制定双Agent能力提升与路由治理方案

2026-03-11 12:05:16 +08:00 · 2026-03-11 12:05:16 +08:00 · 2fd932e395
parent 331e9109e2
commit 2fd932e395
29 changed files with 540 additions and 304 deletions
--- a/.env.example
+++ b/.env.example
@ -1,3 +1,6 @@
+# 唯一配置文件，backend/.env 已移除
+# 所有环境变量在此统一管理
+
 # Docker Compose Runtime Secrets / Config
 POSTGRES_DB=pmflow
 POSTGRES_USER=pmflow
@ -8,37 +11,53 @@ DB_PORT=5435
 REDIS_PORT=6379
 BACKEND_PORT=8000
 FRONTEND_PORT=5174
+PMFLOW_EXPECTED_DB_DIALECT=postgresql

 # Backend Runtime
 CORS_ORIGINS=http://localhost:5174,http://127.0.0.1:5174
 LLM_PROVIDER=glm
 LLM_FALLBACK_PROVIDER=glm
-OPENAI_BASE_URL=https://api.edgefn.net/v1
-OPENAI_API_KEY=
-OPENAI_MODEL=MiniMax-M2.5
 OPENCLAW_ROUTING_ENABLED=false
-OPENCLAW_ROUTING_RULES_PATH=/app/backend/.openclaw/pmflow-routing-rules.json
+OPENCLAW_ROUTING_RULES_PATH=../.openclaw/pmflow-routing-rules.json
 OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS=120
+LLM_DIRECT_EMERGENCY_ENABLED=true
 LLM_MIXED_ROUTING_ENABLED=true
+LLM_ROUTING_DEFAULT_TIER=active
 LLM_LOW_TO_HIGH_ESCALATION_ENABLED=true
+LLM_REQUEST_TIMEOUT=90
+
+# 模型名统一使用裸名，不带供应商前缀
+# GLM-4.7 配 GLM_API_KEY
+# MiniMax-M2.5 配 OPENAI_API_KEY
+GLM_API_KEY=
+GLM_BASE_URL=https://api.edgefn.net/v1
+GLM_MODEL=GLM-4.7
+OPENAI_API_KEY=
+OPENAI_BASE_URL=https://api.edgefn.net/v1
+OPENAI_MODEL=MiniMax-M2.5
+
 LLM_ROUTING_LOW_PROVIDER=glm
 LLM_ROUTING_LOW_MODEL=GLM-4.7
 LLM_ROUTING_HIGH_PROVIDER=openai
 LLM_ROUTING_HIGH_MODEL=MiniMax-M2.5
 LLM_ROUTING_SAFETY_PROVIDER=glm
 LLM_ROUTING_SAFETY_MODEL=GLM-4.7
-LLM_DIRECT_EMERGENCY_ENABLED=true
+
+DEEPSEEK_API_KEY=
+DEEPSEEK_BASE_URL=https://api.deepseek.com/v1
+DEEPSEEK_MODEL=deepseek-chat
+
+OLLAMA_HOST=http://host.docker.internal:11434
+OLLAMA_BASE_URL=http://host.docker.internal:11434
+OLLAMA_MODEL=llama3:latest
+OLLAMA_CHAT_MODEL=llama3:8b
+OLLAMA_EMBEDDING_MODEL=nomic-embed-text
+
+# Adaptive RAG 深入思考模式
 RAG_DEEP_RETRIEVAL_STEPS=5
 RAG_DEEP_VECTOR_TOP_K=12
 RAG_DEEP_CONTEXT_BLOCKS=16
 RAG_DEEP_ANSWER_MAX_TOKENS=1400
-GLM_API_KEY=
-GLM_BASE_URL=https://api.edgefn.net/v1
-GLM_MODEL=GLM-4.7
-DEEPSEEK_API_KEY=
-OLLAMA_HOST=http://host.docker.internal:11434
-OLLAMA_BASE_URL=http://host.docker.internal:11434
-OLLAMA_MODEL=llama3:latest

