Refine AI PMO grounding runtime and consultant flow

CLAUDE.md

@@ -104,6 +104,7 @@ backend/
 - 新增：`backend/alembic/versions/20260311_shift_all_timestamps_to_beijing.py`（全库历史时间 +8h，数据库默认时区切到 Asia/Shanghai）
 - 新增：`backend/alembic/versions/20260327_add_external_qa_followup_state.py`（`qa_sessions` 增加 `external_conversation_key`，`qa_session_states` 增加外部追问待澄清状态字段）
 - 新增：`backend/alembic/versions/20260327_add_qa_record_followup_anchor_fields.py`（`qa_records` 增加外部会话 scope / reply anchor / 路由快照字段，支撑跨 session 追问锚定与元追问追溯）
+- 新增：`backend/alembic/versions/20260331_add_qa_record_runtime_fields.py`（`qa_records` 增加 `answer_mode/grounding_status/policy_version/citations/sources_used/evidence_pack/tool_trace/retrieval_metrics`，用于 Grounding Runtime 持久化与回放）
 - 新增：`backend/alembic/versions/20260312_add_signal_planner_fields.py`（signals 增加 planner 状态字段与索引）
 - 新增：`backend/app/services/signal/signal_planner/signal_strategic_analysis_service.py`（Signal Planner 写回收口）
 - 新增：`backend/app/services/integrations/feishu_outbox/service.py`（复用 reply_hash 的 outbox 入队收口）
@@ -127,6 +128,7 @@ backend/
 - 新增：`backend/tests/test_ai_pmo_task_intent_contract.py`（AI PMO `task.create` 一等 typed intent 契约回归：意图识别、路由归一与执行器分派）
 - 新增：`backend/tests/test_ai_pmo_ingress_mainline.py`（AI PMO 主链入口回归：`/ai-pmo/process-text` 手工问答入口、mixed sidecar 透传、`draft_pending_confirm` 优先分支）
 - 新增：`backend/tests/test_ai_pmo_qa_http_mainline_e2e.py`（AI PMO/QA HTTP 主链集成回归：`/rag/ask` 只读问答、mixed 两段输出、`/ai-pmo/process-text` 草稿待确认）
+- 新增：`backend/tests/test_qa_record_runtime_fields_migration.py`（QA Runtime migration 回归：`qa_records` runtime/evidence 字段新增与回滚）
 - 新增：`backend/tests/test_context_cache_profile_catalog.py`（Context Cache 功能目录契约回归：AI Steward 统一指向 `/rag/ask`）
 - 新增：`backend/tests/test_ai_pmo_tactical_brain_contract.py`（TacticalBrain 执行动作契约回归：`query.answer`/`work_item.resume`/`project_signal.*` 映射）
 - 新增：`backend/app/services/adaptive_rag_service/mixin_answer_mainline.py`（AdaptiveRAG 基础问答主链编排：运行时准备 + 只读主路径执行）
@@ -363,6 +365,11 @@ frontend/
 - 新增：`frontend/src/api/modelRouting.ts`（模型路由 API 封装）
 - 新增：`frontend/src/api/knowledgeGraph.ts`（KG 能力汇总与 low/high 消费验证 API）
 - 更新：`frontend/src/api/aiSteward.ts`（移除 OpenClaw dispatch 封装，新增 LLM 引擎 CRUD / settings 封装）
+- 更新：`frontend/src/api/aiSteward.ts`（新增 AI 管家 QA 封装：`/rag/ask`、QA sessions、feedback、会话改名；浮窗不再直接 `fetch`）
+- 更新：`frontend/src/api/aiPmo.ts`（新增 `processText` 封装，供 AI 管家浮窗把“允许执行”显式交接到 AI PMO Kernel）
+- 新增：`frontend/src/components/ai-steward/QAMarkdownContent.tsx`（AI 管家问答 markdown 渲染组件：统一列表/引用/代码块/长路径换行）
+- 更新：`frontend/src/components/ai-steward/FloatingQAChat.tsx`（回答区改为 markdown 渲染；浮窗顾问动作升级为整组可交互 chips：`深入思考`、`给建议`、`先澄清`、`一起想`、`先劝退`、风险解释、`允许执行` 确认交接、`先别执行` 回填 `next_action`）
+- 更新：`frontend/package.json`（新增 `react-markdown`、`remark-gfm`、`remark-breaks` 依赖，供 AI 管家问答渲染）
 - 更新：`frontend/src/pages/CharterDetailPage.tsx`、`ContractSyncBoardPage.tsx`、`DashboardPage.tsx`、`KnowledgeGraphContextBoardPage.tsx`、`LoginPage.tsx`、`ModelRoutingConfigPage.tsx`、`SchedulerCenterPage.tsx`、`SkillTasksPage.tsx`、`SkillsPage.tsx`、`TaskBriefCenterPage.tsx`、`TaskBriefRunsPage.tsx`、`TaskBriefTemplatePage.tsx`、`UserProfilePage.tsx`、`UserSkillsPage.tsx`、`WorkloadDetailPage.tsx`、`WorkloadsPage.tsx`（统一降级为轻量入口，转发到 `pages/*/` 子目录）
 - 新增：`frontend/src/pages/charter-detail/CharterDetailPage.tsx`、`contract-sync-board/ContractSyncBoardPage.tsx`、`dashboard/DashboardPage.tsx`、`knowledge-graph-context-board/KnowledgeGraphContextBoardPage.tsx`、`login/LoginPage.tsx`、`model-routing-config/ModelRoutingConfigPage.tsx`（剩余核心页面目录化收敛）
 - 新增：`frontend/src/pages/scheduler-center/SchedulerCenterPage.tsx`、`skill-tasks/SkillTasksPage.tsx`、`skills/SkillsPage.tsx`、`task-brief-center/TaskBriefCenterPage.tsx`、`task-brief-runs/TaskBriefRunsPage.tsx`、`task-brief-template/TaskBriefTemplatePage.tsx`（任务/调度/技能页面目录化收敛）
@@ -570,8 +577,11 @@ uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
 - AI PMO 恢复接口（`POST /ai-pmo/work-items/{work_item_id}/resume`）：支持 `action=continue|commit|refresh|cancel`，复用同一闭环恢复入口
 - AI PMO 草稿决策接口（`POST /ai-pmo/work-items/{work_item_id}/draft-decision`）：支持 `action=confirm|revise|cancel`，与飞书“确认/修改/取消”行为一致
 - AI PMO 路由包（`backend/app/api/ai_pmo_api/__init__.py`）只负责聚合与 tags，`/ai-pmo` 前缀只能由子模块共享 router 挂一次，禁止包级与模块级重复叠加前缀导致 `/ai-pmo/ai-pmo/...` 路径
-- RAG 问答主入口契约（`POST /rag/ask`）：HTTP 入口已收敛到 `AIPMOKernelService.process_ingress` 主链；请求侧统一透传 `asked_via/action_mode/message_intent/mixed_signal_part`，由 `ingress_options.rag_request_context` 下传到 query executor 并映射到 `AdaptiveRAGService.answer_question`，返回侧从 ingress `rag_result`/`direct_reply_text` 组装兼容响应
+- RAG 问答主入口契约（`POST /rag/ask`）：HTTP 入口已收敛到 `AIPMOKernelService.process_ingress` 主链；请求侧统一透传 `asked_via/action_mode/message_intent/mixed_signal_part/consultant_mode_preference`，由 `ingress_options.rag_request_context` 下传到 query executor 并映射到 `AdaptiveRAGService.answer_question`，返回侧从 ingress `rag_result`/`direct_reply_text` 组装兼容响应
+- RAG 空回答保护：`POST /rag/ask` 若 `answer` 与 `reply_text` 同时为空，必须返回失败包络，禁止再用“当前暂无可用回答。”之类兜底文案冒充成功结果
+- RAG Grounding Runtime 合同补充：`POST /rag/ask` 的成功返回统一透传 `answer_mode/grounding_status/policy_version/citations/sources_used/evidence_pack/tool_trace/retrieval_metrics`；`AdaptiveRAG` 内部允许通过兼容层把异构 `sources` 映射成稳定 citation 结构，但业务语义判断仍只允许放在 `workflow_semantic_planner`
 - RAG / AI 管家顾问合同：`POST /rag/ask` 返回体会透传 `consultant_dispatch={domain,mode,risk_level,missing_fields,allow_execute,next_action}`；该合同由 `backend/app/services/ai_pmo/consultant_dispatch.py` 统一生成，飞书 mention 回复与 AI 管家问答共享同一份模式判断
+- RAG 顾问模式偏好约束：`consultant_mode_preference` 只允许取 `clarify/advise/challenge/brainstorm/execute`，且仅影响回答风格与 prompt wrapper；它不能替代 `workflow_semantic_planner` 去决定 `query vs execute`
 - RAG / AI Steward 路由包（`backend/app/api/rag_api/__init__.py`、`backend/app/api/ai_steward_api/__init__.py`）不再重复挂包级 prefix；路由 canonical 由 `shared.py` 内的共享 router 统一承担，避免运行中出现 `/rag/rag/*`、`/ai-steward/ai-steward/*`
 - AI PMO 消息台列表接口（`GET /ai-pmo/work-items`）：返回 `latest_policy_gate_code` 与规范化 `approval_requirement{required,reason,message}`（来源于最新 `decision_records` 投影）
 - AI PMO 消息台列表接口补充：返回草稿字段 `draft_content/draft_expires_at/draft_version/confirm_message_id`，用于草稿确认卡片渲染
@@ -650,6 +660,7 @@ uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
 - 外部 QA 若存在显式 `reply_to_message_id`，必须优先按消息级锚定处理：先反查 `feishu_reply_outbox.sent_message_id` 还原被引用的 PMFlow 回复；若消息 ID 无法映射，也要回退到当前入站消息携带的 `closure_turns/closure_dialogue` 恢复上一轮机器人回复，再回退到 session/recent candidates；禁止把“直接引用机器人回复”的消息仍判成无锚点
 - 外部 QA 的“追问歧义”与“模型/上游失败”必须分开：前者返回 `clarification_code`（如 `ambiguous_recent_reply` / `no_anchor_found`）及候选列表，后者继续保留真实 provider/空响应错误，禁止共用模糊兜底文案
 - 外部 QA 的 `qa_records` 必须持久化 `conversation_scope_key/source_message_id/source_chat_id/source_sender_open_id/reply_to_message_id/root_message_id/anchor_kind/anchor_ref/routing_snapshot`，用于回答“为什么没接上上一条”“为什么走了候选澄清”“上一条回复走的是 low 还是 high”
+- `qa_records` 现阶段还必须持久化 Grounding Runtime 字段：`answer_mode/grounding_status/policy_version/citations/sources_used/evidence_pack/tool_trace/retrieval_metrics`；旧 `sources/domains_used/routing_snapshot` 继续保留兼容，禁止只改接口不落库
 - 元追问（如“哪个模型回复的”“为什么那条报错”“你刚才那句什么意思”）必须优先锚定最近一条 PMFlow 回复；对外只允许返回 `planned_tier/executed_tier/route_source/route_rule_name/routing_decision` 这类 low/high 路由语义，禁止暴露真实 provider/model 名称
 - Prompt release 目前只接入只读回复链路（AdaptiveRAG / query.read_only）；写操作 workflow 仍使用代码 prompt 与显式人工 gate，避免训练试验误触执行副作用
 - 训练自动化/影子回放当前只允许覆盖只读问答链路；`project.create/signal.create/signal_source.onboard/task.create` 等写操作 workflow 继续排除在自动训练与自动试验之外
@@ -664,8 +675,10 @@ uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
 - 飞书入口已按职责拆分为“预处理/捕获执行/结果回执”三层：`mixin_receive.py` 仅保留编排，`mixin_receive_prepare.py` 负责传输解析与上下文抽取，`mixin_receive_capture_flow.py` 负责 capture 持久化与回执收尾
 - `ai_orchestration_runs.project_id` 允许为空：用于承载尚未完成项目归属判定的入站异步任务；既有 `signal_source_onboarding` 运行仍按显式 `project_id` 写入
 - `/ai-pmo/process-text` 问答链路在 AdaptiveRAG 内通过 `asked_via=pmflow_manual` 识别为 AI PMO 入口；mixed 场景通过 `message_intent=mixed` 透传触发专用 Prompt 与两段输出格式化
-- `/rag/ask` 已改为经 AI PMO kernel ingress 主链执行（与飞书长连接、`/ai-pmo/process-text` 同链路）；query 执行阶段通过 `rag_request_context` 保留 `session_id/new_session/thinking_mode/action_mode/sync_requirement/message_intent/mixed_signal_part`
+- `/rag/ask` 已改为经 AI PMO kernel ingress 主链执行（与飞书长连接、`/ai-pmo/process-text` 同链路）；query 执行阶段通过 `rag_request_context` 保留 `session_id/new_session/thinking_mode/action_mode/sync_requirement/message_intent/mixed_signal_part/consultant_mode_preference`
 - AdaptiveRAG `answer_question` 入口已拆为“双阶段”：`_prepare_answer_runtime`（会话/上下文/动作模式解析）+ `_execute_base_answer_mainline`（基础问答主路径）；写确认 `ack` gate 仅在 `sync_requirement=true && action_mode in write/read_write && 非 feishu` 时触发
+- 高准确通道不再按 `asked_via` 一刀切强制 deep：`count/list/status/existence` 由 `AdaptiveRAG` retrieval planner 优先回落到 `fast_path`，`snapshot-only + weak evidence` 则要求进入 `clarify` 或 `blocked`；该 planner 只负责 retrieval/runtime 决策，禁止接管 `query vs execute / project.create / signal_source.onboard / followup.*`
+- AI 管家浮窗顾问按钮边界：`深入思考` 与整组 `给建议/先澄清/一起想/先劝退` chips 都属于 `/rag/ask` 内部重答；`允许执行` 必须先确认，再显式交接到 `/ai-pmo/process-text`，禁止把执行动作继续伪装成 RAG 重答；`先别执行` 只允许回填 `next_action` 到输入框，不得偷偷触发写操作
 - AI PMO ingress 执行路径（`waiting_manual/draft_pending_confirm/auto execute`）已统一写入 `operation_execution_ledger`，`operation_code` 与 `execution_result.action_contract` 一致；账本状态统一映射为 `applied/pending_manual/pending_confirmation/failed`
 - `PROMPT_DECOMPOSE_TASKS` 技能标签来源已改为运行时动态注入：优先读取 `skill_tags`（按 `sort_order/name`），为空时才回退默认技能集合
 
@@ -679,6 +692,13 @@ uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
 - [x] AdaptiveRAG 职责已完成首轮收敛：`answer_question` 入口已分层为 `prepare runtime` 与 `base mainline` 两段，写确认逻辑通过 `_should_trigger_write_ack_gate` 独立门控；基础问答主路径不再与写确认判断混写在同一函数体。分层契约由 `test_adaptive_rag_mainline_contract.py` 锁定（read 绕过 ack、write 命中 ack、feishu write 维持主链）
 - [x] 执行层已形成统一执行账本与参数来源约束：ingress 执行路径统一通过 `KernelIngressStorageMixin` 生成 `execution_result.action_contract`（固定 `param_source=decision.execution_plan.actions+workflow_route+execution_result`）并写入 `operation_execution_ledger`（`applied/pending_manual/pending_confirmation/failed`）；`mixin_auto.py`、`mixin_prepare_manual.py` 的 `operation_code` 日志来源已对齐 `action_contract`，回归由 `test_ai_pmo_execution_ledger_contract.py` 锁定
 - [x] QA/RAG 路由契约双轨已收敛：`/ai-steward/rag/ask` 兼容入口已移除，问答统一收敛到 `POST /rag/ask`；参数构造与 tenant 解析收口在 `backend/app/api/rag_ask_contract.py`，前端 `FloatingQAChat` 与 Context Cache API 目录均已对齐，`backend/tests/test_rag_route_contract.py` 与 `backend/tests/test_context_cache_profile_catalog.py` 已锁定回归
+- [x] AI 管家浮窗交互已与 Grounding Runtime 对齐：回答区已切到 markdown 渲染，顾问标签升级为显式交互按钮；`consultant_mode_preference` 只用于 `/rag/ask` 内部重答，`允许执行` 明确交接到 `/ai-pmo/process-text`，避免 query/execute 边界回退到旧的一锅煮
+- [ ] Retrieval source registry / metadata 统一仍在技术债：当前 `doc/url`、`feishu`、`pmflow`、`graph/db` 继续保留异构 metadata，`citations[]/sources_used[]` 先由 response-layer 兼容适配器统一；后续仍需补齐原生 `source_id/chunk_id/permission_scope/updated_at`、统一 source priority 配置并为 reranker 预留接口
+- [ ] 语义治理 baseline 尚未清零：`scripts/semantic_lint_baseline.txt` 仍豁免 `backend/app/services/integrations/feishu_signal_pipeline/shared_core.py`、`backend/app/services/signal/signal_capture_service/shared.py`、`backend/app/api/ai_pmo_api/shared.py` 中的 `*_HINTS/*_RE` 规则常量；`backend/tests/test_semantic_lint.py` 目前只阻止新增散落语义，尚未推动存量归零。后续需把这些兼容残留迁回 `workflow_semantic_planner` 或 prompt catalog，并逐项删除 baseline
+- [ ] 历史 Feishu 队列 / `feishu_items` 兼容链仍未退场：`FeishuSignalPipelineService` 已不再承担线上主入口，但离线回放、单测基座、`ai_logs/project_knowledge_graph` 历史过滤，以及 Alembic / 启动自检仍依赖 `feishu_items`。后续需先迁走历史读链路与回放基座，再为该链路补明确的 deprecated / 下线计划
+- [ ] AdaptiveRAG 向量检索仍保留 legacy fallback：`AdaptiveRAGRetrievalVectorMixin._retrieve_vector_context()` 在非 PostgreSQL 或 PG 查询失败时会回退 `_retrieve_vector_context_legacy()`，对 `knowledge_base_chunks` 做 Python 侧扫描与打分；这使 pgvector 主路径与 legacy 主路径长期双轨。后续需决定是正式声明“只支持 PostgreSQL/pgvector”，还是补齐统一 retrieval contract 与测试矩阵
+- [ ] 历史 Alembic 链幂等性缺陷仍未修复：仓库里已确认 `add_users_table_rbac.py` 在 PostgreSQL 可能重复创建 `userrole` ENUM；当前运行时仍通过直接 `uvicorn` 启动 + `AUTO_CREATE_TABLES=true` 规避全量 upgrade。后续需修补旧 migration，保证 PostgreSQL 上的 Alembic upgrade 可重复执行
+- [ ] 食谱图片上传仍是示例桩实现：`backend/app/services/recipe/service.py` 仍以模拟 URL 代替对象存储，`POST /recipes/upload-image` 也还停留在示例接口形态；若该模块继续保留，需要补齐真实对象存储接入、租户隔离与统一 `{code,message,data}` 响应包络，否则应明确标注为非核心示例模块
 - [x] LLM 运行时仍未完全收敛：`ai_service_core._invoke_with_retry` 与 `stream_chat` 已统一走 `llm_service`；旧 `_get_llm` 路径已显式禁用，并补回归测试锁定
 - [x] QA 主链自动化测试覆盖已补齐：`backend/tests/` 已覆盖问答/RAG 主路径与关键场景；新增 `test_ai_pmo_qa_http_mainline_e2e.py` 通过 HTTP 层锁定 `/rag/ask` 的 `qa_readonly/mixed` 契约以及 `/ai-pmo/process-text` 的 `draft_pending_confirm` 主链返回，配合 `test_ai_pmo_ingress_mainline.py`、`test_ai_pmo_task_intent_contract.py`、`test_ai_pmo_tactical_brain_contract.py` 形成稳定回归基线
 - [x] 飞书交互闭环已补齐：执行场景回执在 `draft_pending_confirm/pending_manual` 下优先生成按钮卡片并经 outbox 异步发送；`card.action.trigger` 已覆盖 `confirm/revise/cancel/continue` 回调分发，分别复用 `DraftConfirmHandler.handle_api_decision` 与 AI PMO `process_ingress` 主链继续执行；回调 HTTP 现在会同步返回 toast 反馈，确保点击按钮后至少有可见结果；`backend/tests/test_feishu_card_action_contract.py` 已补卡片构造与 `continue` 回归契约

backend/alembic/versions/20260331_add_qa_record_runtime_fields.py

@@ -0,0 +1,83 @@
+"""add qa record runtime grounding fields
+
+Revision ID: 20260331_add_qa_record_runtime_fields
+Revises: 20260327_add_reply_iteration_training_tables
+Create Date: 2026-03-31
+"""
+
+from __future__ import annotations
+
+from alembic import op
+import sqlalchemy as sa
+
+try:
+    from sqlalchemy.dialects.postgresql import JSONB
+except Exception:  # pragma: no cover
+    JSONB = sa.JSON
+
+
+revision = "20260331_add_qa_record_runtime_fields"
+down_revision = "20260327_add_reply_iteration_training_tables"
+branch_labels = None
+depends_on = None
+
+
+def _json_type(bind):
+    return JSONB() if bind.dialect.name == "postgresql" else sa.JSON()
+
+
+def _table_exists(bind, table_name: str) -> bool:
+    return sa.inspect(bind).has_table(table_name)
+
+
+def _column_exists(bind, table_name: str, column_name: str) -> bool:
+    inspector = sa.inspect(bind)
+    try:
+        columns = inspector.get_columns(table_name)
+    except Exception:
+        return False
+    return any(str(column.get("name")) == column_name for column in columns)
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+    if not _table_exists(bind, "qa_records"):
+        return
+
+    with op.batch_alter_table("qa_records") as batch_op:
+        if not _column_exists(bind, "qa_records", "answer_mode"):
+            batch_op.add_column(sa.Column("answer_mode", sa.String(length=32), nullable=True))
+        if not _column_exists(bind, "qa_records", "grounding_status"):
+            batch_op.add_column(sa.Column("grounding_status", sa.String(length=32), nullable=True))
+        if not _column_exists(bind, "qa_records", "policy_version"):
+            batch_op.add_column(sa.Column("policy_version", sa.String(length=160), nullable=True))
+        if not _column_exists(bind, "qa_records", "citations"):
+            batch_op.add_column(sa.Column("citations", _json_type(bind), nullable=True))
+        if not _column_exists(bind, "qa_records", "sources_used"):
+            batch_op.add_column(sa.Column("sources_used", _json_type(bind), nullable=True))
+        if not _column_exists(bind, "qa_records", "evidence_pack"):
+            batch_op.add_column(sa.Column("evidence_pack", _json_type(bind), nullable=True))
+        if not _column_exists(bind, "qa_records", "tool_trace"):
+            batch_op.add_column(sa.Column("tool_trace", _json_type(bind), nullable=True))
+        if not _column_exists(bind, "qa_records", "retrieval_metrics"):
+            batch_op.add_column(sa.Column("retrieval_metrics", _json_type(bind), nullable=True))
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+    if not _table_exists(bind, "qa_records"):
+        return
+
+    with op.batch_alter_table("qa_records") as batch_op:
+        for column_name in (
+            "retrieval_metrics",
+            "tool_trace",
+            "evidence_pack",
+            "sources_used",
+            "citations",
+            "policy_version",
+            "grounding_status",
+            "answer_mode",
+        ):
+            if _column_exists(bind, "qa_records", column_name):
+                batch_op.drop_column(column_name)

backend/app/api/rag_api/routes_query.py

@@ -1,8 +1,11 @@
 from __future__ import annotations
 
+from fastapi.responses import JSONResponse
+
 from .shared import *
 from .helpers import _resolve_tenant_id
 from .. import rag_ask_contract
+from ...core.response import ErrorCode, error_response
 from ...services.ai_pmo.ai_pmo_kernel import AIPMOKernelService
 
