "接入 GitHub 监听"},
+                            }
+                        ]
+                    }
+                },
+            },
+            worker_results=[
+                {
+                    "worker_role_key": "tech_architect",
+                    "status": "completed",
+                    "used_fallback": False,
+                }
+            ],
+        )
+
+        self.assertEqual(registry_status, "active")
+        self.assertEqual(onboarding_status, "completed")
+
+    def test_plan_onboarding_accepts_reasoning_only_json(self) -> None:
+        service = SignalSourceOrchestrationService(db=None)  # type: ignore[arg-type]
+
+        with patch(
+            "app.services.signal_source_orchestration_service.llm_service.generate_structured_object",
+            AsyncMock(
+                return_value=StructuredExtractionResult(
+                    payload={
+                        "should_onboard": True,
+                        "summary": "接入 GitHub 仓库信号源。",
+                        "source": {
+                            "source_kind": "github",
+                            "source_name": "核心仓库",
+                            "source_identifier": "github:openai/codex",
+                            "description": "监控 issue 与 PR",
+                        },
+                        "worker_roles": ["pm_dispatcher"],
+                        "openclaw_instruction": "创建接入任务。",
+                    },
+                    source_field="message.reasoning_content",
+                    provider="glm",
+                    model="GLM-4.7",
+                    validation="ok",
+                )
+            ),
+        ):
+            plan, planner_status, diagnostics = self.async_run(
+                service._plan_onboarding(
+                    tenant_id=1,
+                    project=SimpleNamespace(id=3, title="项目A", name="项目A", description="desc"),
+                    signal=SimpleNamespace(id=9, title="新增 GitHub 信号源", summary="监控 issue", signal_type="新需求", recommended_action="enter_only"),
+                    initiating_text="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                    requested_via="pmflow",
+                    response_mode="fast",
+                )
+            )
+
+        self.assertEqual(planner_status, "planned")
+        self.assertTrue(plan["should_onboard"])
+        self.assertEqual(diagnostics["source_field"], "message.reasoning_content")
+
+    def test_plan_onboarding_uses_tool_call_payload_when_content_is_empty(self) -> None:
+        service = SignalSourceOrchestrationService(db=None)  # type: ignore[arg-type]
+
+        with patch(
+            "app.services.signal_source_orchestration_service.llm_service.generate_structured_object",
+            AsyncMock(
+                return_value=StructuredExtractionResult(
+                    payload={
+                        "should_onboard": True,
+                        "summary": "接入 GitHub 仓库信号源。",
+                        "source": {
+                            "source_kind": "github",
+                            "source_name": "核心仓库",
+                            "source_identifier": "github:openai/codex",
+                            "description": "监控 issue 与 PR",
+                        },
+                        "worker_roles": ["pm_dispatcher"],
+                        "openclaw_instruction": "创建接入任务。",
+                    },
+                    source_field="tool_calls.function.arguments",
+                    provider="glm",
+                    model="GLM-4.7",
+                    validation="ok",
+                )
+            ),
+        ):
+            _, planner_status, diagnostics = self.async_run(
+                service._plan_onboarding(
+                    tenant_id=1,
+                    project=SimpleNamespace(id=3, title="项目A", name="项目A", description="desc"),
+                    signal=SimpleNamespace(id=9, title="新增 GitHub 信号源", summary="监控 issue", signal_type="新需求", recommended_action="enter_only"),
+                    initiating_text="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                    requested_via="pmflow",
+                    response_mode="fast",
+                )
+            )
+
+        self.assertEqual(planner_status, "planned")
+        self.assertEqual(diagnostics["source_field"], "tool_calls.function.arguments")
+
+    def test_plan_onboarding_returns_schema_invalid_diagnostics(self) -> None:
+        service = SignalSourceOrchestrationService(db=None)  # type: ignore[arg-type]
+
+        with patch(
+            "app.services.signal_source_orchestration_service.llm_service.generate_structured_object",
+            AsyncMock(
+                return_value=StructuredExtractionResult(
+                    payload=None,
+                    source_field="message.content",
+                    provider="glm",
+                    model="GLM-4.7",
+                    error_code="structured_schema_invalid",
+                    validation="invalid",
+                )
+            ),
+        ):
+            plan, planner_status, diagnostics = self.async_run(
+                service._plan_onboarding(
+                    tenant_id=1,
+                    project=SimpleNamespace(id=3, title="项目A", name="项目A", description="desc"),
+                    signal=SimpleNamespace(id=9, title="新增 GitHub 信号源", summary="监控 issue", signal_type="新需求", recommended_action="enter_only"),
+                    initiating_text="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                    requested_via="pmflow",
+                    response_mode="fast",
+                )
+            )
+
+        self.assertEqual(planner_status, "fallback")
+        self.assertEqual(diagnostics["error_code"], "structured_schema_invalid")
+        self.assertTrue(plan["should_onboard"])
+
+    def test_plan_onboarding_returns_json_not_found_diagnostics(self) -> None:
+        service = SignalSourceOrchestrationService(db=None)  # type: ignore[arg-type]
+
+        with patch(
+            "app.services.signal_source_orchestration_service.llm_service.generate_structured_object",
+            AsyncMock(
+                return_value=StructuredExtractionResult(
+                    payload=None,
+                    source_field="response.reasoning_content",
+                    provider="glm",
+                    model="GLM-4.7",
+                    error_code="structured_json_not_found",
+                    validation="invalid",
+                )
+            ),
+        ):
+            _, planner_status, diagnostics = self.async_run(
+                service._plan_onboarding(
+                    tenant_id=1,
+                    project=SimpleNamespace(id=3, title="项目A", name="项目A", description="desc"),
+                    signal=SimpleNamespace(id=9, title="新增 GitHub 信号源", summary="监控 issue", signal_type="新需求", recommended_action="enter_only"),
+                    initiating_text="新增一个 GitHub 信号源，监控核心仓库 issue。",
+                    requested_via="pmflow",
+                    response_mode="fast",
+                )
+            )
+
+        self.assertEqual(planner_status, "fallback")
+        self.assertEqual(diagnostics["error_code"], "structured_json_not_found")
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/backend/tests/test_sstc_signal_state_machine.py b/backend/tests/test_sstc_signal_state_machine.py
new file mode 100644
index 00000000..a0090368
--- /dev/null
+++ b/backend/tests/test_sstc_signal_state_machine.py
@@ -0,0 +1,162 @@
+from __future__ import annotations
+
+import unittest
+from datetime import datetime
+from types import SimpleNamespace
+
+from fastapi import HTTPException
+
+from app.api.sstc import (
+    _build_signal_created_at_range,
+    _collect_locked_signal_conflicts,
+    _extract_source_attachments,
+    _match_signal_filters,
+    _normalize_signal_status_code,
+)
+from app.services.feishu_integration import FeishuIntegrationService
+
+
+class SignalStateMachineHelpersTest(unittest.TestCase):
+    def test_normalize_signal_status_code_prefers_linked_state(self) -> None:
+        status_code = _normalize_signal_status_code(
+            linked_state="linked",
+            signal_status="待处理",
+            processed=False,
+        )
+        self.assertEqual(status_code, "linked")
+
+    def test_normalize_signal_status_code_uses_closed_for_processed(self) -> None:
+        status_code = _normalize_signal_status_code(
+            linked_state=None,
+            signal_status="",
+            processed=True,
+        )
+        self.assertEqual(status_code, "closed")
+
+    def test_collect_locked_signal_conflicts(self) -> None:
+        charters = [
+            SimpleNamespace(
+                id=101,
+                title="宪章A",
+                content={"signal_ids": [1, 2]},
+                signal_id=None,
+            ),
+            SimpleNamespace(
+                id=102,
+                title="宪章B",
+                content={"signal_ids": [3]},
+                signal_id=5,
+            ),
+        ]
+        conflict_ids, conflict_map = _collect_locked_signal_conflicts(
+            charters=charters,  # type: ignore[arg-type]
+            candidate_signal_ids=[2, 5, 9],
+        )
+        self.assertEqual(conflict_ids, [2, 5])
+        self.assertEqual(conflict_map[2][0]["charter_id"], 101)
+        self.assertEqual(conflict_map[5][0]["charter_id"], 102)
+
+    def test_build_signal_created_at_range_rejects_invalid_order(self) -> None:
+        with self.assertRaises(HTTPException):
+            _build_signal_created_at_range(
+                created_from="2026-03-11",
+                created_to="2026-03-10",
+            )
+
+    def test_match_signal_filters_keyword_and_status(self) -> None:
+        row = {
+            "id": "88",
+            "title": "新增飞书信号源",
+            "content": "支持附件同步",
+            "proposer_name": "老板",
+            "charter_name": "信号治理宪章",
+            "source": "Feishu",
+            "priority": "P1",
+            "status_code": "pending",
+        }
+        matched = _match_signal_filters(
+            row=row,
+            status_filter="pending",
+            keyword="附件",
+            source="Feishu",
+            priority="p1",
+            created_from_utc=datetime(2026, 3, 1, 0, 0, 0),
+            created_to_utc=datetime(2026, 3, 20, 0, 0, 0),
+            created_at_dt=datetime(2026, 3, 10, 12, 0, 0),
+        )
+        self.assertTrue(matched)
+
+    def test_extract_source_attachments_normalizes_and_deduplicates(self) -> None:
+        attachments = _extract_source_attachments(
+            {
+                "attachments": [
+                    {"attachment_type": "image", "image_key": "img_a", "file_name": "截图A"},
+                    {"attachment_type": "img", "image_key": "img_a", "file_name": "截图A重复"},
+                    {"attachment_type": "file", "file_key": "file_b", "file_name": "需求文档.pdf"},
+                ]
+            }
+        )
+        self.assertEqual(len(attachments), 2)
+        self.assertEqual(attachments[0]["attachment_type"], "image")
+        self.assertEqual(attachments[1]["file_name"], "需求文档.pdf")
+
+
+class FeishuMessagePayloadParserTest(unittest.TestCase):
+    def setUp(self) -> None:
+        self.service = FeishuIntegrationService(db=None)  # type: ignore[arg-type]
+
+    def test_extract_message_payload_for_image_message(self) -> None:
+        payload = self.service._extract_message_payload(
+            {
+                "message": {
+                    "message_type": "image",
+                    "content": '{"image_key":"img_001"}',
+                }
+            }
+        )
+        self.assertEqual(payload["message_type"], "image")
+        self.assertEqual(len(payload["attachments"]), 1)
+        self.assertIn("附件消息", str(payload["text"] or ""))
+
+    def test_extract_message_payload_for_file_message(self) -> None:
+        payload = self.service._extract_message_payload(
+            {
+                "message": {
+                    "message_type": "file",
+                    "content": '{"file_key":"file_001","file_name":"立项材料.docx","file_size":1024}',
+                }
+            }
+        )
+        self.assertEqual(len(payload["attachments"]), 1)
+        attachment = payload["attachments"][0]
+        self.assertEqual(attachment["attachment_type"], "file")
+        self.assertEqual(attachment["file_name"], "立项材料.docx")
+
+    def test_extract_message_payload_for_post_with_text_and_image(self) -> None:
+        payload = self.service._extract_message_payload(
+            {
+                "message": {
+                    "message_type": "post",
+                    "content": {
+                        "post": {
+                            "zh_cn": {
+                                "content": [
+                                    [
+                                        {"tag": "text", "text": "请评估这张图的风险点"},
+                                        {"tag": "img", "image_key": "img_post_1"},
+                                    ]
+                                ]
+                            }
+                        }
+                    },
+                }
+            }
+        )
+        self.assertEqual(payload["message_type"], "post")
+        self.assertEqual(len(payload["attachments"]), 1)
+        self.assertIn("请评估这张图的风险点", str(payload["text"] or ""))
+
+
+if __name__ == "__main__":
+    unittest.main()
+
diff --git a/backend/tests/test_task_persistence.py b/backend/tests/test_task_persistence.py
new file mode 100644
index 00000000..54eaf05c
--- /dev/null
+++ b/backend/tests/test_task_persistence.py
@@ -0,0 +1,80 @@
+import unittest
+
+from app.models import SkillTag
+from app.services.task_persistence import (
+    build_task_create_draft,
+    resolve_skill_tag_ids_by_names,
+)
+
+
+class _ScalarRows:
+    def __init__(self, rows: list[object]) -> None:
+        self._rows = rows
+
+    def scalars(self):
+        return self
+
+    def all(self):
+        return self._rows
+
+
+class _FakeDB:
+    def __init__(self, rows: list[object]) -> None:
+        self._rows = rows
+
+    async def execute(self, *_args, **_kwargs):
+        return _ScalarRows(self._rows)
+
+
+class TaskPersistenceTest(unittest.IsolatedAsyncioTestCase):
+    def test_build_task_create_draft_normalizes_fields(self) -> None:
+        draft = build_task_create_draft(
+            title="  补齐统一通知入口  ",
+            required_skill_tag_ids=["1", 2, "2"],
+            priority="P0",
+            status="pending",
+            estimated_hours="4.5",
+            due_date="2026-03-07T10:00:00Z",
+        )
+
+        self.assertEqual(draft.title, "补齐统一通知入口")
+        self.assertEqual(draft.required_skill_tag_ids, [1, 2])
+        self.assertEqual(draft.priority, "high")
+        self.assertEqual(draft.status, "todo")
+        self.assertEqual(draft.estimated_hours, 4.5)
+        self.assertIsNotNone(draft.due_date)
+        self.assertIsNone(draft.due_date.tzinfo)
+
+    async def test_resolve_skill_tag_ids_by_names_preserves_unknowns_for_audit(self) -> None:
+        db = _FakeDB(
+            [
+                SkillTag(id=11, tenant_id=1, name="系统设计", slug="system-design", category="skill"),
+                SkillTag(id=12, tenant_id=1, name="开发", slug="development", category="skill"),
+            ]
+        )
+
+        resolution = await resolve_skill_tag_ids_by_names(
+            db,
+            tenant_id=1,
+            skill_names=["系统设计", "开发", "未知技能"],
+        )
+
+        self.assertEqual(resolution.matched_ids, [11, 12])
+        self.assertEqual(resolution.unmatched_names, ["未知技能"])
+        self.assertEqual(
+            resolution.matched_items,
+            [
+                {
+                    "input_name": "系统设计",
+                    "skill_tag_id": 11,
+                    "skill_tag_name": "系统设计",
+                    "skill_tag_slug": "system-design",
+                },
+                {
+                    "input_name": "开发",
+                    "skill_tag_id": 12,
+                    "skill_tag_name": "开发",
+                    "skill_tag_slug": "development",
+                },
+            ],
+        )
diff --git a/devops/tunnel/setup.sh b/devops/tunnel/setup.sh
new file mode 100755
index 00000000..545746ce
--- /dev/null
+++ b/devops/tunnel/setup.sh
@@ -0,0 +1,105 @@
+#!/bin/bash
+
+# Cloudflare Tunnel Setup Script for PMFlow
+
+set -e
+
+echo "🚀 Starting Cloudflare Tunnel Setup..."
+
+# Check if cloudflared is installed
+if ! command -v cloudflared &> /dev/null; then
+    echo "❌ cloudflared is not installed. Please install it first:"
+    echo "   brew install cloudflared"
+    exit 1
+fi
+
+# Ask for configuration
+read -p "Enter your Frontend domain (e.g., pmflow.com): " FRONTEND_DOMAIN
+read -p "Enter your Backend domain (e.g., api.pmflow.com): " BACKEND_DOMAIN
+read -p "Enter a name for your tunnel (e.g., pmflow-tunnel): " TUNNEL_NAME
+
+if [ -z "$FRONTEND_DOMAIN" ] || [ -z "$BACKEND_DOMAIN" ] || [ -z "$TUNNEL_NAME" ]; then
+    echo "❌ Frontend domain, Backend domain, and Tunnel name are required."
+    exit 1
+fi
+
+echo "Checking authentication..."
+# Check if cert.pem exists in default location
+if [ ! -f "$HOME/.cloudflared/cert.pem" ]; then
+    echo "⚠️  You need to login to Cloudflare first."
+    cloudflared tunnel login
+fi
+
+echo "Creating tunnel '$TUNNEL_NAME'..."
+# Check if tunnel exists, if not create it
+if ! cloudflared tunnel list | grep -q "$TUNNEL_NAME"; then
+    cloudflared tunnel create "$TUNNEL_NAME"
+else
+    echo "ℹ️  Tunnel '$TUNNEL_NAME' already exists."
+fi
+
+echo "Routing DNS..."
+# Route DNS for both domains
+cloudflared tunnel route dns "$TUNNEL_NAME" "$FRONTEND_DOMAIN"
+cloudflared tunnel route dns "$TUNNEL_NAME" "$BACKEND_DOMAIN"
+
+# Get Tunnel ID
+TUNNEL_ID=$(cloudflared tunnel list | grep "$TUNNEL_NAME" | awk '{print $1}')
+
+if [ -z "$TUNNEL_ID" ]; then
+    echo "❌ Could not find Tunnel ID for '$TUNNEL_NAME'"
+    exit 1
+fi
+
+echo "Generating config file..."
+# Create config.yml
+mkdir -p devops/tunnel
+cat > devops/tunnel/config.yml <<EOF
+tunnel: $TUNNEL_ID
+credentials-file: $HOME/.cloudflared/$TUNNEL_ID.json
+
+ingress:
+  - hostname: $FRONTEND_DOMAIN
+    service: http://localhost:5174
+  - hostname: $BACKEND_DOMAIN
+    service: http://localhost:8000
+  - service: http_status:404
+EOF
+
+echo "✅ Tunnel configuration created at devops/tunnel/config.yml"
+
+# Update .env file in frontend
+ENV_FILE="frontend/.env"
+if [ ! -f "$ENV_FILE" ]; then
+    touch "$ENV_FILE"
+fi
+
+# Update VITE_APP_DOMAIN and VITE_API_BASE_URL
+if [[ "$OSTYPE" == "darwin"* ]]; then
+    sed -i '' "s|VITE_APP_DOMAIN=.*|VITE_APP_DOMAIN=$FRONTEND_DOMAIN|" "$ENV_FILE"
+    # Check if VITE_API_BASE_URL exists, if so replace it, else append
+    if grep -q "VITE_API_BASE_URL=" "$ENV_FILE"; then
+        sed -i '' "s|VITE_API_BASE_URL=.*|VITE_API_BASE_URL=https://$BACKEND_DOMAIN|" "$ENV_FILE"
+    else
+        echo "VITE_API_BASE_URL=https://$BACKEND_DOMAIN" >> "$ENV_FILE"
+    fi
+else
+    sed -i "s|VITE_APP_DOMAIN=.*|VITE_APP_DOMAIN=$FRONTEND_DOMAIN|" "$ENV_FILE"
+    if grep -q "VITE_API_BASE_URL=" "$ENV_FILE"; then
+        sed -i "s|VITE_API_BASE_URL=.*|VITE_API_BASE_URL=https://$BACKEND_DOMAIN|" "$ENV_FILE"
+    else
+        echo "VITE_API_BASE_URL=https://$BACKEND_DOMAIN" >> "$ENV_FILE"
+    fi
+fi
+
+if ! grep -q "VITE_APP_DOMAIN=" "$ENV_FILE"; then
+    echo "VITE_APP_DOMAIN=$FRONTEND_DOMAIN" >> "$ENV_FILE"
+fi
+
+echo "✅ Updated $ENV_FILE with:"
+echo "   VITE_APP_DOMAIN=$FRONTEND_DOMAIN"
+echo "   VITE_API_BASE_URL=https://$BACKEND_DOMAIN"
+
+echo "🎉 Setup complete! You can now run the tunnel with:"
+echo "   ./devops/tunnel/start.sh"
+
diff --git a/devops/tunnel/start-quick.sh b/devops/tunnel/start-quick.sh
new file mode 100755
index 00000000..ef2177ae
--- /dev/null
+++ b/devops/tunnel/start-quick.sh
@@ -0,0 +1,67 @@
+#!/bin/bash
+
+# Cloudflare Quick Tunnel (No Domain Required)
+# Generates a random stable-ish URL for the session
+
+echo "🚀 Starting Quick Tunnel..."
+echo "⚠️  Note: This uses a random trycloudflare.com URL."
+echo "⚠️  For a fixed custom domain, you must PURCHASE a domain and configure it via setup.sh"
+
+# Kill existing tunnels
+pkill -f "cloudflared tunnel" >/dev/null 2>&1 || true
+
+echo "🌐 Starting Frontend Tunnel (port 5174)..."
+cloudflared tunnel --url http://localhost:5174 > logs/tunnel-frontend.log 2>&1 &
+PID_FRONT=$!
+
+echo "🌐 Starting Backend Tunnel (port 8000)..."
+cloudflared tunnel --url http://localhost:8000 > logs/tunnel-backend.log 2>&1 &
+PID_BACK=$!
+
+echo "⏳ Waiting for tunnels to initialize..."
+sleep 5
+
+URL_FRONT=$(grep -o 'https://.*\.trycloudflare.com' logs/tunnel-frontend.log | head -1)
+URL_BACK=$(grep -o 'https://.*\.trycloudflare.com' logs/tunnel-backend.log | head -1)
+
+check_url() {
+    local url="$1"
+    local max_retry="$2"
+    local retry=1
+    while [ "$retry" -le "$max_retry" ]; do
+        if curl -kfsS --max-time 8 "$url" >/dev/null 2>&1; then
+            return 0
+        fi
+        sleep 2
+        retry=$((retry + 1))
+    done
+    return 1
+}
+
+if [ -n "$URL_FRONT" ] && [ -n "$URL_BACK" ]; then
+    FRONT_OK=false
+    BACK_OK=false
+    if check_url "${URL_FRONT}/login" 25 || check_url "${URL_FRONT}" 25; then
+        FRONT_OK=true
+    fi
+    if check_url "${URL_BACK}/healthz" 25; then
+        BACK_OK=true
+    fi
+
+    echo ""
+    if [ "$FRONT_OK" = true ] && [ "$BACK_OK" = true ]; then
+        echo "✅ Tunnels Started and reachable!"
+    else
+        echo "⚠️ Tunnels started, but health check not fully ready yet. Retry in 10-30s."
+    fi
+    echo "----------------------------------------"
+    echo "Frontend: $URL_FRONT"
+    echo "Backend:  $URL_BACK"
+    echo "----------------------------------------"
+    echo "ℹ️ Frontend tunnel can call backend via /api proxy; no need to set VITE_API_BASE_URL."
+    echo "----------------------------------------"
+else
+    echo "❌ Failed to get tunnel URLs. Check logs/tunnel-*.log"
+fi
+
+wait $PID_FRONT $PID_BACK
diff --git a/devops/tunnel/start.sh b/devops/tunnel/start.sh
new file mode 100755
index 00000000..51eb4cbc
--- /dev/null
+++ b/devops/tunnel/start.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+
+# Cloudflare Tunnel Start Script for PMFlow
+
+CONFIG_FILE="devops/tunnel/config.yml"
+SETUP_SCRIPT="devops/tunnel/setup.sh"
+
+echo "🚀 Checking Cloudflare Tunnel configuration..."
+
+# Check if config.yml exists
+if [ ! -f "$CONFIG_FILE" ]; then
+    echo "⚠️  Named Tunnel Configuration not found."
+    echo "   If you have a domain, you can set up a stable tunnel."
+    echo "   If not, you can use Quick Tunnels (random URLs)."
+    echo ""
+    read -p "Do you have a domain and want to set up a named tunnel? (y/N): " SETUP_CHOICE
+    
+    if [[ "$SETUP_CHOICE" =~ ^[Yy]$ ]]; then
+        if [ -f "$SETUP_SCRIPT" ]; then
+            # Run setup script and check its exit code
+            if ! bash "$SETUP_SCRIPT"; then
+                echo "❌ Setup failed or was cancelled."
+                exit 1
+            fi
+        else
+            echo "❌ Setup script not found at $SETUP_SCRIPT"
+            exit 1
+        fi
+    else
+        echo "🔄 Falling back to Quick Tunnels..."
+        bash devops/tunnel/start-quick.sh
+        exit 0
+    fi
+fi
+
+# Re-check config after potential setup
+if [ -f "$CONFIG_FILE" ]; then
+    echo "✅ Configuration found."
+    echo "🌐 Starting tunnel for Frontend and Backend..."
+    
+    # Run the tunnel using the config file
+    cloudflared tunnel run --config "$CONFIG_FILE"
+else
+    echo "❌ Failed to establish configuration."
+    exit 1
+fi
diff --git a/frontend/src/components/ai-steward/OpenClawCommanderPanel.tsx b/frontend/src/components/ai-steward/OpenClawCommanderPanel.tsx
new file mode 100644
index 00000000..27464b2d
--- /dev/null
+++ b/frontend/src/components/ai-steward/OpenClawCommanderPanel.tsx
@@ -0,0 +1,390 @@
+import { useEffect, useMemo, useState } from "react";
+import {
+  AlertCircle,
+  Bot,
+  CheckCircle2,
+  Loader2,
+  Play,
+  RefreshCcw,
+  ShieldAlert,
+  Wand2,
+} from "lucide-react";
+
+import { aiStewardApi, type OpenClawCommandDispatchResponse } from "@/api/aiSteward";
+import { projectsApi, type ProjectSummary } from "@/api/projects";
+import { Badge } from "@/components/ui/badge";
+import { Button } from "@/components/ui/button";
+import { Card, CardContent, CardDescription, CardHeader, CardTitle } from "@/components/ui/card";
+import { Input } from "@/components/ui/input";
+import { Label } from "@/components/ui/label";
+import { Switch } from "@/components/ui/switch";
+import { Textarea } from "@/components/ui/textarea";
+import { formatProjectNameWithId, toZhStatus } from "@/lib/display";
+
+const EXAMPLE_COMMANDS = [
+  "把当前未分配的高优先级任务分给最合适的人，并通知相关成员。",
+  "根据当前负载重新平衡本项目任务，先 dry run 给出调整方案。",
+  "检查 blocked 任务，能恢复推进的直接切状态，不能恢复的通知负责人补充信息。",
+];
+
+const MANAGER_ROLES = new Set(["BOSS", "ADMIN"]);
+
+function resultTone(status: string): string {
+  const key = String(status || "").trim().toLowerCase();
+  if (key === "applied") return "bg-green-50 text-green-700 border-green-200 dark:bg-green-900/20 dark:text-green-300 dark:border-green-800";
+  if (key === "planned") return "bg-blue-50 text-blue-700 border-blue-200 dark:bg-blue-900/20 dark:text-blue-300 dark:border-blue-800";
+  if (key === "skipped") return "bg-amber-50 text-amber-700 border-amber-200 dark:bg-amber-900/20 dark:text-amber-300 dark:border-amber-800";
+  return "bg-red-50 text-red-700 border-red-200 dark:bg-red-900/20 dark:text-red-300 dark:border-red-800";
+}
+
+function renderOperationLabel(operationCode: string): string {
+  const code = String(operationCode || "").trim().toLowerCase();
+  if (code === "task.create") return "任务创建";
+  if (code === "task.assign") return "任务分配";
+  if (code === "task.prioritize") return "优先级调整";
+  if (code === "task.status_transition") return "状态迁移";
+  if (code === "notification.create") return "通知创建";
+  if (code === "personnel.workload_update") return "工时容量调整";
+  return operationCode || "-";
+}
+
+interface OpenClawCommanderPanelProps {
+  currentUserRole?: string | null;
+}
+
+export function OpenClawCommanderPanel({ currentUserRole }: OpenClawCommanderPanelProps) {
+  const [projects, setProjects] = useState<ProjectSummary[]>([]);
+  const [projectLoading, setProjectLoading] = useState(false);
+  const [projectId, setProjectId] = useState("");
+  const [instruction, setInstruction] = useState("");
+  const [dryRun, setDryRun] = useState(true);
+  const [maxOperations, setMaxOperations] = useState("6");
+  const [dispatching, setDispatching] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const [result, setResult] = useState<OpenClawCommandDispatchResponse | null>(null);
+
+  const canDispatch = MANAGER_ROLES.has(String(currentUserRole || "").trim().toUpperCase());
+  const selectedProject = useMemo(
+    () => projects.find((item) => String(item.id) === String(projectId)) || null,
+    [projectId, projects]
+  );
+
+  const loadProjects = async () => {
+    setProjectLoading(true);
+    try {
+      const payload = await projectsApi.getSummary();
+      const items = payload.projects || [];
+      setProjects(items);
+      if (!projectId && items[0]?.id) {
+        setProjectId(String(items[0].id));
+      }
+    } catch (loadError) {
+      console.error("load commander projects failed", loadError);
+      setError("加载项目列表失败");
+    } finally {
+      setProjectLoading(false);
+    }
+  };
+
+  useEffect(() => {
+    if (!canDispatch) {
+      return;
+    }
+    void loadProjects();
+  }, [canDispatch]);
+
+  const handleDispatch = async () => {
+    const parsedProjectId = Number(projectId);
+    const parsedMaxOperations = Number(maxOperations);
+
+    if (!Number.isInteger(parsedProjectId) || parsedProjectId <= 0) {
+      setError("请选择有效项目");
+      return;
+    }
+    if (!instruction.trim()) {
+      setError("请输入调度指令");
+      return;
+    }
+    if (!Number.isInteger(parsedMaxOperations) || parsedMaxOperations < 1 || parsedMaxOperations > 20) {
+      setError("max operations 仅支持 1-20");
+      return;
+    }
+
+    setDispatching(true);
+    setError(null);
+    try {
+      const payload = await aiStewardApi.dispatchCommand({
+        project_id: parsedProjectId,
+        instruction: instruction.trim(),
+        dry_run: dryRun,
+        max_operations: parsedMaxOperations,
+      });
+      setResult(payload);
+      if (!dryRun) {
+        void loadProjects();
+      }
+    } catch (dispatchError: any) {
+      const text = typeof dispatchError?.message === "string" ? dispatchError.message : "指挥调度失败";
+      setError(text);
+    } finally {
+      setDispatching(false);
+    }
+  };
+
+  if (!canDispatch) {
+    return (
+      <Card className="border-amber-200 dark:border-amber-800">
+        <CardContent className="py-10 text-center space-y-3">
+          <ShieldAlert className="w-8 h-8 mx-auto text-amber-500" />
+          <div className="text-lg font-semibold">Commander 仅开放给管理角色</div>
+          <div className="text-sm text-slate-600 dark:text-slate-400">
+            当前接口要求 `BOSS / ADMIN`，普通成员仍可通过任务看板或 AI 问答拿建议，但不能直接下发执行。
+          </div>
+        </CardContent>
+      </Card>
+    );
+  }
+
+  return (
+    <div className="space-y-6">
+      <Card className="border-slate-200 dark:border-slate-800">
+        <CardHeader>
+          <div className="flex items-center gap-3">
+            <div className="w-11 h-11 rounded-2xl bg-slate-900 text-white dark:bg-slate-100 dark:text-slate-900 flex items-center justify-center">
+              <Bot className="w-5 h-5" />
+            </div>
+            <div>
+              <CardTitle>OpenClaw Commander</CardTitle>
+              <CardDescription>直接下达自然语言指令，让指挥层规划并驱动 AI 管家执行。</CardDescription>
+            </div>
+          </div>
+        </CardHeader>
+        <CardContent className="grid gap-6 lg:grid-cols-[320px_minmax(0,1fr)]">
+          <div className="space-y-4">
+            <div className="flex items-center justify-between">
+              <Label>项目范围</Label>
+              <Button variant="outline" size="sm" onClick={() => void loadProjects()} disabled={projectLoading}>
+                {projectLoading ? <Loader2 className="w-4 h-4 animate-spin" /> : <RefreshCcw className="w-4 h-4" />}
+              </Button>
+            </div>
+            <div className="space-y-2 max-h-[460px] overflow-y-auto pr-1">
+              {projects.length === 0 ? (
+                <div className="rounded-lg border border-dashed border-slate-300 p-4 text-sm text-slate-500 dark:border-slate-700">
+                  {projectLoading ? "正在加载项目..." : "暂无可调度项目"}
+                </div>
+              ) : (
+                projects.map((project) => {
+                  const active = String(project.id) === String(projectId);
+                  return (
+                    <button
+                      key={project.id}
+                      type="button"
+                      onClick={() => setProjectId(String(project.id))}
+                      className={`w-full rounded-xl border p-4 text-left transition-colors ${
+                        active
+                          ? "border-slate-900 bg-slate-900 text-white dark:border-slate-100 dark:bg-slate-100 dark:text-slate-900"
+                          : "border-slate-200 bg-white hover:border-slate-300 dark:border-slate-800 dark:bg-slate-950 dark:hover:border-slate-700"
+                      }`}
+                    >
+                      <div className="flex items-start justify-between gap-3">
+                        <div className="font-medium">{formatProjectNameWithId(project)}</div>
+                        <Badge variant={active ? "secondary" : "outline"}>{toZhStatus(project.status)}</Badge>
+                      </div>
+                      <div className={`mt-2 text-xs ${active ? "text-slate-200 dark:text-slate-700" : "text-slate-500 dark:text-slate-400"}`}>
+                        任务 {project.total_tasks} / 完成 {project.completed_tasks} / 进行中 {project.in_progress_tasks}
+                      </div>
+                      {project.description ? (
+                        <div className={`mt-2 line-clamp-2 text-xs ${active ? "text-slate-200 dark:text-slate-700" : "text-slate-500 dark:text-slate-400"}`}>
+                          {project.description}
+                        </div>
+                      ) : null}
+                    </button>
+                  );
+                })
+              )}
+            </div>
+          </div>
+
+          <div className="space-y-4">
+            <div className="grid gap-4 md:grid-cols-[minmax(0,1fr)_160px]">
+              <div className="space-y-2">
+                <Label htmlFor="openclaw-project-id">项目 ID</Label>
+                <Input
+                  id="openclaw-project-id"
+                  value={projectId}
+                  onChange={(event) => setProjectId(event.target.value)}
+                  placeholder="例如 1"
+                />
+              </div>
+              <div className="space-y-2">
+                <Label htmlFor="openclaw-max-ops">Max Operations</Label>
+                <Input
+                  id="openclaw-max-ops"
+                  value={maxOperations}
+                  onChange={(event) => setMaxOperations(event.target.value)}
+                  inputMode="numeric"
+                  placeholder="6"
+                />
+              </div>
+            </div>
+
+            {selectedProject ? (
+              <div className="rounded-xl border border-slate-200 bg-slate-50 p-4 text-sm dark:border-slate-800 dark:bg-slate-900/60">
+                <div className="font-medium">{formatProjectNameWithId(selectedProject)}</div>
+                <div className="mt-1 text-slate-600 dark:text-slate-400">
+                  Active {selectedProject.in_progress_tasks} / Pending {selectedProject.pending_tasks} / Blocked 由执行结果返回。
+                </div>
+              </div>
+            ) : null}
+
+            <div className="space-y-2">
+              <Label htmlFor="openclaw-instruction">调度指令</Label>
+              <Textarea
+                id="openclaw-instruction"
+                rows={6}
+                value={instruction}
+                onChange={(event) => setInstruction(event.target.value)}
+                placeholder="例如：把当前未分配的高优先级任务分给最合适的人，并给新负责人发通知。"
+              />
+              <div className="flex flex-wrap gap-2">
+                {EXAMPLE_COMMANDS.map((item) => (
+                  <button
+                    key={item}
+                    type="button"
+                    className="rounded-full border border-slate-200 px-3 py-1 text-xs text-slate-600 transition-colors hover:border-slate-300 hover:text-slate-900 dark:border-slate-800 dark:text-slate-300 dark:hover:border-slate-700 dark:hover:text-slate-100"
+                    onClick={() => setInstruction(item)}
+                  >
+                    {item}
+                  </button>
+                ))}
+              </div>
+            </div>
+
+            <div className="flex flex-wrap items-center justify-between gap-4 rounded-xl border border-slate-200 p-4 dark:border-slate-800">
+              <div className="space-y-1">
+                <div className="text-sm font-medium">Dry Run</div>
+                <div className="text-xs text-slate-500 dark:text-slate-400">
+                  开启后只生成计划，不落库执行。
+                </div>
+              </div>
+              <Switch checked={dryRun} onCheckedChange={setDryRun} />
+            </div>
+
+            {error ? (
+              <div className="rounded-xl border border-red-200 bg-red-50 px-4 py-3 text-sm text-red-700 dark:border-red-800 dark:bg-red-900/20 dark:text-red-300">
+                <div className="flex items-center gap-2">
+                  <AlertCircle className="w-4 h-4" />
+                  {error}
+                </div>
+              </div>
+            ) : null}
+
+            <Button onClick={handleDispatch} disabled={dispatching || projectLoading} className="w-full bg-slate-900 hover:bg-slate-800 dark:bg-slate-100 dark:text-slate-900 dark:hover:bg-slate-200">
+              {dispatching ? (
+                <>
+                  <Loader2 className="mr-2 w-4 h-4 animate-spin" />
+                  Commander 执行中...