 @router.post("/ask")
@@ -27,9 +30,13 @@ async def rag_ask(
             status_code=400,
             detail=str(ingress_result.get("reason") or "ai_pmo_kernel_not_handled"),
         )
-    return success_response(
-        rag_ask_contract.build_rag_response_payload(ingress_result=ingress_result),
+    payload = rag_ask_contract.build_rag_response_payload(ingress_result=ingress_result)
+    if not str(payload.get("answer") or "").strip() and not str(payload.get("reply_text") or "").strip():
+        return JSONResponse(
+            status_code=500,
+            content=error_response(ErrorCode.INTERNAL_ERROR, "RAG 返回空回答，未生成有效响应。"),
         )
+    return success_response(payload)
 
 @router.get("/manifest")
 async def rag_manifest(

backend/app/api/rag_ask_contract.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from typing import Any
+from uuid import uuid4
 
 from ..models import User
 from ..schemas.rag import RAGAskRequest
@@ -28,6 +29,7 @@ def build_answer_question_kwargs(request: RAGAskRequest, current_user: User) ->
         "sync_requirement": bool(request.sync_requirement),
         "message_intent": request.message_intent,
         "mixed_signal_part": request.mixed_signal_part,
+        "consultant_mode_preference": request.consultant_mode_preference,
     }
 
 
@@ -49,11 +51,28 @@ def _resolve_sender_name(current_user: User) -> str:
     return sender or "PMFlow用户"
 
 
+def _normalize_dict_list(value: Any) -> list[dict[str, Any]]:
+    if not isinstance(value, list):
+        return []
+    return [item for item in value if isinstance(item, dict)]
+
+
+def _default_evidence_pack(*, answer: str, tool_trace: list[dict[str, Any]]) -> dict[str, Any]:
+    return {
+        "summary": {
+            "conclusion": str(answer or "").strip(),
+            "evidence": [],
+            "uncertainties": [],
+        },
+        "tool_trace": tool_trace,
+    }
+
+
 def build_kernel_ingress_kwargs(request: RAGAskRequest, current_user: User) -> dict[str, Any]:
     rag_request_context = build_answer_question_kwargs(request=request, current_user=current_user)
     return {
         "source_system": str(request.asked_via or "pmflow").strip() or "pmflow",
-        "source_message_id": None,
+        "source_message_id": f"rag-ask-{uuid4().hex}",
         "message_text": request.question,
         "sender": _resolve_sender_name(current_user),
         "channel_ref": request.source_chat_id,
@@ -80,7 +99,7 @@ def build_rag_response_payload(*, ingress_result: dict[str, Any]) -> dict[str, A
         payload["consultant_dispatch"] = consultant_dispatch
 
     answer = str(payload.get("answer") or result.get("direct_reply_text") or "").strip()
-    payload["answer"] = answer or "当前暂无可用回答。"
+    payload["answer"] = answer
 
     if "reply_required" not in payload:
         payload["reply_required"] = bool(result.get("direct_reply_required"))
@@ -90,6 +109,22 @@ def build_rag_response_payload(*, ingress_result: dict[str, Any]) -> dict[str, A
         payload["trace_id"] = str(result.get("ai_trace_id") or "").strip() or None
     if "processing_stage" not in payload:
         payload["processing_stage"] = "qa"
+    if "tool_trace" not in payload:
+        payload["tool_trace"] = []
+    payload["tool_trace"] = _normalize_dict_list(payload.get("tool_trace"))
+    payload["citations"] = _normalize_dict_list(payload.get("citations"))
+    payload["sources_used"] = _normalize_dict_list(payload.get("sources_used"))
+    payload["answer_mode"] = str(payload.get("answer_mode") or "retrieval_normal").strip() or "retrieval_normal"
+    payload["grounding_status"] = str(payload.get("grounding_status") or "partially_grounded").strip() or "partially_grounded"
+    payload["policy_version"] = (
+        str(payload.get("policy_version") or "retrieval-v2.1|routing:unknown|prompt:unknown").strip()
+        or "retrieval-v2.1|routing:unknown|prompt:unknown"
+    )
+    evidence_pack = payload.get("evidence_pack") if isinstance(payload.get("evidence_pack"), dict) else None
+    payload["evidence_pack"] = evidence_pack or _default_evidence_pack(
+        answer=answer,
+        tool_trace=payload["tool_trace"],
+    )
 
     payload["work_item_id"] = _safe_int(result.get("work_item_id"))
     payload["ingress_status"] = str(result.get("status") or "").strip() or None

backend/app/models/qa_record.py

@@ -40,3 +40,11 @@ class QARecord(Base):
     anchor_kind: Mapped[Optional[str]] = mapped_column(VARCHAR(32), nullable=True)
     anchor_ref: Mapped[Optional[str]] = mapped_column(VARCHAR(128), nullable=True)
     routing_snapshot: Mapped[Optional[Any]] = mapped_column(JSONB, nullable=True)
+    answer_mode: Mapped[Optional[str]] = mapped_column(VARCHAR(32), nullable=True)
+    grounding_status: Mapped[Optional[str]] = mapped_column(VARCHAR(32), nullable=True)
+    policy_version: Mapped[Optional[str]] = mapped_column(VARCHAR(160), nullable=True)
+    citations: Mapped[Optional[Any]] = mapped_column(JSONB, nullable=True)
+    sources_used: Mapped[Optional[Any]] = mapped_column(JSONB, nullable=True)
+    evidence_pack: Mapped[Optional[Any]] = mapped_column(JSONB, nullable=True)
+    tool_trace: Mapped[Optional[Any]] = mapped_column(JSONB, nullable=True)
+    retrieval_metrics: Mapped[Optional[Any]] = mapped_column(JSONB, nullable=True)

backend/app/schemas/rag.py

@@ -24,6 +24,13 @@ class RAGAskRequest(BaseModel):
     _ASKED_VIA_ALIAS: ClassVar[dict[str, str]] = {
         "ai-pmo-process-text": "/ai-pmo/process-text",
     }
+    _ALLOWED_CONSULTANT_MODES: ClassVar[set[str]] = {
+        "clarify",
+        "advise",
+        "challenge",
+        "brainstorm",
+        "execute",
+    }
     question: str = Field(..., min_length=1, max_length=4000)
     project_id: Optional[int] = Field(default=None)
     asked_via: str = Field(default="pmflow", min_length=1, max_length=32)
@@ -35,6 +42,7 @@ class RAGAskRequest(BaseModel):
     sync_requirement: bool = Field(default=True)
     message_intent: Optional[str] = Field(default=None, max_length=32)
     mixed_signal_part: Optional[str] = Field(default=None, max_length=2000)
+    consultant_mode_preference: Optional[str] = Field(default=None, max_length=16)
 
     @field_validator("asked_via", mode="before")
     @classmethod
@@ -64,6 +72,19 @@ class RAGAskRequest(BaseModel):
         text = str(value).strip()
         return text or None
 
+    @field_validator("consultant_mode_preference", mode="before")
+    @classmethod
+    def _normalize_consultant_mode_preference(cls, value: object) -> Optional[str]:
+        if value is None:
+            return None
+        text = str(value).strip().lower()
+        if not text:
+            return None
+        if text not in cls._ALLOWED_CONSULTANT_MODES:
+            allowed = ", ".join(sorted(cls._ALLOWED_CONSULTANT_MODES))
+            raise ValueError(f"consultant_mode_preference must be one of: {allowed}")
+        return text
+
 
 class KnowledgeDocCreateRequest(BaseModel):
     title: str = Field(..., min_length=1, max_length=255)

backend/app/services/adaptive_rag_service/mixin_answer_entry.py

@@ -29,20 +29,7 @@ class AdaptiveRAGAnswerEntryMixin:
         message_intent: Optional[str],
     ) -> bool:
         intent = str(message_intent or "").strip().lower()
-        if intent == "mixed":
-            return True
-        channel = str(asked_via or "").strip().lower()
-        return channel in {
-            "feishu",
-            "lark",
-            "feishu_group",
-            "feishu_dm",
-            "pmflow_manual",
-            "/ai-pmo/process-text",
-            "ai-pmo-process-text",
-            "ai_pmo",
-            "ai_pmo_ui",
-        }
+        return intent == "mixed"
 
     @staticmethod
     def _should_trigger_write_ack_gate(
@@ -80,12 +67,11 @@ class AdaptiveRAGAnswerEntryMixin:
         sync_requirement: bool = True,
         message_intent: Optional[str] = None,
         mixed_signal_part: Optional[str] = None,
+        consultant_mode_preference: Optional[str] = None,
     ) -> dict[str, Any]:
         started = time.perf_counter()
         clean_q = str(question or "").strip()
         active_thinking_mode = self._normalize_thinking_mode(thinking_mode)
-        if self._is_high_accuracy_channel(asked_via):
-            active_thinking_mode = THINKING_MODE_DEEP
         deep_thinking = active_thinking_mode == THINKING_MODE_DEEP
         enforce_scene_prompt_only = self._should_enforce_scene_prompt_only(
             asked_via=asked_via,
@@ -186,4 +172,5 @@ class AdaptiveRAGAnswerEntryMixin:
             message_intent=message_intent,
             mixed_signal_part=mixed_signal_part,
             anchor_resolution=anchor_resolution,
+            consultant_mode_preference=consultant_mode_preference,
         )

backend/app/services/adaptive_rag_service/mixin_answer_mainline.py

@@ -5,6 +5,80 @@ from .shared import *
 
 
 class AdaptiveRAGAnswerMainlineMixin:
+    def _plan_query_runtime(
+        self,
+        *,
+        question: str,
+        effective_question: str,
+        query_type: str,
+        requested_thinking_mode: str,
+        enforce_scene_prompt_only: bool,
+        consultant_mode_preference: Optional[str] = None,
+    ) -> QueryPlannerDecision:
+        normalize_thinking_mode = getattr(
+            self,
+            "_normalize_thinking_mode",
+            lambda value: THINKING_MODE_DEEP if str(value or "").strip().lower() == THINKING_MODE_DEEP else THINKING_MODE_NORMAL,
+        )
+        normalized_thinking_mode = normalize_thinking_mode(requested_thinking_mode)
+        normalized_query_type = str(query_type or QUERY_TYPE_UNKNOWN).strip().lower()
+        normalized_consultant_mode = str(consultant_mode_preference or "").strip().lower()
+        fast_path_candidate = normalized_query_type in {QUERY_TYPE_COUNT, QUERY_TYPE_LIST, QUERY_TYPE_STATUS}
+        normalized_question = str(question or "").strip()
+        normalized_effective = str(effective_question or "").strip()
+        combined = f"{normalized_question} {normalized_effective}".strip()
+        is_existence_query = any(token in combined for token in ("有没有", "是否存在", "有无", "存在"))
+        if is_existence_query:
+            fast_path_candidate = True
+
+        is_engine_query = bool(getattr(self, "_is_engine_process_query", lambda _text: False)(combined))
+        is_explicit_deep = bool(getattr(self, "_is_explicit_deep_request", lambda _text: False)(combined))
+        if enforce_scene_prompt_only:
+            retrieval_mode = ANSWER_MODE_RETRIEVAL_DEEP if normalized_thinking_mode == THINKING_MODE_DEEP else ANSWER_MODE_RETRIEVAL_NORMAL
+            return QueryPlannerDecision(
+                answer_mode=retrieval_mode,
+                thinking_mode=THINKING_MODE_DEEP if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else THINKING_MODE_NORMAL,
+                allow_fast_path=False,
+                retrieval_answer_mode=retrieval_mode,
+                budget=RetrievalBudget(
+                    max_steps=DEEP_RETRIEVAL_STEPS if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else AGENT_MAX_RETRIEVAL_STEPS,
+                    vector_top_k=DEEP_VECTOR_TOP_K if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else DEFAULT_TOP_K,
+                ),
+                reason="scene_prompt_only",
+            )
+
+        if fast_path_candidate and not is_engine_query and not is_explicit_deep:
+            if normalized_consultant_mode:
+                return QueryPlannerDecision(
+                    answer_mode=ANSWER_MODE_RETRIEVAL_NORMAL,
+                    thinking_mode=THINKING_MODE_NORMAL,
+                    allow_fast_path=False,
+                    retrieval_answer_mode=ANSWER_MODE_RETRIEVAL_NORMAL,
+                    budget=RetrievalBudget(max_steps=AGENT_MAX_RETRIEVAL_STEPS, vector_top_k=DEFAULT_TOP_K),
+                    reason=f"consultant_mode:{normalized_consultant_mode}",
+                )
+            return QueryPlannerDecision(
+                answer_mode=ANSWER_MODE_FAST_PATH,
+                thinking_mode=THINKING_MODE_NORMAL,
+                allow_fast_path=True,
+                retrieval_answer_mode=ANSWER_MODE_RETRIEVAL_NORMAL,
+                budget=RetrievalBudget(max_steps=AGENT_MAX_RETRIEVAL_STEPS, vector_top_k=DEFAULT_TOP_K),
+                reason="deterministic_query_type",
+            )
+
+        retrieval_mode = ANSWER_MODE_RETRIEVAL_DEEP if (normalized_thinking_mode == THINKING_MODE_DEEP or is_engine_query or is_explicit_deep) else ANSWER_MODE_RETRIEVAL_NORMAL
+        return QueryPlannerDecision(
+            answer_mode=retrieval_mode,
+            thinking_mode=THINKING_MODE_DEEP if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else THINKING_MODE_NORMAL,
+            allow_fast_path=False,
+            retrieval_answer_mode=retrieval_mode,
+            budget=RetrievalBudget(
+                max_steps=DEEP_RETRIEVAL_STEPS if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else AGENT_MAX_RETRIEVAL_STEPS,
+                vector_top_k=DEEP_VECTOR_TOP_K if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else DEFAULT_TOP_K,
+            ),
+            reason="requested_thinking_mode" if retrieval_mode == ANSWER_MODE_RETRIEVAL_DEEP else "default_retrieval",
+        )
+
     async def _prepare_answer_runtime(
         self,
         *,
@@ -135,9 +209,23 @@ class AdaptiveRAGAnswerMainlineMixin:
         message_intent: Optional[str],
         mixed_signal_part: Optional[str],
         anchor_resolution: Optional[FollowupAnchorResolution] = None,
+        consultant_mode_preference: Optional[str] = None,
     ) -> dict[str, Any]:
+        planner = self._plan_query_runtime(
+            question=clean_q,
+            effective_question=effective_q,
+            query_type=query_type,
+            requested_thinking_mode=active_thinking_mode,
+            enforce_scene_prompt_only=enforce_scene_prompt_only,
+            consultant_mode_preference=consultant_mode_preference,
+        )
+        active_thinking_mode = planner.thinking_mode
+        deep_thinking = planner.retrieval_answer_mode == ANSWER_MODE_RETRIEVAL_DEEP
+        deep_retrieval_steps = int(planner.budget.max_steps or deep_retrieval_steps)
+        deep_vector_top_k = int(planner.budget.vector_top_k or deep_vector_top_k)
+
         forced_deep_reason = ""
-        if not deep_thinking:
+        if not deep_thinking and not planner.allow_fast_path:
             forced_deep_reason = self._forced_deep_reason(
                 question=clean_q,
                 effective_question=effective_q,
@@ -211,7 +299,7 @@ class AdaptiveRAGAnswerMainlineMixin:
             user_id=asked_by,
         )
 
-        if not deep_thinking and not enforce_scene_prompt_only:
+        if planner.allow_fast_path:
             deterministic_result = await self._run_deterministic_query(
                 question=effective_q,
                 project_id=project_id,
@@ -233,10 +321,11 @@ class AdaptiveRAGAnswerMainlineMixin:
                     effective_question=effective_q,
                     rewrite_meta=rewrite_meta,
                     thinking_mode=active_thinking_mode,
+                    answer_mode=ANSWER_MODE_FAST_PATH,
                 )
 
         direct_answer = None
-        if not deep_thinking and not enforce_scene_prompt_only:
+        if planner.allow_fast_path:
             direct_answer = self._build_snapshot_direct_answer(
                 effective_q,
                 system_snapshot=system_snapshot,
@@ -262,7 +351,7 @@ class AdaptiveRAGAnswerMainlineMixin:
             return direct_result
 
         is_snapshot_direct_question = self._is_snapshot_direct_question(effective_q)
-        if not deep_thinking and not enforce_scene_prompt_only and not is_snapshot_direct_question:
+        if planner.allow_fast_path 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(
@@ -280,6 +369,7 @@ class AdaptiveRAGAnswerMainlineMixin:
                     effective_question=effective_q,
                     rewrite_meta=rewrite_meta,
                     thinking_mode=active_thinking_mode,
+                    answer_mode=ANSWER_MODE_FAST_PATH,
                 )
 
         return await self._build_retrieval_response(
@@ -299,7 +389,7 @@ class AdaptiveRAGAnswerMainlineMixin:
             thinking_mode=thinking_mode,
             active_thinking_mode=active_thinking_mode,
             deep_thinking=deep_thinking,
-            forced_deep_reason=forced_deep_reason,
+            forced_deep_reason=forced_deep_reason or str(planner.reason or ""),
             deep_retrieval_steps=deep_retrieval_steps,
             deep_vector_top_k=deep_vector_top_k,
             system_snapshot=system_snapshot,
@@ -307,4 +397,6 @@ class AdaptiveRAGAnswerMainlineMixin:
             started=started,
             message_intent=message_intent,
             mixed_signal_part=mixed_signal_part,
+            answer_mode=planner.retrieval_answer_mode,
+            consultant_mode_preference=consultant_mode_preference,
         )