+                </>
+              ) : (
+                <>
+                  <Play className="mr-2 w-4 h-4" />
+                  {dryRun ? "生成调度计划" : "执行调度指令"}
+                </>
+              )}
+            </Button>
+          </div>
+        </CardContent>
+      </Card>
+
+      {result ? (
+        <Card className="border-slate-200 dark:border-slate-800">
+          <CardHeader>
+            <div className="flex flex-wrap items-center gap-3">
+              <CardTitle className="mr-auto">执行结果</CardTitle>
+              <Badge variant="outline">{result.planner_status}</Badge>
+              <Badge variant="secondary">{result.prompt_role.role_name || result.prompt_role.role_key || "default"}</Badge>
+              <Badge variant={result.dry_run ? "outline" : "default"}>{result.dry_run ? "Dry Run" : "Executed"}</Badge>
+            </div>
+            <CardDescription>{result.summary || "OpenClaw 已完成规划。"}</CardDescription>
+          </CardHeader>
+          <CardContent className="space-y-4">
+            <div className="grid gap-3 md:grid-cols-4">
+              <div className="rounded-xl border border-slate-200 p-4 dark:border-slate-800">
+                <div className="text-xs text-slate-500 dark:text-slate-400">Trace</div>
+                <div className="mt-1 break-all text-sm font-medium">{result.trace_id}</div>
+              </div>
+              <div className="rounded-xl border border-slate-200 p-4 dark:border-slate-800">
+                <div className="text-xs text-slate-500 dark:text-slate-400">项目成员</div>
+                <div className="mt-1 text-sm font-medium">{result.member_count}</div>
+              </div>
+              <div className="rounded-xl border border-slate-200 p-4 dark:border-slate-800">
+                <div className="text-xs text-slate-500 dark:text-slate-400">计划操作</div>
+                <div className="mt-1 text-sm font-medium">{result.planned_operations}</div>
+              </div>
+              <div className="rounded-xl border border-slate-200 p-4 dark:border-slate-800">
+                <div className="text-xs text-slate-500 dark:text-slate-400">可信度</div>
+                <div className="mt-1 text-sm font-medium">{Math.round((result.confidence || 0) * 100)}%</div>
+              </div>
+            </div>
+
+            <div className="grid gap-3 md:grid-cols-4">
+              <div className="rounded-xl border border-green-200 bg-green-50 p-4 dark:border-green-800 dark:bg-green-900/20">
+                <div className="text-xs text-green-700 dark:text-green-300">Applied</div>
+                <div className="mt-1 text-2xl font-semibold text-green-800 dark:text-green-100">{result.execution.applied_count}</div>
+              </div>
+              <div className="rounded-xl border border-blue-200 bg-blue-50 p-4 dark:border-blue-800 dark:bg-blue-900/20">
+                <div className="text-xs text-blue-700 dark:text-blue-300">Requested</div>
+                <div className="mt-1 text-2xl font-semibold text-blue-800 dark:text-blue-100">{result.execution.requested_count}</div>
+              </div>
+              <div className="rounded-xl border border-amber-200 bg-amber-50 p-4 dark:border-amber-800 dark:bg-amber-900/20">
+                <div className="text-xs text-amber-700 dark:text-amber-300">Skipped</div>
+                <div className="mt-1 text-2xl font-semibold text-amber-800 dark:text-amber-100">{result.execution.skipped_count}</div>
+              </div>
+              <div className="rounded-xl border border-red-200 bg-red-50 p-4 dark:border-red-800 dark:bg-red-900/20">
+                <div className="text-xs text-red-700 dark:text-red-300">Failed</div>
+                <div className="mt-1 text-2xl font-semibold text-red-800 dark:text-red-100">{result.execution.failed_count}</div>
+              </div>
+            </div>
+
+            <div className="space-y-3">
+              {result.execution.results.length === 0 ? (
+                <div className="rounded-xl border border-dashed border-slate-300 p-5 text-sm text-slate-500 dark:border-slate-700">
+                  当前没有可执行 operation，通常意味着上下文不足，或者指令被安全约束拦住了。
+                </div>
+              ) : (
+                result.execution.results.map((item) => (
+                  <div key={`${item.index}-${item.operation_code}-${item.idempotency_key || "no-key"}`} className="rounded-xl border border-slate-200 p-4 dark:border-slate-800">
+                    <div className="flex flex-wrap items-center gap-2">
+                      <Badge variant="outline">#{item.index}</Badge>
+                      <div className="font-medium">{renderOperationLabel(item.operation_code)}</div>
+                      <div className={`rounded-full border px-2.5 py-0.5 text-xs ${resultTone(item.status)}`}>
+                        {item.status}
+                      </div>
+                      {item.entity_type && item.entity_id ? (
+                        <Badge variant="secondary">
+                          {item.entity_type}:{item.entity_id}
+                        </Badge>
+                      ) : null}
+                    </div>
+                    <div className="mt-2 flex items-start gap-2 text-sm text-slate-600 dark:text-slate-300">
+                      {item.status === "applied" ? <CheckCircle2 className="mt-0.5 w-4 h-4 text-green-500" /> : <Wand2 className="mt-0.5 w-4 h-4 text-slate-400" />}
+                      <div>{item.message}</div>
+                    </div>
+                    <pre className="mt-3 overflow-x-auto rounded-lg bg-slate-950 p-3 text-xs text-slate-100">
+                      {JSON.stringify(item.payload, null, 2)}
+                    </pre>
+                    {item.trace_id ? (
+                      <div className="mt-2 text-xs text-slate-500 dark:text-slate-400">trace: {item.trace_id}</div>
+                    ) : null}
+                  </div>
+                ))
+              )}
+            </div>
+          </CardContent>
+        </Card>
+      ) : null}
+    </div>
+  );
+}
+
+export default OpenClawCommanderPanel;
diff --git a/frontend/src/pages/InboxPage.tsx b/frontend/src/pages/InboxPage.tsx
new file mode 100644
index 00000000..aed99a60
--- /dev/null
+++ b/frontend/src/pages/InboxPage.tsx
@@ -0,0 +1,899 @@
+import { useMemo, useState } from "react";
+import { useNavigate, useSearchParams } from "react-router-dom";
+import { useQuery, useQueryClient } from "@tanstack/react-query";
+import {
+  BellRing,
+  CheckCircle2,
+  Filter,
+  LoaderCircle,
+  ShieldAlert,
+  SlidersHorizontal,
+  Sparkles,
+} from "lucide-react";
+import { aiPmoApi } from "@/api/aiPmo";
+import { notificationsApi } from "@/api/notifications";
+import { Badge } from "@/components/ui/badge";
+import { Button } from "@/components/ui/button";
+import { Card, CardContent, CardDescription, CardHeader, CardTitle } from "@/components/ui/card";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import { Input } from "@/components/ui/input";
+import { Switch } from "@/components/ui/switch";
+import { Tabs, TabsList, TabsTrigger } from "@/components/ui/tabs";
+import { useToast } from "@/components/ui/toast";
+import { buildInboxView, getInboxSectionLabel } from "@/lib/inbox";
+import { getApiErrorMessage } from "@/lib/apiError";
+import { cn } from "@/lib/utils";
+import { toZhPriority } from "@/lib/display";
+import { parseBackendDate } from "@/lib/datetime";
+import type { InboxActionDescriptor, InboxDisplayKind, InboxItemViewModel, InboxSectionKey } from "@/types/inbox";
+
+const POLL_INTERVAL_MS = 15000;
+
+const STATUS_LABELS: Record<InboxDisplayKind, string> = {
+  action_required: "需你决策",
+  auto_running: "AI 已接手",
+  auto_done: "AI 执行完成",
+  info: "仅通知",
+};
+
+const STATUS_STYLES: Record<InboxDisplayKind, string> = {
+  action_required: "border-amber-500/40 bg-amber-500/10 text-amber-700 dark:text-amber-300",
+  auto_running: "border-sky-500/40 bg-sky-500/10 text-sky-700 dark:text-sky-300",
+  auto_done: "border-emerald-500/40 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300",
+  info: "border-slate-400/40 bg-slate-500/10 text-slate-700 dark:text-slate-300",
+};
+
+const METRIC_CARD_STYLES = [
+  "bg-blue-50 text-blue-700",
+  "bg-sky-50 text-sky-700",
+  "bg-emerald-50 text-emerald-700",
+  "bg-amber-50 text-amber-700",
+  "bg-violet-50 text-violet-700",
+];
+
+const PRIORITY_OPTIONS = [
+  { value: "all", label: "全部优先级" },
+  { value: "high", label: "高优先级" },
+  { value: "medium", label: "中优先级" },
+  { value: "low", label: "低优先级" },
+];
+
+function matchesPriority(item: InboxItemViewModel, priorityFilter: string): boolean {
+  if (priorityFilter === "all") return true;
+  if (priorityFilter === "high") return item.priority_rank >= 3;
+  if (priorityFilter === "medium") return item.priority_rank === 2;
+  return item.priority_rank <= 1;
+}
+
+function buildScopedSections(items: InboxItemViewModel[]) {
+  const sectionKeys: InboxSectionKey[] = ["all", "action_required", "auto_running", "auto_done", "info"];
+  return sectionKeys.map((key) => ({
+    key,
+    label: getInboxSectionLabel(key),
+    count: key === "all" ? items.length : items.filter((item) => item.display_kind === key).length,
+  }));
+}
+
+function isToday(value?: string | null) {
+  const date = parseBackendDate(value);
+  if (!date) return false;
+  const now = new Date();
+  return (
+    date.getFullYear() === now.getFullYear() &&
+    date.getMonth() === now.getMonth() &&
+    date.getDate() === now.getDate()
+  );
+}
+
+function formatDetailTime(value?: string | null) {
+  const date = parseBackendDate(value);
+  if (!date) return value || "-";
+  return date.toLocaleString("zh-CN", { hour12: false });
+}
+
+function formatJsonBlock(value: unknown): string {
+  if (!value) return "暂无";
+  try {
+    return JSON.stringify(value, null, 2);
+  } catch {
+    return String(value);
+  }
+}
+
+function buildActionKey(item: InboxItemViewModel, action: InboxActionDescriptor): string {
+  return `${item.id}:${action.kind}:${action.label}`;
+}
+
+function renderMiniList(title: string, description: string, items: InboxItemViewModel[], onOpen: (item: InboxItemViewModel) => void) {
+  return (
+    <Card className="rounded-2xl border-0 bg-white shadow-[0_8px_22px_rgba(15,23,42,0.08)]">
+      <CardHeader className="pb-3">
+        <CardTitle className="text-base text-slate-900">{title}</CardTitle>
+        <CardDescription className="text-slate-500">{description}</CardDescription>
+      </CardHeader>
+      <CardContent className="space-y-3">
+        {items.length === 0 ? (
+          <div className="rounded-xl border border-dashed border-slate-200 bg-slate-50 px-3 py-4 text-sm text-slate-500">
+            当前为空
+          </div>
+        ) : (
+          items.map((item) => (
+            <button
+              key={item.id}
+              type="button"
+              className="w-full rounded-xl border border-slate-100 bg-slate-50 px-3 py-3 text-left transition hover:border-slate-200 hover:bg-white"
+              onClick={() => onOpen(item)}
+            >
+              <div className="flex items-start justify-between gap-3">
+                <span className="line-clamp-1 text-sm font-semibold text-slate-900">{item.title}</span>
+                <Badge variant="outline" className={cn("text-[11px]", STATUS_STYLES[item.display_kind])}>
+                  {STATUS_LABELS[item.display_kind]}
+                </Badge>
+              </div>
+              <p className="mt-1 line-clamp-2 text-xs text-slate-500">{item.summary}</p>
+            </button>
+          ))
+        )}
+      </CardContent>
+    </Card>
+  );
+}
+
+function InboxCard({
+  item,
+  onAction,
+  pendingActionKey,
+}: {
+  item: InboxItemViewModel;
+  onAction: (item: InboxItemViewModel, action: InboxActionDescriptor) => void;
+  pendingActionKey: string | null;
+}) {
+  const stageCards = [
+    { label: "当前阶段", value: item.stage_text || STATUS_LABELS[item.display_kind] },
+    { label: "预计下一步", value: item.next_step_text || item.ai_plan || "-" },
+    { label: "撤销状态", value: item.can_undo ? `可撤销，截止 ${item.undo_deadline || "-"}` : "不可撤销" },
+  ];
+  const primaryActionKey = buildActionKey(item, item.primary_action);
+  const secondaryActionKey = buildActionKey(item, item.secondary_action);
+  const weakActionKey = buildActionKey(item, item.weak_action);
+  const primaryPending = pendingActionKey === primaryActionKey;
+  const secondaryPending = pendingActionKey === secondaryActionKey;
+  const weakPending = pendingActionKey === weakActionKey;
+
+  return (
+    <Card className="rounded-2xl border-0 bg-white shadow-[0_8px_22px_rgba(15,23,42,0.08)] transition hover:-translate-y-0.5 hover:shadow-[0_14px_30px_rgba(15,23,42,0.12)]">
+      <CardHeader className="space-y-3 pb-3">
+        <div className="flex flex-wrap items-start justify-between gap-3">
+          <div className="flex flex-wrap items-center gap-2">
+            <Badge variant="outline" className={cn("text-[11px]", STATUS_STYLES[item.display_kind])}>
+              {STATUS_LABELS[item.display_kind]}
+            </Badge>
+            <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+              {toZhPriority(item.priority)}
+            </span>
+            {item.project_id && (
+              <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                {item.project_name || `项目 #${item.project_id}`}
+              </span>
+            )}
+            {item.dedupe_count > 1 && (
+              <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                重复 {item.dedupe_count} 次
+              </span>
+            )}
+          </div>
+          <div className="text-right">
+            <div className="text-xs text-slate-400">{item.updated_at || item.created_at || "-"}</div>
+          </div>
+        </div>
+        <div className="flex items-start justify-between gap-3">
+          <div className="space-y-1">
+            <CardTitle className="text-base text-slate-900">{item.title}</CardTitle>
+            <CardDescription className="line-clamp-2 text-slate-500">{item.summary}</CardDescription>
+          </div>
+          <BellRing className="mt-0.5 h-4 w-4 shrink-0 text-slate-400" />
+        </div>
+      </CardHeader>
+      <CardContent className="space-y-4">
+        {item.ai_plan && (
+          <div className="rounded-xl bg-slate-50 px-4 py-3">
+            <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">AI 已做/将做什么</div>
+            <p className="mt-2 text-sm leading-6 text-slate-700">{item.ai_plan}</p>
+          </div>
+        )}
+
+        {item.blocking_reason && (
+          <div className="rounded-xl border border-amber-100 bg-amber-50 px-4 py-3">
+            <div className="text-xs font-semibold tracking-[0.18em] text-amber-700">为什么需要你</div>
+            <p className="mt-2 text-sm leading-6 text-amber-900/80">{item.blocking_reason}</p>
+          </div>
+        )}
+
+        {(item.stage_text || item.next_step_text || item.can_undo) && (
+          <div className="grid gap-3 md:grid-cols-3">
+            {stageCards.map((stage) => (
+              <div key={stage.label} className="rounded-xl bg-slate-50 px-4 py-3">
+                <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">{stage.label}</div>
+                <p className="mt-2 text-sm text-slate-700">{stage.value}</p>
+              </div>
+            ))}
+          </div>
+        )}
+
+        <div className="flex flex-wrap items-center gap-2">
+          <Button
+            className="bg-blue-600 text-white hover:bg-blue-700"
+            disabled={primaryPending}
+            onClick={() => onAction(item, item.primary_action)}
+          >
+            {primaryPending ? "处理中…" : item.primary_action.label}
+          </Button>
+          <Button
+            variant="outline"
+            className="border-slate-200 bg-white text-slate-700 hover:bg-slate-50"
+            disabled={secondaryPending}
+            onClick={() => onAction(item, item.secondary_action)}
+          >
+            {secondaryPending ? "加载中…" : item.secondary_action.label}
+          </Button>
+          {item.weak_action.kind === "none" ? (
+            <span className="px-3 py-2 text-sm text-slate-400">{item.weak_action.label}</span>
+          ) : (
+            <Button
+              variant="ghost"
+              className="text-slate-500 hover:bg-slate-50 hover:text-slate-700"
+              disabled={weakPending}
+              onClick={() => onAction(item, item.weak_action)}
+            >
+              {weakPending ? "处理中…" : item.weak_action.label}
+            </Button>
+          )}
+        </div>
+      </CardContent>
+    </Card>
+  );
+}
+
+export default function InboxPage() {
+  const navigate = useNavigate();
+  const queryClient = useQueryClient();
+  const { toast } = useToast();
+  const [searchParams, setSearchParams] = useSearchParams();
+  const [notificationDetailItem, setNotificationDetailItem] = useState<InboxItemViewModel | null>(null);
+  const [workItemDetailItem, setWorkItemDetailItem] = useState<InboxItemViewModel | null>(null);
+  const [pendingActionKey, setPendingActionKey] = useState<string | null>(null);
+  const [priorityFilter, setPriorityFilter] = useState("all");
+  const [onlyDecision, setOnlyDecision] = useState(false);
+  const [onlyUnread, setOnlyUnread] = useState(false);
+  const activeSection = (searchParams.get("section") as InboxSectionKey) || "all";
+  const projectFilter = searchParams.get("project_id") || "all";
+  const searchValue = searchParams.get("q") || "";
+
+  const notificationsQuery = useQuery({
+    queryKey: ["notifications", "inbox"],
+    queryFn: () => notificationsApi.list({ limit: 60 }),
+    refetchInterval: POLL_INTERVAL_MS,
+  });
+
+  const workItemsQuery = useQuery({
+    queryKey: ["ai-pmo-work-items", "inbox"],
+    queryFn: () => aiPmoApi.listWorkItems({ limit: 60 }),
+    refetchInterval: POLL_INTERVAL_MS,
+  });
+
+  const workItemDetailQuery = useQuery({
+    queryKey: ["ai-pmo-work-item-detail", workItemDetailItem?.source_work_item?.id || null],
+    queryFn: () => aiPmoApi.getWorkItem(Number(workItemDetailItem?.source_work_item?.id)),
+    enabled: Boolean(workItemDetailItem?.source_work_item?.id),
+  });
+
+  const inboxData = useMemo(() => {
+    return buildInboxView({
+      notifications: notificationsQuery.data || { items: [], unread_count: 0 },
+      workItems: workItemsQuery.data || { items: [] },
+    });
+  }, [notificationsQuery.data, workItemsQuery.data]);
+
+  const projectOptions = useMemo(() => {
+    const map = new Map<string, string>();
+    for (const item of inboxData.items) {
+      if (!item.project_id) continue;
+      map.set(String(item.project_id), item.project_name || `项目 #${item.project_id}`);
+    }
+    return Array.from(map.entries()).map(([value, label]) => ({ value, label }));
+  }, [inboxData.items]);
+
+  const projectScopedItems = useMemo(() => {
+    return inboxData.items.filter((item) => {
+      if (projectFilter === "all") return true;
+      return String(item.project_id || "") === projectFilter;
+    });
+  }, [inboxData.items, projectFilter]);
+
+  const filteredItems = useMemo(() => {
+    const query = searchValue.trim().toLowerCase();
+    return projectScopedItems.filter((item) => {
+      if (activeSection !== "all" && item.display_kind !== activeSection) return false;
+      if (onlyDecision && item.display_kind !== "action_required") return false;
+      if (onlyUnread && item.is_read) return false;
+      if (!matchesPriority(item, priorityFilter)) return false;
+      if (!query) return true;
+      return `${item.title} ${item.summary} ${item.ai_plan || ""} ${item.project_name || ""}`.toLowerCase().includes(query);
+    });
+  }, [activeSection, onlyDecision, onlyUnread, priorityFilter, projectScopedItems, searchValue]);
+
+  const scopedSections = useMemo(() => buildScopedSections(projectScopedItems), [projectScopedItems]);
+  const scopedLeaderSummary = useMemo(() => {
+    return {
+      decision_count: projectScopedItems.filter((item) => item.display_kind === "action_required").length,
+      auto_running_count: projectScopedItems.filter((item) => item.display_kind === "auto_running").length,
+      auto_done_today: projectScopedItems.filter(
+        (item) => item.display_kind === "auto_done" && isToday(item.updated_at || item.created_at)
+      ).length,
+      high_priority_blocked_count: projectScopedItems.filter(
+        (item) => item.display_kind === "action_required" && item.priority_rank >= 3 && Boolean(item.blocking_reason)
+      ).length,
+      external_authorization_count: projectScopedItems.filter(
+        (item) =>
+          item.display_kind === "action_required" &&
+          /(github|gitlab|jira|接入|授权|跨系统|外部系统)/i.test(`${item.title} ${item.summary} ${item.blocking_reason || ""}`)
+      ).length,
+      top_decisions: projectScopedItems.filter((item) => item.display_kind === "action_required").slice(0, 3),
+      recent_completions: projectScopedItems.filter((item) => item.display_kind === "auto_done").slice(0, 3),
+    };
+  }, [projectScopedItems]);
+
+  const setSection = (section: string) => {
+    const next = new URLSearchParams(searchParams);
+    next.set("section", section);
+    setSearchParams(next);
+  };
+
+  const setProject = (projectId: string) => {
+    const next = new URLSearchParams(searchParams);
+    if (projectId === "all") next.delete("project_id");
+    else next.set("project_id", projectId);
+    setSearchParams(next);
+  };
+
+  const setSearch = (value: string) => {
+    const next = new URLSearchParams(searchParams);
+    if (!value.trim()) next.delete("q");
+    else next.set("q", value.trim());
+    setSearchParams(next);
+  };
+
+  const refreshAll = async () => {
+    await Promise.all([
+      queryClient.invalidateQueries({ queryKey: ["notifications", "inbox"] }),
+      queryClient.invalidateQueries({ queryKey: ["ai-pmo-work-items", "inbox"] }),
+    ]);
+  };
+
+  const markNotificationRead = async (item: InboxItemViewModel) => {
+    if (item.source_kind === "notification" && item.source_notification && !item.source_notification.is_read) {
+      try {
+        await notificationsApi.markRead(Number(item.source_notification.id));
+      } catch {
+        // 不阻断后续跳转。
+      }
+    }
+  };
+
+  const openNotificationDetail = async (item: InboxItemViewModel) => {
+    await markNotificationRead(item);
+    setNotificationDetailItem(item);
+    await refreshAll();
+  };
+
+  const openWorkItemDetail = async (item: InboxItemViewModel) => {
+    setWorkItemDetailItem(item);
+    await refreshAll();
+  };
+
+  const openTarget = async (item: InboxItemViewModel, options?: { forceDirect?: boolean }) => {
+    if (!options?.forceDirect && item.source_kind === "notification" && item.display_kind === "info") {
+      await openNotificationDetail(item);
+      return;
+    }
+
+    await markNotificationRead(item);
+
+    if (item.target_url) {
+      if (item.target_url.startsWith("/")) navigate(item.target_url);
+      else window.location.assign(item.target_url);
+      await refreshAll();
+      return;
+    }
+    if (item.source_kind === "ai_pmo_work_item") {
+      await openWorkItemDetail(item);
+      return;
+    }
+    toast({ title: "当前卡片暂无可打开的详情入口", variant: "error" });
+  };
+
+  const handleAction = async (item: InboxItemViewModel, action: InboxActionDescriptor) => {
+    const actionKey = buildActionKey(item, action);
+    setPendingActionKey(actionKey);
+    try {
+      if (item.source_kind === "notification" && action.label === "查看详情") {
+        await openNotificationDetail(item);
+        return;
+      }
+      if (action.kind === "open_target") {
+        await openTarget(item);
+        return;
+      }
+      if (action.kind === "open_detail") {
+        if (item.source_kind === "notification") {
+          await openNotificationDetail(item);
+          return;
+        }
+        await openWorkItemDetail(item);
+        return;
+      }
+      if (action.kind === "confirm" && item.source_kind === "ai_pmo_work_item" && item.source_work_item) {
+        const result = await aiPmoApi.resumeWorkItem(item.source_work_item.id, {
+          project_id: item.project_id ?? undefined,
+        });
+        await refreshAll();
+        const followupReply = String(result.followup_reply || "").trim();
+        if (result.status === "succeeded") {
+          toast({ title: followupReply || "AI PMO 已继续执行并完成该工作项", variant: "success" });
+        } else if (result.status === "pending_manual") {
+          toast({ title: followupReply || "继续执行后仍需补充信息或人工确认", variant: "default" });
+        } else {
+          toast({ title: followupReply || "继续执行失败，请查看详情", variant: "error" });
+        }
+        setWorkItemDetailItem({
+          ...item,
+          display_kind:
+            result.status === "succeeded"
+              ? "auto_done"
+              : result.status === "pending_manual"
+                ? "action_required"
+                : "info",
+          stage_text: String(result.target_state || item.stage_text || "").trim() || item.stage_text,
+          next_step_text: String(result.followup_reply || item.next_step_text || "").trim() || item.next_step_text,
+        });
+        return;
+      }
+      if (action.kind === "mark_read" && item.source_kind === "notification" && item.source_notification) {
+        await notificationsApi.markRead(item.source_notification.id);
+        await refreshAll();
+        return;
+      }
+      if (action.kind === "none") {
+        if (item.source_kind === "ai_pmo_work_item") {
+          await openWorkItemDetail(item);
+          return;
+        }
+        toast({ title: action.label, variant: "default" });
+      }
+    } catch (error) {
+      toast({ title: getApiErrorMessage(error), variant: "error" });
+    } finally {
+      setPendingActionKey((current) => (current === actionKey ? null : current));
+    }
+  };
+
+  const markAllRead = async () => {
+    try {
+      await notificationsApi.markAllRead();
+      toast({ title: "站内通知已全部设为已读", variant: "success" });
+      await refreshAll();
+    } catch (error) {
+      toast({ title: getApiErrorMessage(error, "全部已读失败"), variant: "error" });
+    }
+  };
+
+  const isLoading = notificationsQuery.isLoading || workItemsQuery.isLoading;
+
+  return (
+    <>
+      <div className="space-y-6 pb-16">
+        <section className="rounded-2xl bg-[#F8F9FA] p-4 md:p-5">
+          <div className="flex flex-wrap items-start justify-between gap-3">
+            <div>
+              <h1 className="text-2xl font-bold tracking-tight text-slate-900">消息台</h1>
+              <p className="mt-1 text-slate-500">AI 决策收件箱 · 先看摘要，再处理真正需要你拍板的事项</p>
+              <p className="mt-1 text-xs text-slate-500">
+                当前结果：{filteredItems.length} 条
+                {projectFilter !== "all" ? `（已筛选项目 #${projectFilter}）` : "（全租户汇总）"}
+              </p>
+            </div>
+            <div className="flex flex-wrap gap-2">
+              <Button
+                variant="outline"
+                className="border-slate-200 bg-white text-slate-700 hover:bg-slate-50"
+                onClick={markAllRead}
+                disabled={notificationsQuery.data?.unread_count === 0}
+              >
+                全部已读
+              </Button>
+              <Button className="bg-blue-600 text-white hover:bg-blue-700" onClick={() => setSection("action_required")}>
+                <SlidersHorizontal className="mr-2 h-4 w-4" />
+                先看待决策
+              </Button>
+            </div>
+          </div>
+
+          <div className="mt-4 rounded-2xl bg-white p-4 shadow-[0_8px_22px_rgba(15,23,42,0.08)]">
+            <div className="flex flex-wrap items-center gap-2">
+              <span className="rounded-full bg-blue-50 px-3 py-1 text-xs font-semibold text-blue-700">
+                AI 决策收件箱
+              </span>
+              <p className="text-sm font-medium text-slate-700">铃铛只看摘要，完整工作流统一沉到 Inbox</p>
+            </div>
+            <CardDescription className="mt-2 text-slate-500">
+              只保留真正可执行的事项；查询、闲聊、健康提示默认不进入消息台。
+            </CardDescription>
+            <div className="mt-4 grid gap-3 md:grid-cols-2 xl:grid-cols-5">
+              {[
+                { icon: ShieldAlert, label: "需你决策", value: scopedLeaderSummary.decision_count },
+                { icon: LoaderCircle, label: "AI处理中", value: scopedLeaderSummary.auto_running_count },
+                { icon: CheckCircle2, label: "今日自动完成", value: scopedLeaderSummary.auto_done_today },
+                { icon: Sparkles, label: "高优先级阻塞", value: scopedLeaderSummary.high_priority_blocked_count },
+                { icon: Filter, label: "待授权外部接入", value: scopedLeaderSummary.external_authorization_count },
+              ].map(({ icon: Icon, label, value }, index) => (
+                <div key={label} className="rounded-xl bg-slate-50 px-3 py-3">
+                  <div className="flex items-center justify-between gap-3">
+                    <div>
+                      <div className="text-xs font-semibold text-slate-500">{label}</div>
+                      <div className="mt-2 text-2xl font-bold text-slate-900">{value}</div>
+                    </div>
+                    <span className={cn("inline-flex h-9 w-9 items-center justify-center rounded-full", METRIC_CARD_STYLES[index])}>
+                      <Icon className="h-4 w-4" />
+                    </span>
+                  </div>
+                </div>
+              ))}
+            </div>
+          </div>
+          <div className="mt-4 grid gap-4 lg:grid-cols-2">
+            {renderMiniList("最需要你拍板的 3 件事", "优先看高风险、上下文不完整或需要授权的事项。", scopedLeaderSummary.top_decisions, (item) =>
+              void handleAction(item, item.primary_action)
+            )}
+            {renderMiniList("AI 刚完成的 3 件事", "自动结果会优先给出产物和影响范围。", scopedLeaderSummary.recent_completions, (item) =>
+              void handleAction(item, item.primary_action)
+            )}
+          </div>
+        </section>
+
+        <section className="rounded-2xl bg-[#F8F9FA] p-4 md:p-5">
+          <div className="rounded-2xl bg-white p-4 shadow-[0_8px_22px_rgba(15,23,42,0.08)]">
+            <div className="flex flex-wrap items-center gap-3">
+              <div className="min-w-[220px] flex-1">
+                <Input
+                  placeholder="搜索标题、摘要、项目"
+                  defaultValue={searchValue}
+                  className="border-slate-200 bg-white text-slate-700 placeholder:text-slate-400"
+                  onBlur={(event) => setSearch(event.target.value)}
+                  onKeyDown={(event) => {
+                    if (event.key === "Enter") setSearch((event.target as HTMLInputElement).value);
+                  }}
+                />
+              </div>
+              <select
+                value={projectFilter}
+                onChange={(event) => setProject(event.target.value)}
+                className="h-10 rounded-md border border-slate-200 bg-white px-3 text-sm text-slate-700"
+              >
+                <option value="all">全部项目</option>
+                {projectOptions.map((option) => (
+                  <option key={option.value} value={option.value}>
+                    {option.label}
+                  </option>
+                ))}
+              </select>
+              <select
+                value={priorityFilter}
+                onChange={(event) => setPriorityFilter(event.target.value)}
+                className="h-10 rounded-md border border-slate-200 bg-white px-3 text-sm text-slate-700"
+              >
+                {PRIORITY_OPTIONS.map((option) => (
+                  <option key={option.value} value={option.value}>
+                    {option.label}
+                  </option>
+                ))}
+              </select>
+              <label className="inline-flex items-center gap-2 rounded-full bg-slate-50 px-3 py-2 text-sm text-slate-600">
+                <Switch checked={onlyDecision} onCheckedChange={setOnlyDecision} />
+                仅需决策
+              </label>
+              <label className="inline-flex items-center gap-2 rounded-full bg-slate-50 px-3 py-2 text-sm text-slate-600">
+                <Switch checked={onlyUnread} onCheckedChange={setOnlyUnread} />
+                仅未读
+              </label>
+            </div>
+
+            <Tabs value={activeSection} onValueChange={setSection}>
+              <TabsList className="mt-4 grid h-auto w-full grid-cols-2 gap-2 rounded-2xl bg-slate-50 p-2 md:grid-cols-5">
+                {scopedSections.map((section) => (
+                  <TabsTrigger
+                    key={section.key}
+                    value={section.key}
+                    className="rounded-xl border border-transparent text-slate-600 data-[state=active]:border-slate-200 data-[state=active]:bg-white data-[state=active]:text-slate-900"
+                  >
+                    {section.label} {section.count}
+                  </TabsTrigger>
+                ))}
+              </TabsList>
+            </Tabs>
+          </div>
+        </section>
+
+        {isLoading ? (
+          <section className="rounded-2xl bg-[#F8F9FA] p-6">
+            <div className="flex items-center gap-2 rounded-xl bg-white px-4 py-3 text-slate-500 shadow-[0_4px_14px_rgba(15,23,42,0.06)]">
+              <span className="inline-block h-4 w-4 animate-spin rounded-full border-2 border-blue-600 border-t-transparent" />
+              加载消息台中…
+            </div>
+          </section>
+        ) : filteredItems.length === 0 ? (
+          <section className="rounded-2xl bg-[#F8F9FA] p-6">
+            <Card className="rounded-2xl border-0 bg-white shadow-[0_8px_22px_rgba(15,23,42,0.08)]">
+              <CardContent className="py-12 text-center text-slate-500">
+                <BellRing className="mx-auto mb-3 h-12 w-12 opacity-50" />
+                <p className="font-medium text-slate-700">当前没有匹配收件项</p>
+                <p className="mt-1 text-sm">请调整筛选条件后重试</p>
+              </CardContent>
+            </Card>
+          </section>
+        ) : (
+          <section className="rounded-2xl bg-[#F8F9FA] p-4 md:p-5">
+            <div className="grid gap-4 xl:grid-cols-2">
+              {filteredItems.map((item) => (
+                <InboxCard
+                  key={item.id}
+                  item={item}
+                  pendingActionKey={pendingActionKey}
+                  onAction={(target, action) => void handleAction(target, action)}
+                />
+              ))}
+            </div>
+          </section>
+        )}
+      </div>
+
+      <Dialog
+        open={Boolean(workItemDetailItem)}
+        onOpenChange={(open) => {
+          if (!open) setWorkItemDetailItem(null);
+        }}
+      >
+        <DialogContent className="max-h-[90vh] w-[720px] max-w-[95vw] overflow-y-auto">
+          <DialogHeader>
+            <DialogTitle>{workItemDetailItem?.title || "工作项详情"}</DialogTitle>
+            <DialogDescription>
+              {workItemDetailItem?.project_name || (workItemDetailItem?.project_id ? `项目 #${workItemDetailItem.project_id}` : "未绑定项目")}
+              {" · "}
+              {formatDetailTime(workItemDetailItem?.updated_at || workItemDetailItem?.created_at)}
+            </DialogDescription>
+          </DialogHeader>
+
+          {workItemDetailItem && (
+            <div className="space-y-4">
+              <div className="flex flex-wrap items-center gap-2">
+                <Badge variant="outline" className={cn("text-[11px]", STATUS_STYLES[workItemDetailItem.display_kind])}>
+                  {STATUS_LABELS[workItemDetailItem.display_kind]}
+                </Badge>
+                <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                  {toZhPriority(workItemDetailItem.priority)}
+                </span>
+                {workItemDetailItem.source_work_item?.intent && (
+                  <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                    {workItemDetailItem.source_work_item.intent}
+                  </span>
+                )}
+                {workItemDetailItem.source_work_item?.source_system && (
+                  <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                    {workItemDetailItem.source_work_item.source_system}
+                  </span>
+                )}
+              </div>
+
+              <div className="rounded-xl border border-slate-200 bg-slate-50 px-4 py-3">
+                <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">工作项摘要</div>
+                <p className="mt-2 whitespace-pre-wrap text-sm leading-6 text-slate-700">
+                  {workItemDetailItem.summary || "暂无摘要"}
+                </p>
+              </div>
+
+              {workItemDetailItem.source_work_item?.blocking_reason && (
+                <div className="rounded-xl border border-amber-100 bg-amber-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-amber-700">阻塞原因</div>
+                  <p className="mt-2 whitespace-pre-wrap text-sm leading-6 text-amber-900/80">
+                    {workItemDetailItem.source_work_item.blocking_reason}
+                  </p>
+                </div>
+              )}
+
+              {workItemDetailItem.source_work_item?.ai_plan && (
+                <div className="rounded-xl border border-slate-200 bg-white px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">AI 执行计划</div>
+                  <p className="mt-2 whitespace-pre-wrap text-sm leading-6 text-slate-700">
+                    {workItemDetailItem.source_work_item.ai_plan}
+                  </p>
+                </div>
+              )}
+
+              <div className="grid gap-3 md:grid-cols-2">
+                <div className="rounded-xl bg-slate-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">当前阶段</div>
+                  <p className="mt-2 text-sm text-slate-700">
+                    {workItemDetailItem.source_work_item?.stage_text || STATUS_LABELS[workItemDetailItem.display_kind]}
+                  </p>
+                </div>
+                <div className="rounded-xl bg-slate-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">下一步</div>
+                  <p className="mt-2 whitespace-pre-wrap text-sm text-slate-700">
+                    {workItemDetailItem.source_work_item?.next_step_text || "暂无"}
+                  </p>
+                </div>
+              </div>
+
+              {workItemDetailQuery.isLoading ? (
+                <div className="rounded-xl border border-slate-200 bg-slate-50 px-4 py-3 text-sm text-slate-500">
+                  加载工作项详情中…
+                </div>
+              ) : workItemDetailQuery.data ? (
+                <>
+                  <div className="grid gap-3 md:grid-cols-3">
+                    <div className="rounded-xl bg-slate-50 px-4 py-3">
+                      <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">决策记录</div>
+                      <p className="mt-2 text-sm text-slate-700">{workItemDetailQuery.data.decision_records.length} 条</p>
+                    </div>
+                    <div className="rounded-xl bg-slate-50 px-4 py-3">
+                      <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">执行计划</div>
+                      <p className="mt-2 text-sm text-slate-700">{workItemDetailQuery.data.plan_bundles.length} 条</p>
+                    </div>
+                    <div className="rounded-xl bg-slate-50 px-4 py-3">
+                      <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">回执记录</div>
+                      <p className="mt-2 text-sm text-slate-700">{workItemDetailQuery.data.reply_receipts.length} 条</p>
+                    </div>
+                  </div>
+
+                  <div className="rounded-xl border border-slate-200 bg-slate-50 px-4 py-3">
+                    <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">执行结果</div>
+                    <pre className="mt-2 overflow-x-auto whitespace-pre-wrap break-all text-xs leading-6 text-slate-700">
+                      {formatJsonBlock(workItemDetailQuery.data.execution_result)}
+                    </pre>
+                  </div>
+
+                  <div className="rounded-xl border border-slate-200 bg-slate-50 px-4 py-3">
+                    <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">复盘摘要</div>