backend/app/services/adaptive_rag_service/mixin_answer_payload.py

@@ -5,6 +5,194 @@ from .shared import *
 
 
 class AdaptiveRAGAnswerPayloadMixin:
+    @staticmethod
+    def _safe_source_metadata(raw: Any) -> dict[str, Any]:
+        return raw if isinstance(raw, dict) else {}
+
+    @staticmethod
+    def _source_ref_value(value: Any) -> Any:
+        if value is None:
+            return None
+        if isinstance(value, (str, int, float, bool)):
+            text = str(value).strip()
+            return text or None
+        if isinstance(value, list):
+            rows = [str(item).strip() for item in value if str(item).strip()]
+            return rows or None
+        if isinstance(value, dict):
+            try:
+                return json.loads(json.dumps(value, ensure_ascii=False))
+            except Exception:
+                return dict(value)
+        text = str(value).strip()
+        return text or None
+
+    @staticmethod
+    def _normalize_answer_mode(value: Optional[str]) -> str:
+        raw = str(value or "").strip().lower()
+        if raw in {
+            ANSWER_MODE_FAST_PATH,
+            ANSWER_MODE_RETRIEVAL_NORMAL,
+            ANSWER_MODE_RETRIEVAL_DEEP,
+            ANSWER_MODE_CLARIFY,
+            ANSWER_MODE_BLOCKED,
+        }:
+            return raw
+        return ANSWER_MODE_RETRIEVAL_NORMAL
+
+    def _normalize_grounding_status(
+        self,
+        value: Optional[str],
+        *,
+        sources: Optional[list[dict[str, Any]]] = None,
+        missing_fields: Optional[list[str]] = None,
+        answer_mode: Optional[str] = None,
+        reason_code: Optional[str] = None,
+    ) -> str:
+        raw = str(value or "").strip().lower()
+        if raw in {
+            GROUNDING_STATUS_GROUNDED,
+            GROUNDING_STATUS_PARTIALLY_GROUNDED,
+            GROUNDING_STATUS_INSUFFICIENT_EVIDENCE,
+            GROUNDING_STATUS_BLOCKED,
+        }:
+            return raw
+
+        normalized_mode = self._normalize_answer_mode(answer_mode)
+        if normalized_mode == ANSWER_MODE_BLOCKED:
+            return GROUNDING_STATUS_BLOCKED
+        if normalized_mode == ANSWER_MODE_CLARIFY:
+            return GROUNDING_STATUS_INSUFFICIENT_EVIDENCE
+
+        source_rows = sources if isinstance(sources, list) else []
+        snapshot_only = False
+        if hasattr(self, "_is_snapshot_only_sources"):
+            try:
+                snapshot_only = bool(self._is_snapshot_only_sources(source_rows))
+            except Exception:
+                snapshot_only = False
+        if not snapshot_only and source_rows:
+            snapshot_only = all(str(item.get("source") or "").strip().lower() == "system" for item in source_rows if isinstance(item, dict))
+
+        missing = [str(item).strip() for item in (missing_fields or []) if str(item).strip()]
+        normalized_reason = str(reason_code or "").strip().lower()
+        if normalized_reason == "insufficient_evidence":
+            return GROUNDING_STATUS_INSUFFICIENT_EVIDENCE
+        if snapshot_only or missing:
+            return GROUNDING_STATUS_PARTIALLY_GROUNDED
+        if source_rows:
+            return GROUNDING_STATUS_GROUNDED
+        return GROUNDING_STATUS_PARTIALLY_GROUNDED
+
+    def _build_citations_from_sources(self, sources: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        normalized: list[dict[str, Any]] = []
+        seen: set[tuple[str, str, str, str]] = set()
+        for item in (sources or [])[:8]:
+            if not isinstance(item, dict):
+                continue
+            metadata = self._safe_source_metadata(item.get("metadata"))
+            source_type = self._source_ref_value(
+                metadata.get("source_type")
+                or item.get("source")
+                or metadata.get("table")
+                or metadata.get("mode")
+            )
+            source_label = self._source_ref_value(item.get("label") or metadata.get("source_label"))
+            source_id = self._source_ref_value(
+                metadata.get("source_id")
+                or metadata.get("doc_id")
+                or metadata.get("knowledge_doc_id")
+                or metadata.get("message_id")
+                or metadata.get("url")
+                or metadata.get("table")
+                or metadata.get("matched_project_ids")
+            )
+            chunk_id = self._source_ref_value(metadata.get("chunk_id") or metadata.get("chunk_index"))
+            target_url = self._source_ref_value(
+                metadata.get("target_url")
+                or metadata.get("doc_url")
+                or metadata.get("url")
+                or metadata.get("permalink")
+            )
+            project_id = self._source_ref_value(metadata.get("project_id"))
+            updated_at = self._source_ref_value(
+                metadata.get("updated_at")
+                or metadata.get("doc_updated_at")
+                or metadata.get("source_updated_at")
+            )
+            citation = {
+                "source_type": source_type,
+                "source_label": source_label,
+                "source_id": source_id,
+                "chunk_id": chunk_id,
+                "target_url": target_url,
+                "project_id": project_id,
+                "updated_at": updated_at,
+                "confidence": round(self._to_float(item.get("score"), 0.0), 4),
+            }
+            key = (
+                str(source_type or ""),
+                str(source_id or ""),
+                str(chunk_id or ""),
+                str(target_url or ""),
+            )
+            if key in seen:
+                continue
+            seen.add(key)
+            normalized.append(citation)
+        return normalized
+
+    @staticmethod
+    def _build_sources_used_from_citations(citations: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        rows: list[dict[str, Any]] = []
+        for item in citations:
+            if not isinstance(item, dict):
+                continue
+            rows.append(
+                {
+                    "source_type": item.get("source_type"),
+                    "source_label": item.get("source_label"),
+                    "source_id": item.get("source_id"),
+                    "chunk_id": item.get("chunk_id"),
+                    "target_url": item.get("target_url"),
+                    "project_id": item.get("project_id"),
+                    "updated_at": item.get("updated_at"),
+                }
+            )
+        return rows
+
+    def _build_evidence_pack(
+        self,
+        *,
+        answer: str,
+        citations: list[dict[str, Any]],
+        tool_trace: list[dict[str, Any]],
+        missing_fields: Optional[list[str]] = None,
+        grounding_status: Optional[str] = None,
+        reason_code: Optional[str] = None,
+    ) -> dict[str, Any]:
+        evidence_labels: list[str] = []
+        for item in citations[:5]:
+            if not isinstance(item, dict):
+                continue
+            label = str(item.get("source_label") or item.get("source_id") or item.get("source_type") or "").strip()
+            if label:
+                evidence_labels.append(label)
+        uncertainties = [str(item).strip() for item in (missing_fields or []) if str(item).strip()]
+        normalized_grounding = str(grounding_status or "").strip().lower()
+        normalized_reason = str(reason_code or "").strip()
+        if normalized_grounding in {GROUNDING_STATUS_INSUFFICIENT_EVIDENCE, GROUNDING_STATUS_BLOCKED} and normalized_reason:
+            if normalized_reason not in uncertainties:
+                uncertainties.append(normalized_reason)
+        return {
+            "summary": {
+                "conclusion": str(answer or "").strip(),
+                "evidence": evidence_labels,
+                "uncertainties": uncertainties,
+            },
+            "tool_trace": tool_trace if isinstance(tool_trace, list) else [],
+        }
+
     def _build_response_payload(
         self,
         *,
@@ -40,6 +228,13 @@ class AdaptiveRAGAnswerPayloadMixin:
         anchor_kind: Optional[str] = None,
         meta_followup_kind: Optional[str] = None,
         reason_code: Optional[str] = None,
+        answer_mode: Optional[str] = None,
+        grounding_status: Optional[str] = None,
+        policy_version: Optional[str] = None,
+        citations: Optional[list[dict[str, Any]]] = None,
+        sources_used: Optional[list[dict[str, Any]]] = None,
+        evidence_pack: Optional[dict[str, Any]] = None,
+        prompt_variant_id: Optional[str] = None,
     ) -> dict[str, Any]:
         evidence: list[dict[str, Any]] = []
         for item in sources[:8]:
@@ -53,6 +248,37 @@ class AdaptiveRAGAnswerPayloadMixin:
                     "metadata": item.get("metadata") if isinstance(item.get("metadata"), dict) else {},
                 }
             )
+        normalized_tool_trace = tool_trace if isinstance(tool_trace, list) else []
+        normalized_citations = citations if isinstance(citations, list) else self._build_citations_from_sources(sources)
+        normalized_sources_used = (
+            sources_used
+            if isinstance(sources_used, list)
+            else self._build_sources_used_from_citations(normalized_citations)
+        )
+        normalized_answer_mode = self._normalize_answer_mode(answer_mode)
+        normalized_grounding_status = self._normalize_grounding_status(
+            grounding_status,
+            sources=sources,
+            missing_fields=missing_fields,
+            answer_mode=normalized_answer_mode,
+            reason_code=reason_code,
+        )
+        normalized_policy_version = str(
+            policy_version
+            or self._build_policy_version(prompt_variant_id=prompt_variant_id)
+        ).strip() or self._build_policy_version(prompt_variant_id=prompt_variant_id)
+        normalized_evidence_pack = (
+            evidence_pack
+            if isinstance(evidence_pack, dict)
+            else self._build_evidence_pack(
+                answer=answer,
+                citations=normalized_citations,
+                tool_trace=normalized_tool_trace,
+                missing_fields=missing_fields,
+                grounding_status=normalized_grounding_status,
+                reason_code=reason_code,
+            )
+        )
         return {
             "answer": str(answer or "").strip(),
             "evidence": evidence,
@@ -71,7 +297,7 @@ class AdaptiveRAGAnswerPayloadMixin:
                 "vector_timeout": bool(vector_timeout),
                 "graph_timeout": bool(graph_timeout),
             },
-            "tool_trace": tool_trace if isinstance(tool_trace, list) else [],
+            "tool_trace": normalized_tool_trace,
             "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),
@@ -93,6 +319,12 @@ class AdaptiveRAGAnswerPayloadMixin:
             "anchor_kind": str(anchor_kind or "").strip() or None,
             "meta_followup_kind": str(meta_followup_kind or "").strip() or None,
             "reason_code": str(reason_code or "").strip() or None,
+            "answer_mode": normalized_answer_mode,
+            "grounding_status": normalized_grounding_status,
+            "policy_version": normalized_policy_version,
+            "citations": normalized_citations,
+            "sources_used": normalized_sources_used,
+            "evidence_pack": normalized_evidence_pack,
         }
 
     @staticmethod

backend/app/services/adaptive_rag_service/mixin_answer_precheck.py

@@ -52,6 +52,24 @@ class AdaptiveRAGAnswerPrecheckMixin:
                 },
             }
         ]
+        runtime_fields = self._build_qa_record_runtime_fields(
+            answer=answer,
+            sources=direct_sources,
+            missing_fields=[],
+            answer_mode=ANSWER_MODE_FAST_PATH,
+            grounding_status=GROUNDING_STATUS_GROUNDED,
+            prompt_variant_id="meta_followup",
+            reason_code=reason_code,
+            routing_snapshot=routing_snapshot,
+            retrieval_metrics={
+                "db_query_ms": 0,
+                "vector_search_ms": 0,
+                "graph_query_ms": 0,
+                "db_timeout": False,
+                "vector_timeout": False,
+                "graph_timeout": False,
+            },
+        )
         active_session_id = int(session_id) if session_id is not None else (
             int(context_session_id) if context_session_id is not None else None
         )
@@ -73,6 +91,7 @@ class AdaptiveRAGAnswerPrecheckMixin:
                 anchor_ref=str(meta_followup_payload.get("anchor_ref") or "").strip() or None,
                 routing_snapshot=routing_snapshot,
             ),
+            **runtime_fields,
         )
         self.db.add(qa)
         await self._update_session_state_after_turn(
@@ -114,6 +133,14 @@ class AdaptiveRAGAnswerPrecheckMixin:
             anchor_kind=anchor_kind,
             meta_followup_kind=meta_followup_kind,
             reason_code=reason_code,
+            answer_mode=ANSWER_MODE_FAST_PATH,
+            grounding_status=GROUNDING_STATUS_GROUNDED,
+            policy_version=runtime_fields["policy_version"],
+            citations=runtime_fields["citations"],
+            sources_used=runtime_fields["sources_used"],
+            evidence_pack=runtime_fields["evidence_pack"],
+            tool_trace=runtime_fields["tool_trace"],
+            prompt_variant_id="meta_followup",
         )
 
     async def _maybe_return_clarification_response(
@@ -164,6 +191,23 @@ class AdaptiveRAGAnswerPrecheckMixin:
                 "metadata": {"mode": "clarification", "reason": clarification_code},
             }
         ]
+        runtime_fields = self._build_qa_record_runtime_fields(
+            answer=clarification_answer,
+            sources=clarify_sources,
+            missing_fields=missing_fields,
+            answer_mode=ANSWER_MODE_CLARIFY,
+            grounding_status=GROUNDING_STATUS_INSUFFICIENT_EVIDENCE,
+            prompt_variant_id="clarification",
+            reason_code=clarification_code,
+            retrieval_metrics={
+                "db_query_ms": 0,
+                "vector_search_ms": 0,
+                "graph_query_ms": 0,
+                "db_timeout": False,
+                "vector_timeout": False,
+                "graph_timeout": False,
+            },
+        )
         active_session_id = int(session_id) if session_id is not None else (
             int(context_session_id) if context_session_id is not None else None
         )
@@ -181,6 +225,7 @@ class AdaptiveRAGAnswerPrecheckMixin:
             session_id=active_session_id,
             turn_index=turn_index,
             **self._build_qa_record_external_fields(),
+            **runtime_fields,
         )
         self.db.add(qa)
         if clarification_code == "ambiguous_recent_reply" and clarification_candidates:
@@ -233,6 +278,14 @@ class AdaptiveRAGAnswerPrecheckMixin:
             clarification_code=clarification_code,
             clarification_candidates=clarification_candidates,
             reason_code=clarification_code,
+            answer_mode=ANSWER_MODE_CLARIFY,
+            grounding_status=GROUNDING_STATUS_INSUFFICIENT_EVIDENCE,
+            policy_version=runtime_fields["policy_version"],
+            citations=runtime_fields["citations"],
+            sources_used=runtime_fields["sources_used"],
+            evidence_pack=runtime_fields["evidence_pack"],
+            tool_trace=runtime_fields["tool_trace"],
+            prompt_variant_id="clarification",
         )
 
     async def _maybe_return_snapshot_direct_response(
@@ -274,6 +327,22 @@ class AdaptiveRAGAnswerPrecheckMixin:
                     "metadata": {"mode": "snapshot_direct", "topic": "current_user"},
                 }
             )
+        runtime_fields = self._build_qa_record_runtime_fields(
+            answer=direct_answer,
+            sources=direct_sources,
+            missing_fields=[],
+            answer_mode=ANSWER_MODE_FAST_PATH,
+            grounding_status=GROUNDING_STATUS_PARTIALLY_GROUNDED,
+            prompt_variant_id="snapshot_direct",
+            retrieval_metrics={
+                "db_query_ms": 0,
+                "vector_search_ms": 0,
+                "graph_query_ms": 0,
+                "db_timeout": False,
+                "vector_timeout": False,
+                "graph_timeout": False,
+            },
+        )
 
         qa = QARecord(
             tenant_id=self.tenant_id,
@@ -289,6 +358,7 @@ class AdaptiveRAGAnswerPrecheckMixin:
             session_id=int(session_id) if session_id is not None else (int(context_session_id) if context_session_id is not None else None),
             turn_index=turn_index,
             **self._build_qa_record_external_fields(),
+            **runtime_fields,
         )
         self.db.add(qa)
         await self._update_session_state_after_turn(
@@ -326,4 +396,12 @@ class AdaptiveRAGAnswerPrecheckMixin:
             confidence=0.99,
             missing_fields=[],
             thinking_mode=active_thinking_mode,
+            answer_mode=ANSWER_MODE_FAST_PATH,
+            grounding_status=GROUNDING_STATUS_PARTIALLY_GROUNDED,
+            policy_version=runtime_fields["policy_version"],
+            citations=runtime_fields["citations"],
+            sources_used=runtime_fields["sources_used"],
+            evidence_pack=runtime_fields["evidence_pack"],
+            tool_trace=runtime_fields["tool_trace"],
+            prompt_variant_id="snapshot_direct",
         )

backend/app/services/adaptive_rag_service/mixin_answer_retrieval.py

@@ -32,6 +32,8 @@ class AdaptiveRAGAnswerRetrievalMixin:
         started: float,
         message_intent: Optional[str] = None,
         mixed_signal_part: Optional[str] = None,
+        answer_mode: str = ANSWER_MODE_RETRIEVAL_NORMAL,
+        consultant_mode_preference: Optional[str] = None,
     ) -> dict[str, Any]:
         domains = self._route_domains(effective_q)
         if str(asked_via or "").strip().lower() == "feishu" and source_chat_id and "feishu" not in domains:
@@ -65,6 +67,7 @@ class AdaptiveRAGAnswerRetrievalMixin:
             },
         )
         final_domains = [str(x) for x in (search_meta.get("domains_used") or domains)]
+        prompt_variant_id = self._peek_answer_prompt_variant()
 
         max_score = retrieved[0].score if retrieved else 0.0
         answer_text = await self._generate_answer(
@@ -82,7 +85,9 @@ class AdaptiveRAGAnswerRetrievalMixin:
             message_intent=message_intent,
             mixed_signal_part=mixed_signal_part,
             query_type=query_type,
+            consultant_mode_preference=consultant_mode_preference,
         )
+        prompt_variant_id = self._peek_answer_prompt_variant()
         routing_snapshot = self._peek_last_answer_routing_snapshot()
         used_sources = self._dedupe_sources(retrieved)
         if not used_sources:
@@ -145,6 +150,7 @@ class AdaptiveRAGAnswerRetrievalMixin:
                 message_intent=message_intent,
                 mixed_signal_part=mixed_signal_part,
                 query_type=query_type,
+                consultant_mode_preference=consultant_mode_preference,
             )
             deep_sources = self._dedupe_sources(deep_retrieved)
             if not deep_sources:
@@ -203,6 +209,37 @@ class AdaptiveRAGAnswerRetrievalMixin:
             used_sources = deep_sources
             deep_thinking = True
             active_thinking_mode = THINKING_MODE_DEEP
+            answer_mode = ANSWER_MODE_RETRIEVAL_DEEP
+            prompt_variant_id = self._peek_answer_prompt_variant()
+            routing_snapshot = self._peek_last_answer_routing_snapshot()
+
+        evidence_guard_triggered = False
+        clarification_due_to_evidence = False
+        if not evidence_guard_triggered and self._is_snapshot_only_sources(used_sources) and float(max_score or 0.0) < 0.45:
+            clarification_due_to_evidence = True
+            answer_text = (
+                f"当前关于“{clean_q[:80]}”的证据不足，现阶段只命中系统快照。"
+                "请补充更明确的对象、范围或文档入口，我再给你可核验的结论。"
+            )
+            used_sources = [
+                {
+                    "source": "system",
+                    "label": "Evidence Guard",
+                    "score": 0.91,
+                    "metadata": {"mode": "clarification", "reason": "insufficient_evidence"},
+                }
+            ]
+            max_score = min(float(max_score or 0.0), 0.4)
+            if "system_context" not in final_domains:
+                final_domains.append("system_context")
+            tool_trace.append(
+                {
+                    "step": 97,
+                    "tool": "clarification_guard",
+                    "reason": "snapshot_only_low_relevance",
+                }
+            )
+            answer_mode = ANSWER_MODE_CLARIFY
 
         evidence_guard_triggered = self._should_block_answer_with_guard(
             question=clean_q,
@@ -230,8 +267,53 @@ class AdaptiveRAGAnswerRetrievalMixin:
                     "reason": "insufficient_call_chain_evidence",
                 }
             )
+            answer_mode = ANSWER_MODE_BLOCKED
 
         elapsed = int((time.perf_counter() - started) * 1000)
+        grounding_verdict = GroundingVerdict(
+            status=GROUNDING_STATUS_GROUNDED,
+            missing_fields=[],
+            reason_code=None,
+        )
+        if evidence_guard_triggered:
+            grounding_verdict = GroundingVerdict(
+                status=GROUNDING_STATUS_BLOCKED,
+                missing_fields=["call_chain_evidence", "target_scope"],
+                reason_code="insufficient_evidence",
+            )
+        elif clarification_due_to_evidence:
+            grounding_verdict = GroundingVerdict(
+                status=GROUNDING_STATUS_INSUFFICIENT_EVIDENCE,
+                missing_fields=["high_relevance_evidence"],
+                reason_code="insufficient_evidence",
+            )
+        elif self._is_snapshot_only_sources(used_sources) or float(max_score or 0.0) < 0.45:
+            grounding_verdict = GroundingVerdict(
+                status=GROUNDING_STATUS_PARTIALLY_GROUNDED,
+                missing_fields=["high_relevance_evidence"] if float(max_score or 0.0) < 0.45 else [],
+                reason_code=None,
+            )
+
+        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),
+        }
+        runtime_fields = self._build_qa_record_runtime_fields(
+            answer=answer_text,
+            sources=used_sources,
+            missing_fields=grounding_verdict.missing_fields,
+            answer_mode=answer_mode,
+            grounding_status=grounding_verdict.status,
+            prompt_variant_id=prompt_variant_id,
+            reason_code=grounding_verdict.reason_code,
+            routing_snapshot=routing_snapshot,
+            tool_trace=tool_trace,
+            retrieval_metrics=retrieval_metrics,
+        )
         qa = QARecord(
             tenant_id=self.tenant_id,
             question=clean_q,
@@ -250,16 +332,13 @@ class AdaptiveRAGAnswerRetrievalMixin:
                 anchor_ref=None,
                 routing_snapshot=routing_snapshot,
             ),
+            **runtime_fields,
         )
         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"]))
+        missing_fields = list(grounding_verdict.missing_fields)
         await self._update_session_state_after_turn(
             session_state=state_row,
             question=clean_q,
@@ -290,7 +369,7 @@ class AdaptiveRAGAnswerRetrievalMixin:
             db_timeout=db_timeout,
             vector_timeout=vector_timeout,
             graph_timeout=graph_timeout,
-            tool_trace=tool_trace,
+            tool_trace=runtime_fields["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),
@@ -299,5 +378,12 @@ class AdaptiveRAGAnswerRetrievalMixin:
             confidence=confidence,
             missing_fields=missing_fields,
             thinking_mode=active_thinking_mode,
-            reason_code="answer_generated",
+            reason_code=grounding_verdict.reason_code or ("answer_generated" if answer_mode not in {ANSWER_MODE_CLARIFY, ANSWER_MODE_BLOCKED} else "insufficient_evidence"),
+            answer_mode=answer_mode,
+            grounding_status=grounding_verdict.status,
+            policy_version=runtime_fields["policy_version"],
+            citations=runtime_fields["citations"],
+            sources_used=runtime_fields["sources_used"],
+            evidence_pack=runtime_fields["evidence_pack"],
+            prompt_variant_id=prompt_variant_id,
         )

backend/app/services/adaptive_rag_service/mixin_core_intent.py

@@ -11,6 +11,7 @@ class AdaptiveRAGCoreIntentMixin:
         self.llm = LLMService()
         self.context_cache = context_cache_service
         self._active_external_qa_context: dict[str, Any] = {}
+        self._last_answer_prompt_variant_id = "unknown"
 
     @staticmethod
     def _normalize_thinking_mode(value: Optional[str]) -> str:
@@ -94,12 +95,14 @@ class AdaptiveRAGCoreIntentMixin:
         route_source = str(payload.get("route_source") or "").strip().lower()
         route_rule_name = str(payload.get("route_rule_name") or "").strip()
         routing_decision = str(payload.get("routing_decision") or "").strip()
+        routing_config_version = str(payload.get("routing_config_version") or "").strip()
         normalized = {
             "planned_tier": planned_tier if planned_tier in {"low", "high"} else None,
             "executed_tier": executed_tier if executed_tier in {"low", "high"} else None,
             "route_source": route_source or None,
             "route_rule_name": route_rule_name or None,
             "routing_decision": routing_decision or None,
+            "routing_config_version": routing_config_version or None,
         }
         return {key: value for key, value in normalized.items() if value is not None}
 