+                    <pre className="mt-2 overflow-x-auto whitespace-pre-wrap break-all text-xs leading-6 text-slate-700">
+                      {formatJsonBlock(workItemDetailQuery.data.review_summary)}
+                    </pre>
+                  </div>
+                </>
+              ) : (
+                <div className="rounded-xl border border-rose-100 bg-rose-50 px-4 py-3 text-sm text-rose-700">
+                  工作项详情加载失败，请稍后重试。
+                </div>
+              )}
+            </div>
+          )}
+
+          <DialogFooter>
+            <Button variant="outline" onClick={() => setWorkItemDetailItem(null)}>
+              关闭
+            </Button>
+            {workItemDetailItem?.target_url && (
+              <Button
+                className="bg-blue-600 text-white hover:bg-blue-700"
+                onClick={() => {
+                  void openTarget(workItemDetailItem, { forceDirect: true });
+                  setWorkItemDetailItem(null);
+                }}
+              >
+                进入项目
+              </Button>
+            )}
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+
+      <Dialog
+        open={Boolean(notificationDetailItem)}
+        onOpenChange={(open) => {
+          if (!open) setNotificationDetailItem(null);
+        }}
+      >
+        <DialogContent className="max-h-[90vh] w-[560px] max-w-[95vw] overflow-y-auto">
+          <DialogHeader>
+            <DialogTitle>{notificationDetailItem?.title || "通知详情"}</DialogTitle>
+            <DialogDescription>
+              {notificationDetailItem?.project_name || (notificationDetailItem?.project_id ? `项目 #${notificationDetailItem.project_id}` : "未绑定项目")}
+              {" · "}
+              {formatDetailTime(notificationDetailItem?.updated_at || notificationDetailItem?.created_at)}
+            </DialogDescription>
+          </DialogHeader>
+
+          {notificationDetailItem && (
+            <div className="space-y-4">
+              <div className="flex flex-wrap items-center gap-2">
+                <Badge variant="outline" className={cn("text-[11px]", STATUS_STYLES[notificationDetailItem.display_kind])}>
+                  {STATUS_LABELS[notificationDetailItem.display_kind]}
+                </Badge>
+                <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                  {toZhPriority(notificationDetailItem.priority)}
+                </span>
+                {notificationDetailItem.source_notification?.source && (
+                  <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                    {notificationDetailItem.source_notification.source}
+                  </span>
+                )}
+                {notificationDetailItem.source_work_item?.source_system && (
+                  <span className="rounded-full bg-slate-100 px-3 py-1 text-xs font-semibold text-slate-600">
+                    {notificationDetailItem.source_work_item.source_system}
+                  </span>
+                )}
+              </div>
+
+              <div className="rounded-xl border border-slate-200 bg-slate-50 px-4 py-3">
+                <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">通知摘要</div>
+                <p className="mt-2 whitespace-pre-wrap text-sm leading-6 text-slate-700">
+                  {notificationDetailItem.source_notification?.content || notificationDetailItem.summary || "暂无摘要"}
+                </p>
+              </div>
+
+              {notificationDetailItem.ai_plan && (
+                <div className="rounded-xl bg-slate-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">AI 已做/将做什么</div>
+                  <p className="mt-2 whitespace-pre-wrap text-sm leading-6 text-slate-700">{notificationDetailItem.ai_plan}</p>
+                </div>
+              )}
+
+              {notificationDetailItem.blocking_reason && (
+                <div className="rounded-xl border border-amber-100 bg-amber-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-amber-700">当前阻塞</div>
+                  <p className="mt-2 whitespace-pre-wrap text-sm leading-6 text-amber-900/80">{notificationDetailItem.blocking_reason}</p>
+                </div>
+              )}
+
+              <div className="grid gap-3 md:grid-cols-2">
+                <div className="rounded-xl bg-slate-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">当前阶段</div>
+                  <p className="mt-2 text-sm text-slate-700">{notificationDetailItem.stage_text || STATUS_LABELS[notificationDetailItem.display_kind]}</p>
+                </div>
+                <div className="rounded-xl bg-slate-50 px-4 py-3">
+                  <div className="text-xs font-semibold tracking-[0.18em] text-slate-500">建议下一步</div>
+                  <p className="mt-2 text-sm text-slate-700">{notificationDetailItem.next_step_text || "先确认通知内容，再进入结果页处理。"}</p>
+                </div>
+              </div>
+            </div>
+          )}
+
+          <DialogFooter>
+            <Button variant="outline" onClick={() => setNotificationDetailItem(null)}>
+              关闭
+            </Button>
+            {notificationDetailItem?.target_url && (
+              <Button
+                className="bg-blue-600 text-white hover:bg-blue-700"
+                onClick={() => {
+                  const currentItem = notificationDetailItem;
+                  setNotificationDetailItem(null);
+                  if (currentItem) {
+                    void openTarget(currentItem, { forceDirect: true });
+                  }
+                }}
+              >
+                进入结果页
+              </Button>
+            )}
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+    </>
+  );
+}
diff --git a/frontend/src/types/inbox.ts b/frontend/src/types/inbox.ts
new file mode 100644
index 00000000..383243e2
--- /dev/null
+++ b/frontend/src/types/inbox.ts
@@ -0,0 +1,70 @@
+import type { AIPMOWorkItem } from "@/api/aiPmo";
+import type { NotificationItem } from "@/api/notifications";
+
+export type InboxSourceKind = "notification" | "ai_pmo_work_item";
+export type InboxDisplayKind = "action_required" | "auto_running" | "auto_done" | "info";
+export type InboxSectionKey = "all" | InboxDisplayKind;
+export type InboxActionKind =
+  | "confirm"
+  | "open_detail"
+  | "open_target"
+  | "mark_read"
+  | "none";
+
+export interface InboxActionDescriptor {
+  kind: InboxActionKind;
+  label: string;
+}
+
+export interface InboxItemViewModel {
+  id: string;
+  source_id: string;
+  source_kind: InboxSourceKind;
+  display_kind: InboxDisplayKind;
+  title: string;
+  summary: string;
+  ai_plan?: string | null;
+  blocking_reason?: string | null;
+  primary_action: InboxActionDescriptor;
+  secondary_action: InboxActionDescriptor;
+  weak_action: InboxActionDescriptor;
+  project_id?: number | null;
+  project_name?: string | null;
+  priority: string;
+  priority_rank: number;
+  is_read: boolean;
+  dedupe_count: number;
+  can_undo: boolean;
+  undo_deadline?: string | null;
+  target_url?: string | null;
+  created_at?: string | null;
+  updated_at?: string | null;
+  stage_text?: string | null;
+  next_step_text?: string | null;
+  source_notification?: NotificationItem;
+  source_work_item?: AIPMOWorkItem;
+}
+
+export interface InboxSection {
+  key: InboxSectionKey;
+  label: string;
+  count: number;
+  items: InboxItemViewModel[];
+}
+
+export interface LeaderSummary {
+  decision_count: number;
+  auto_running_count: number;
+  auto_done_today: number;
+  high_priority_blocked_count: number;
+  external_authorization_count: number;
+  top_decisions: InboxItemViewModel[];
+  recent_completions: InboxItemViewModel[];
+}
+
+export interface InboxBuildResult {
+  items: InboxItemViewModel[];
+  sections: InboxSection[];
+  previewSections: InboxSection[];
+  leaderSummary: LeaderSummary;
+}
diff --git a/frontend/tsconfig.tsbuildinfo b/frontend/tsconfig.tsbuildinfo
new file mode 100644
index 00000000..e005b0b4
--- /dev/null
+++ b/frontend/tsconfig.tsbuildinfo
@@ -0,0 +1 @@
+{"root":["./src/main.tsx","./src/vite-env.d.ts","./src/components/skillradarchart.tsx","./src/components/themeinit.tsx","./src/components/ai-steward/floatingqachat.tsx","./src/components/ai-steward/openclawcommanderpanel.tsx","./src/components/ai-steward/promptrolesettings.tsx","./src/components/ai-steward/qahistorypanel.tsx","./src/components/budget/budgetcontrolpanel.tsx","./src/components/command/commanditems.ts","./src/components/layout/layout.tsx","./src/components/layout/sidebar.tsx","./src/components/layout/topbar.tsx","./src/components/model-capability/costoptimizationpanel.tsx","./src/components/model-capability/experimentsyncpanel.tsx","./src/components/model-capability/modelcapabilitymatrix.tsx","./src/components/model-capability/modelupdateimpactpanel.tsx","./src/components/sstc/aivisibilitypanel.tsx","./src/components/sstc/sstcchainview.tsx","./src/components/sstc/sstcglobalheader.tsx","./src/components/sstc/sstcriskfloatpanel.tsx","./src/components/sstc/sstcstagenav.tsx","./src/components/sstc/sstcworkspace.tsx","./src/components/sstc/signalcontenthovercard.tsx","./src/components/sstc/chartermodules.ts","./src/components/ui/badge.tsx","./src/components/ui/button.tsx","./src/components/ui/card.tsx","./src/components/ui/dialog.tsx","./src/components/ui/dropdown-menu.tsx","./src/components/ui/input.tsx","./src/components/ui/label.tsx","./src/components/ui/switch.tsx","./src/components/ui/table.tsx","./src/components/ui/tabs.tsx","./src/components/ui/textarea.tsx","./src/components/ui/toast.tsx","./src/components/users/userformdialog.tsx","./src/hooks/usetheme.ts","./src/lib/api.ts","./src/lib/apierror.ts","./src/lib/datetime.ts","./src/lib/display.ts","./src/lib/feishueventmode.ts","./src/lib/inbox.ts","./src/lib/rbac.ts","./src/lib/sessionguard.ts","./src/lib/title.ts","./src/lib/utils.ts","./src/lib/zhcn.ts","./src/pages/aiconfigpage.tsx","./src/pages/aistewardpage.tsx","./src/pages/charterdetailpage.tsx","./src/pages/dashboardpage.tsx","./src/pages/feishuconfigpage.tsx","./src/pages/inboxpage.tsx","./src/pages/loginpage.tsx","./src/pages/projectboardpage.tsx","./src/pages/projectspage.tsx","./src/pages/recipespage.tsx","./src/pages/schedulercenterpage.tsx","./src/pages/skilltaskspage.tsx","./src/pages/skillspage.tsx","./src/pages/userprofilepage.tsx","./src/pages/userskillspage.tsx","./src/pages/userspage.tsx","./src/pages/workloaddetailpage.tsx","./src/pages/workloadspage.tsx","./src/router/index.tsx","./src/stores/auth.ts","./src/types/inbox.ts"],"version":"5.6.3"}
\ No newline at end of file
diff --git a/scripts/openclaw_routing_lint.py b/scripts/openclaw_routing_lint.py
new file mode 100644
index 00000000..beee4a07
--- /dev/null
+++ b/scripts/openclaw_routing_lint.py
@@ -0,0 +1,176 @@
+#!/usr/bin/env python3
+"""OpenClaw 路由规则静态校验（PMFlow 三引擎版）。"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any, Optional
+
+VALID_PROVIDERS = {"glm", "openai", "deepseek", "ollama"}
+VALID_TIERS = {"low", "high", "safety"}
+
+LOW_REQUIRED = {
+    "feishu_level_classification",
+    "feishu_signal_extraction",
+    "rag_intent_classification",
+    "classify_intent",
+    "feishu_humanized_reply",
+}
+
+HIGH_REQUIRED = {
+    "adaptive_rag_answer",
+    "feishu_project_prefill",
+    "feishu_task_orchestration",
+    "generate_task_breakdown",
+    "generate_charter_from_signal",
+    "decompose_tasks_from_charter",
+    "recommend_assignees_for_task",
+    "evaluate_model_requirement",
+    "calculate_alignment_score",
+    "detect_cognitive_conflicts",
+    "rag_db_project_semantic_retrieval",
+    "workload_optimize_prompt",
+    "analyze_cost_roi",
+}
+
+
+def _rule_matches_function(rule: dict[str, Any], function_name: str) -> bool:
+    if bool(rule.get("match_all", False)):
+        return True
+    fn = str(function_name or "").strip().lower()
+    if not fn:
+        return False
+    exact = str(rule.get("function_exact") or "").strip().lower()
+    if exact and fn == exact:
+        return True
+    prefix = str(rule.get("function_prefix") or "").strip().lower()
+    if prefix in {"*", "all"}:
+        return True
+    if prefix and fn.startswith(prefix):
+        return True
+    match_any = rule.get("match_any")
+    if isinstance(match_any, list):
+        for hint in match_any:
+            token = str(hint or "").strip().lower()
+            if token and token in fn:
+                return True
+    return False
+
+
+def _resolve_primary_tier(function_name: str, rules: list[dict[str, Any]]) -> Optional[str]:
+    scored: list[tuple[int, int, str]] = []
+    for idx, row in enumerate(rules):
+        if not bool(row.get("enabled", True)):
+            continue
+        tier = str(row.get("tier") or "").strip().lower()
+        if tier not in {"low", "high"}:
+            continue
+        if not _rule_matches_function(row, function_name):
+            continue
+        scored.append((int(row.get("priority") or 0), -idx, tier))
+    if not scored:
+        return None
+    scored.sort(reverse=True)
+    return scored[0][2]
+
+
+def _lint(path: Path) -> tuple[bool, dict[str, Any]]:
+    if not path.exists():
+        return False, {"error": f"rules file not found: {path}"}
+    try:
+        payload = json.loads(path.read_text(encoding="utf-8"))
+    except Exception as exc:
+        return False, {"error": f"invalid json: {exc}"}
+    if not isinstance(payload, dict):
+        return False, {"error": "root must be object"}
+
+    rules_raw = payload.get("rules")
+    if not isinstance(rules_raw, list):
+        return False, {"error": "rules must be list"}
+
+    rules: list[dict[str, Any]] = []
+    rule_names: list[str] = []
+    problems: list[str] = []
+    for idx, row in enumerate(rules_raw, start=1):
+        if not isinstance(row, dict):
+            problems.append(f"rule#{idx}: must be object")
+            continue
+        name = str(row.get("name") or "").strip()
+        if not name:
+            problems.append(f"rule#{idx}: missing name")
+            continue
+        tier = str(row.get("tier") or "").strip().lower()
+        if tier and tier not in VALID_TIERS:
+            problems.append(f"rule#{idx}:{name} unsupported tier={tier}")
+        provider = str(row.get("provider") or "").strip().lower()
+        if provider and provider not in VALID_PROVIDERS:
+            problems.append(f"rule#{idx}:{name} unsupported provider={provider}")
+        if row.get("match_any") is not None and not isinstance(row.get("match_any"), list):
+            problems.append(f"rule#{idx}:{name} match_any must be list")
+        rules.append(row)
+        rule_names.append(name)
+
+    for fn in sorted(LOW_REQUIRED):
+        actual_tier = _resolve_primary_tier(fn, rules)
+        if actual_tier != "low":
+            problems.append(f"function {fn} expected tier=low, actual={actual_tier}")
+    for fn in sorted(HIGH_REQUIRED):
+        actual_tier = _resolve_primary_tier(fn, rules)
+        if actual_tier != "high":
+            problems.append(f"function {fn} expected tier=high, actual={actual_tier}")
+
+    safety_rules = [
+        row
+        for row in rules
+        if bool(row.get("enabled", True)) and str(row.get("tier") or "").strip().lower() == "safety"
+    ]
+    if not safety_rules:
+        problems.append("missing safety tier rule")
+    else:
+        for fn in sorted(LOW_REQUIRED | HIGH_REQUIRED):
+            if not any(_rule_matches_function(row, fn) for row in safety_rules):
+                problems.append(f"function {fn} not covered by safety tier rule")
+                break
+
+    model_hints = {
+        "low": "zai/glm-4-flash",
+        "high": "minimax/MiniMax-M2.5",
+        "safety": "deepseek/deepseek-chat",
+    }
+    for tier, expected_model in model_hints.items():
+        tier_rows = [
+            row for row in rules
+            if bool(row.get("enabled", True)) and str(row.get("tier") or "").strip().lower() == tier
+        ]
+        if tier_rows and not any(str(row.get("model") or "").strip() == expected_model for row in tier_rows):
+            problems.append(f"tier={tier} missing expected model hint: {expected_model}")
+
+    return len(problems) == 0, {
+        "path": str(path),
+        "version": payload.get("version"),
+        "rule_count": len(rules),
+        "rule_names": sorted(rule_names),
+        "required_low_count": len(LOW_REQUIRED),
+        "required_high_count": len(HIGH_REQUIRED),
+        "problems": problems,
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="OpenClaw routing lint")
+    parser.add_argument(
+        "--input",
+        default=".openclaw/pmflow-routing-rules.json",
+        help="Rules file path",
+    )
+    args = parser.parse_args()
+
+    ok, report = _lint(Path(args.input))
+    print(json.dumps(report, ensure_ascii=False, indent=2))
+    raise SystemExit(0 if ok else 2)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/start_legacy_local.sh b/scripts/start_legacy_local.sh
new file mode 100755
index 00000000..b69e157e
--- /dev/null
+++ b/scripts/start_legacy_local.sh
@@ -0,0 +1,163 @@
+#!/bin/bash
+# PMFlow 启动器：后端 + 前端 + Cloudflare Quick Tunnel（自动输出临时公网地址）
+
+set -euo pipefail
+
+PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$PROJECT_DIR"
+
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+RED='\033[0;31m'
+NC='\033[0m'
+
+die() {
+    echo -e "${RED}❌ $1${NC}"
+    exit 1
+}
+
+kill_port() {
+    local port="$1"
+    lsof -tiTCP:"$port" -sTCP:LISTEN 2>/dev/null | xargs kill -9 2>/dev/null || true
+}
+
+wait_http_ok() {
+    local url="$1"
+    local max_retry="$2"
+    local retry=1
+    while [ "$retry" -le "$max_retry" ]; do
+        if curl -fsS "$url" >/dev/null 2>&1; then
+            return 0
+        fi
+        sleep 1
+        retry=$((retry + 1))
+    done
+    return 1
+}
+
+wait_remote_ok() {
+    local url="$1"
+    local max_retry="$2"
+    local retry=1
+    while [ "$retry" -le "$max_retry" ]; do
+        if curl -kfsS --max-time 8 "$url" >/dev/null 2>&1; then
+            return 0
+        fi
+        sleep 2
+        retry=$((retry + 1))
+    done
+    return 1
+}
+
+echo "🚀 PMFlow 启动器"
+echo "===================="
+
+[ -d "venv" ] || die "虚拟环境不存在，请先创建: python -m venv venv"
+[ -d "backend" ] || die "缺少 backend 目录"
+[ -d "frontend" ] || die "缺少 frontend 目录"
+command -v cloudflared >/dev/null 2>&1 || die "未安装 cloudflared"
+command -v npm >/dev/null 2>&1 || die "未安装 npm"
+
+mkdir -p logs
+rm -f logs/backend.pid logs/frontend.pid logs/tunnel.pid logs/tunnel-frontend.pid logs/tunnel-backend.pid logs/public_url.txt
+rm -f logs/tunnel.log logs/tunnel-backend.log logs/tunnel-frontend.log
+
+echo "🧹 清理旧进程与端口..."
+pkill -f "cloudflared tunnel --url http://127.0.0.1:5174" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel --url http://localhost:5174" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel --url http://127.0.0.1:8000" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel --url http://localhost:8000" >/dev/null 2>&1 || true
+kill_port 8000
+kill_port 5173
+kill_port 5174
+sleep 1
+
+echo -e "${GREEN}▶️ 启动后端 (8000)...${NC}"
+(
+    cd backend
+    nohup "$PROJECT_DIR/venv/bin/uvicorn" app.main:app --host 0.0.0.0 --port 8000 --reload --reload-dir app > "$PROJECT_DIR/logs/backend.log" 2>&1 &
+    echo $! > "$PROJECT_DIR/logs/backend.pid"
+)
+BACKEND_PID="$(cat logs/backend.pid)"
+wait_http_ok "http://127.0.0.1:8000/healthz" 15 || die "后端启动失败，查看 logs/backend.log"
+echo -e "${GREEN}✅ 后端已启动 (PID: ${BACKEND_PID})${NC}"
+
+echo -e "${GREEN}▶️ 启动前端 (5174)...${NC}"
+(
+    cd frontend
+    nohup npm run dev -- --host 0.0.0.0 --port 5174 > "$PROJECT_DIR/logs/frontend.log" 2>&1 &
+    echo $! > "$PROJECT_DIR/logs/frontend.pid"
+)
+FRONTEND_PID="$(cat logs/frontend.pid)"
+wait_http_ok "http://127.0.0.1:5174/login" 20 || wait_http_ok "http://127.0.0.1:5174" 10 || die "前端启动失败，查看 logs/frontend.log"
+echo -e "${GREEN}✅ 前端已启动 (PID: ${FRONTEND_PID})${NC}"
+
+echo -e "${GREEN}▶️ 启动 Cloudflare Quick Tunnels...${NC}"
+# Start frontend tunnel
+nohup cloudflared tunnel --url http://127.0.0.1:5174 > logs/tunnel-frontend.log 2>&1 &
+TUNNEL_FRONT_PID=$!
+echo "$TUNNEL_FRONT_PID" > logs/tunnel-frontend.pid
+
+# Start backend tunnel
+nohup cloudflared tunnel --url http://127.0.0.1:8000 > logs/tunnel-backend.log 2>&1 &
+TUNNEL_BACK_PID=$!