@@ -135,6 +138,97 @@ class AdaptiveRAGCoreIntentMixin:
             "routing_snapshot": normalized_snapshot,
         }
 
+    def _remember_answer_prompt_variant(self, prompt_variant_id: Optional[str]) -> None:
+        normalized = str(prompt_variant_id or "").strip() or "unknown"
+        self._last_answer_prompt_variant_id = normalized[:64]
+
+    def _peek_answer_prompt_variant(self) -> str:
+        return str(getattr(self, "_last_answer_prompt_variant_id", None) or "").strip() or "unknown"
+
+    def _build_policy_version(
+        self,
+        *,
+        routing_snapshot: Optional[dict[str, Any]] = None,
+        prompt_variant_id: Optional[str] = None,
+        retrieval_policy_version: Optional[str] = None,
+    ) -> str:
+        snapshot = self._normalize_routing_snapshot(routing_snapshot or self._peek_last_answer_routing_snapshot())
+        routing_version = str(snapshot.get("routing_config_version") or "").strip() or "unknown"
+        prompt_variant = str(prompt_variant_id or self._peek_answer_prompt_variant() or "").strip() or "unknown"
+        retrieval_version = str(retrieval_policy_version or RETRIEVAL_POLICY_VERSION).strip() or RETRIEVAL_POLICY_VERSION
+        return f"{retrieval_version}|routing:{routing_version}|prompt:{prompt_variant}"
+
+    def _build_qa_record_runtime_fields(
+        self,
+        *,
+        answer: str,
+        sources: list[dict[str, Any]],
+        missing_fields: Optional[list[str]] = None,
+        answer_mode: Optional[str] = None,
+        grounding_status: Optional[str] = None,
+        policy_version: Optional[str] = None,
+        prompt_variant_id: Optional[str] = None,
+        citations: Optional[list[dict[str, Any]]] = None,
+        sources_used: Optional[list[dict[str, Any]]] = None,
+        evidence_pack: Optional[dict[str, Any]] = None,
+        tool_trace: Optional[list[dict[str, Any]]] = None,
+        retrieval_metrics: Optional[dict[str, Any]] = None,
+        reason_code: Optional[str] = None,
+        routing_snapshot: Optional[dict[str, Any]] = None,
+    ) -> dict[str, Any]:
+        normalized_citations = (
+            citations
+            if isinstance(citations, list)
+            else self._build_citations_from_sources(sources)
+        )
+        normalized_sources_used = (
+            sources_used
+            if isinstance(sources_used, list)
+            else self._build_sources_used_from_citations(normalized_citations)
+        )
+        normalized_tool_trace = tool_trace if isinstance(tool_trace, list) else []
+        normalized_metrics = retrieval_metrics if isinstance(retrieval_metrics, dict) else {}
+        normalized_answer_mode = self._normalize_answer_mode(answer_mode)
+        normalized_grounding_status = self._normalize_grounding_status(
+            grounding_status,
+            sources=sources,
+            missing_fields=missing_fields,
+            answer_mode=normalized_answer_mode,
+            reason_code=reason_code,
+        )
+        normalized_policy_version = str(
+            policy_version
+            or self._build_policy_version(
+                routing_snapshot=routing_snapshot,
+                prompt_variant_id=prompt_variant_id,
+            )
+        ).strip() or self._build_policy_version(
+            routing_snapshot=routing_snapshot,
+            prompt_variant_id=prompt_variant_id,
+        )
+        normalized_evidence_pack = (
+            evidence_pack
+            if isinstance(evidence_pack, dict)
+            else self._build_evidence_pack(
+                answer=answer,
+                citations=normalized_citations,
+                tool_trace=normalized_tool_trace,
+                missing_fields=missing_fields,
+                grounding_status=normalized_grounding_status,
+                reason_code=reason_code,
+            )
+        )
+        return {
+            "answer_mode": normalized_answer_mode,
+            "grounding_status": normalized_grounding_status,
+            "policy_version": normalized_policy_version,
+            "citations": normalized_citations,
+            "sources_used": normalized_sources_used,
+            "evidence_pack": normalized_evidence_pack,
+            "tool_trace": normalized_tool_trace,
+            "retrieval_metrics": normalized_metrics,
+        }
+
     @staticmethod
     def _external_session_id_from_key(conversation_key: str) -> Optional[int]:
         key = str(conversation_key or "").strip()

backend/app/services/adaptive_rag_service/mixin_deep_guard_answer.py

@@ -7,6 +7,50 @@ from .shared import *
 
 
 class AdaptiveRAGDeepGuardAnswerMixin:
+    @staticmethod
+    def _normalize_consultant_mode_preference(value: Optional[str]) -> Optional[str]:
+        text = str(value or "").strip().lower()
+        if text in {"clarify", "advise", "challenge", "brainstorm", "execute"}:
+            return text
+        return None
+
+    @classmethod
+    def _apply_consultant_mode_preference(
+        cls,
+        *,
+        scene_key: str,
+        prompt: str,
+        consultant_mode_preference: Optional[str],
+    ) -> tuple[str, str]:
+        mode = cls._normalize_consultant_mode_preference(consultant_mode_preference)
+        if not mode:
+            return scene_key, prompt
+
+        label_map = {
+            "clarify": "先澄清",
+            "advise": "给建议",
+            "challenge": "先劝退",
+            "brainstorm": "一起想",
+            "execute": "可执行",
+        }
+        instruction_map = {
+            "clarify": "请优先指出证据不足或边界不清的地方，并给出最多 3 个必须先补充的澄清问题。",
+            "advise": "请优先给出稳妥建议，明确推荐方案、原因、取舍和下一步。",
+            "challenge": "请优先挑战隐含假设、范围膨胀和高风险点，不要把推断包装成已验证事实。",
+            "brainstorm": "请优先给出 2-3 个可选方向，分别说明适用场景、代价和收敛建议。",
+            "execute": "请只说明可执行前提、风险和下一步，不得把未发生的写操作表述为已完成。",
+        }
+        label = label_map.get(mode, mode)
+        supplement = instruction_map.get(mode, "")
+        wrapped_prompt = str(prompt or "").rstrip()
+        if supplement:
+            wrapped_prompt = (
+                f"{wrapped_prompt}\n\n"
+                f"补充要求（顾问模式：{label}）：\n"
+                f"{supplement}"
+            )
+        return f"{scene_key}_consultant_{mode}", wrapped_prompt
+
     @staticmethod
     def _compact_member_workload_user_snapshot(user_snapshot: Optional[dict[str, Any]]) -> dict[str, Any]:
         if not isinstance(user_snapshot, dict):
@@ -171,6 +215,7 @@ class AdaptiveRAGDeepGuardAnswerMixin:
         user_snapshot_text: str,
         relevance_max: float,
         query_type: Optional[str] = None,
+        consultant_mode_preference: Optional[str] = None,
     ) -> tuple[str, str]:
         normalized_question = re.sub(r"\s+", "", str(question or "").lower())
         member_keywords = ("谁", "成员", "负责人", "负载")
@@ -180,7 +225,7 @@ class AdaptiveRAGDeepGuardAnswerMixin:
             filtered_context_blocks = self._filter_member_workload_context_blocks(context_blocks_text)
             compact_user_snapshot = self._compact_member_workload_user_snapshot(user_snapshot)
             compact_user_snapshot_text = json.dumps(compact_user_snapshot or {}, ensure_ascii=False)[:6000]
-            return (
+            scene_key, prompt = (
                 "member_workload",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_MEMBER_WORKLOAD_QA",
@@ -191,11 +236,16 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     context_blocks=filtered_context_blocks,
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
         if str(message_intent or "").strip().lower() == "mixed":
             safe_signal_part = str(mixed_signal_part or "").strip() or "需求信息待补充"
             merged_message = f"{question}\n{safe_signal_part}".strip()
-            return (
+            scene_key, prompt = (
                 "mixed",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_FEISHU_MIXED_QA",
@@ -208,9 +258,14 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     context_blocks=context_blocks_text or "无",
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
         if self._is_project_status_question(question, query_type):
-            return (
+            scene_key, prompt = (
                 "project_status",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_PROJECT_STATUS_QA",
@@ -222,6 +277,11 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     context_blocks=context_blocks_text or "无",
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
         if self._contains_any_keyword(normalized_question, trend_keywords):
             explicit_project_context_markers = (
@@ -250,7 +310,7 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     },
                     ensure_ascii=False,
                 )[:16000]
-                return (
+                scene_key, prompt = (
                     "external_trend",
                     await self._render_scene_prompt(
                         prompt_key="PROMPT_EXTERNAL_TREND_QA",
@@ -260,7 +320,12 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                         external_results=context_blocks_text or "无",
                     ),
                 )
-            return (
+                return self._apply_consultant_mode_preference(
+                    scene_key=scene_key,
+                    prompt=prompt,
+                    consultant_mode_preference=consultant_mode_preference,
+                )
+            scene_key, prompt = (
                 "industry_trend",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_INDUSTRY_TREND_QA",
@@ -268,9 +333,14 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     question=question,
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
         if self._is_decision_choice_question(question):
-            return (
+            scene_key, prompt = (
                 "decision_choice",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_DECISION_QA",
@@ -278,6 +348,11 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     question=question,
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
         if self._is_ai_pmo_process_entry(asked_via):
             world_model_text = json.dumps(
@@ -287,7 +362,7 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                 },
                 ensure_ascii=False,
             )[:18000]
-            return (
+            scene_key, prompt = (
                 "ai_pmo",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_AI_PMO_QA",
@@ -298,10 +373,15 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     context_blocks=context_blocks_text or "无",
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
         normalized_asked_via = str(asked_via or "").strip().lower()
         if normalized_asked_via in {"feishu", "lark", "feishu_group", "feishu_dm"}:
-            return (
+            scene_key, prompt = (
                 "feishu_direct",
                 await self._render_scene_prompt(
                     prompt_key="PROMPT_FEISHU_DIRECT_QA",
@@ -314,8 +394,13 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                     context_blocks=context_blocks_text or "无",
                 ),
             )
+            return self._apply_consultant_mode_preference(
+                scene_key=scene_key,
+                prompt=prompt,
+                consultant_mode_preference=consultant_mode_preference,
+            )
 
-        return (
+        scene_key, prompt = (
             "default",
             await self._render_scene_prompt(
                 prompt_key="PROMPT_RAG_ANSWER",
@@ -329,6 +414,11 @@ class AdaptiveRAGDeepGuardAnswerMixin:
                 relevance_max=f"{max(0.0, float(relevance_max or 0.0)):.3f}",
             ),
         )
+        return self._apply_consultant_mode_preference(
+            scene_key=scene_key,
+            prompt=prompt,
+            consultant_mode_preference=consultant_mode_preference,
+        )
 
     def _build_snapshot_direct_answer(
         self,
@@ -534,6 +624,7 @@ class AdaptiveRAGDeepGuardAnswerMixin:
         message_intent: Optional[str] = None,
         mixed_signal_part: Optional[str] = None,
         query_type: Optional[str] = None,
+        consultant_mode_preference: Optional[str] = None,
     ) -> str:
         active_thinking_mode = self._normalize_thinking_mode(thinking_mode)
         deep_thinking = active_thinking_mode == THINKING_MODE_DEEP
@@ -582,7 +673,11 @@ class AdaptiveRAGDeepGuardAnswerMixin:
             user_snapshot_text=user_snapshot_text,
             relevance_max=relevance_max,
             query_type=query_type,
+            consultant_mode_preference=consultant_mode_preference,
         )
+        remember_prompt_variant = getattr(self, "_remember_answer_prompt_variant", None)
+        if callable(remember_prompt_variant):
+            remember_prompt_variant(scene_key)
         if deep_thinking:
             prompt = (
                 f"{prompt}\n\n"
@@ -613,7 +708,7 @@ class AdaptiveRAGDeepGuardAnswerMixin:
         text = str(answer or "").strip()
         if not text:
             raise ValueError("adaptive_rag_answer_empty_response")
-        if scene_key == "mixed":
+        if scene_key == "mixed" or str(scene_key).startswith("mixed_consultant_"):
             return self._format_mixed_two_block_output(
                 answer_text=text,
                 question_part=question,

backend/app/services/adaptive_rag_service/mixin_fast_answer_semantic_p1.py

@@ -166,6 +166,7 @@ class AdaptiveRAGFastAnswerSemanticMixinP1:
                 ],
                 "db_query_ms": db_query_ms,
                 "response_time_ms": response_ms,
+                "prompt_variant_id": "glm_semantic",
             }
         except Exception:
             return None

backend/app/services/adaptive_rag_service/mixin_fast_answer_semantic_p2.py

@@ -22,6 +22,7 @@ class AdaptiveRAGFastAnswerSemanticMixinP2:
         effective_question: Optional[str] = None,
         rewrite_meta: Optional[dict[str, Any]] = None,
         thinking_mode: str = THINKING_MODE_NORMAL,
+        answer_mode: str = ANSWER_MODE_FAST_PATH,
     ) -> dict[str, Any]:
         started = time.perf_counter()
         answer_text = str(fast_result.get("answer") or "未命中结果")
@@ -42,6 +43,24 @@ class AdaptiveRAGFastAnswerSemanticMixinP2:
         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)
+        prompt_variant_id = str(fast_result.get("prompt_variant_id") or "deterministic").strip() or "deterministic"
+        runtime_fields = self._build_qa_record_runtime_fields(
+            answer=answer_text,
+            sources=sources,
+            missing_fields=missing_fields,
+            answer_mode=answer_mode,
+            grounding_status=GROUNDING_STATUS_GROUNDED,
+            prompt_variant_id=prompt_variant_id,
+            reason_code=str(fast_result.get("reason_code") or "").strip() or None,
+            retrieval_metrics={
+                "db_query_ms": int(fast_result.get("db_query_ms") or 0),
+                "vector_search_ms": 0,
+                "graph_query_ms": 0,
+                "db_timeout": False,
+                "vector_timeout": False,
+                "graph_timeout": False,
+            },
+        )
 
         qa = QARecord(
             tenant_id=self.tenant_id,
@@ -57,6 +76,7 @@ class AdaptiveRAGFastAnswerSemanticMixinP2:
             session_id=active_session_id_int,
             turn_index=active_turn_index,
             **self._build_qa_record_external_fields(),
+            **runtime_fields,
         )
         self.db.add(qa)
         if state_row is None and active_session_id_int is not None:
@@ -100,4 +120,12 @@ class AdaptiveRAGFastAnswerSemanticMixinP2:
             confidence=confidence,
             missing_fields=list(missing_fields or []),
             thinking_mode=thinking_mode,
+            answer_mode=answer_mode,
+            grounding_status=GROUNDING_STATUS_GROUNDED,
+            policy_version=runtime_fields["policy_version"],
+            citations=runtime_fields["citations"],
+            sources_used=runtime_fields["sources_used"],
+            evidence_pack=runtime_fields["evidence_pack"],
+            tool_trace=runtime_fields["tool_trace"],
+            prompt_variant_id=prompt_variant_id,
         )

backend/app/services/adaptive_rag_service/mixin_requirement_ack_flow.py

@@ -102,7 +102,26 @@ class AdaptiveRAGRequirementAckFlowMixin:
         ]
 
         qa_id: Optional[int] = None
+        missing_fields = [] if resolved_project_id is not None else ["project_id"]
         try:
+            runtime_fields = self._build_qa_record_runtime_fields(
+                answer=answer_text,
+                sources=sources,
+                missing_fields=missing_fields,
+                answer_mode=ANSWER_MODE_FAST_PATH,
+                grounding_status=GROUNDING_STATUS_GROUNDED if action_taken in {"signal_created", "signal_deduped"} else GROUNDING_STATUS_PARTIALLY_GROUNDED,
+                prompt_variant_id="requirement_ack",
+                reason_code=str(action_taken or "").strip() or None,
+                retrieval_metrics={
+                    "db_query_ms": 0,
+                    "vector_search_ms": 0,
+                    "graph_query_ms": 0,
+                    "db_timeout": False,
+                    "vector_timeout": False,
+                    "graph_timeout": False,
+                },
+                tool_trace=[],
+            )
             qa = QARecord(
                 tenant_id=self.tenant_id,
                 question=question,
@@ -117,6 +136,7 @@ class AdaptiveRAGRequirementAckFlowMixin:
                 session_id=session_id,
                 turn_index=int(turn_index or 1),
                 **self._build_qa_record_external_fields(),
+                **runtime_fields,
             )
             self.db.add(qa)
             await self._update_session_state_after_turn(
@@ -146,7 +166,6 @@ class AdaptiveRAGRequirementAckFlowMixin:
             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,
@@ -171,4 +190,8 @@ class AdaptiveRAGRequirementAckFlowMixin:
             ack_latency_ms=ack_latency_ms,
             trace_id=trace_id,
             processing_stage="ack_only",
+            answer_mode=ANSWER_MODE_FAST_PATH,
+            grounding_status=GROUNDING_STATUS_GROUNDED if action_taken in {"signal_created", "signal_deduped"} else GROUNDING_STATUS_PARTIALLY_GROUNDED,
+            policy_version=self._build_policy_version(prompt_variant_id="requirement_ack"),
+            prompt_variant_id="requirement_ack",
         )

backend/app/services/adaptive_rag_service/shared_part_02.py

@@ -54,6 +54,16 @@ QUERY_TYPE_LIST = "list"
 QUERY_TYPE_STATUS = "status"
 QUERY_TYPE_EXPLAIN = "explain"
 QUERY_TYPE_UNKNOWN = "unknown"
+ANSWER_MODE_FAST_PATH = "fast_path"
+ANSWER_MODE_RETRIEVAL_NORMAL = "retrieval_normal"
+ANSWER_MODE_RETRIEVAL_DEEP = "retrieval_deep"
+ANSWER_MODE_CLARIFY = "clarify"
+ANSWER_MODE_BLOCKED = "blocked"
+GROUNDING_STATUS_GROUNDED = "grounded"
+GROUNDING_STATUS_PARTIALLY_GROUNDED = "partially_grounded"
+GROUNDING_STATUS_INSUFFICIENT_EVIDENCE = "insufficient_evidence"
+GROUNDING_STATUS_BLOCKED = "blocked"
+RETRIEVAL_POLICY_VERSION = "retrieval-v2.1"
 ACTION_MODE_AUTO = "auto"
 ACTION_MODE_READ = "read"
 ACTION_MODE_WRITE = "write"
@@ -147,3 +157,26 @@ class RetrievedItem:
     score: float
     source_label: str
     metadata: dict[str, Any]
+
+
+@dataclass
+class RetrievalBudget:
+    max_steps: int
+    vector_top_k: int
+
+
+@dataclass
+class QueryPlannerDecision:
+    answer_mode: str
+    thinking_mode: str
+    allow_fast_path: bool
+    retrieval_answer_mode: str
+    budget: RetrievalBudget
+    reason: Optional[str] = None
+
+
+@dataclass
+class GroundingVerdict:
+    status: str
+    missing_fields: list[str]
+    reason_code: Optional[str] = None

backend/app/services/ai_pmo/ai_pmo_kernel/ai_pmo_typed_executor_common.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import hashlib
 import json
+import re
 from datetime import timedelta
 from typing import Any, Optional
 

backend/app/services/ai_pmo/ai_pmo_typed_executor_execute/mixin_execute_query_project.py

@@ -43,6 +43,7 @@ class ExecuteQueryProjectMixin:
             sync_requirement = bool(rag_request_context.get("sync_requirement"))
             message_intent = str(rag_request_context.get("message_intent") or "").strip() or None
             mixed_signal_part = str(rag_request_context.get("mixed_signal_part") or "").strip() or None
+            consultant_mode_preference = str(rag_request_context.get("consultant_mode_preference") or "").strip().lower() or None
         else:
             thread_context = metadata.get("thread_context") if isinstance(metadata.get("thread_context"), dict) else None
             source_sender_open_id = str(metadata.get("source_sender_open_id") or "").strip() or None
@@ -71,6 +72,7 @@ class ExecuteQueryProjectMixin:
             sync_requirement = False
             message_intent = None
             mixed_signal_part = None
+            consultant_mode_preference = None
         if isinstance(rag_request_context, dict):
             source_sender_open_id = str(rag_request_context.get("source_sender_open_id") or "").strip() or None
             thread_context = rag_request_context.get("thread_context") if isinstance(rag_request_context.get("thread_context"), dict) else None
@@ -88,6 +90,7 @@ class ExecuteQueryProjectMixin:
             "sync_requirement": sync_requirement,
             "message_intent": message_intent,
             "mixed_signal_part": mixed_signal_part,
+            "consultant_mode_preference": consultant_mode_preference,
         }
         if session_id is not None:
             rag_kwargs["session_id"] = int(session_id)
@@ -98,6 +101,17 @@ class ExecuteQueryProjectMixin:
         try:
             rag_result = await rag_service.answer_question(**rag_kwargs)
         except Exception as exc:
+            rollback = getattr(self.db, "rollback", None)
+            if callable(rollback):
+                try:
+                    await rollback()
+                except Exception:
+                    logger.warning(
+                        "ai_pmo_query_intent_rollback_failed",
+                        tenant_id=self.tenant_id,
+                        source_system=source_system,
+                        source_message_id=source_message_id,
+                    )
             return {
                 "status": "failed_safe",
                 "error_message": self._build_query_error_message(exc),

backend/app/services/ai_pmo/consultant_dispatch.py

@@ -63,6 +63,48 @@ _MISSING_FIELD_LABELS = {
     "project_scope": "项目范围",
 }
 
+_LONG_REPLY_SECTION_HINTS = (
+    "结论：",
+    "依据：",
+    "不确定项：",
+    "事实：",
+    "建议：",
+    "待确认项：",
+)
+
+_DOMAIN_NEXT_ACTION_FALLBACKS: dict[str, dict[str, str]] = {
+    DOMAIN_RAG: {
+        MODE_ADVISE: "继续追问具体模块、入口函数或文件路径，我会把调用链收窄到那一段。",
+        MODE_CLARIFY: "请先限定范围：前端入口、/rag/ask、Kernel 或 AdaptiveRAG 其中一段。",
+        MODE_CHALLENGE: "先确认你要的是 RAG 链路、业务执行链路，还是完整系统拓扑，再继续追查。",
+        MODE_BRAINSTORM: "告诉我你最想看的那一段，我给你 2-3 条追查路径。",
+    },
+    DOMAIN_PROJECT: {
+        MODE_ADVISE: "继续补充项目目标、时间边界或负责人，我会把建议收敛成更具体方案。",
+        MODE_CLARIFY: "请先补充归属项目、目标范围或关键约束，我再继续推进。",
+        MODE_CHALLENGE: "先确认项目范围和优先级，避免一次推进过多内容。",
+        MODE_BRAINSTORM: "告诉我目标、约束和时间窗口，我给你 2-3 个项目推进方向。",
+    },
+    DOMAIN_TASK: {
+        MODE_ADVISE: "继续补充任务目标、负责人或截止时间，我会给出更具体建议。",
+        MODE_CLARIFY: "请先补充任务标题、归属项目或截止时间，我再继续推进。",
+        MODE_CHALLENGE: "先确认任务边界和优先级，避免一次塞进过多要求。",
+        MODE_BRAINSTORM: "告诉我任务目标和约束，我给你拆成 2-3 条可执行路径。",
+    },
+    DOMAIN_SIGNAL: {
+        MODE_ADVISE: "继续补充信号背景、归属项目或影响范围，我会给出更具体建议。",
+        MODE_CLARIFY: "请先补充归属项目、标题或信号背景，我再继续推进。",
+        MODE_CHALLENGE: "先确认这个信号是否值得入库，再决定是否继续推进执行。",
+        MODE_BRAINSTORM: "告诉我信号来源、影响对象和目标，我给你列 2-3 种处理路径。",
+    },
+    DOMAIN_GENERIC: {
+        MODE_ADVISE: "继续补充你最关心的对象、范围或约束，我会给出更具体建议。",
+        MODE_CLARIFY: "请先补充关键对象、范围或约束，我再继续推进。",
+        MODE_CHALLENGE: "先确认目标边界和优先级，再决定是否继续推进。",
+        MODE_BRAINSTORM: "告诉我目标和约束，我给你列 2-3 个可选方向。",
+    },
+}
+
 
 def _normalize_missing_fields(value: Any) -> list[str]:
     if not isinstance(value, list):
@@ -75,6 +117,27 @@ def _normalize_missing_fields(value: Any) -> list[str]:
     return fields
 
 
+def _normalize_action_text(text: str) -> str:
+    return re.sub(r"\s+", " ", str(text or "")).strip()
+
+
+def _looks_like_long_answer(text: str) -> bool:
+    raw = str(text or "")
+    normalized = _normalize_action_text(raw)
+    if not normalized:
+        return False
+    if "\n" in raw or len(normalized) > 120:
+        return True
+    return any(token in normalized for token in _LONG_REPLY_SECTION_HINTS)
+
+
+def _domain_next_action(domain: str, mode: str) -> str:
+    normalized_domain = str(domain or "").strip().lower() or DOMAIN_GENERIC
+    normalized_mode = str(mode or "").strip().lower() or MODE_ADVISE
+    domain_map = _DOMAIN_NEXT_ACTION_FALLBACKS.get(normalized_domain) or _DOMAIN_NEXT_ACTION_FALLBACKS[DOMAIN_GENERIC]
+    return str(domain_map.get(normalized_mode) or _DOMAIN_NEXT_ACTION_FALLBACKS[DOMAIN_GENERIC][MODE_ADVISE]).strip()
+
+
 def _resolve_domain(*, intent_name: str, primary_operation_code: Optional[str], classification: str) -> str:
     normalized_op = str(primary_operation_code or "").strip().lower()
     normalized_intent = str(intent_name or "").strip().lower()
@@ -162,11 +225,11 @@ def _derive_mode(
 
 
 def _build_clarify_next_action(*, missing_fields: list[str], direct_reply_text: str) -> str:
-    if direct_reply_text:
-        return direct_reply_text[:240]
+    if direct_reply_text and not _looks_like_long_answer(direct_reply_text):
+        return _normalize_action_text(direct_reply_text)[:120]
     labels = [_MISSING_FIELD_LABELS.get(item, item) for item in missing_fields]
     suffix = "、".join([label for label in labels if label]) or "关键信息"
-    return f"请先补充 {suffix}，我再继续推进。"
+    return f"请先补充 {suffix}，我再继续推进。"[:120]
 
 
 def _build_next_action(
@@ -183,13 +246,13 @@ def _build_next_action(
             direct_reply_text=direct_reply_text,
         )
     if mode == MODE_CHALLENGE:
-        if direct_reply_text:
-            return direct_reply_text[:240]
+        if direct_reply_text and not _looks_like_long_answer(direct_reply_text):
+            return _normalize_action_text(direct_reply_text)[:120]
         if str(risk_level or "").strip().lower() == "high":
-            return "建议先拆分范围，再确认是否继续推进高风险动作。"
-        return "建议先确认目标边界与优先级，再决定是否继续推进。"
+            return "建议先拆分范围，再确认是否继续推进高风险动作。"[:120]
+        return _domain_next_action(domain=domain, mode=mode)
     if mode == MODE_BRAINSTORM:
-        return "请补充目标用户、核心场景和约束，我会继续给出 2-3 个可选方向。"
+        return _domain_next_action(domain=domain, mode=mode)
     if mode == MODE_EXECUTE:
         if domain == DOMAIN_PROJECT:
             return "可继续查看项目结果，或补充负责人与排期信息。"
@@ -198,9 +261,9 @@ def _build_next_action(
         if domain == DOMAIN_SIGNAL:
             return "可继续确认归属项目，或查看后续自动处理结果。"
         return "可继续查看结果，或补充更多执行细节。"
-    if direct_reply_text:
-        return direct_reply_text[:240]
-    return "你可以继续追问细节，或让我给出更具体的建议。"
+    if direct_reply_text and not _looks_like_long_answer(direct_reply_text):
+        return _normalize_action_text(direct_reply_text)[:120]
+    return _domain_next_action(domain=domain, mode=mode)
 
 
 def build_consultant_dispatch_contract(

backend/app/services/feedback_center/mappers.py

@@ -198,6 +198,8 @@ def map_qa_feedback_to_record_payload(*, qa_feedback: Any, qa_record: Any) -> di
         "question": getattr(qa_record, "question", None),
         "answer": getattr(qa_record, "answer", None),
         "sources": _jsonable(getattr(qa_record, "sources", None) or []),
+        "citations": _jsonable(getattr(qa_record, "citations", None) or []),
+        "sources_used": _jsonable(getattr(qa_record, "sources_used", None) or []),
     }
     after_value = {
         "rating": rating,
@@ -228,6 +230,14 @@ def map_qa_feedback_to_record_payload(*, qa_feedback: Any, qa_record: Any) -> di
             "answer": getattr(qa_record, "answer", None),
             "sources": _jsonable(getattr(qa_record, "sources", None) or []),
             "domains_used": _jsonable(getattr(qa_record, "domains_used", None) or []),
+            "answer_mode": getattr(qa_record, "answer_mode", None),
+            "grounding_status": getattr(qa_record, "grounding_status", None),
+            "policy_version": getattr(qa_record, "policy_version", None),
+            "citations": _jsonable(getattr(qa_record, "citations", None) or []),
+            "sources_used": _jsonable(getattr(qa_record, "sources_used", None) or []),
+            "evidence_pack": _jsonable(getattr(qa_record, "evidence_pack", None) or {}),
+            "tool_trace": _jsonable(getattr(qa_record, "tool_trace", None) or []),
+            "retrieval_metrics": _jsonable(getattr(qa_record, "retrieval_metrics", None) or {}),
         },
         "ai_confidence": None,
         "reason_candidates": reason_candidates,

backend/app/services/feedback_center/reason_catalog.py

@@ -95,6 +95,62 @@ _REASON_DEFINITIONS: list[dict[str, Any]] = [
         "field_names": {"qa_feedback"},
         "confidence": 0.86,
     },
+    {
+        "code": "wrong_fast_path",
+        "label": "误走快速路径",
+        "description": "问题本应进入更重的检索路径，却被过早走了 fast path。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.9,
+    },
+    {
+        "code": "should_have_gone_deep",
+        "label": "应该进入深检索",
+        "description": "当前问题需要更深层的证据检索与合成，但 runtime 未升级。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.91,
+    },
+    {
+        "code": "wrong_source_priority",
+        "label": "来源优先级错误",
+        "description": "检索时对事实源的优先级排序不合理，导致引用偏差。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.88,
+    },
+    {
+        "code": "insufficient_evidence",
+        "label": "证据不足",
+        "description": "回答缺少足够内部证据支撑，应澄清或拒答。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.94,
+    },
+    {
+        "code": "good_retrieval_bad_synthesis",
+        "label": "检索好但合成差",
+        "description": "证据命中基本正确，但最终答案组织、归纳或结论表达有偏差。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.83,
+    },
+    {
+        "code": "over_deep_latency",
+        "label": "过度深检索延迟",
+        "description": "问题本可快速回答，但进入过深链路导致延迟偏高。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.84,
+    },
+    {
+        "code": "permission_scope_miss",
+        "label": "权限范围漏检",
+        "description": "检索没有正确命中用户应有权限范围内的事实源。",
+        "source_types": {"qa_feedback"},
+        "field_names": {"qa_feedback"},
+        "confidence": 0.87,
+    },
     {
         "code": "strategic_judgement_mismatch",
         "label": "战略判断偏差",
@@ -325,6 +381,22 @@ def build_reason_candidates(
                 0.78,
                 "回答缺少关键信息或上下文不够完整。",
             )
+            grounding_reason_map = {
+                "wrong_fast_path": ("wrong_fast_path", "误走快速路径", 0.91, "问题本应进入更重的检索路径，却被过早走了 fast path。"),
+                "fast_path": ("wrong_fast_path", "误走快速路径", 0.89, "问题本应进入更重的检索路径，却被过早走了 fast path。"),
+                "should_have_gone_deep": ("should_have_gone_deep", "应该进入深检索", 0.92, "当前问题需要更深层的证据检索与合成，但 runtime 未升级。"),
+                "deep": ("should_have_gone_deep", "应该进入深检索", 0.88, "当前问题需要更深层的证据检索与合成，但 runtime 未升级。"),
+                "wrong_source_priority": ("wrong_source_priority", "来源优先级错误", 0.88, "检索时对事实源的优先级排序不合理。"),
+                "insufficient_evidence": ("insufficient_evidence", "证据不足", 0.94, "回答缺少足够内部证据支撑，应澄清或拒答。"),
+                "good_retrieval_bad_synthesis": ("good_retrieval_bad_synthesis", "检索好但合成差", 0.83, "证据命中基本正确，但最终答案组织或结论表达有偏差。"),
+                "over_deep_latency": ("over_deep_latency", "过度深检索延迟", 0.84, "问题本可快速回答，但进入过深链路导致延迟偏高。"),
+                "permission_scope_miss": ("permission_scope_miss", "权限范围漏检", 0.87, "检索没有正确命中用户应有权限范围内的事实源。"),
+            }
+            for token in error_types:
+                mapping = grounding_reason_map.get(token)
+                if mapping is None:
+                    continue
+                add(mapping[0], mapping[1], mapping[2], mapping[3])
         else:
             add("manual_review_needed", "需要人工复核", 0.72, "QA 反馈建议保留人工复核以确认样本质量。")
     elif source_key == "signal_feedback":

backend/app/services/llm_service/mixin_prompt_profile_runtime.py

@@ -177,6 +177,7 @@ class LLMPromptProfileRuntimeMixin:
             "route_source": meta.get("route_source") or None,
             "route_rule_name": meta.get("route_rule_name") or None,
             "routing_decision": meta.get("routing_decision") or None,
+            "routing_config_version": meta.get("routing_config_version") or None,
         }
         async with AsyncSessionLocal() as session:
             for attempt in range(2):

backend/tests/test_adaptive_rag_external_followup_contract.py

@@ -29,6 +29,11 @@ def _use_sqlite_json_types() -> None:
             if column.name in {
                 "sources",
                 "domains_used",
+                "citations",
+                "sources_used",
+                "evidence_pack",
+                "tool_trace",
+                "retrieval_metrics",
                 "last_entity_ids",
                 "rewrite_history",
                 "pending_clarification",

backend/tests/test_adaptive_rag_mainline_contract.py

@@ -33,10 +33,12 @@ class _FakeAdaptiveRAGService(
             "context_session_id": 101,
             "conversation": [],
             "state_row": None,
+            "resolved_question": clean_q,
             "effective_q": clean_q,
             "rewrite_meta": {"mode": "direct"},
             "query_type": "explain",
             "effective_action_mode": action_mode,
+            "anchor_resolution": None,
         }
 
     async def _handle_ack_first_requirement(self, **kwargs):
@@ -85,25 +87,43 @@ async def test_answer_entry_write_mode_uses_requirement_ack_gate() -> None:
 
 
 @pytest.mark.asyncio
-async def test_answer_entry_high_accuracy_channel_forces_deep_mode() -> None:
+async def test_answer_entry_high_accuracy_channel_respects_requested_normal_mode() -> None:
     service = _FakeAdaptiveRAGService()
 
     result = await service.answer_question(
         question="最近有哪些项目？",
         asked_by=1,
         asked_via="pmflow",
+        thinking_mode="normal",
         sync_requirement=True,
     )
 
     assert result["processing_stage"] == "qa"
     assert service.ack_calls == 0
     assert service.base_calls == 1
-    assert service.base_kwargs["active_thinking_mode"] == "deep"
-    assert service.base_kwargs["deep_thinking"] is True
+    assert service.base_kwargs["active_thinking_mode"] == "normal"
+    assert service.base_kwargs["deep_thinking"] is False
     assert int(service.base_kwargs["deep_retrieval_steps"]) == DEEP_RETRIEVAL_STEPS
     assert int(service.base_kwargs["deep_vector_top_k"]) == DEEP_VECTOR_TOP_K
 
 
+@pytest.mark.asyncio
+async def test_answer_entry_forwards_consultant_mode_preference_to_mainline() -> None:
+    service = _FakeAdaptiveRAGService()
+
+    result = await service.answer_question(
+        question="这个方案怎么做更稳妥？",
+        asked_by=1,
+        asked_via="pmflow",
+        thinking_mode="normal",
+        consultant_mode_preference="advise",
+    )
+
+    assert result["processing_stage"] == "qa"
+    assert service.base_calls == 1
+    assert service.base_kwargs["consultant_mode_preference"] == "advise"
+
+
 @pytest.mark.asyncio
 async def test_answer_entry_mixed_request_skips_requirement_ack_gate() -> None:
     service = _FakeAdaptiveRAGService()
@@ -120,8 +140,8 @@ async def test_answer_entry_mixed_request_skips_requirement_ack_gate() -> None:
     assert result["processing_stage"] == "qa"
     assert service.ack_calls == 0
     assert service.base_calls == 1
-    assert service.base_kwargs["active_thinking_mode"] == "deep"
-    assert service.base_kwargs["deep_thinking"] is True
+    assert service.base_kwargs["active_thinking_mode"] == "normal"
+    assert service.base_kwargs["deep_thinking"] is False
 
 
 @pytest.mark.asyncio
@@ -139,12 +159,49 @@ async def test_answer_entry_feishu_write_mode_still_goes_base_mainline() -> None
     assert result["processing_stage"] == "qa"
     assert service.ack_calls == 0
     assert service.base_calls == 1
-    assert service.base_kwargs["active_thinking_mode"] == "deep"
-    assert service.base_kwargs["deep_thinking"] is True
+    assert service.base_kwargs["active_thinking_mode"] == "normal"
+    assert service.base_kwargs["deep_thinking"] is False
     assert int(service.base_kwargs["deep_retrieval_steps"]) == DEEP_RETRIEVAL_STEPS
     assert int(service.base_kwargs["deep_vector_top_k"]) == DEEP_VECTOR_TOP_K
 
 
+def test_query_runtime_planner_prefers_fast_path_for_list_question_even_if_requested_deep() -> None:
+    class _FakePlannerService(AdaptiveRAGAnswerMainlineMixin):
+        pass
+
+    service = _FakePlannerService()
+    decision = service._plan_query_runtime(
+        question="现在有哪些项目？",
+        effective_question="现在有哪些项目？",
+        query_type="list",
+        requested_thinking_mode="deep",
+        enforce_scene_prompt_only=False,
+    )
+
+    assert decision.answer_mode == "fast_path"
+    assert decision.thinking_mode == "normal"
+    assert decision.allow_fast_path is True
+
+
+def test_query_runtime_planner_bypasses_fast_path_when_consultant_mode_preference_present() -> None:
+    class _FakePlannerService(AdaptiveRAGAnswerMainlineMixin):
+        pass
+
+    service = _FakePlannerService()
+    decision = service._plan_query_runtime(
+        question="现在有哪些项目？",
+        effective_question="现在有哪些项目？",
+        query_type="list",
+        requested_thinking_mode="normal",
+        enforce_scene_prompt_only=False,
+        consultant_mode_preference="advise",
+    )
+
+    assert decision.answer_mode == "retrieval_normal"
+    assert decision.thinking_mode == "normal"
+    assert decision.allow_fast_path is False
+
+
 @pytest.mark.asyncio
 async def test_deterministic_query_prefers_project_created_at_answer() -> None:
     calls: list[str] = []
@@ -207,7 +264,7 @@ async def test_deep_guard_answer_raises_on_empty_llm_response() -> None:
             _ = project_id
             return {}
 
-        def _select_scene_prompt(self, **kwargs):
+        async def _select_scene_prompt(self, **kwargs):
             _ = kwargs
             return "external_trend", "请回答问题"
 
@@ -289,3 +346,38 @@ async def test_deep_guard_answer_prefers_member_workload_direct_answer() -> None
 
     assert "杨宇当前负载265.56%" in result
     assert "严重超载" in result
+
+
+@pytest.mark.asyncio
+async def test_select_scene_prompt_applies_consultant_mode_preference_wrapper() -> None:
+    class _FakeService(AdaptiveRAGDeepGuardAnswerMixin):
+        def __init__(self) -> None:
+            self.db = object()
+            self.tenant_id = 1
+
+        async def _render_scene_prompt(self, *, prompt_key: str, fallback_template: str, **kwargs):
+            _ = (fallback_template, kwargs)
+            return f"PROMPT::{prompt_key}"
+
+    service = _FakeService()
+
+    scene_key, prompt = await service._select_scene_prompt(
+        question="这个方案怎么做更稳妥？",
+        project_id=7,
+        asked_via="pmflow",
+        message_intent=None,
+        mixed_signal_part=None,
+        conversation_text="无",
+        context_blocks_text="无",
+        system_snapshot={},
+        user_snapshot={},
+        system_snapshot_text="{}",
+        user_snapshot_text="{}",
+        relevance_max=0.61,
+        query_type="explain",
+        consultant_mode_preference="advise",
+    )
+
+    assert scene_key == "default_consultant_advise"
+    assert "PROMPT::PROMPT_RAG_ANSWER" in prompt
+    assert "稳妥建议" in prompt

backend/tests/test_ai_pmo_ingress_mainline.py

@@ -6,11 +6,13 @@ from fastapi import HTTPException
 
 from app.api.ai_pmo_api import routes_core
 from app.api.ai_pmo_api.schemas import PMOProcessTextRequest
+from app.services.ai_pmo.consultant_dispatch import build_consultant_dispatch_contract
 from app.services.ai_pmo.ai_pmo_kernel.mixins.intent.ai_pmo_kernel_mixins_intent_rules import (
     KernelIntentRulesMixin,
 )
 from app.services.ai_pmo.ai_pmo_kernel.ai_agent_kernel import gate_execution
 from app.services.ai_pmo.ai_pmo_kernel.ai_pmo_reply_renderer import ReplyRenderer
+from app.services.ai_pmo.ai_pmo_kernel import ai_pmo_typed_executor_mixins_parse
 from app.services.ai_pmo.ai_pmo_kernel_ingress_flow import mixin_entry
 from app.services.ai_pmo.ai_pmo_kernel_ingress_flow import mixin_prepare_manual
 from app.services.ai_pmo.ai_pmo_typed_executor_execute import mixin_execute_query_project
@@ -499,6 +501,7 @@ async def test_execute_query_intent_honors_rag_request_context(monkeypatch) -> N
                 "sync_requirement": True,
                 "message_intent": "mixed",
                 "mixed_signal_part": "新增任务确认流",
+                "consultant_mode_preference": "challenge",
             }
         },
     )
@@ -525,6 +528,7 @@ async def test_execute_query_intent_honors_rag_request_context(monkeypatch) -> N
     assert answer_kwargs["sync_requirement"] is True
     assert answer_kwargs["message_intent"] == "mixed"
     assert answer_kwargs["mixed_signal_part"] == "新增任务确认流"
+    assert answer_kwargs["consultant_mode_preference"] == "challenge"
 
 
 @pytest.mark.asyncio
@@ -577,6 +581,91 @@ async def test_execute_query_intent_feishu_fallback_uses_thread_key(monkeypatch)
     )
 