+echo "$TUNNEL_BACK_PID" > logs/tunnel-backend.pid
+
+TEMP_URL_FRONT=""
+TEMP_URL_BACK=""
+
+for _ in {1..25}; do
+    if [ -z "$TEMP_URL_FRONT" ]; then
+        TEMP_URL_FRONT="$(grep -aEo 'https://[a-z0-9-]+\.trycloudflare\.com' logs/tunnel-frontend.log | head -n 1 || true)"
+    fi
+    if [ -z "$TEMP_URL_BACK" ]; then
+        TEMP_URL_BACK="$(grep -aEo 'https://[a-z0-9-]+\.trycloudflare\.com' logs/tunnel-backend.log | head -n 1 || true)"
+    fi
+    
+    if [ -n "$TEMP_URL_FRONT" ] && [ -n "$TEMP_URL_BACK" ]; then
+        break
+    fi
+    sleep 1
+done
+
+if [ -n "$TEMP_URL_FRONT" ] && [ -n "$TEMP_URL_BACK" ]; then
+    FRONT_OK=false
+    BACK_OK=false
+    if wait_remote_ok "${TEMP_URL_FRONT}/login" 25 || wait_remote_ok "${TEMP_URL_FRONT}" 25; then
+        FRONT_OK=true
+    fi
+    if wait_remote_ok "${TEMP_URL_BACK}/healthz" 25; then
+        BACK_OK=true
+    fi
+
+    echo "$TEMP_URL_FRONT" > logs/public_url.txt
+    echo "$TEMP_URL_BACK" >> logs/public_url.txt
+
+    if [ "$FRONT_OK" = true ] && [ "$BACK_OK" = true ]; then
+        echo -e "${GREEN}✅ Tunnels 已启动且公网可访问${NC}"
+    else
+        echo -e "${YELLOW}⚠️ Tunnel 已启动，但公网健康检查未全部通过（请稍后再试）${NC}"
+    fi
+    echo -e "🌐 前端公网地址(临时): ${YELLOW}${TEMP_URL_FRONT}${NC}"
+    echo -e "🌐 后端公网地址(临时): ${YELLOW}${TEMP_URL_BACK}${NC}"
+    echo -e "${YELLOW}ℹ️ 前端公网默认通过 /api 代理访问本地后端，无需手动修改 VITE_API_BASE_URL。${NC}"
+else
+    echo -e "${YELLOW}⚠️ Tunnel 已启动但部分地址未解析，请查看 logs/tunnel-*.log${NC}"
+fi
+
+LAN_IP="$(ipconfig getifaddr en0 2>/dev/null || echo '本机IP')"
+echo
+echo -e "${GREEN}🎉 启动完成${NC}"
+echo "===================="
+echo "本地前端: http://127.0.0.1:5174"
+echo "本地后端: http://127.0.0.1:8000"
+echo "局域网前端: http://${LAN_IP}:5174"
+echo "API 文档: http://127.0.0.1:8000/docs"
+if [ -n "$TEMP_URL_FRONT" ]; then
+    echo "前端公网: ${TEMP_URL_FRONT}"
+fi
+if [ -n "$TEMP_URL_BACK" ]; then
+    echo "后端公网: ${TEMP_URL_BACK}"
+fi
+echo "日志: logs/backend.log | logs/frontend.log | logs/tunnel-*.log"
+echo "停止: ./stop.sh"
diff --git a/scripts/stop_legacy_local.sh b/scripts/stop_legacy_local.sh
new file mode 100755
index 00000000..6f3ace82
--- /dev/null
+++ b/scripts/stop_legacy_local.sh
@@ -0,0 +1,79 @@
+#!/bin/bash
+# Neural Sync 停止器 - 停止所有服务
+
+PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$PROJECT_DIR"
+
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+NC='\033[0m'
+
+echo "🛑 停止 Neural Sync 服务"
+echo "===================="
+
+# 从 PID 文件读取并停止
+if [ -f "logs/backend.pid" ]; then
+    BACKEND_PID=$(cat logs/backend.pid)
+    if ps -p $BACKEND_PID > /dev/null 2>&1; then
+        kill $BACKEND_PID 2>/dev/null
+        echo -e "${GREEN}✅ 后端已停止 (PID: $BACKEND_PID)${NC}"
+    fi
+    rm logs/backend.pid
+fi
+
+if [ -f "logs/frontend.pid" ]; then
+    FRONTEND_PID=$(cat logs/frontend.pid)
+    if ps -p $FRONTEND_PID > /dev/null 2>&1; then
+        kill $FRONTEND_PID 2>/dev/null
+        echo -e "${GREEN}✅ 前端已停止 (PID: $FRONTEND_PID)${NC}"
+    fi
+    rm logs/frontend.pid
+fi
+
+if [ -f "logs/tunnel.pid" ]; then
+    TUNNEL_PID=$(cat logs/tunnel.pid)
+    if ps -p $TUNNEL_PID > /dev/null 2>&1; then
+        kill $TUNNEL_PID 2>/dev/null
+        echo -e "${GREEN}✅ Cloudflare Tunnel 已停止 (PID: $TUNNEL_PID)${NC}"
+    fi
+    rm logs/tunnel.pid
+fi
+
+if [ -f "logs/tunnel-frontend.pid" ]; then
+    TUNNEL_FRONT_PID=$(cat logs/tunnel-frontend.pid)
+    if ps -p $TUNNEL_FRONT_PID > /dev/null 2>&1; then
+        kill $TUNNEL_FRONT_PID 2>/dev/null
+        echo -e "${GREEN}✅ 前端 Tunnel 已停止 (PID: $TUNNEL_FRONT_PID)${NC}"
+    fi
+    rm logs/tunnel-frontend.pid
+fi
+
+if [ -f "logs/tunnel-backend.pid" ]; then
+    TUNNEL_BACK_PID=$(cat logs/tunnel-backend.pid)
+    if ps -p $TUNNEL_BACK_PID > /dev/null 2>&1; then
+        kill $TUNNEL_BACK_PID 2>/dev/null
+        echo -e "${GREEN}✅ 后端 Tunnel 已停止 (PID: $TUNNEL_BACK_PID)${NC}"
+    fi
+    rm logs/tunnel-backend.pid
+fi
+
+# Fallback: kill by name just in case
+pkill -f "cloudflared tunnel --url http://127.0.0.1:5174" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel --url http://localhost:5174" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel --url http://127.0.0.1:8000" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel --url http://localhost:8000" >/dev/null 2>&1 || true
+pkill -f "cloudflared tunnel run" >/dev/null 2>&1 || true
+
+# 清理端口 (8000 后端, 5173/5174 前端)
+lsof -ti:8000 2>/dev/null | xargs kill -9 2>/dev/null || true
+lsof -ti:5173 2>/dev/null | xargs kill -9 2>/dev/null || true
+lsof -ti:5174 2>/dev/null | xargs kill -9 2>/dev/null || true
+
+# 停止 Ollama（可选）
+read -p "是否同时停止 Ollama? (y/N) " -n 1 -r
+echo
+if [[ $REPLY =~ ^[Yy]$ ]]; then
+    pkill -9 -f "ollama serve" 2>/dev/null && echo -e "${GREEN}✅ Ollama 已停止${NC}" || true
+fi
+
+echo -e "${GREEN}🎉 所有服务已停止${NC}"
diff --git a/scripts/tunnel_frontend_watch.sh b/scripts/tunnel_frontend_watch.sh
new file mode 100755
index 00000000..f2b669bb
--- /dev/null
+++ b/scripts/tunnel_frontend_watch.sh
@@ -0,0 +1,340 @@
+#!/usr/bin/env bash
+set -uo pipefail
+
+PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_DIR"
+
+mkdir -p logs
+
+PATH="/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin:${PATH:-}"
+URL_PATTERN='https://[a-z0-9-]+\.trycloudflare\.com'
+WATCH_LOG="logs/tunnel-frontend-watch.log"
+WATCH_STDOUT_LOG="logs/tunnel-frontend-watch.stdout"
+WATCH_PID_FILE="logs/tunnel-frontend-watch.pid"
+RUN_LOG="logs/tunnel-frontend.log"
+TUNNEL_PID_FILE="logs/tunnel-frontend.pid"
+PUBLIC_URL_FILE="logs/public_url.txt"
+ACTIVE_CHILD_PID=""
+
+is_pid_running() {
+  local pid="${1:-}"
+  [[ -n "$pid" ]] || return 1
+  ps -p "$pid" >/dev/null 2>&1
+}
+
+log_watch() {
+  local message="$1"
+  local line="[$(date '+%Y-%m-%d %H:%M:%S')] [watch] ${message}"
+  printf '%s\n' "$line" | tee -a "$WATCH_LOG"
+}
+
+log_watch_err() {
+  local message="$1"
+  local line="[$(date '+%Y-%m-%d %H:%M:%S')] [watch] ${message}"
+  printf '%s\n' "$line" | tee -a "$WATCH_LOG" >&2
+}
+
+read_frontend_url() {
+  local line=""
+  line="$(rg '^frontend=' "$PUBLIC_URL_FILE" 2>/dev/null | tail -n 1 || true)"
+  if [[ -n "$line" ]]; then
+    echo "${line#frontend=}"
+    return 0
+  fi
+  sed -n '1p' "$PUBLIC_URL_FILE" 2>/dev/null | rg '^https?://' -N || true
+}
+
+read_backend_url() {
+  local line=""
+  line="$(rg '^backend=' "$PUBLIC_URL_FILE" 2>/dev/null | tail -n 1 || true)"
+  if [[ -n "$line" ]]; then
+    echo "${line#backend=}"
+    return 0
+  fi
+  sed -n '2p' "$PUBLIC_URL_FILE" 2>/dev/null | rg '^https?://' -N || true
+}
+
+write_frontend_url() {
+  local frontend_url="$1"
+  local backend_url=""
+  backend_url="$(read_backend_url)"
+  : > "$PUBLIC_URL_FILE"
+  echo "frontend=$frontend_url" >> "$PUBLIC_URL_FILE"
+  if [[ -n "$backend_url" ]]; then
+    echo "backend=$backend_url" >> "$PUBLIC_URL_FILE"
+  fi
+}
+
+find_cloudflared_bin() {
+  local bin
+  bin="$(command -v cloudflared || true)"
+  if [[ -n "$bin" ]]; then
+    echo "$bin"
+    return 0
+  fi
+  for candidate in /opt/homebrew/bin/cloudflared /usr/local/bin/cloudflared; do
+    if [[ -x "$candidate" ]]; then
+      echo "$candidate"
+      return 0
+    fi
+  done
+  echo ""
+}
+
+extract_tunnel_url() {
+  local logfile="$1"
+  rg -o "$URL_PATTERN" "$logfile" 2>/dev/null \
+    | rg -v "^https://api\\.trycloudflare\\.com$" \
+    | tail -n 1
+}
+
+probe_local_origin() {
+  local port="$1"
+  local code=""
+  code="$(curl --noproxy '*' -s -m 3 -o /dev/null -w "%{http_code}" "http://127.0.0.1:${port}/" || true)"
+  [[ "$code" =~ ^[0-9]{3}$ ]] || return 1
+  [[ "$code" -ge 200 && "$code" -lt 500 ]]
+}
+
+probe_tunnel_url() {
+  local url="$1"
+  local code=""
+  code="$(curl --noproxy '*' -s -m 8 -o /dev/null -w "%{http_code}" "${url}/" || true)"
+  [[ "$code" =~ ^[0-9]{3}$ ]] || return 1
+  [[ "$code" -ge 200 && "$code" -lt 500 ]]
+}
+
+probe_tunnel_url_stable() {
+  local url="$1"
+  local checks="${2:-2}"
+  local current=1
+  while [[ "$current" -le "$checks" ]]; do
+    probe_tunnel_url "$url" || return 1
+    if [[ "$current" -lt "$checks" ]]; then
+      sleep 1
+    fi
+    current=$((current + 1))
+  done
+}
+
+resolve_frontend_port() {
+  local preferred="${PMFLOW_FRONTEND_TUNNEL_PORT:-${FRONTEND_PORT:-5174}}"
+  local candidates=("$preferred" 5174 5175 5176 5177 5178)
+  local seen=" "
+  local port=""
+  for port in "${candidates[@]}"; do
+    [[ "$seen" == *" ${port} "* ]] && continue
+    seen="${seen}${port} "
+    if probe_local_origin "$port"; then
+      echo "$port"
+      return 0
+    fi
+  done
+  echo "$preferred"
+}
+
+is_watch_running() {
+  if [[ ! -f "$WATCH_PID_FILE" ]]; then
+    return 1
+  fi
+  local pid
+  pid="$(cat "$WATCH_PID_FILE" 2>/dev/null || true)"
+  if [[ -z "$pid" ]]; then
+    return 1
+  fi
+  if ps -p "$pid" >/dev/null 2>&1; then
+    return 0
+  fi
+  return 1
+}
+
+cleanup_children() {
+  if [[ -n "${ACTIVE_CHILD_PID:-}" ]] && is_pid_running "$ACTIVE_CHILD_PID"; then
+    kill "$ACTIVE_CHILD_PID" >/dev/null 2>&1 || true
+  fi
+  if [[ -f "$TUNNEL_PID_FILE" ]]; then
+    local pid
+    pid="$(cat "$TUNNEL_PID_FILE" 2>/dev/null || true)"
+    if [[ -n "$pid" ]] && is_pid_running "$pid"; then
+      kill "$pid" >/dev/null 2>&1 || true
+    fi
+  fi
+}
+
+stop_watch() {
+  local stopped=0
+  if [[ -f "$WATCH_PID_FILE" ]]; then
+    local wpid
+    wpid="$(cat "$WATCH_PID_FILE" 2>/dev/null || true)"
+    if [[ -n "$wpid" ]] && is_pid_running "$wpid"; then
+      kill "$wpid" >/dev/null 2>&1 || true
+      stopped=1
+    fi
+  fi
+  cleanup_children
+  rm -f "$WATCH_PID_FILE" "$TUNNEL_PID_FILE"
+  if [[ "$stopped" -eq 1 ]]; then
+    echo "[前端隧道守护] 已停止"
+  else
+    echo "[前端隧道守护] 未运行"
+  fi
+}
+
+print_status() {
+  local tunnel_pid=""
+  if [[ -f "$TUNNEL_PID_FILE" ]]; then
+    tunnel_pid="$(cat "$TUNNEL_PID_FILE" 2>/dev/null || true)"
+  fi
+  if is_watch_running; then
+    local wpid cpid
+    wpid="$(cat "$WATCH_PID_FILE" 2>/dev/null || true)"
+    cpid="${tunnel_pid:-$(cat "$TUNNEL_PID_FILE" 2>/dev/null || true)}"
+    echo "[前端隧道守护] 运行中 watch_pid=${wpid:-unknown} tunnel_pid=${cpid:-unknown}"
+  else
+    echo "[前端隧道守护] 未运行"
+  fi
+  local frontend_url backend_url
+  frontend_url="$(read_frontend_url)"
+  backend_url="$(read_backend_url)"
+  if [[ -n "$frontend_url" ]]; then
+    echo "前端地址: $frontend_url"
+    if is_pid_running "$tunnel_pid"; then
+      echo "前端健康: tunnel进程运行中"
+    else
+      echo "前端健康: tunnel进程未运行"
+    fi
+  else
+    echo "前端地址: (空)"
+  fi
+  if [[ -n "$backend_url" ]]; then
+    echo "后端地址: $backend_url"
+  else
+    echo "后端地址: (未配置/未暴露)"
+  fi
+}
+
+start_watch_background() {
+  if is_watch_running; then
+    print_status
+    return 0
+  fi
+  nohup bash "$0" run >"$WATCH_STDOUT_LOG" 2>&1 &
+  local pid="$!"
+  echo "$pid" >"$WATCH_PID_FILE"
+  echo "[前端隧道守护] 已启动 pid=$pid"
+  sleep 1
+  tail -n 5 "$WATCH_STDOUT_LOG" 2>/dev/null || true
+}
+
+run_watch_loop() {
+  local cf_bin
+  cf_bin="$(find_cloudflared_bin)"
+  if [[ -z "$cf_bin" ]]; then
+    log_watch_err "cloudflared_not_found"
+    exit 1
+  fi
+
+  echo "$$" >"$WATCH_PID_FILE"
+  log_watch "start bin=$cf_bin"
+  trap 'cleanup_children; rm -f "$WATCH_PID_FILE"; log_watch "exit code=$?"' EXIT INT TERM
+
+  while true; do
+    local frontend_port tunnel_protocol child_pid published current_url url
+    local strict_public_health public_health_interval public_health_failure_threshold
+    local unhealthy_count last_health_check_at
+    frontend_port="$(resolve_frontend_port)"
+    tunnel_protocol="${PMFLOW_TUNNEL_PROTOCOL_FRONTEND:-${PMFLOW_TUNNEL_PROTOCOL:-http2}}"
+    strict_public_health="${PMFLOW_TUNNEL_STRICT_PUBLIC_HEALTH:-0}"
+    public_health_interval="${PMFLOW_TUNNEL_PUBLIC_HEALTH_INTERVAL_SECONDS:-30}"
+    public_health_failure_threshold="${PMFLOW_TUNNEL_PUBLIC_HEALTH_FAILURE_THRESHOLD:-5}"
+    : > "$RUN_LOG"
+    log_watch "origin=http://127.0.0.1:${frontend_port} protocol=${tunnel_protocol}"
+
+    env -u HTTPS_PROXY -u HTTP_PROXY -u ALL_PROXY -u https_proxy -u http_proxy -u all_proxy \
+      NO_PROXY="*" \
+      "$cf_bin" tunnel --protocol "$tunnel_protocol" --url "http://127.0.0.1:${frontend_port}" >"$RUN_LOG" 2>&1 &
+    child_pid="$!"
+    ACTIVE_CHILD_PID="$child_pid"
+    echo "$child_pid" > "$TUNNEL_PID_FILE"
+
+    published="0"
+    current_url=""
+    for _ in {1..90}; do
+      url="$(extract_tunnel_url "$RUN_LOG" || true)"
+      if [[ -n "$url" ]]; then
+        if [[ "$strict_public_health" == "1" ]] && ! probe_tunnel_url_stable "$url"; then
+          :
+        else
+          if [[ "$url" != "$current_url" ]]; then
+            write_frontend_url "$url"
+            log_watch "url=$url"
+          fi
+          current_url="$url"
+          published="1"
+          break
+        fi
+      fi
+      if ! is_pid_running "$child_pid"; then
+        break
+      fi
+      sleep 1
+    done
+
+    unhealthy_count=0
+    last_health_check_at=0
+    while is_pid_running "$child_pid"; do
+      if [[ "$strict_public_health" == "1" && "$published" == "1" && -n "$current_url" ]]; then
+        if (( $(date +%s) - last_health_check_at >= public_health_interval )); then
+          last_health_check_at="$(date +%s)"
+          if probe_tunnel_url "$current_url"; then
+            unhealthy_count=0
+          else
+            unhealthy_count=$((unhealthy_count + 1))
+            log_watch_err "health_check_failed url=${current_url} count=${unhealthy_count}"
+          fi
+          if [[ "$unhealthy_count" -ge "$public_health_failure_threshold" ]]; then
+            log_watch_err "url_unhealthy_restart url=${current_url}"
+            kill "$child_pid" >/dev/null 2>&1 || true
+            break
+          fi
+        fi
+      fi
+      sleep 5
+    done
+
+    wait "$child_pid" || true
+    ACTIVE_CHILD_PID=""
+    if [[ "$published" != "1" ]]; then
+      log_watch_err "exited_before_url"
+    else
+      log_watch_err "exited_restart"
+    fi
+    sleep 2
+  done
+}
+
+cmd="${1:-run}"
+case "$cmd" in
+  run)
+    run_watch_loop
+    ;;
+  start)
+    start_watch_background
+    ;;
+  stop)
+    stop_watch
+    ;;
+  status)
+    print_status
+    ;;
+  url)
+    front="$(read_frontend_url)"
+    back="$(read_backend_url)"
+    echo "前端地址: ${front:-未就绪}"
+    echo "后端地址: ${back:-未配置/未暴露}"
+    ;;
+  *)
+    echo "Usage: $0 {run|start|stop|status|url}"
+    exit 1
+    ;;
+esac
diff --git a/tests/__init__.py b/tests/__init__.py
new file mode 100644
index 00000000..d4839a6b
--- /dev/null
+++ b/tests/__init__.py
@@ -0,0 +1 @@
+# Tests package
diff --git a/tests/test_adaptive_rag_ack_first.py b/tests/test_adaptive_rag_ack_first.py
new file mode 100644
index 00000000..4d31aa2a
--- /dev/null
+++ b/tests/test_adaptive_rag_ack_first.py
@@ -0,0 +1,83 @@
+from backend.app.services.adaptive_rag_service import (
+    ACTION_MODE_READ,
+    ACTION_MODE_READ_WRITE,
+    ACTION_MODE_WRITE,
+    AdaptiveRAGService,
+)
+
+
+def test_resolve_action_mode_as_write_for_requirement_command():
+    mode = AdaptiveRAGService._resolve_action_mode(
+        requested_mode="auto",
+        question="给健身项目新增一个信号：各类健身动作消耗的大概卡路里",
+        effective_question="给健身项目新增一个信号：各类健身动作消耗的大概卡路里",
+    )
+    assert mode == ACTION_MODE_WRITE
+
+
+def test_resolve_action_mode_as_read_write_for_requirement_question():
+    mode = AdaptiveRAGService._resolve_action_mode(
+        requested_mode="auto",
+        question="要不要给健身项目新增一个信号：各类健身动作消耗的大概卡路里？",
+        effective_question="要不要给健身项目新增一个信号：各类健身动作消耗的大概卡路里？",
+    )
+    assert mode == ACTION_MODE_READ_WRITE
+
+
+def test_resolve_action_mode_keeps_read_for_plain_question():
+    mode = AdaptiveRAGService._resolve_action_mode(
+        requested_mode="auto",
+        question="健身项目现在有哪些阻塞项？",
+        effective_question="健身项目现在有哪些阻塞项？",
+    )
+    assert mode == ACTION_MODE_READ
+
+
+def test_build_requirement_ack_text_variants():
+    created = AdaptiveRAGService._build_requirement_ack_text(
+        action_taken="signal_created",
+        signal_id=123,
+        project_id=5,
+        error_reason="",
+    )
+    deduped = AdaptiveRAGService._build_requirement_ack_text(
+        action_taken="signal_deduped",
+        signal_id=123,
+        project_id=5,
+        error_reason="",
+    )
+    failed = AdaptiveRAGService._build_requirement_ack_text(
+        action_taken="signal_failed",
+        signal_id=None,
+        project_id=5,
+        error_reason="db unavailable",
+    )
+    assert created == "已入库：已为项目#5创建信号 #123，并进入待处理队列。"
+    assert deduped == "已入库：检测到重复需求，已复用项目#5的信号 #123。"
+    assert "未完成入库" in failed
+
+
+def test_extract_ack_metadata_from_sources():
+    meta = AdaptiveRAGService._extract_ack_metadata_from_sources(
+        [
+            {
+                "source": "pmflow",
+                "label": "Signal Queue",
+                "score": 0.98,
+                "metadata": {
+                    "action_taken": "signal_created",
+                    "signal_id": 77,
+                    "trace_id": "rag-ack-1",
+                    "processing_stage": "ack_only",
+                    "ack_latency_ms": 320,
+                },
+            }
+        ],
+        answer="已入库：已为项目#5创建信号 #77，并进入待处理队列。",
+    )
+    assert meta["action_taken"] == "signal_created"
+    assert meta["signal_id"] == 77
+    assert meta["reply_required"] is True
+    assert meta["processing_stage"] == "ack_only"
+    assert meta["ack_latency_ms"] == 320
+    assert meta["reply_text"] == "已入库：已为项目#5创建信号 #77，并进入待处理队列。"
diff --git a/tests/test_adaptive_rag_deep_policy.py b/tests/test_adaptive_rag_deep_policy.py
new file mode 100644
index 00000000..8656cb22
--- /dev/null
+++ b/tests/test_adaptive_rag_deep_policy.py
@@ -0,0 +1,48 @@
+from backend.app.services.adaptive_rag_service import AdaptiveRAGService
+
+
+def test_feishu_no_explicit_deep_should_not_force():
+    should = AdaptiveRAGService._should_force_deep_for_channel(
+        asked_via="feishu",
+        question="不是啊，我说的是 ai 引擎驱动方式",
+        effective_question="不是啊，我说的是 ai 引擎驱动方式",
+        reason="correction_turn",
+    )
+    assert not should
+
+
+def test_feishu_explicit_deep_can_force():
+    should = AdaptiveRAGService._should_force_deep_for_channel(
+        asked_via="feishu",
+        question="请深入分析调用链并给证据链",
+        effective_question="请深入分析调用链并给证据链",
+        reason="engine_process_query",
+    )
+    assert should
+
+
+def test_pmflow_force_deep_kept():
+    should = AdaptiveRAGService._should_force_deep_for_channel(
+        asked_via="pmflow",
+        question="不是啊，我说的是 ai 引擎驱动方式",
+        effective_question="不是啊，我说的是 ai 引擎驱动方式",
+        reason="correction_turn",
+    )
+    assert should
+
+
+def test_feishu_auto_retry_requires_explicit_deep():
+    allow_plain = AdaptiveRAGService._allow_auto_deep_retry_for_channel(
+        asked_via="feishu",
+        question="这个问题不太对",
+        effective_question="这个问题不太对",
+        reason="snapshot_only",
+    )
+    allow_explicit = AdaptiveRAGService._allow_auto_deep_retry_for_channel(
+        asked_via="feishu",
+        question="请深度分析这个调用链",
+        effective_question="请深度分析这个调用链",
+        reason="snapshot_only",
+    )
+    assert not allow_plain
+    assert allow_explicit
diff --git a/tests/test_adaptive_rag_external_conversation.py b/tests/test_adaptive_rag_external_conversation.py
new file mode 100644
index 00000000..d47fe458
--- /dev/null
+++ b/tests/test_adaptive_rag_external_conversation.py
@@ -0,0 +1,52 @@
+import pytest
+
+from backend.app.services.adaptive_rag_service import AdaptiveRAGService
+
+
+class _FakeContextCache:
+    async def get_session_messages(self, *, tenant_id: int, session_id: int):
+        if tenant_id == 1 and session_id == 1500000001:
+            return [
+                {"role": "user", "content": "上一轮问题"},