 
+@pytest.mark.asyncio
+async def test_execute_query_intent_rolls_back_failed_rag_transaction(monkeypatch) -> None:
+    class FakeAdaptiveRAGService:
+        def __init__(self, db, tenant_id: int):
+            _ = (db, tenant_id)
+
+        async def answer_question(self, **kwargs):
+            _ = kwargs
+            raise RuntimeError("rag failed before response")
+
+    class FakeDB:
+        def __init__(self) -> None:
+            self.rollback_calls = 0
+
+        async def rollback(self) -> None:
+            self.rollback_calls += 1
+
+    class FakeExecutor(mixin_execute_query_project.ExecuteQueryProjectMixin):
+        def __init__(self) -> None:
+            self.db = FakeDB()
+            self.tenant_id = 3
+
+        @staticmethod
+        def _build_query_error_message(exc: Exception) -> str:
+            return f"query failed: {exc}"
+
+    monkeypatch.setattr(mixin_execute_query_project, "AdaptiveRAGService", FakeAdaptiveRAGService)
+    executor = FakeExecutor()
+    intent = SimpleNamespace(project_id=12, metadata={})
+
+    result = await executor._execute_query_intent(
+        intent=intent,
+        message_text="现在完整调用链是什么？",
+        operator_user_id=999,
+        source_system="pmflow_manual",
+        source_message_id="msg-rollback-1",
+    )
+
+    assert result["status"] == "failed_safe"
+    assert "rag failed" in str(result["error_message"] or "")
+    assert executor.db.rollback_calls == 1
+
+
+def test_typed_executor_common_exports_regex_helpers_to_query_mixins() -> None:
+    assert mixin_execute_query_project.ExecuteQueryProjectMixin._extract_task_priority("p0 紧急任务") == "high"
+    assert (
+        ai_pmo_typed_executor_mixins_parse.TypedExecutorParseMixin._extract_requested_signal_count(
+            message_text="请帮我分析 3 个信号",
+            metadata={},
+        )
+        == 3
+    )
+    assert (
+        ai_pmo_typed_executor_mixins_parse.TypedExecutorParseMixin._requires_strict_signal_analysis(
+            message_text="先分析后入库 2 个信号",
+            requested_count=2,
+        )
+        is True
+    )
+
+
+def test_consultant_dispatch_uses_short_next_action_for_query_answers() -> None:
+    dispatch = build_consultant_dispatch_contract(
+        intent_name="query.answer",
+        message_text="PMFlow 现在完整调用链是什么？",
+        source_system="pmflow_manual",
+        status="answered",
+        classification="question",
+        primary_operation_code="query.answer",
+        risk_level="low",
+        missing_fields=[],
+        direct_reply_text=(
+            "结论：PMFlow 当前完整调用链无法完整还原。\n\n"
+            "依据：前端 `/ai-steward` 页面通过 `/rag/ask` 发起请求，"
+            "并继续进入 `AdaptiveRAGService.upsert_pmflow_record_async`。"
+        ),
+    )
+
+    assert dispatch["allow_execute"] is False
+    assert dispatch["mode"] == "advise"
+    assert dispatch["next_action"] == "继续追问具体模块、入口函数或文件路径，我会把调用链收窄到那一段。"
+    assert "结论：" not in dispatch["next_action"]
+    assert "/rag/ask" not in dispatch["next_action"]
+
+
 @pytest.mark.asyncio
 async def test_low_tier_rag_failure_retries_high_tier(monkeypatch) -> None:
     captured: dict[str, object] = {"logs": []}

backend/tests/test_ai_pmo_qa_http_mainline_e2e.py

@@ -48,7 +48,23 @@ async def test_http_rag_ask_readonly_mainline_contract(monkeypatch) -> None:
                 "intent": "query.read_only",
                 "direct_reply_required": True,
                 "direct_reply_text": "当前主链状态正常。",
-                "rag_result": {"answer": "当前主链状态正常。", "processing_stage": "qa"},
+                "rag_result": {
+                    "answer": "当前主链状态正常。",
+                    "processing_stage": "qa",
+                    "answer_mode": "fast_path",
+                    "grounding_status": "grounded",
+                    "policy_version": "retrieval-v2.1|routing:v3|prompt:deterministic",
+                    "citations": [],
+                    "sources_used": [],
+                    "evidence_pack": {
+                        "summary": {
+                            "conclusion": "当前主链状态正常。",
+                            "evidence": [],
+                            "uncertainties": [],
+                        },
+                        "tool_trace": [],
+                    },
+                },
             }
 
     monkeypatch.setattr(routes_query, "AIPMOKernelService", FakeKernel)
@@ -68,6 +84,11 @@ async def test_http_rag_ask_readonly_mainline_contract(monkeypatch) -> None:
     payload = response.json()
     assert payload["code"] == 0
     assert "当前主链状态正常" in str((payload.get("data") or {}).get("answer") or "")
+    assert payload["data"]["answer_mode"] == "fast_path"
+    assert payload["data"]["grounding_status"] == "grounded"
+    assert payload["data"]["policy_version"] == "retrieval-v2.1|routing:v3|prompt:deterministic"
+    assert payload["data"]["citations"] == []
+    assert payload["data"]["sources_used"] == []
     ingress_kwargs = captured["ingress_kwargs"]
     assert isinstance(ingress_kwargs, dict)
     assert ingress_kwargs["source_system"] == "pmflow_manual"

backend/tests/test_feedback_center_service.py

@@ -6,7 +6,7 @@ from types import SimpleNamespace
 
 import pytest
 import pytest_asyncio
-from sqlalchemy import select
+from sqlalchemy import JSON, select
 from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
 
 os.environ["DATABASE_URL"] = "sqlite+aiosqlite:///:memory:"
@@ -18,14 +18,29 @@ from app.services.feedback_center.mappers import (
     map_signal_feedback_to_record_payload,
     map_task_override_to_record_payloads,
 )
-from app.services.feedback_center.reason_catalog import build_reason_candidates
+from app.services.feedback_center.reason_catalog import build_reason_candidates, list_reason_catalog
 import app.services.feedback_center.service as feedback_center_service_module
 from app.services.feedback_center.service import FeedbackCenterService
 
 
+def _use_sqlite_json_types() -> None:
+    for table in (FeedbackCenterRecord.__table__, FeedbackCenterRecordRevision.__table__):
+        for column in table.columns:
+            if column.name in {
+                "before_value",
+                "after_value",
+                "ai_snapshot",
+                "reason_candidates",
+                "before_snapshot",
+                "after_snapshot",
+            }:
+                column.type = JSON()
+
+
 @pytest_asyncio.fixture
 async def db_session():
     engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    _use_sqlite_json_types()
     async with engine.begin() as conn:
         await conn.run_sync(
             lambda sync_conn: Base.metadata.create_all(
@@ -268,6 +283,19 @@ def test_reason_catalog_is_deterministic_for_field_diff() -> None:
     assert first[0]["code"] == "effort_estimation_bias"
 
 
+def test_reason_catalog_exposes_grounding_retrieval_codes_for_qa_feedback() -> None:
+    catalog = list_reason_catalog(source_type="qa_feedback")
+    codes = {item["code"] for item in catalog}
+
+    assert "wrong_fast_path" in codes
+    assert "should_have_gone_deep" in codes
+    assert "wrong_source_priority" in codes
+    assert "insufficient_evidence" in codes
+    assert "good_retrieval_bad_synthesis" in codes
+    assert "over_deep_latency" in codes
+    assert "permission_scope_miss" in codes
+
+
 def test_source_mappers_generate_unified_payloads() -> None:
     qa_payload = map_qa_feedback_to_record_payload(
         qa_feedback=SimpleNamespace(

backend/tests/test_qa_record_runtime_fields_migration.py

@@ -0,0 +1,93 @@
+from __future__ import annotations
+
+import importlib.util
+from pathlib import Path
+
+import sqlalchemy as sa
+from alembic.migration import MigrationContext
+from alembic.operations import Operations
+
+
+def _load_migration_module():
+    path = (
+        Path(__file__).resolve().parents[1]
+        / "alembic"
+        / "versions"
+        / "20260331_add_qa_record_runtime_fields.py"
+    )
+    spec = importlib.util.spec_from_file_location("qa_record_runtime_fields_migration", path)
+    module = importlib.util.module_from_spec(spec)
+    assert spec is not None
+    assert spec.loader is not None
+    spec.loader.exec_module(module)
+    return module
+
+
+def _create_qa_record_table(metadata: sa.MetaData) -> None:
+    sa.Table(
+        "qa_records",
+        metadata,
+        sa.Column("id", sa.Integer(), primary_key=True, autoincrement=True),
+        sa.Column("tenant_id", sa.BigInteger(), nullable=False),
+        sa.Column("created_at", sa.TIMESTAMP(), nullable=False),
+        sa.Column("updated_at", sa.TIMESTAMP(), nullable=False),
+        sa.Column("question", sa.Text(), nullable=False),
+        sa.Column("answer", sa.Text(), nullable=False),
+        sa.Column("sources", sa.JSON(), nullable=True),
+        sa.Column("intent", sa.String(length=20), nullable=False),
+        sa.Column("domains_used", sa.JSON(), nullable=True),
+        sa.Column("response_time_ms", sa.Integer(), nullable=True),
+        sa.Column("asked_by", sa.BigInteger(), nullable=True),
+        sa.Column("asked_via", sa.String(length=20), nullable=False),
+        sa.Column("project_id", sa.BigInteger(), nullable=True),
+        sa.Column("session_id", sa.Integer(), nullable=True),
+        sa.Column("turn_index", sa.Integer(), nullable=False),
+        sa.Column("source_message_id", sa.String(length=128), nullable=True),
+        sa.Column("source_chat_id", sa.String(length=128), nullable=True),
+        sa.Column("source_sender_open_id", sa.String(length=128), nullable=True),
+        sa.Column("reply_to_message_id", sa.String(length=128), nullable=True),
+        sa.Column("root_message_id", sa.String(length=128), nullable=True),
+        sa.Column("conversation_scope_key", sa.String(length=160), nullable=True),
+        sa.Column("anchor_kind", sa.String(length=32), nullable=True),
+        sa.Column("anchor_ref", sa.String(length=128), nullable=True),
+        sa.Column("routing_snapshot", sa.JSON(), nullable=True),
+    )
+
+
+def test_qa_record_runtime_fields_migration_adds_and_removes_columns() -> None:
+    migration = _load_migration_module()
+    engine = sa.create_engine("sqlite:///:memory:")
+    metadata = sa.MetaData()
+    _create_qa_record_table(metadata)
+    metadata.create_all(engine)
+
+    with engine.begin() as conn:
+        context = MigrationContext.configure(conn)
+        migration.op = Operations(context)
+        migration.upgrade()
+
+        inspector = sa.inspect(conn)
+        record_columns = {column["name"] for column in inspector.get_columns("qa_records")}
+
+        assert "answer_mode" in record_columns
+        assert "grounding_status" in record_columns
+        assert "policy_version" in record_columns
+        assert "citations" in record_columns
+        assert "sources_used" in record_columns
+        assert "evidence_pack" in record_columns
+        assert "tool_trace" in record_columns
+        assert "retrieval_metrics" in record_columns
+
+        migration.downgrade()
+
+        inspector = sa.inspect(conn)
+        record_columns = {column["name"] for column in inspector.get_columns("qa_records")}
+
+        assert "answer_mode" not in record_columns
+        assert "grounding_status" not in record_columns
+        assert "policy_version" not in record_columns
+        assert "citations" not in record_columns
+        assert "sources_used" not in record_columns
+        assert "evidence_pack" not in record_columns
+        assert "tool_trace" not in record_columns
+        assert "retrieval_metrics" not in record_columns

backend/tests/test_rag_route_contract.py

@@ -1,3 +1,4 @@
+import json
 from types import SimpleNamespace
 
 import pytest
@@ -18,11 +19,13 @@ def test_rag_ask_request_normalizes_contract_fields() -> None:
         message_intent="MIXED",
         mixed_signal_part="新增任务确认流",
         action_mode="read",
+        consultant_mode_preference="ADVise",
     )
     assert req.asked_via == "/ai-pmo/process-text"
     assert req.message_intent == "mixed"
     assert req.mixed_signal_part == "新增任务确认流"
     assert req.action_mode == "read"
+    assert req.consultant_mode_preference == "advise"
 
 
 def test_rag_ask_request_rejects_unknown_asked_via() -> None:
@@ -44,6 +47,7 @@ def test_build_answer_question_kwargs_maps_fields() -> None:
         sync_requirement=0,
         message_intent="mixed",
         mixed_signal_part="新增任务确认流",
+        consultant_mode_preference="challenge",
     )
     kwargs = rag_ask_contract.build_answer_question_kwargs(
         request=request,
@@ -63,9 +67,29 @@ def test_build_answer_question_kwargs_maps_fields() -> None:
         "sync_requirement": False,
         "message_intent": "mixed",
         "mixed_signal_part": "新增任务确认流",
+        "consultant_mode_preference": "challenge",
     }
 
 
+def test_build_kernel_ingress_kwargs_assigns_unique_source_message_id_for_rag() -> None:
+    request = RAGAskRequest(question="现在进展怎么样")
+    kwargs_a = rag_ask_contract.build_kernel_ingress_kwargs(
+        request=request,
+        current_user=SimpleNamespace(id=1, tenant_id=1, username="admin"),
+    )
+    kwargs_b = rag_ask_contract.build_kernel_ingress_kwargs(
+        request=request,
+        current_user=SimpleNamespace(id=1, tenant_id=1, username="admin"),
+    )
+
+    source_message_id_a = str(kwargs_a["source_message_id"] or "")
+    source_message_id_b = str(kwargs_b["source_message_id"] or "")
+
+    assert source_message_id_a.startswith("rag-ask-")
+    assert source_message_id_b.startswith("rag-ask-")
+    assert source_message_id_a != source_message_id_b
+
+
 @pytest.mark.asyncio
 async def test_rag_ask_route_forwards_contract_fields(monkeypatch) -> None:
     captured: dict[str, object] = {}
@@ -94,6 +118,7 @@ async def test_rag_ask_route_forwards_contract_fields(monkeypatch) -> None:
         action_mode="read",
         message_intent="mixed",
         mixed_signal_part="新增任务确认流",
+        consultant_mode_preference="clarify",
     )
     current_user = SimpleNamespace(id=101, tenant_id=9)
     response = await routes_query.rag_ask(
@@ -116,6 +141,7 @@ async def test_rag_ask_route_forwards_contract_fields(monkeypatch) -> None:
     assert rag_request_context["action_mode"] == "read"
     assert rag_request_context["message_intent"] == "mixed"
     assert rag_request_context["mixed_signal_part"] == "新增任务确认流"
+    assert rag_request_context["consultant_mode_preference"] == "clarify"
 
 
 def test_ai_steward_router_does_not_expose_legacy_rag_ask_route() -> None:
@@ -167,10 +193,12 @@ async def test_rag_ask_route_uses_shared_contract_builder(monkeypatch) -> None:
 
     assert response_rag["code"] == 0
     assert captured_tenants == [88]
-    assert captured_kwargs == [
-        {
+    assert len(captured_kwargs) == 1
+    forwarded = captured_kwargs[0]
+    assert str(forwarded["source_message_id"]).startswith("rag-ask-")
+    assert forwarded == {
         "source_system": "pmflow",
-            "source_message_id": None,
+        "source_message_id": forwarded["source_message_id"],
         "message_text": "q1",
         "sender": "PMFlow用户",
         "channel_ref": None,
@@ -186,7 +214,6 @@ async def test_rag_ask_route_uses_shared_contract_builder(monkeypatch) -> None:
             }
         },
     }
-    ]
 
 
 def test_build_rag_response_payload_preserves_rag_fields() -> None:
@@ -202,6 +229,19 @@ def test_build_rag_response_payload_preserves_rag_fields() -> None:
                 "answer": "final answer",
                 "session_id": 18,
                 "sources": [{"source": "snapshot"}],
+                "answer_mode": "retrieval_deep",
+                "grounding_status": "grounded",
+                "policy_version": "gr-v2.1|routing-v3|prompt:feishu_direct",
+                "citations": [{"source_type": "doc", "source_id": "doc-1"}],
+                "sources_used": [{"source_type": "doc", "source_id": "doc-1"}],
+                "evidence_pack": {
+                    "summary": {
+                        "conclusion": "final answer",
+                        "evidence": ["doc-1"],
+                        "uncertainties": [],
+                    },
+                    "tool_trace": [{"tool": "vector_search"}],
+                },
                 "processing_stage": "qa",
             },
         }
@@ -213,6 +253,40 @@ def test_build_rag_response_payload_preserves_rag_fields() -> None:
     assert payload["trace_id"] == "ai-pmo-321"
     assert payload["reply_required"] is True
     assert payload["reply_text"] == "fallback"
+    assert payload["answer_mode"] == "retrieval_deep"
+    assert payload["grounding_status"] == "grounded"
+    assert payload["policy_version"] == "gr-v2.1|routing-v3|prompt:feishu_direct"
+    assert payload["citations"] == [{"source_type": "doc", "source_id": "doc-1"}]
+    assert payload["sources_used"] == [{"source_type": "doc", "source_id": "doc-1"}]
+    assert payload["evidence_pack"]["summary"]["conclusion"] == "final answer"
+    assert payload["evidence_pack"]["tool_trace"] == [{"tool": "vector_search"}]
+
+
+def test_build_rag_response_payload_adds_grounding_contract_defaults() -> None:
+    payload = rag_ask_contract.build_rag_response_payload(
+        ingress_result={
+            "status": "answered",
+            "intent": "query.read_only",
+            "direct_reply_text": "fallback",
+            "rag_result": {
+                "answer": "final answer",
+            },
+        }
+    )
+
+    assert payload["answer_mode"] == "retrieval_normal"
+    assert payload["grounding_status"] == "partially_grounded"
+    assert payload["policy_version"] == "retrieval-v2.1|routing:unknown|prompt:unknown"
+    assert payload["citations"] == []
+    assert payload["sources_used"] == []
+    assert payload["evidence_pack"] == {
+        "summary": {
+            "conclusion": "final answer",
+            "evidence": [],
+            "uncertainties": [],
+        },
+        "tool_trace": [],
+    }
 
 
 def test_build_rag_response_payload_falls_back_to_top_level_consultant_dispatch() -> None:
@@ -239,3 +313,55 @@ def test_build_rag_response_payload_falls_back_to_top_level_consultant_dispatch(
     )
 
     assert payload["consultant_dispatch"] == consultant_dispatch
+
+
+def test_build_rag_response_payload_does_not_inject_fake_answer_text() -> None:
+    payload = rag_ask_contract.build_rag_response_payload(
+        ingress_result={
+            "status": "answered",
+            "intent": "query.read_only",
+            "direct_reply_text": "",
+            "rag_result": {
+                "processing_stage": "qa",
+            },
+        }
+    )
+
+    assert payload["answer"] == ""
+
+
+@pytest.mark.asyncio
+async def test_rag_ask_route_rejects_empty_rag_answer(monkeypatch) -> None:
+    class FakeKernel:
+        def __init__(self, db, tenant_id: int):
+            _ = db
+            _ = tenant_id
+
+        async def process_ingress(self, **kwargs):
+            _ = kwargs
+            return {
+                "handled": True,
+                "status": "answered",
+                "intent": "query.read_only",
+                "direct_reply_text": "",
+                "rag_result": {
+                    "processing_stage": "qa",
+                },
+            }
+
+    monkeypatch.setattr(routes_query, "AIPMOKernelService", FakeKernel)
+
+    response = await routes_query.rag_ask(
+        request=RAGAskRequest(question="空回答测试"),
+        current_user=SimpleNamespace(id=1, tenant_id=1),
+        db=object(),
+        tenant_id=1,
+    )
+
+    assert getattr(response, "status_code", None) == 500
+    body = json.loads(response.body.decode("utf-8"))
+    assert body == {
+        "code": 500,
+        "message": "RAG 返回空回答，未生成有效响应。",
+        "data": None,
+    }

backend/tests/test_reply_iteration_api.py

@@ -65,6 +65,11 @@ def _use_sqlite_json_types() -> None:
                 "sources",
                 "domains_used",
                 "routing_snapshot",
+                "citations",
+                "sources_used",
+                "evidence_pack",
+                "tool_trace",
+                "retrieval_metrics",
                 "evidence",
                 "source_snapshot",
                 "tags",

backend/tests/test_reply_iteration_service.py

@@ -44,6 +44,11 @@ def _use_sqlite_json_types() -> None:
                 "sources",
                 "domains_used",
                 "routing_snapshot",
+                "citations",
+                "sources_used",
+                "evidence_pack",
+                "tool_trace",
+                "retrieval_metrics",
                 "evidence",
                 "source_snapshot",
                 "tags",

frontend/package.json

@@ -22,8 +22,11 @@
     "lucide-react": "^0.460.0",
     "react": "^18.3.1",
     "react-dom": "^18.3.1",
+    "react-markdown": "^10.1.0",
     "react-router-dom": "^6.28.0",
     "recharts": "^2.13.0",
+    "remark-breaks": "^4.0.0",
+    "remark-gfm": "^4.0.1",
     "tailwind-merge": "^2.5.0",
     "tailwindcss-animate": "^1.0.7",
     "zustand": "^4.5.0"

frontend/src/api/aiPmo.ts

@@ -150,6 +150,26 @@ export interface AIPMOCadenceSnapshot {
   created_at?: string | null;
 }
 