+                {"role": "assistant", "content": "上一轮回答"},
+            ]
+        return []
+
+    @staticmethod
+    def messages_to_turns(messages):
+        turns = []
+        q = None
+        for row in messages:
+            role = row.get("role")
+            content = row.get("content")
+            if role == "user":
+                q = content
+                continue
+            if role == "assistant" and q:
+                turns.append({"question": q, "answer": content})
+                q = None
+        return turns
+
+
+def test_external_session_id_stable():
+    sid1 = AdaptiveRAGService._external_session_id_from_key("feishu:oc_xxx:ou_yyy")
+    sid2 = AdaptiveRAGService._external_session_id_from_key("feishu:oc_xxx:ou_yyy")
+    assert sid1 == sid2
+    assert sid1 is not None
+    assert 1500000000 <= int(sid1) < 2000000000
+
+
+@pytest.mark.asyncio
+async def test_load_recent_conversation_for_external_actor():
+    service = AdaptiveRAGService(db=None, tenant_id=1)  # type: ignore[arg-type]
+    service.context_cache = _FakeContextCache()  # type: ignore[assignment]
+
+    rows = await service._load_recent_conversation(
+        asked_by=None,
+        asked_via="feishu",
+        project_id=None,
+        session_id=None,
+        external_session_id=1500000001,
+        limit=10,
+    )
+    assert rows == [{"question": "上一轮问题", "answer": "上一轮回答"}]
diff --git a/tests/test_dispatcher.py b/tests/test_dispatcher.py
new file mode 100644
index 00000000..adf8b885
--- /dev/null
+++ b/tests/test_dispatcher.py
@@ -0,0 +1,264 @@
+"""
+Neural Sync - Role Dispatcher Unit Tests
+"""
+
+import pytest
+from backend.app.dispatcher import (
+    Role,
+    RoutingRule,
+    RoleRouter,
+    RoutingResult,
+    route_signal,
+    get_router,
+)
+
+
+class TestRole:
+    """测试 Role 枚举"""
+
+    def test_role_values(self):
+        """测试角色值"""
+        assert Role.BOSS.value == "boss"
+        assert Role.AI_ENGINEER.value == "ai_engineer"
+        assert Role.BACKEND.value == "backend"
+        assert Role.FRONTEND.value == "frontend"
+
+
+class TestRoleRouter:
+    """测试 RoleRouter"""
+
+    def test_router_initialization(self):
+        """测试路由器初始化"""
+        router = RoleRouter()
+        assert len(router.ROLE_RULES) == 4
+        assert router._compiled_patterns is not None
+
+    def test_get_role_info(self):
+        """测试获取角色信息"""
+        router = RoleRouter()
+        boss_info = router.get_role_info(Role.BOSS)
+        assert boss_info["role"] == "boss"
+        assert boss_info["name"] == "战略决策"
+        assert "keywords" in boss_info
+
+    def test_get_all_roles(self):
+        """测试获取所有角色"""
+        router = RoleRouter()
+        roles = router.get_all_roles()
+        assert len(roles) == 4
+        # 应该按优先级降序排列
+        assert roles[0]["priority"] >= roles[1]["priority"]
+
+    def test_route_to_boss_by_cost(self):
+        """测试根据成本路由到老板"""
+        router = RoleRouter()
+        analysis = {"resource_cost": {"dev_days": 10}}
+        role, confidence, details = router.route("需要重构系统架构", analysis)
+        assert role == Role.BOSS
+        assert confidence == 1.0
+        assert "threshold_check" in details.get("matched_rules", [])
+
+    def test_route_to_boss_by_risk(self):
+        """测试根据风险路由到老板"""
+        router = RoleRouter()
+        analysis = {"risk_level": "CRITICAL"}
+        role, confidence, details = router.route("修复安全漏洞", analysis)
+        assert role == Role.BOSS
+        assert confidence == 1.0
+
+    def test_route_to_ai_engineer(self):
+        """测试路由到 AI 工程师"""
+        router = RoleRouter()
+        role, confidence, details = router.route("优化 prompt 提高模型准确率", {})
+        assert role == Role.AI_ENGINEER
+        assert confidence > 0
+
+    def test_route_to_backend(self):
+        """测试路由到后端开发"""
+        router = RoleRouter()
+        role, confidence, details = router.route("使用 FastAPI 重构 API 接口", {})
+        assert role == Role.BACKEND
+        assert confidence > 0
+
+    def test_route_to_frontend(self):
+        """测试路由到前端开发"""
+        router = RoleRouter()
+        role, confidence, details = router.route("使用 Vue 3 重构仪表盘组件", {})
+        assert role == Role.FRONTEND
+        assert confidence > 0
+
+    def test_route_with_multiple_keywords(self):
+        """测试多个关键词匹配"""
+        router = RoleRouter()
+        content = "使用 FastAPI 和 PostgreSQL 优化数据库查询性能"
+        role, confidence, details = router.route(content, {})
+        assert role == Role.BACKEND
+        # 多个关键词应该提高置信度
+        assert confidence > 0.5
+
+    def test_route_with_pattern_matching(self):
+        """测试正则模式匹配"""
+        router = RoleRouter()
+        content = "创建新的 component.tsx 文件"
+        role, confidence, details = router.route(content, {})
+        # 应该匹配到前端（tsx 模式）
+        assert role == Role.FRONTEND
+
+    def test_route_default_to_ai_engineer(self):
+        """测试默认路由到 AI 工程师"""
+        router = RoleRouter()
+        # 没有明显关键词的内容
+        role, confidence, details = router.route("这是一些普通的描述文字", {})
+        assert role == Role.AI_ENGINEER
+
+    def test_route_with_low_confidence(self):
+        """测试低置信度路由"""
+        router = RoleRouter()
+        role, confidence, details = router.route("处理用户反馈", {})
+        assert confidence < 0.5  # 应该是较低的置信度
+
+    def test_route_tie_breaker_by_priority(self):
+        """测试优先级平局决胜"""
+        router = RoleRouter()
+        # 同时匹配前端和后端关键词，但后端优先级应该更高
+        content = "修复 API 接口并更新 UI 组件"
+        role, confidence, details = router.route(content, {})
+        # 应该根据优先级选择
+        assert role in [Role.BACKEND, Role.FRONTEND]
+
+    def test_route_details(self):
+        """测试路由详情"""
+        router = RoleRouter()
+        role, confidence, details = router.route("使用 FastAPI 创建 REST API", {})
+
+        assert "matched_rules" in details
+        assert "scores" in details
+        assert "reasons" in details
+        assert "final_scores" in details
+
+    def test_should_escalate_to_boss(self):
+        """测试是否应该升级到老板"""
+        router = RoleRouter()
+
+        # 测试成本阈值
+        assert router._should_escalate_to_boss({"resource_cost": {"dev_days": 5}})
+        assert router._should_escalate_to_boss({"resource_cost": {"dev_days": 10}})
+
+        # 测试风险阈值
+        assert router._should_escalate_to_boss({"risk_level": "CRITICAL"})
+        assert router._should_escalate_to_boss({"risk_level": "HIGH"})
+        assert not router._should_escalate_to_boss({"risk_level": "MEDIUM"})
+
+        # 测试架构变更
+        assert router._should_escalate_to_boss({"is_architecture_change": True})
+
+        # 测试不需要升级
+        assert not router._should_escalate_to_boss({})
+
+
+class TestRoutingResult:
+    """测试 RoutingResult"""
+
+    def test_routing_result_creation(self):
+        """测试创建路由结果"""
+        result = RoutingResult(
+            role="boss",
+            confidence=0.95,
+            details={"matched_rules": ["threshold_check"]},
+            signal_id="test-123",
+        )
+        assert result.role == "boss"
+        assert result.confidence == 0.95
+        assert result.signal_id == "test-123"
+        assert result.timestamp is not None
+
+
+class TestGlobalRouter:
+    """测试全局路由器"""
+
+    def test_get_router_singleton(self):
+        """测试获取单例"""
+        router1 = get_router()
+        router2 = get_router()
+        assert router1 is router2
+
+
+class TestConvenienceFunctions:
+    """测试便捷函数"""
+
+    def test_route_signal(self):
+        """测试路由信号便捷函数"""
+        result = route_signal("优化 prompt 提高准确率", signal_id="signal-1")
+
+        assert isinstance(result, RoutingResult)
+        assert result.signal_id == "signal-1"
+        assert result.role in ["boss", "ai_engineer", "backend", "frontend"]
+        assert 0 <= result.confidence <= 1
+
+
+class TestRealWorldScenarios:
+    """真实场景测试"""
+
+    def test_security_issue(self):
+        """测试安全问题 - 应该路由到老板"""
+        result = route_signal("发现严重的 SQL 注入漏洞，需要紧急修复", {"risk_level": "CRITICAL"})
+        assert result.role == "boss"
+
+    def test_large_feature_request(self):
+        """测试大功能需求 - 应该路由到老板"""
+        result = route_signal(
+            "重写整个用户认证系统，支持 OAuth2 和 SAML",
+            {"resource_cost": {"dev_days": 15}},
+        )
+        assert result.role == "boss"
+
+    def test_llm_optimization(self):
+        """测试 LLM 优化 - 应该路由到 AI 工程师"""
+        result = route_signal("优化 embedding 模型提高检索准确率")
+        assert result.role == "ai_engineer"
+
+    def test_api_development(self):
+        """测试 API 开发 - 应该路由到后端"""
+        result = route_signal("创建新的 REST API 端点用于用户管理")
+        assert result.role == "backend"
+
+    def test_frontend_component(self):
+        """测试前端组件 - 应该路由到前端"""
+        result = route_signal("使用 Vue 3 创建新的数据可视化组件")
+        assert result.role == "frontend"
+
+    def test_ambiguous_request(self):
+        """测试模糊需求 - 应该有默认行为"""
+        result = route_signal("优化系统性能")
+        # 应该有一个合理的默认路由
+        assert result.role in ["boss", "ai_engineer", "backend", "frontend"]
+        # 置信度应该较低
+        assert result.confidence < 0.7
+
+
+class TestEdgeCases:
+    """边缘情况测试"""
+
+    def test_empty_content(self):
+        """测试空内容"""
+        result = route_signal("")
+        assert result.role in ["boss", "ai_engineer", "backend", "frontend"]
+
+    def test_chinese_content(self):
+        """测试中文内容"""
+        result = route_signal("使用 Vue 3 开发前端组件")
+        assert result.role == "frontend"
+
+    def test_mixed_language_content(self):
+        """测试混合语言内容"""
+        result = route_signal("创建新的 FastAPI endpoint 用于数据处理")
+        assert result.role == "backend"
+
+    def test_technical_jargon(self):
+        """测试技术术语"""
+        result = route_signal("Implement RAG pipeline with vector database")
+        assert result.role == "ai_engineer"
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])
diff --git a/tests/test_events.py b/tests/test_events.py
new file mode 100644
index 00000000..cdb3f1ba
--- /dev/null
+++ b/tests/test_events.py
@@ -0,0 +1,340 @@
+"""
+Neural Sync - Event Bus Unit Tests
+"""
+
+import pytest
+import asyncio
+from datetime import datetime
+from backend.app.events import (
+    EventType,
+    Event,
+    EventBus,
+    get_event_bus,
+    event_bus_context,
+    publish_event,
+    subscribe_event,
+)
+
+
+@pytest.fixture
+async def event_bus():
+    """创建事件总线 fixture"""
+    bus = EventBus()
+    await bus.start()
+    yield bus
+    await bus.stop()
+
+
+@pytest.fixture
+async def event_bus_context_fixture():
+    """创建事件总线上下文 fixture"""
+    async with event_bus_context() as bus:
+        yield bus
+
+
+class TestEvent:
+    """测试 Event 数据类"""
+
+    def test_create_event(self):
+        """测试创建事件"""
+        event = Event(
+            type=EventType.SIGNAL_INGESTED,
+            data={"signal_id": "123", "content": "Test signal"},
+            source="test",
+        )
+        assert event.type == EventType.SIGNAL_INGESTED
+        assert event.data["signal_id"] == "123"
+        assert event.source == "test"
+        assert isinstance(event.timestamp, float)
+
+    def test_event_to_dict(self):
+        """测试事件转字典"""
+        event = Event(
+            type=EventType.SIGNAL_ANALYZED,
+            data={"score": 0.95},
+            correlation_id="test-123",
+        )
+        event_dict = event.to_dict()
+        assert event_dict["type"] == "signal.analyzed"
+        assert event_dict["data"]["score"] == 0.95
+        assert event_dict["correlation_id"] == "test-123"
+
+
+class TestEventBus:
+    """测试 EventBus"""
+
+    @pytest.mark.asyncio
+    async def test_start_stop(self, event_bus):
+        """测试启动和停止"""
+        assert event_bus._running is True
+        await event_bus.stop()
+        assert event_bus._running is False
+
+    @pytest.mark.asyncio
+    async def test_subscribe_and_publish(self, event_bus):
+        """测试订阅和发布"""
+        received_events = []
+
+        async def handler(event: Event):
+            received_events.append(event)
+
+        # 订阅事件
+        unsubscribe = event_bus.subscribe(EventType.SIGNAL_INGESTED, handler)
+
+        # 发布事件
+        event = Event(
+            type=EventType.SIGNAL_INGESTED, data={"signal_id": "123"}
+        )
+        await event_bus.publish(event)
+
+        # 等待异步处理
+        await asyncio.sleep(0.1)
+
+        assert len(received_events) == 1
+        assert received_events[0].data["signal_id"] == "123"
+
+        # 取消订阅
+        unsubscribe()
+
+        # 发布另一个事件
+        await event_bus.publish(event)
+        await asyncio.sleep(0.1)
+
+        # 应该没有新事件
+        assert len(received_events) == 1
+
+    @pytest.mark.asyncio
+    async def test_wildcard_subscription(self, event_bus):
+        """测试通配符订阅"""
+        received_events = []
+
+        async def handler(event: Event):
+            received_events.append(event.type)
+
+        event_bus.subscribe_wildcard(handler)
+
+        # 发布不同类型的事件
+        await event_bus.publish(
+            Event(type=EventType.SIGNAL_INGESTED, data={})
+        )
+        await event_bus.publish(
+            Event(type=EventType.TASK_ASSIGNED, data={})
+        )
+        await asyncio.sleep(0.1)
+
+        assert len(received_events) == 2
+        assert EventType.SIGNAL_INGESTED in received_events
+        assert EventType.TASK_ASSIGNED in received_events
+
+    @pytest.mark.asyncio
+    async def test_event_filter(self, event_bus):
+        """测试事件过滤"""
+        received_events = []
+
+        async def handler(event: Event):
+            received_events.append(event)
+
+        # 只接收 signal_id > 100 的事件
+        def filter_func(event: Event):
+            return event.data.get("signal_id", 0) > 100
+
+        event_bus.subscribe(
+            EventType.SIGNAL_INGESTED, handler, filter_func=filter_func
+        )
+
+        # 发布事件
+        await event_bus.publish(
+            Event(type=EventType.SIGNAL_INGESTED, data={"signal_id": 50})
+        )
+        await event_bus.publish(
+            Event(type=EventType.SIGNAL_INGESTED, data={"signal_id": 150})
+        )
+        await asyncio.sleep(0.1)
+
+        # 只应该收到一个事件
+        assert len(received_events) == 1
+        assert received_events[0].data["signal_id"] == 150
+
+    @pytest.mark.asyncio
+    async def test_publish_and_wait(self, event_bus):
+        """测试发布并等待处理完成"""
+        results = []
+
+        async def handler1(event: Event):
+            results.append("handler1")
+            await asyncio.sleep(0.1)
+            return "result1"
+
+        async def handler2(event: Event):
+            results.append("handler2")
+            return "result2"
+
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, handler1)
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, handler2)
+
+        event = Event(type=EventType.SIGNAL_INGESTED, data={})
+        handler_results = await event_bus.publish_and_wait(event, timeout=1.0)
+
+        assert len(results) == 2
+        assert len(handler_results) == 2
+
+    @pytest.mark.asyncio
+    async def test_metrics(self, event_bus):
+        """测试事件指标"""
+        async def handler(event: Event):
+            pass
+
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, handler)
+
+        await event_bus.publish(
+            Event(type=EventType.SIGNAL_INGESTED, data={})
+        )
+        await asyncio.sleep(0.1)
+
+        metrics = event_bus.metrics
+        assert metrics["published"] >= 1
+        assert metrics["delivered"] >= 1
+
+    @pytest.mark.asyncio
+    async def test_event_history(self, event_bus):
+        """测试事件历史"""
+        event_bus._max_history = 5
+
+        # 发布超过历史限制的事件
+        for i in range(10):
+            await event_bus.publish(
+                Event(type=EventType.SIGNAL_INGESTED, data={"index": i})
+            )
+
+        history = event_bus.history
+        assert len(history) == 5
+        # 最新的事件应该在历史中
+        assert history[-1]["data"]["index"] == 9
+
+    @pytest.mark.asyncio
+    async def test_subscriber_count(self, event_bus):
+        """测试订阅者计数"""
+        async def handler1(event: Event):
+            pass
+
+        async def handler2(event: Event):
+            pass
+
+        # 特定事件订阅
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, handler1)
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, handler2)
+
+        # 通配符订阅
+        event_bus.subscribe_wildcard(handler1)
+
+        assert event_bus.get_subscriber_count(EventType.SIGNAL_INGESTED) == 2
+        assert event_bus.get_subscriber_count() == 3  # 包括通配符订阅
+
+    @pytest.mark.asyncio
+    async def test_handler_error_handling(self, event_bus):
+        """测试处理器错误处理"""
+        received_normal = []
+
+        async def failing_handler(event: Event):
+            raise ValueError("Test error")
+
+        async def normal_handler(event: Event):
+            received_normal.append(event)
+
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, failing_handler)
+        event_bus.subscribe(EventType.SIGNAL_INGESTED, normal_handler)
+
+        await event_bus.publish(
+            Event(type=EventType.SIGNAL_INGESTED, data={})
+        )
+        await asyncio.sleep(0.1)
+
+        # 正常处理器应该仍然被调用
+        assert len(received_normal) == 1
+        # 错误应该被记录
+        assert event_bus.metrics["failed"] >= 1
+
+
+class TestGlobalEventBus:
+    """测试全局事件总线"""
+
+    @pytest.mark.asyncio
+    async def test_get_event_bus_singleton(self):
+        """测试获取单例"""
+        bus1 = get_event_bus()
+        bus2 = get_event_bus()
+        assert bus1 is bus2
+
+    @pytest.mark.asyncio
+    async def test_convenience_functions(self):
+        """测试便捷函数"""
+        received = []
+
+        async def handler(event: Event):
+            received.append(event)
+
+        # 使用便捷函数订阅
+        unsubscribe = subscribe_event(EventType.SIGNAL_INGESTED, handler)
+
+        # 使用便捷函数发布
+        await publish_event(EventType.SIGNAL_INGESTED, {"test": True})
+        await asyncio.sleep(0.1)
+
+        assert len(received) == 1
+
+        # 取消订阅
+        unsubscribe()
+
+    @pytest.mark.asyncio
+    async def test_event_bus_context(self):
+        """测试事件总线上下文管理器"""
+        async with event_bus_context() as bus:
+            assert bus._running is True
+
+            received = []
+
+            async def handler(event: Event):
+                received.append(event)
+
+            bus.subscribe(EventType.SIGNAL_INGESTED, handler)
+            await bus.publish(Event(type=EventType.SIGNAL_INGESTED, data={}))
+            await asyncio.sleep(0.1)
+
+            assert len(received) == 1
+
+        # 退出上下文后应该停止
+        assert bus._running is False
+
+
+@pytest.mark.asyncio
+async def test_multiple_event_types():
+    """测试多种事件类型"""
+    bus = EventBus()
+    await bus.start()
+
+    try:
+        counters = {event_type: 0 for event_type in EventType}
+
+        async def count_handler(event: Event):
+            counters[event.type] += 1
+
+        # 订阅所有事件类型
+        for event_type in EventType:
+            bus.subscribe(event_type, count_handler)
+
+        # 发布各种事件
+        await bus.publish(Event(type=EventType.SIGNAL_INGESTED, data={}))
+        await bus.publish(Event(type=EventType.TASK_ASSIGNED, data={}))
+        await bus.publish(Event(type=EventType.CODE_COMMIT, data={}))
+        await asyncio.sleep(0.1)
+
+        assert counters[EventType.SIGNAL_INGESTED] == 1
+        assert counters[EventType.TASK_ASSIGNED] == 1
+        assert counters[EventType.CODE_COMMIT] == 1
+
+    finally:
+        await bus.stop()
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])
diff --git a/tests/test_experiment_sync_flow.py b/tests/test_experiment_sync_flow.py
new file mode 100644
index 00000000..1402b081
--- /dev/null
+++ b/tests/test_experiment_sync_flow.py
@@ -0,0 +1,118 @@
+import pytest
+
+from backend.app.services.experiment_sync_service import ExperimentSyncService
+
+
+class _FakeExperiment:
+    def __init__(self):
+        self.id = 11
+        self.tenant_id = 1
+        self.title = "exp"
+        self.status = "running"
+        self.target_metric = "acc"
+        self.current_score = 0.9
+        self.linked_context_nodes = [101]
+        self.project_id = 1
+
+
+class _FakeTruthNode:
+    def __init__(self):
+        self.id = 101
+        self.tenant_id = 1
+        self.project_id = 1
+        self.content = {}
+        self.version = 1
+
+
+class _FakeSyncRecord:
+    def __init__(self):
+        self.id = 999
+        self.tenant_id = 1
+        self.experiment_id = 11
+        self.truth_node_id = 101
+        self.sync_type = "result_sync"
+        self.confidence_score = 0.9
+        self.auto_applied = True
+        self.apply_status = "applied"
+        self.sync_data = {"previous_truth_content": {}}
+        self.reviewed_at = None
+
+
+class _FakeResult:
+    def __init__(self, value):
+        self._value = value
+
+    def scalar_one_or_none(self):
+        return self._value
+
+    def all(self):
+        if isinstance(self._value, list):
+            return self._value
+        return []
+
+
+class _FakeDB:
+    def __init__(self):
+        self.experiment = _FakeExperiment()
+        self.truth_node = _FakeTruthNode()
+        self.sync_record = _FakeSyncRecord()
+        self.added = []
+
+    async def execute(self, query):
+        sql = str(query)
+        if "experiment_truth_node_syncs" in sql:
+            return _FakeResult(self.sync_record)
+        if "experiments" in sql:
+            return _FakeResult(self.experiment)
+        if "SELECT truth_nodes.id," in sql:
+            return _FakeResult(self.truth_node)
+        if "SELECT truth_nodes.id" in sql:
+            return _FakeResult([(101,)])
+        if "FROM truth_nodes" in sql:
+            return _FakeResult(self.truth_node)
+        return _FakeResult(self.truth_node)
+
+    def add(self, obj):
+        self.added.append(obj)
+        if hasattr(obj, "id") and getattr(obj, "id", None) is None:
+            setattr(obj, "id", 1001)
+
+    async def commit(self):
+        return None
+
+    async def refresh(self, obj):
+        if hasattr(obj, "id") and getattr(obj, "id", None) is None:
+            setattr(obj, "id", 1001)
+        return None
+
+
+@pytest.mark.asyncio
+async def test_sync_create_and_review_and_rollback():
+    db = _FakeDB()
+    service = ExperimentSyncService(db=db)  # type: ignore[arg-type]
+
+    created = await service.auto_sync_experiment_result(
+        experiment_id=11,
+        truth_node_id=101,
+        tenant_id=1,
+        project_id=1,
+        auto_apply_threshold=0.8,
+    )
+    assert created["auto_applied"] is True
+    assert created["apply_status"] == "applied"
+
+    reviewed = await service.review_sync(
+        sync_id=999,
+        tenant_id=1,
+        reviewer_id=1,
+        action="approve",
+    )
+    assert reviewed["apply_status"] in ("applied",)
+
+    rolled = await service.rollback_sync(
+        sync_id=999,
+        tenant_id=1,
+        reviewer_id=1,
+    )
+    assert rolled["rolled_back"] is True
+    assert rolled["apply_status"] == "requires_review"
diff --git a/tests/test_feishu_pipeline_fallbacks.py b/tests/test_feishu_pipeline_fallbacks.py
new file mode 100644
index 00000000..9d598e2f
--- /dev/null
+++ b/tests/test_feishu_pipeline_fallbacks.py
@@ -0,0 +1,51 @@
+import pytest
+
+from backend.app.services.feishu_signal_pipeline import FeishuSignalPipelineService
+
+
+@pytest.mark.asyncio
+async def test_classify_level_fallback_when_llm_empty(monkeypatch):
+    service = FeishuSignalPipelineService(db=None, tenant_id=1)
+
+    async def _fake_generate(*args, **kwargs):
+        return ""
+
+    monkeypatch.setattr(service.llm, "generate", _fake_generate)
+    result = await service._classify_level(message="我们需要新增导出按钮", projects=[])
+
+    assert result["level"] == "signal"
+    assert 0.0 <= float(result["project_score"]) <= 1.0
+    assert str(result["level_reasoning"]).strip()
+
+
+@pytest.mark.asyncio
+async def test_extract_signal_fastpath_for_simple_request(monkeypatch):
+    service = FeishuSignalPipelineService(db=None, tenant_id=1)
+
+    async def _should_not_be_called(*args, **kwargs):
+        raise AssertionError("LLM should not be called on simple signal request fastpath")
+
+    monkeypatch.setattr(service.llm, "generate", _should_not_be_called)
+    result = await service._extract_signal(
+        "新增一个导出按钮，支持飞书同步",
+        projects=[{"id": 3, "name": "飞书协同平台"}],
+    )
+
+    assert result["type"] == "新需求"
+    assert result["priority"] in {"P1", "P2", "P3"}
+    assert str(result["summary"]).strip()
+
+
+@pytest.mark.asyncio
+async def test_extract_project_prefill_fallback_when_llm_empty(monkeypatch):
+    service = FeishuSignalPipelineService(db=None, tenant_id=1)
+
+    async def _fake_generate(*args, **kwargs):
+        return ""
+
+    monkeypatch.setattr(service.llm, "generate", _fake_generate)
+    result = await service._extract_project_prefill("我们要做一个飞书需求同步项目")
+
+    assert str(result["suggested_name"]).strip()
+    assert str(result["suggested_description"]).strip()
+    assert str(result["suggested_owner"]).strip()
diff --git a/tests/test_feishu_question_requirement_routing.py b/tests/test_feishu_question_requirement_routing.py
new file mode 100644
index 00000000..383225be
--- /dev/null
+++ b/tests/test_feishu_question_requirement_routing.py
@@ -0,0 +1,140 @@
+import pytest
+
+from backend.app.services.adaptive_rag_service import AdaptiveRAGService
+from backend.app.services.feishu_signal_pipeline import FeishuSignalPipelineService
+
+
+def test_requirement_like_question_detected():
+    assert FeishuSignalPipelineService._is_requirement_like_question("这个接口要不要支持飞书同步？")
+    assert FeishuSignalPipelineService._is_requirement_like_question("能不能新增一个导出按钮？")
+    assert FeishuSignalPipelineService._is_requirement_like_question("是否需要补充需求优先级字段？")
+
+
+def test_non_requirement_question_not_detected():
+    assert not FeishuSignalPipelineService._is_requirement_like_question("今天几号？")
+    assert not FeishuSignalPipelineService._is_requirement_like_question("你是谁？")
+    assert not FeishuSignalPipelineService._is_requirement_like_question("这个现在怎么样？")
+
+
+def test_question_to_statement_normalization():
+    normalized = FeishuSignalPipelineService._extract_requirement_statement_from_question("这个接口要不要支持飞书同步？")
+    assert "支持飞书同步" in normalized
+    assert "？" not in normalized
+
+    normalized2 = FeishuSignalPipelineService._extract_requirement_statement_from_question("能不能新增一个导出按钮？")
+    assert "新增一个导出按钮" in normalized2
+    assert "？" not in normalized2
+
+
+def test_build_rag_conversation_key():
+    key = FeishuSignalPipelineService._build_rag_conversation_key(
+        source_chat_id="oc_123",
+        source_sender_open_id="ou_abc",
+        sender="张三",
+    )
+    assert key == "feishu:oc_123:ou_abc"
+
+    key2 = FeishuSignalPipelineService._build_rag_conversation_key(
+        source_chat_id="oc_123",
+        source_sender_open_id=None,
+        sender=" 张 三 ",
+    )
+    assert key2 == "feishu:oc_123:张三"
+
+
+def test_openclaw_command_detection():
+    assert FeishuSignalPipelineService._is_openclaw_business_command("帮我创建一个项目，财务冲账协同")
+    assert FeishuSignalPipelineService._is_openclaw_business_command("把这个项目归档")
+    assert not FeishuSignalPipelineService._is_openclaw_business_command("今天几号？")
+
+
+def test_build_openclaw_command_reply_for_project_create():
+    reply = FeishuSignalPipelineService._build_openclaw_command_reply(
+        {
+            "execution": {
+                "applied_count": 1,
+                "skipped_count": 0,
+                "failed_count": 0,
+                "results": [
+                    {
+                        "operation_code": "project.create",
+                        "status": "applied",
+                        "entity_type": "project",
+                        "entity_id": "8",
+                        "result": {
+                            "project_id": 8,
+                            "title": "财务冲账协同",
+                            "status": "active",
+                        },
+                    }
+                ],
+            }
+        }
+    )
+
+    assert reply == "已执行完成：项目#8「财务冲账协同」已创建，状态 active。"
+
+
+def test_build_openclaw_command_reply_for_existing_project():
+    reply = FeishuSignalPipelineService._build_openclaw_command_reply(
+        {
+            "execution": {
+                "applied_count": 0,
+                "skipped_count": 1,
+                "failed_count": 0,
+                "results": [
+                    {
+                        "operation_code": "project.create",
+                        "status": "skipped",
+                        "message": "project_already_exists",
+                        "entity_type": "project",
+                        "entity_id": "8",
+                        "result": {
+                            "project_id": 8,
+                            "title": "财务冲账协同",
+                            "status": "active",
+                        },
+                    }
+                ],
+            }
+        }
+    )
+
+    assert reply == "未重复执行：项目#8「财务冲账协同」已存在，状态 active。"
+
+
+def test_adaptive_rag_intent_rule_detects_project_queries():
+    service = AdaptiveRAGService(db=None, tenant_id=1)
+
+    result = service._classify_intent_rule("现在有哪些项目")
+    assert result["intent"] == "question"
+    assert result["confidence"] >= 0.86
+
+    result2 = service._classify_intent_rule("一个 AI 客服工具，这个项目有吗")
+    assert result2["intent"] == "question"
+    assert result2["confidence"] >= 0.86
+
+
+@pytest.mark.asyncio
+async def test_deterministic_project_lookup_answers_from_db(monkeypatch):
+    service = FeishuSignalPipelineService(db=None, tenant_id=1)
+
+    async def _fake_write_processing_log(**kwargs):
+        return None
+
+    monkeypatch.setattr(service, "_write_processing_log", _fake_write_processing_log)
+    result = await service._maybe_answer_project_lookup_question(
+        message="现在有哪些项目",
+        feishu_message_id="probe-lookup-1",
+        process_started_at=0.0,
+        projects=[
+            {"id": 8, "title": "财务冲账协同", "name": "财务冲账协同", "status": "active"},
+            {"id": 5, "title": "食谱需求", "name": "食谱需求", "status": "active"},
+        ],
+    )
+
+    assert result is not None
+    assert result["status"] == "answered"
+    assert result["classification"] == "question"
+    assert result["created_project_id"] is None
+    assert "#8 财务冲账协同(active)" in result["direct_reply_text"]
diff --git a/tests/test_llm_fail_fast.py b/tests/test_llm_fail_fast.py
new file mode 100644
index 00000000..88d5ab2d
--- /dev/null
+++ b/tests/test_llm_fail_fast.py
@@ -0,0 +1,101 @@
+import httpx
+
+from backend.app.services.llm import LLMService
+
+
+def _build_http_status_error(status_code: int) -> httpx.HTTPStatusError:
+    request = httpx.Request("POST", "https://example.com/v1/chat/completions")
+    response = httpx.Response(status_code, request=request)
+    return httpx.HTTPStatusError(
+        message=f"status={status_code}",
+        request=request,
+        response=response,
+    )
+
+
+def test_fail_fast_for_auth_status_errors():
+    assert LLMService._is_fail_fast_exception(_build_http_status_error(401))
+    assert LLMService._is_fail_fast_exception(_build_http_status_error(403))
+    assert not LLMService._is_fail_fast_exception(_build_http_status_error(500))
+
+
+def test_fail_fast_for_empty_response_markers():
+    assert LLMService._is_fail_fast_exception(ValueError("primary_engine_empty_response"))
+    assert LLMService._is_fail_fast_exception(ValueError("LLM 返回空内容"))
+    assert LLMService._is_fail_fast_exception(ValueError("provider_circuit_open:openai:gpt-4o"))
+
+
+def test_provider_circuit_breaker_opens_after_threshold():
+    service = LLMService()
+    circuit_key = service._provider_circuit_key(
+        provider="openai",
+        base_url="https://example.com/v1",
+        model="gpt-4o-mini",
+    )
+    LLMService._provider_circuit_state.pop(circuit_key, None)
+
+    service._record_provider_failure(circuit_key, ValueError("LLM 返回空内容"))
+    assert not service._is_provider_circuit_open(circuit_key)
+
+    service._record_provider_failure(circuit_key, ValueError("LLM 返回空内容"))
+    assert service._is_provider_circuit_open(circuit_key)
+
+    service._record_provider_success(circuit_key)
+    assert not service._is_provider_circuit_open(circuit_key)
+
+
+def test_extract_structured_text_from_reasoning_returns_json_candidate():
+    service = LLMService()
+
+    text = service._extract_structured_text_from_reasoning(
+        message={"reasoning_content": "先分析，再输出结果：{\"ok\": true, \"items\": [1, 2]}"},
+        choice={},
+        response={},
+    )
+
+    assert text == '{"ok": true, "items": [1, 2]}'
+
+
+def test_extract_structured_text_from_reasoning_ignores_non_json_reasoning():
+    service = LLMService()
+
+    text = service._extract_structured_text_from_reasoning(
+        message={"reasoning_content": "先分析需求，再决定是否返回。"},
+        choice={},
+        response={},
+    )
+
+    assert text == ""
+
+
+def test_prefer_fast_local_fallback_for_signal_source_functions():
+    assert LLMService._prefer_fast_local_fallback("signal_source_onboarding_plan")
+    assert LLMService._prefer_fast_local_fallback("openclaw_command_dispatch")
+    assert LLMService._prefer_fast_local_fallback("signal_source_ai_worker_execution")
+    assert not LLMService._prefer_fast_local_fallback("adaptive_rag_answer_normal")
+
+
+def test_request_timeout_for_fast_local_fallback_functions(monkeypatch):
+    monkeypatch.delenv("LLM_FAST_LOCAL_FALLBACK_TIMEOUT", raising=False)
+    service = LLMService()
+    service.request_timeout = 90.0
+
+    assert service._request_timeout_for_function("signal_source_onboarding_plan") == 8.0
+    assert service._request_timeout_for_function("openclaw_command_dispatch") == 8.0
+    assert service._request_timeout_for_function("adaptive_rag_answer_normal") == 90.0
+
+
+def test_request_timeout_for_fast_local_fallback_respects_runtime_cap(monkeypatch):
+    monkeypatch.setenv("LLM_FAST_LOCAL_FALLBACK_TIMEOUT", "20")
+    service = LLMService()
+    service.request_timeout = 8.0
+
+    assert service._request_timeout_for_function("signal_source_ai_worker_execution") == 8.0
+
+
+def test_fast_local_fallback_functions_skip_provider_circuit_breaker():
+    service = LLMService()
+
+    assert not service._use_provider_circuit_breaker("signal_source_onboarding_plan")
+    assert not service._use_provider_circuit_breaker("openclaw_command_dispatch")
+    assert service._use_provider_circuit_breaker("adaptive_rag_answer_normal")
diff --git a/tests/test_model_capability_context_validation.py b/tests/test_model_capability_context_validation.py
new file mode 100644
index 00000000..7895359b
--- /dev/null
+++ b/tests/test_model_capability_context_validation.py
@@ -0,0 +1,56 @@
+import pytest
+
+from backend.app.services.cost_optimization_service import CostOptimizationService
+from backend.app.services.model_capability_service import ModelCapabilityService
+
+
+class _FakeTruthNode:
+    def __init__(self, node_id: int, tenant_id: int, project_id: int):
+        self.id = node_id
+        self.tenant_id = tenant_id
+        self.project_id = project_id
+        self.title = "t"
+        self.content = {}
+
+
+class _FakeResult:
+    def __init__(self, value):
+        self._value = value
+
+    def scalar_one_or_none(self):
+        return self._value
+
+
+class _FakeDB:
+    def __init__(self, truth_node):
+        self.truth_node = truth_node
+
+    async def execute(self, query):
+        _ = query
+        return _FakeResult(self.truth_node)
+
+
+@pytest.mark.asyncio
+async def test_validate_truth_node_in_project_rejects_cross_project():
+    truth_node = _FakeTruthNode(node_id=1, tenant_id=1, project_id=2)
+    service = ModelCapabilityService(db=_FakeDB(truth_node), ai_service=None)  # type: ignore[arg-type]
+
+    with pytest.raises(ValueError, match="不属于项目"):
+        await service.validate_truth_node_in_project(
+            truth_node_id=1,
+            tenant_id=1,
+            project_id=1,
+        )
+
+
+@pytest.mark.asyncio
+async def test_cost_opt_truth_node_validation_rejects_cross_project():
+    truth_node = _FakeTruthNode(node_id=1, tenant_id=1, project_id=3)
+    service = CostOptimizationService(db=_FakeDB(truth_node))  # type: ignore[arg-type]
+
+    with pytest.raises(ValueError, match="不属于项目"):
+        await service.validate_truth_node_in_project(
+            truth_node_id=1,
+            tenant_id=1,
+            project_id=1,
+        )
diff --git a/tests/test_signal_capture_recent_cache.py b/tests/test_signal_capture_recent_cache.py
new file mode 100644
index 00000000..4546aa3f
--- /dev/null
+++ b/tests/test_signal_capture_recent_cache.py
@@ -0,0 +1,68 @@
+import pytest
+
+from backend.app.services.adaptive_rag_service import AdaptiveRAGService
+from backend.app.services.feishu_signal_pipeline import FeishuSignalPipelineService
+from backend.app.services.signal_capture import ProactiveSignalCaptureService
+from backend.app.services.context_cache import context_cache_service
+
+
+@pytest.mark.asyncio
+async def test_sync_feishu_recent_cache_skip_when_rag_result_present(monkeypatch):
+    service = ProactiveSignalCaptureService(db=None)
+    calls: list[dict] = []
+
+    async def _fake_append_session_messages(**kwargs):
+        calls.append(kwargs)
+
+    monkeypatch.setattr(context_cache_service, "append_session_messages", _fake_append_session_messages)
+
+    await service._sync_feishu_recent_cache(
+        tenant_id=1,
+        group_chat_id="oc-chat-1",
+        sender="测试用户",
+        sender_open_id=None,
+        message="这个需求要不要支持飞书同步？",
+        pipeline_result={"status": "captured", "rag_result": {"answer": "已回答"}},
+    )
+
+    assert calls == []
+
+
+@pytest.mark.asyncio
+async def test_sync_feishu_recent_cache_write_external_session(monkeypatch):
+    service = ProactiveSignalCaptureService(db=None)
+    calls: list[dict] = []
+
+    async def _fake_append_session_messages(**kwargs):
+        calls.append(kwargs)
+
+    monkeypatch.setattr(context_cache_service, "append_session_messages", _fake_append_session_messages)
+
+    await service._sync_feishu_recent_cache(
+        tenant_id=1,
+        group_chat_id="oc-chat-2",
+        sender="测试用户",
+        sender_open_id=None,
+        message="需要支持飞书同步",
+        pipeline_result={"status": "captured"},
+    )
+
+    assert len(calls) == 1
+    call = calls[0]
+    assert call["tenant_id"] == 1
+    assert call["question"] == "需要支持飞书同步"
+    assert "已收到并记录该需求" in call["answer"]
+
+    key = FeishuSignalPipelineService._build_rag_conversation_key(
+        source_chat_id="oc-chat-2",
+        source_sender_open_id=None,
+        sender="测试用户",
+    )
+    expected_sid = AdaptiveRAGService._external_session_id_from_key(key)
+    assert call["session_id"] == expected_sid
+
+
+def test_build_recent_cache_ack_text_variants():
+    assert ProactiveSignalCaptureService._build_recent_cache_ack_text({"status": "captured"}) == "已收到并记录该需求，已进入信号处理队列。"
+    assert ProactiveSignalCaptureService._build_recent_cache_ack_text({"status": "ignored"}) == "已收到该消息，本次未进入需求处理队列。"
+    assert "系统处理中出现异常" in ProactiveSignalCaptureService._build_recent_cache_ack_text({"status": "error"})