+export interface AIPMOProcessTextRequest {
+  message_text: string;
+  sender?: string | null;
+  source_message_id?: string | null;
+  project_id?: number | null;
+}
+
+export interface AIPMOProcessTextResult {
+  handled?: boolean;
+  status?: string | null;
+  intent?: string | null;
+  direct_reply_required?: boolean;
+  direct_reply_text?: string | null;
+  followup_reply?: string | null;
+  consultant_dispatch?: Record<string, unknown> | null;
+  work_item_id?: number | null;
+  signal_id?: number | null;
+  ai_trace_id?: string | null;
+}
+
 export interface AIPMOProgramView {
   summary: {
     program_count: number;
@@ -169,6 +189,12 @@ export interface AIPMOProgramView {
 }
 
 export const aiPmoApi = {
+  async processText(body: AIPMOProcessTextRequest): Promise<AIPMOProcessTextResult> {
+    const { data } = await api.post<ApiEnvelope<AIPMOProcessTextResult>>("/ai-pmo/process-text", body, {
+      timeout: 15000,
+    });
+    return unwrap(data);
+  },
   async listWorkItems(params?: { limit?: number; state?: string; projectId?: number }): Promise<AIPMOWorkItemListPayload> {
     const { data } = await api.get<ApiEnvelope<AIPMOWorkItemListPayload>>("/ai-pmo/work-items", {
       params: {

frontend/src/api/aiSteward.ts

@@ -279,7 +279,162 @@ export interface LLMEngineConnectionTestResponse {
   response_preview: string;
 }
 
+export type QAFeedbackRating = "positive" | "negative";
+export type QAHelpfulness = "fully" | "partially" | "could_be_better";
+export type QAThinkingMode = "normal" | "deep";
+export type QAActionTaken = "signal_created" | "signal_deduped" | "signal_failed";
+
+export interface QAMessageSource {
+  source: string;
+  label: string;
+  score: number;
+  metadata?: Record<string, unknown>;
+}
+
+export interface QAConsultantDispatch {
+  domain: string;
+  mode: string;
+  risk_level: string;
+  missing_fields: string[];
+  allow_execute: boolean;
+  next_action: string;
+}
+
+export interface QASessionSummary {
+  id: number;
+  title: string;
+  message_count: number;
+  last_message_at?: string | null;
+  rating_summary?: {
+    positive: number;
+    negative: number;
+    unrated: number;
+    corrected: number;
+  };
+  last_asked_via?: "pmflow" | "feishu";
+  preview_question?: string;
+}
+
+export interface QASessionMessage {
+  id: number;
+  question: string;
+  answer: string;
+  sources: QAMessageSource[];
+  response_time_ms?: number;
+  turn_index: number;
+  created_at?: string | null;
+  feedback?: {
+    id: number;
+    rating: QAFeedbackRating;
+    helpfulness?: QAHelpfulness;
+    error_types?: string[];
+    correct_answer?: string;
+  } | null;
+  thinking_mode?: QAThinkingMode;
+  action_taken?: QAActionTaken | null;
+  signal_id?: number | null;
+  reply_required?: boolean;
+  reply_text?: string | null;
+  ack_latency_ms?: number | null;
+  trace_id?: string | null;
+  processing_stage?: string | null;
+  consultant_dispatch?: QAConsultantDispatch | null;
+}
+
+export interface QASessionDetail {
+  session: {
+    id: number;
+    title: string;
+    message_count: number;
+    started_at?: string | null;
+    last_message_at?: string | null;
+  };
+  messages: QASessionMessage[];
+}
+
+export interface QAListSessionsResponse {
+  items: QASessionSummary[];
+  page: number;
+  total_pages: number;
+}
+
+export interface QAStats {
+  total_sessions: number;
+  total_qa: number;
+  rated_count: number;
+  rated_ratio: number;
+  average_helpfulness: number;
+  exportable_pairs: number;
+}
+
+export interface QAAskRequest {
+  question: string;
+  asked_via: "pmflow";
+  thinking_mode: QAThinkingMode;
+  new_session: boolean;
+  action_mode: "auto";
+  sync_requirement: boolean;
+  session_id?: number;
+  consultant_mode_preference?: "clarify" | "advise" | "challenge" | "brainstorm" | "execute";
+}
+
+export interface QAAskResponse {
+  answer: string;
+  session_id?: number | null;
+  thinking_mode?: QAThinkingMode;
+  sources?: QAMessageSource[];
+  response_time_ms?: number | null;
+  action_taken?: QAActionTaken | null;
+  signal_id?: number | null;
+  reply_required?: boolean;
+  reply_text?: string | null;
+  ack_latency_ms?: number | null;
+  trace_id?: string | null;
+  processing_stage?: string | null;
+  consultant_dispatch?: QAConsultantDispatch | null;
+}
+
 export const aiStewardApi = {
+  async askRag(body: QAAskRequest): Promise<QAAskResponse> {
+    const raw = await apiRequest<ApiEnvelope<QAAskResponse>>("/rag/ask", {
+      method: "POST",
+      body: JSON.stringify(body),
+    });
+    return unwrap(raw);
+  },
+  async listQASessions(params?: { page?: number; pageSize?: number }): Promise<QAListSessionsResponse> {
+    const query = new URLSearchParams();
+    query.set("page", String(params?.page ?? 1));
+    query.set("page_size", String(params?.pageSize ?? 20));
+    const raw = await apiRequest<ApiEnvelope<QAListSessionsResponse>>(`/rag/qa-sessions?${query.toString()}`);
+    return unwrap(raw);
+  },
+  async getQASession(sessionId: number): Promise<QASessionDetail> {
+    const raw = await apiRequest<ApiEnvelope<QASessionDetail>>(`/rag/qa-sessions/${sessionId}`);
+    return unwrap(raw);
+  },
+  async renameQASession(sessionId: number, title: string): Promise<{ title: string }> {
+    const raw = await apiRequest<ApiEnvelope<{ title: string }>>(`/rag/qa-sessions/${sessionId}/title`, {
+      method: "PATCH",
+      body: JSON.stringify({ title }),
+    });
+    return unwrap(raw);
+  },
+  async submitQAFeedback(
+    qaRecordId: number,
+    body: {
+      rating: QAFeedbackRating;
+      helpfulness?: QAHelpfulness;
+      error_types?: string[];
+      correct_answer?: string;
+    },
+  ): Promise<Record<string, unknown>> {
+    const raw = await apiRequest<ApiEnvelope<Record<string, unknown>>>(`/rag/qa-records/${qaRecordId}/feedback`, {
+      method: "POST",
+      body: JSON.stringify(body),
+    });
+    return unwrap(raw);
+  },
   async listCognitiveConflicts(projectId?: number): Promise<CognitiveConflict[]> {
     const query = typeof projectId === "number" ? `?project_id=${projectId}` : "";
     return await apiRequest<CognitiveConflict[]>(`/ai-steward/cognitive-conflicts${query}`);

frontend/src/components/ai-steward/FloatingQAChat.tsx

@@ -1,18 +1,30 @@
 import { useEffect, useMemo, useRef, useState } from "react";
 import { AlertCircle, Loader2, MessageSquare, Sparkles, ThumbsDown, ThumbsUp, X } from "lucide-react";
 
+import { aiPmoApi } from "@/api/aiPmo";
+import { aiStewardApi } from "@/api/aiSteward";
+import { QAMarkdownContent } from "@/components/ai-steward/QAMarkdownContent";
 import { Button } from "@/components/ui/button";
 import { Input } from "@/components/ui/input";
 import { Textarea } from "@/components/ui/textarea";
 import { Badge } from "@/components/ui/badge";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
 import { Switch } from "@/components/ui/switch";
-import { apiBaseURL, getAuthToken } from "@/lib/api";
 import { formatBeijingDate, formatBeijingTime, parseBackendDate } from "@/lib/datetime";
+import { cn } from "@/lib/utils";
 
 type FeedbackRating = "positive" | "negative";
 type Helpfulness = "fully" | "partially" | "could_be_better";
 type ThinkingMode = "normal" | "deep";
 type ActionTaken = "signal_created" | "signal_deduped" | "signal_failed";
+type ConsultantModePreference = "clarify" | "advise" | "challenge" | "brainstorm" | "execute";
 type MessageSource = { source: string; label: string; score: number; metadata?: Record<string, unknown> };
 type ConsultantDispatch = {
   domain: string;
@@ -84,18 +96,23 @@ const ERROR_OPTIONS = ["答案错误", "来源错误", "不完整", "答非所
 const LAST_SEEN_KEY = "pmflow_qa_chat_last_seen";
 const THINKING_MODE_KEY = "pmflow_qa_chat_thinking_mode";
 const QUICK_QUESTIONS = ["你是谁", "系统架构是什么", "最近有哪些项目", "当前阻塞项有哪些"];
-const CONSULTANT_MODE_LABELS: Record<string, string> = {
-  clarify: "先澄清",
-  advise: "给建议",
-  challenge: "先劝退",
-  brainstorm: "一起想",
-  execute: "可执行",
-};
+const CONSULTANT_MODE_ACTIONS: ReadonlyArray<{
+  value: ConsultantModePreference;
+  label: string;
+  tone: "advise" | "clarify" | "brainstorm" | "challenge";
+}> = [
+  { value: "advise", label: "给建议", tone: "advise" },
+  { value: "clarify", label: "先澄清", tone: "clarify" },
+  { value: "brainstorm", label: "一起想", tone: "brainstorm" },
+  { value: "challenge", label: "先劝退", tone: "challenge" },
+];
 const CONSULTANT_RISK_LABELS: Record<string, string> = {
   low: "低风险",
   medium: "中风险",
   high: "高风险",
 };
+const ACTION_CHIP_BASE =
+  "h-8 rounded-full border px-3 text-[11px] font-medium shadow-sm transition-all duration-150";
 
 function normalizeRequestError(error: unknown, fallback = "请求失败，请稍后重试。"): string {
   const raw = error instanceof Error ? error.message : String(error || "");
@@ -138,6 +155,14 @@ function normalizeConsultantDispatch(value: unknown): ConsultantDispatch | null
   };
 }
 
+function normalizeConsultantModePreference(value: unknown): ConsultantModePreference | null {
+  const mode = String(value || "").trim().toLowerCase();
+  if (mode === "clarify" || mode === "advise" || mode === "challenge" || mode === "brainstorm" || mode === "execute") {
+    return mode;
+  }
+  return null;
+}
+
 function looksLikeRequirementWrite(text: string): boolean {
   const clean = String(text || "").trim();
   if (!clean) return false;
@@ -167,20 +192,6 @@ function getStageBadge(message: Pick<SessionMessage, "processing_stage" | "reply
   return null;
 }
 
-function getConsultantModeBadge(
-  consultantDispatch?: ConsultantDispatch | null,
-): { label: string; variant: "secondary" | "destructive" | "outline" } | null {
-  const mode = String(consultantDispatch?.mode || "").trim().toLowerCase();
-  if (!mode) return null;
-  if (mode === "challenge") {
-    return { label: CONSULTANT_MODE_LABELS[mode] || mode, variant: "destructive" };
-  }
-  if (mode === "clarify" || mode === "execute") {
-    return { label: CONSULTANT_MODE_LABELS[mode] || mode, variant: "secondary" };
-  }
-  return { label: CONSULTANT_MODE_LABELS[mode] || mode, variant: "outline" };
-}
-
 function getConsultantRiskBadge(
   consultantDispatch?: ConsultantDispatch | null,
 ): { label: string; variant: "secondary" | "destructive" | "outline" } | null {
@@ -195,127 +206,79 @@ function getConsultantRiskBadge(
   return { label: CONSULTANT_RISK_LABELS[riskLevel] || riskLevel, variant: "secondary" };
 }
 
-async function fetchApi<T>(path: string, init?: RequestInit): Promise<T | null> {
-  const token = getAuthToken();
-  if (!token) return null;
-  try {
-    const res = await fetch(`${apiBaseURL}${path}`, {
-      ...init,
-      headers: {
-        "Content-Type": "application/json",
-        Authorization: `Bearer ${token}`,
-        ...(init?.headers || {}),
-      },
-    });
-    if (!res.ok) return null;
-    const payload = await res.json().catch(() => ({}));
-    return (payload?.data ?? payload) as T;
-  } catch (error) {
-    console.error(`[FloatingQAChat] request failed: ${path}`, error);
-    return null;
+function buildRiskExplanation(consultantDispatch?: ConsultantDispatch | null): string {
+  const riskLevel = String(consultantDispatch?.risk_level || "").trim().toLowerCase();
+  const missingCount = consultantDispatch?.missing_fields?.length || 0;
+  if (riskLevel === "high") {
+    return missingCount > 0
+      ? "当前被判为高风险，且还缺少关键输入，系统不会建议直接推进执行。"
+      : "当前被判为高风险，建议先重新确认范围、影响面和回滚路径。";
   }
+  if (riskLevel === "medium") {
+    return missingCount > 0
+      ? "当前风险中等，建议先补齐缺失信息，再决定是否继续推进。"
+      : "当前风险中等，适合先补充证据或确认边界，再继续下一步。";
+  }
+  if (consultantDispatch?.allow_execute) {
+    return "当前被判为低风险，且信息较完整，可以在确认后交接给 AI PMO 执行流。";
+  }
+  return "当前被判为低风险，但系统仍建议先按下一步补齐信息，而不是直接执行。";
 }
 
-async function askRAG(
-  question: string,
-  sessionId?: number | null,
-  thinkingMode: ThinkingMode = "normal",
-  newSession = false,
-): Promise<{
-  answer: string;
-  session_id?: number | null;
-  thinking_mode?: ThinkingMode;
-  sources?: MessageSource[];
-  response_time_ms?: number | null;
-  action_taken?: ActionTaken | null;
-  signal_id?: number | null;
-  reply_required?: boolean;
-  reply_text?: string | null;
-  ack_latency_ms?: number | null;
-  trace_id?: string | null;
-  processing_stage?: string | null;
-  consultant_dispatch?: ConsultantDispatch | null;
-}> {
-  const token = getAuthToken();
-  if (!token) {
-    throw new Error("登录状态已失效，请重新登录后提问。");
+function getActionChipClasses(
+  tone: "deep" | "advise" | "clarify" | "brainstorm" | "challenge" | "risk-low" | "risk-medium" | "risk-high" | "execute" | "next",
+  active = false,
+): string {
+  const sharedInactive = "bg-white/90 text-slate-700 hover:bg-white hover:text-slate-900";
+  switch (tone) {
+    case "deep":
+      return cn(
+        ACTION_CHIP_BASE,
+        active
+          ? "border-violet-500 bg-violet-50 text-violet-700"
+          : `${sharedInactive} border-violet-200 hover:border-violet-300 hover:text-violet-700`,
+      );
+    case "advise":
+      return cn(
+        ACTION_CHIP_BASE,
+        active
+          ? "border-emerald-500 bg-emerald-50 text-emerald-700"
+          : `${sharedInactive} border-emerald-200 hover:border-emerald-300 hover:text-emerald-700`,
+      );
+    case "clarify":
+      return cn(
+        ACTION_CHIP_BASE,
+        active
+          ? "border-amber-400 bg-amber-50 text-amber-800"
+          : `${sharedInactive} border-amber-200 hover:border-amber-300 hover:text-amber-800`,
+      );
+    case "brainstorm":
+      return cn(
+        ACTION_CHIP_BASE,
+        active
+          ? "border-sky-500 bg-sky-50 text-sky-700"
+          : `${sharedInactive} border-sky-200 hover:border-sky-300 hover:text-sky-700`,
+      );
+    case "challenge":
+      return cn(
+        ACTION_CHIP_BASE,
+        active
+          ? "border-rose-500 bg-rose-50 text-rose-700"
+          : `${sharedInactive} border-rose-200 hover:border-rose-300 hover:text-rose-700`,
+      );
+    case "risk-high":
+      return cn(ACTION_CHIP_BASE, "border-rose-300 bg-rose-50 text-rose-700 hover:border-rose-400");
+    case "risk-medium":
+      return cn(ACTION_CHIP_BASE, "border-amber-300 bg-amber-50 text-amber-800 hover:border-amber-400");
+    case "risk-low":
+      return cn(ACTION_CHIP_BASE, "border-emerald-300 bg-emerald-50 text-emerald-700 hover:border-emerald-400");
+    case "execute":
+      return cn(ACTION_CHIP_BASE, "border-slate-900 bg-slate-900 text-white hover:bg-slate-800");
+    case "next":
+      return cn(ACTION_CHIP_BASE, "border-slate-300 bg-slate-100 text-slate-800 hover:border-slate-400 hover:bg-slate-200");
+    default:
+      return cn(ACTION_CHIP_BASE, sharedInactive);
   }
-
-  const payloadObj: {
-    question: string;
-    asked_via: "pmflow";
-    thinking_mode: ThinkingMode;
-    new_session: boolean;
-    action_mode: "auto";
-    sync_requirement: boolean;
-    session_id?: number;
-  } = {
-    question,
-    asked_via: "pmflow",
-    thinking_mode: thinkingMode,
-    new_session: Boolean(newSession),
-    action_mode: "auto",
-    sync_requirement: true,
-  };
-  if (typeof sessionId === "number" && Number.isFinite(sessionId) && sessionId > 0) {
-    payloadObj.session_id = Math.trunc(sessionId);
-  }
-  const payload = JSON.stringify(payloadObj);
-  let res: Response;
-  try {
-    res = await fetch(`${apiBaseURL}/rag/ask`, {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/json",
-        Authorization: `Bearer ${token}`,
-      },
-      body: payload,
-    });
-  } catch (error) {
-    throw new Error(normalizeRequestError(error, "提问失败，请稍后重试。"));
-  }
-
-  if (!res.ok) {
-    const body = await res.json().catch(() => ({}));
-    const detail = String(body?.detail || body?.message || "").trim();
-    throw new Error(detail || `请求失败（${res.status}）`);
-  }
-
-  const body = await res.json().catch(() => ({}));
-  const data = (body?.data ?? body) as {
-    answer?: string;
-    session_id?: number | null;
-    thinking_mode?: string;
-    sources?: unknown;
-    response_time_ms?: number | null;
-    action_taken?: ActionTaken | null;
-    signal_id?: number | null;
-    reply_required?: boolean;
-    reply_text?: string | null;
-    ack_latency_ms?: number | null;
-    trace_id?: string | null;
-    processing_stage?: string | null;
-    consultant_dispatch?: unknown;
-  };
-  if (!data || !String(data.answer || "").trim()) {
-    throw new Error("服务返回为空，未生成有效回答。");
-  }
-  const resolvedThinkingMode: ThinkingMode = String(data.thinking_mode || thinkingMode).toLowerCase() === "deep" ? "deep" : "normal";
-  return {
-    answer: String(data.answer),
-    session_id: data.session_id ?? null,
-    thinking_mode: resolvedThinkingMode,
-    sources: normalizeSources(data.sources),
-    response_time_ms: data.response_time_ms ?? null,
-    action_taken: data.action_taken ?? null,
-    signal_id: data.signal_id ?? null,
-    reply_required: Boolean(data.reply_required),
-    reply_text: data.reply_text ?? null,
-    ack_latency_ms: data.ack_latency_ms ?? null,
-    trace_id: data.trace_id ?? null,
-    processing_stage: data.processing_stage ?? null,
-    consultant_dispatch: normalizeConsultantDispatch(data.consultant_dispatch),
-  };
 }
 
 function groupSessions(items: SessionSummary[]) {
@@ -384,6 +347,9 @@ export function FloatingQAChat() {
   const [errorTypes, setErrorTypes] = useState<Record<number, string[]>>({});
   const [correctAnswer, setCorrectAnswer] = useState<Record<number, string>>({});
   const [submittingFeedbackId, setSubmittingFeedbackId] = useState<number | null>(null);
+  const [actionBusyKey, setActionBusyKey] = useState<string | null>(null);
+  const [expandedRiskMessageId, setExpandedRiskMessageId] = useState<number | null>(null);
+  const [executeDialogMessage, setExecuteDialogMessage] = useState<SessionMessage | null>(null);
   const [showOlder, setShowOlder] = useState(false);
   const [unread, setUnread] = useState(false);
   const [renameEditing, setRenameEditing] = useState(false);
@@ -395,19 +361,40 @@ export function FloatingQAChat() {
   const messageViewportRef = useRef<HTMLDivElement | null>(null);
   const messageEndRef = useRef<HTMLDivElement | null>(null);
   const forceStickBottomRef = useRef(false);
+  const questionInputRef = useRef<HTMLTextAreaElement | null>(null);
+
+  const normalizeSessionMessage = (message: SessionMessage): SessionMessage => ({
+    ...message,
+    question: String(message.question || ""),
+    answer: String(message.answer || ""),
+    sources: normalizeSources(message.sources),
+    consultant_dispatch: normalizeConsultantDispatch(message.consultant_dispatch),
+  });
+
+  const normalizeSessionDetail = (detail: SessionDetail): SessionDetail => ({
+    session: detail.session,
+    messages: Array.isArray(detail.messages) ? detail.messages.map(normalizeSessionMessage) : [],
+  });
+
+  const appendLocalMessage = (message: Omit<SessionMessage, "id" | "turn_index" | "created_at" | "feedback">) => {
+    const localId = localMessageIdRef.current;
+    localMessageIdRef.current -= 1;
+    setLocalMessages((prev) => [
+      ...prev,
+      {
+        id: localId,
+        turn_index: prev.length + 1,
+        created_at: new Date().toISOString(),
+        feedback: null,
+        ...message,
+      },
+    ]);
+  };
 
   const loadSessions = async (page = 1, append = false) => {
     setHistoryLoading(true);
     try {
-      const data = await fetchApi<{
-        items: SessionSummary[];
-        page: number;
-        total_pages: number;
-      }>(`/rag/qa-sessions?page=${page}&page_size=20`);
-      if (!data) {
-        setHistoryApiAvailable(false);
-        return;
-      }
+      const data = await aiStewardApi.listQASessions({ page, pageSize: 20 });
       setHistoryApiAvailable(true);
 
       setHistoryPage(Number(data.page || page));
@@ -437,6 +424,9 @@ export function FloatingQAChat() {
       if (!activeSessionIdRef.current && !draftNewSessionRef.current && Array.isArray(data.items) && data.items.length > 0) {
         setActiveSessionId(data.items[0].id);
       }
+    } catch (error) {
+      console.error("[FloatingQAChat] loadSessions failed", error);
+      setHistoryApiAvailable(false);
     } finally {
       setHistoryLoading(false);
     }
@@ -449,17 +439,16 @@ export function FloatingQAChat() {
       setDetailLoading(true);
     }
     try {
-      const data = await fetchApi<SessionDetail>(`/rag/qa-sessions/${sessionId}`);
-      if (!data) {
-        setHistoryApiAvailable(false);
-        return;
-      }
+      const data = normalizeSessionDetail(await aiStewardApi.getQASession(sessionId));
       setHistoryApiAvailable(true);
       setSessionDetail(data);
       if (!preserveLocal) {
         setLocalMessages([]);
       }
       setRenameTitle(String(data.session?.title || ""));
+    } catch (error) {
+      console.error("[FloatingQAChat] loadSessionDetail failed", error);
+      setHistoryApiAvailable(false);
     } finally {
       if (!silent) {
         setDetailLoading(false);
@@ -487,34 +476,52 @@ export function FloatingQAChat() {
     setRenameTitle("");
   };
 
-  const askQuestion = async () => {
-    const clean = question.trim();
+  const runQuestionRequest = async ({
+    prompt,
+    requestThinkingModeOverride,
+    consultantModePreference,
+    clearInput = false,
+  }: {
+    prompt: string;
+    requestThinkingModeOverride?: ThinkingMode;
+    consultantModePreference?: ConsultantModePreference | null;
+    clearInput?: boolean;
+  }) => {
+    const clean = prompt.trim();
     if (!clean) return;
     const currentSessionId = draftNewSession ? null : activeSessionIdRef.current;
-    const currentThinkingMode: ThinkingMode = thinkingMode;
+    const currentThinkingMode: ThinkingMode = requestThinkingModeOverride || thinkingMode;
     forceStickBottomRef.current = true;
     setAskError("");
     setRequestThinkingMode(currentThinkingMode);
     setAsking(true);
     try {
-      const data = await askRAG(clean, currentSessionId, currentThinkingMode, draftNewSession);
+      const data = await aiStewardApi.askRag({
+        question: clean,
+        asked_via: "pmflow",
+        thinking_mode: currentThinkingMode,
+        new_session: Boolean(draftNewSession),
+        action_mode: "auto",
+        sync_requirement: true,
+        ...(typeof currentSessionId === "number" && Number.isFinite(currentSessionId) && currentSessionId > 0
+          ? { session_id: Math.trunc(currentSessionId) }
+          : {}),
+        ...(consultantModePreference ? { consultant_mode_preference: consultantModePreference } : {}),
+      });
+      if (!String(data.answer || "").trim() && !String(data.reply_text || "").trim()) {
+        throw new Error("服务返回为空，未生成有效回答。");
+      }
+      if (clearInput) {
         setQuestion("");
+      }
       setSessionDetail(null);
       draftNewSessionRef.current = false;
       setDraftNewSession(false);
-      const localId = localMessageIdRef.current;
-      localMessageIdRef.current -= 1;
-      setLocalMessages((prev) => [
-        ...prev,
-        {
-          id: localId,
+      appendLocalMessage({
         question: clean,
         answer: String(data.reply_text || data.answer || ""),
-          sources: data.sources?.length ? data.sources : [{ source: "rag", label: "AI 管家", score: 1 }],
+        sources: data.sources?.length ? normalizeSources(data.sources) : [{ source: "rag", label: "AI 管家", score: 1 }],
         response_time_ms: data.response_time_ms ?? undefined,
-          turn_index: prev.length + 1,
-          created_at: new Date().toISOString(),
-          feedback: null,
         thinking_mode: data.thinking_mode || currentThinkingMode,
         action_taken: data.action_taken ?? null,
         signal_id: data.signal_id ?? null,
@@ -523,9 +530,8 @@ export function FloatingQAChat() {
         ack_latency_ms: data.ack_latency_ms ?? null,
         trace_id: data.trace_id ?? null,
         processing_stage: data.processing_stage ?? null,
-          consultant_dispatch: data.consultant_dispatch ?? null,
-        },
-      ]);
+        consultant_dispatch: normalizeConsultantDispatch(data.consultant_dispatch),
+      });
       const returnedSessionId = Number(data.session_id || 0) || null;
       if (historyApiAvailable) {
         await loadSessions(1, false);
@@ -536,7 +542,7 @@ export function FloatingQAChat() {
           setActiveSessionId(currentSessionId);
           await refreshActiveSession(currentSessionId, { silent: true, preserveLocal: true });
         } else if (latestSessionAtRef.current) {
-          const fallback = await fetchApi<{ items: SessionSummary[] }>("/rag/qa-sessions?page=1&page_size=1");
+          const fallback = await aiStewardApi.listQASessions({ page: 1, pageSize: 1 });
           const target = fallback?.items?.[0]?.id || null;
           if (target) {
             setActiveSessionId(target);
@@ -553,17 +559,22 @@ export function FloatingQAChat() {
     }
   };
 
+  const askQuestion = async () => {
+    await runQuestionRequest({
+      prompt: question,
+      requestThinkingModeOverride: thinkingMode,
+      clearInput: true,
+    });
+  };
+
   const submitFeedback = async (messageId: number, rating: FeedbackRating) => {
     setSubmittingFeedbackId(messageId);
     try {
-      await fetchApi(`/rag/qa-records/${messageId}/feedback`, {
-        method: "POST",
-        body: JSON.stringify({
+      await aiStewardApi.submitQAFeedback(messageId, {
         rating,
         helpfulness: rating === "positive" ? helpfulness[messageId] || "fully" : undefined,
         error_types: rating === "negative" ? errorTypes[messageId] || [] : [],
         correct_answer: rating === "negative" ? correctAnswer[messageId] || undefined : undefined,
-        }),
       });
       setFeedbackMode((prev) => ({ ...prev, [messageId]: null }));
       await Promise.all([loadSessions(1, false), refreshActiveSession()]);
@@ -576,15 +587,83 @@ export function FloatingQAChat() {
     if (!activeSessionId) return;
     const clean = renameTitle.trim();
     if (!clean) return;
-    const data = await fetchApi<{ title: string }>(`/rag/qa-sessions/${activeSessionId}/title`, {
-      method: "PATCH",
-      body: JSON.stringify({ title: clean }),
-    });
-    if (!data) return;
+    await aiStewardApi.renameQASession(activeSessionId, clean);
     setRenameEditing(false);
     await Promise.all([loadSessions(1, false), refreshActiveSession(activeSessionId)]);
   };
 
+  const handleDeepRerun = async (message: SessionMessage) => {
+    setActionBusyKey(`deep-${message.id}`);
+    try {
+      await runQuestionRequest({
+        prompt: message.question,
+        requestThinkingModeOverride: "deep",
+      });
+    } finally {
+      setActionBusyKey(null);
+    }
+  };
+
+  const handleConsultantModeRerun = async (message: SessionMessage, modeValue: ConsultantModePreference) => {
+    setActionBusyKey(`mode-${message.id}-${modeValue}`);
+    try {
+      await runQuestionRequest({
+        prompt: message.question,
+        requestThinkingModeOverride: message.thinking_mode || thinkingMode,
+        consultantModePreference: modeValue,
+      });
+    } finally {
+      setActionBusyKey(null);
+    }
+  };
+
+  const handleToggleRiskExplanation = (messageId: number) => {
+    setExpandedRiskMessageId((prev) => (prev === messageId ? null : messageId));
+  };
+
+  const handleUseNextAction = (consultantDispatch?: ConsultantDispatch | null) => {
+    const nextAction = String(consultantDispatch?.next_action || "").trim();
+    if (!nextAction) return;
+    setMode("chat");
+    setQuestion(nextAction);
+    setAskError("");
+    window.setTimeout(() => {
+      questionInputRef.current?.focus();
+      questionInputRef.current?.setSelectionRange(nextAction.length, nextAction.length);
+    }, 0);
+  };
+
+  const handleExecuteViaKernel = async () => {
+    if (!executeDialogMessage) return;
+    setActionBusyKey(`execute-${executeDialogMessage.id}`);
+    try {
+      const result = await aiPmoApi.processText({
+        message_text: executeDialogMessage.question,
+      });
+      const replyText = String(result.direct_reply_text || result.followup_reply || "已交接到 AI PMO 执行流继续处理。").trim();
+      appendLocalMessage({
+        question: executeDialogMessage.question,
+        answer: replyText || "已交接到 AI PMO 执行流继续处理。",
+        sources: [{ source: "ai_pmo", label: "AI PMO Kernel", score: 1 }],
+        thinking_mode: executeDialogMessage.thinking_mode || "normal",
+        action_taken: null,
+        signal_id: result.signal_id ?? null,
+        reply_required: Boolean(result.direct_reply_required),
+        reply_text: result.direct_reply_text ?? result.followup_reply ?? null,
+        trace_id: result.ai_trace_id ?? null,
+        processing_stage: "execute_handoff",
+        consultant_dispatch: normalizeConsultantDispatch(result.consultant_dispatch),
+      });
+      setMode("chat");
+      setExecuteDialogMessage(null);
+    } catch (error) {
+      const message = normalizeRequestError(error, "执行交接失败，请稍后重试。");
+      setAskError(message);
+    } finally {
+      setActionBusyKey(null);
+    }
+  };
+
   const grouped = useMemo(() => groupSessions(sessions), [sessions]);
   const requestLooksLikeWrite = useMemo(() => looksLikeRequirementWrite(question), [question]);
   const displayMessages = useMemo(() => {
@@ -831,11 +910,9 @@ export function FloatingQAChat() {
                       const actionBadge = getActionBadge(msg.action_taken);
                       const stageBadge = getStageBadge(msg);
                       const consultantDispatch = msg.consultant_dispatch ?? null;
-                      const consultantModeBadge = getConsultantModeBadge(consultantDispatch);
+                      const consultantModePreference = normalizeConsultantModePreference(consultantDispatch?.mode);
                       const consultantRiskBadge = getConsultantRiskBadge(consultantDispatch);
-                      const consultantMissingFields = consultantDispatch?.missing_fields?.length
-                        ? consultantDispatch.missing_fields.join(" / ")
-                        : "";
+                      const consultantMissingFields = consultantDispatch?.missing_fields?.filter(Boolean) ?? [];
                       return (
                         <div key={msg.id} className="space-y-2">
                           <div className="flex justify-end">
@@ -846,7 +923,7 @@ export function FloatingQAChat() {
 
                           <div className="flex justify-start">
                             <div className="max-w-[85%] rounded-xl bg-slate-100 px-3 py-2 text-sm text-slate-900">
-                              <div className="whitespace-pre-wrap">{msg.answer}</div>
+                              <QAMarkdownContent content={msg.answer} />
                               <div className="mt-2 flex flex-wrap items-center gap-2 text-[11px] text-slate-600">
                                 <span>
                                   来源：{msg.sources?.length ? msg.sources.map((s) => s.label).join(" / ") : "无"}
@@ -858,28 +935,107 @@ export function FloatingQAChat() {
                                 {actionBadge ? <Badge variant={actionBadge.variant}>{actionBadge.label}</Badge> : null}
                                 {msg.signal_id ? <Badge variant="outline">signal #{msg.signal_id}</Badge> : null}
                                 {stageBadge ? <Badge variant="outline">{stageBadge}</Badge> : null}
-                                <Badge variant="secondary">{msg.thinking_mode === "deep" ? "深入思考" : "标准回答"}</Badge>
-                                {consultantModeBadge ? (
-                                  <Badge variant={consultantModeBadge.variant}>{consultantModeBadge.label}</Badge>
+                              </div>
+                              <div className="mt-3 rounded-2xl border border-slate-200/80 bg-slate-50/90 p-2">
+                                <div className="mb-2 text-[11px] font-medium text-slate-500">顾问动作</div>
+                                <div className="flex flex-wrap items-center gap-2">
+                                <Button
+                                  size="sm"
+                                  variant="ghost"
+                                  className={getActionChipClasses("deep", msg.thinking_mode === "deep")}
+                                  disabled={asking || actionBusyKey === `deep-${msg.id}`}
+                                  onClick={() => handleDeepRerun(msg)}
+                                >
+                                  {actionBusyKey === `deep-${msg.id}` ? "处理中..." : "深入思考"}
+                                </Button>
+                                {CONSULTANT_MODE_ACTIONS.map((item) => {
+                                  const busyKey = `mode-${msg.id}-${item.value}`;
+                                  return (
+                                  <Button
+                                    key={`${msg.id}-${item.value}`}
+                                    size="sm"
+                                    variant="ghost"
+                                    className={getActionChipClasses(item.tone, consultantModePreference === item.value)}
+                                    disabled={asking || actionBusyKey === busyKey}
+                                    onClick={() => handleConsultantModeRerun(msg, item.value)}
+                                  >
+                                    {actionBusyKey === busyKey ? "处理中..." : item.label}
+                                  </Button>
+                                  );
+                                })}
+                                {consultantRiskBadge && consultantDispatch ? (
+                                  <Button
+                                    size="sm"
+                                    variant="ghost"
+                                    className={getActionChipClasses(
+                                      consultantRiskBadge.label === "高风险"
+                                        ? "risk-high"
+                                        : consultantRiskBadge.label === "中风险"
+                                          ? "risk-medium"
+                                          : "risk-low",
+                                    )}
+                                    onClick={() => handleToggleRiskExplanation(msg.id)}
+                                  >
+                                    {consultantRiskBadge.label}
+                                  </Button>
                                 ) : null}
-                                {consultantRiskBadge ? (
-                                  <Badge variant={consultantRiskBadge.variant}>{consultantRiskBadge.label}</Badge>
+                                {consultantDispatch?.allow_execute ? (
+                                    <Button
+                                      size="sm"
+                                      variant="ghost"
+                                      className={getActionChipClasses("execute")}
+                                      disabled={Boolean(actionBusyKey)}
+                                      onClick={() => setExecuteDialogMessage(msg)}
+                                    >
+                                      允许执行
+                                    </Button>
                                 ) : null}
-                                {consultantDispatch ? (
-                                  <Badge variant={consultantDispatch.allow_execute ? "secondary" : "outline"}>
-                                    {consultantDispatch.allow_execute ? "允许执行" : "先别执行"}
-                                  </Badge>
+                                {consultantDispatch?.next_action ? (
+                                    <Button
+                                      size="sm"
+                                      variant="ghost"
+                                      className={getActionChipClasses("next")}
+                                      onClick={() => handleUseNextAction(consultantDispatch)}
+                                    >
+                                      先别执行
+                                    </Button>
                                 ) : null}
                                 </div>
+                              </div>
+                              {consultantDispatch && expandedRiskMessageId === msg.id ? (
+                                <div className="mt-3 rounded-md border border-amber-200 bg-amber-50/80 px-3 py-2 text-xs leading-6 text-amber-900">
+                                  {buildRiskExplanation(consultantDispatch)}
+                                </div>
+                              ) : null}
                               {consultantDispatch ? (
-                                <div className="mt-3 rounded-md border border-slate-200 bg-white/80 px-3 py-2 text-xs text-slate-700">
-                                  {consultantMissingFields ? (
-                                    <div>缺失信息：{consultantMissingFields}</div>
+                                <div className="mt-3 overflow-hidden rounded-xl border border-slate-200 bg-white/90">
+                                  {consultantMissingFields.length ? (
+                                    <div className="border-b border-slate-200/80 px-3 py-3">
+                                      <div className="text-[11px] font-medium uppercase tracking-[0.08em] text-slate-500">
+                                        缺失信息
+                                      </div>
+                                      <div className="mt-2 flex flex-wrap gap-1.5">
+                                        {consultantMissingFields.map((field) => (
+                                          <span
+                                            key={`${msg.id}-${field}`}
+                                            className="rounded-full border border-amber-200 bg-amber-50 px-2.5 py-1 text-[11px] text-amber-800"
+                                          >
+                                            {field}
+                                          </span>
+                                        ))}
+                                      </div>
+                                    </div>
                                   ) : null}
                                   {consultantDispatch.next_action ? (
-                                    <div className={consultantMissingFields ? "mt-2" : undefined}>
-                                      <span className="font-medium">下一步：</span>
-                                      {consultantDispatch.next_action}
+                                    <div className="px-3 py-3">
+                                      <div className="text-[11px] font-medium uppercase tracking-[0.08em] text-slate-500">
+                                        下一步建议
+                                      </div>
+                                      <QAMarkdownContent
+                                        content={consultantDispatch.next_action}
+                                        compact
+                                        className="mt-2 text-xs leading-6 text-slate-800"
+                                      />
                                     </div>
                                   ) : null}
                                 </div>
@@ -1046,6 +1202,7 @@ export function FloatingQAChat() {
                       />
                     </div>
                     <Textarea
+                      ref={questionInputRef}
                       rows={2}
                       placeholder="输入问题..."
                       value={question}
@@ -1091,6 +1248,28 @@ export function FloatingQAChat() {
           )}
         </div>
       </div>
+
+      <Dialog open={Boolean(executeDialogMessage)} onOpenChange={(openValue) => (!openValue ? setExecuteDialogMessage(null) : undefined)}>
+        <DialogContent className="sm:max-w-md">
+          <DialogHeader>
+            <DialogTitle>确认交接执行</DialogTitle>
+            <DialogDescription>
+              这一步会把当前问题交给 AI PMO Kernel，由统一语义规划和执行门禁继续处理，而不是继续停留在 RAG 问答链路里。
+            </DialogDescription>
+          </DialogHeader>
+          <div className="rounded-lg border bg-slate-50 px-3 py-3 text-sm leading-6 text-slate-700 [overflow-wrap:anywhere]">
+            {executeDialogMessage?.question}
+          </div>
+          <DialogFooter>
+            <Button variant="ghost" onClick={() => setExecuteDialogMessage(null)} disabled={actionBusyKey?.startsWith("execute-")}>
+              取消
+            </Button>
+            <Button onClick={handleExecuteViaKernel} disabled={actionBusyKey?.startsWith("execute-")}>
+              {actionBusyKey?.startsWith("execute-") ? "执行中..." : "继续执行"}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
     </>
   );
 }

frontend/src/components/ai-steward/QAMarkdownContent.tsx

@@ -0,0 +1,73 @@
+import ReactMarkdown from "react-markdown";
+import remarkBreaks from "remark-breaks";
+import remarkGfm from "remark-gfm";
+
+import { cn } from "@/lib/utils";
+
+type QAMarkdownContentProps = {
+  content: string;
+  className?: string;
+  compact?: boolean;
+};
+
+export function QAMarkdownContent({ content, className, compact = false }: QAMarkdownContentProps) {
+  const paragraphClassName = compact ? "mb-2 last:mb-0" : "mb-3 last:mb-0";
+  const listClassName = compact
+    ? "mb-2 ml-5 list-disc space-y-1 last:mb-0"
+    : "mb-3 ml-5 list-disc space-y-1 last:mb-0";
+  const orderedListClassName = compact
+    ? "mb-2 ml-5 list-decimal space-y-1 last:mb-0"
+    : "mb-3 ml-5 list-decimal space-y-1 last:mb-0";
+  const blockquoteClassName = compact
+    ? "mb-2 border-l-2 border-slate-300 bg-white/70 px-3 py-2 text-slate-700 last:mb-0"
+    : "mb-3 border-l-2 border-slate-300 bg-white/70 px-3 py-2 text-slate-700 last:mb-0";
+  const preClassName = compact
+    ? "mb-2 overflow-x-auto rounded-lg bg-slate-900 p-3 text-slate-100 last:mb-0"
+    : "mb-3 overflow-x-auto rounded-lg bg-slate-900 p-3 text-slate-100 last:mb-0";
+
+  return (
+    <div className={cn("text-sm leading-7 text-slate-900 [overflow-wrap:anywhere]", className)}>
+      <ReactMarkdown
+        remarkPlugins={[remarkGfm, remarkBreaks]}
+        skipHtml
+        components={{
+          p: ({ children }) => <p className={paragraphClassName}>{children}</p>,
+          ul: ({ children }) => <ul className={listClassName}>{children}</ul>,
+          ol: ({ children }) => <ol className={orderedListClassName}>{children}</ol>,
+          li: ({ children }) => <li className="[overflow-wrap:anywhere]">{children}</li>,
+          blockquote: ({ children }) => <blockquote className={blockquoteClassName}>{children}</blockquote>,
+          a: ({ children, href }) => (
+            <a
+              href={href}
+              target="_blank"
+              rel="noreferrer"
+              className="break-all text-slate-700 underline decoration-slate-300 underline-offset-2 hover:text-slate-950"
+            >
+              {children}
+            </a>
+          ),
+          code: ({ children, className: codeClassName }) => {
+            const isBlock = Boolean(codeClassName);
+            if (isBlock) {
+              return (
+                <code className={cn("block whitespace-pre-wrap break-words font-mono text-[12px] leading-6", codeClassName)}>
+                  {children}
+                </code>
+              );
+            }
+            return (
+              <code className="break-all rounded bg-slate-200/80 px-1.5 py-0.5 font-mono text-[12px] text-slate-900">
+                {children}
+              </code>
+            );
+          },
+          pre: ({ children }) => <pre className={preClassName}>{children}</pre>,
+          hr: () => <hr className="my-4 border-slate-200" />,
+          strong: ({ children }) => <strong className="font-semibold text-slate-950">{children}</strong>,
+        }}
+      >
+        {String(content || "").trim()}
+      </ReactMarkdown>
+    </div>
+  );
+}