 # Feishu long connection retry policy (1s -> 2s -> 4s, max 3 attempts)
 FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS=1
--- a/.openclaw/pmflow-routing-rules.json
+++ b/.openclaw/pmflow-routing-rules.json
@ -1,6 +1,6 @@
 {
  "version": "v2",
-  "updated_at": "2026-03-05T00:00:00Z",
+  "updated_at": "2026-03-11T00:00:00Z",
  "source_of_truth": "openclaw",
  "description": "PMFlow 三引擎统一路由：所有主链路统一经 OpenClaw OpenAI-Compatible 网关。",
  "rules": [
@ -9,8 +9,9 @@
      "enabled": true,
      "priority": 300,
      "tier": "low",
-      "provider": "openai",
-      "model": "zai/glm-4-flash",
+      "provider": "glm",
+      "model": "GLM-4.7",
+      "api_env": "GLM_API_KEY",
      "match_any": [
        "feishu_level_classification",
        "feishu_signal_extraction",
@ -26,7 +27,8 @@
      "priority": 220,
      "tier": "high",
      "provider": "openai",
-      "model": "minimax/MiniMax-M2.5",
+      "model": "MiniMax-M2.5",
+      "api_env": "OPENAI_API_KEY",
      "match_any": [
        "adaptive_rag_answer",
        "feishu_project_prefill",
@ -49,7 +51,8 @@
      "priority": 200,
      "tier": "high",
      "provider": "openai",
-      "model": "minimax/MiniMax-M2.5",
+      "model": "MiniMax-M2.5",
+      "api_env": "OPENAI_API_KEY",
      "match_any": [
        "analyze_food_calories",
        "feishu_new_project_brief",
@ -67,7 +70,8 @@
      "priority": 50,
      "tier": "safety",
      "provider": "openai",
-      "model": "deepseek/deepseek-chat",
+      "model": "deepseek-chat",
+      "api_env": "OPENAI_API_KEY",
      "match_all": true
    }
  ]
--- a/CLAUDE.md
+++ b/CLAUDE.md
@ -50,6 +50,9 @@ backend/
 ├── Dockerfile
 └── tests/
 ```
+- 新增：`backend/app/api/llm_providers.py`（AI 管家 LLM provider preset 配置读取接口）
+- 移除：`backend/.env.example`（配置模板统一到根目录 `.env.example`）
+- 移除：`backend/.openclaw/pmflow-routing-rules.json`（路由规则统一到根目录 `.openclaw/`）

 ## 5. 前端结构（`frontend/`）
 ```text
@ -140,6 +143,7 @@ uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

 ## 9. 新增代码的参考模式
 - 新增后端接口：参考 `backend/app/api/sstc.py` 的结构
+- 配置读取类接口：参考 `backend/app/api/llm_providers.py`（`GET /ai-steward/llm/providers`）
 - 新增前端 API 调用：参考 `frontend/src/api/sstc.ts`
 - 新增数据库表：参考 `alembic/versions/` 最新 migration
 - 新增飞书事件处理：参考 `backend/app/services/feishu_integration.py`
@ -156,3 +160,6 @@ uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
 - World Model 静态层已接入 Redis 缓存（key: `wm:static:{tenant_id}:{project_id}`，TTL 600s）
 - 写操作后主动 invalidate，动态层不缓存
 - AI PMO 意图识别已启用 LLM 主路径（`AI_PMO_INTENT_LLM_ENABLED`，默认开启）；当模型超时/不可用/结构化解析失败时自动降级到规则链路，不中断主流程
+- 环境变量以根目录 `.env` / `.env.example` 为唯一配置源；`backend/.env` 已移除且不再读取
+- OpenClaw 路由规则统一由根目录 `.openclaw/pmflow-routing-rules.json` 管理，容器通过 `OPENCLAW_ROUTING_RULES_PATH` 相对路径 `../.openclaw/pmflow-routing-rules.json` 读取
+- 模型命名统一使用裸模型名（如 `GLM-4.7`、`MiniMax-M2.5`），不使用供应商前缀模型名
--- a/DEPLOY_TENCENT_CLOUD.md
+++ b/DEPLOY_TENCENT_CLOUD.md
@ -120,11 +120,11 @@ source venv/bin/activate
 pip install -r requirements.txt

 # 配置环境变量
-cp .env.example .env
-nano .env
+cp /opt/neural-sync/.env.example /opt/neural-sync/.env
+nano /opt/neural-sync/.env
 ```

-编辑 `.env` 文件:
+编辑项目根目录 `.env` 文件:

 ```env
 # 数据库 - PostgreSQL
@ -227,7 +227,7 @@ Type=simple
 User=root
 WorkingDirectory=/opt/neural-sync/backend
 Environment="PATH=/opt/neural-sync/backend/venv/bin"
-EnvironmentFile=/opt/neural-sync/backend/.env
+EnvironmentFile=/opt/neural-sync/.env
 ExecStart=/opt/neural-sync/backend/venv/bin/uvicorn app.main:app --host 127.0.0.1 --port 8000
 Restart=always
 RestartSec=10
--- a/backend/.env.example
+++ b/backend/.env.example
@ -1,96 +0,0 @@
-# Database Configuration
-DATABASE_URL=postgresql+asyncpg://pmflow:pmflow_pass@db:5432/pmflow
-# 可选：启动方言守卫（postgresql / sqlite）。配置后若不一致将拒绝启动。
-PMFLOW_EXPECTED_DB_DIALECT=postgresql
-# For PostgreSQL with pgvector:
-# DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5435/neural_sync
-# For local SQLite quick-start:
-# DATABASE_URL=sqlite+aiosqlite:///./neural_sync.db
-
-# Redis Configuration
-REDIS_URL=redis://:pmflow_redis_pass@redis:6379/0
-
-# Knowledge Graph Incremental Scheduler
-KG_INCREMENTAL_INTERVAL_MINUTES=5
-KG_INCREMENTAL_DELTA_LIMIT=1000
-KG_INCREMENTAL_EVENT_BATCH=200
-KG_CODE_SCAN_INTERVAL_HOURS=6
-
-# ========================================
-# LLM / OpenClaw Control Plane
-# ========================================
-# 单一控制面：PMFlow 仅调用 OpenClaw 网关（OpenAI-Compatible）
-LLM_PROVIDER=openai
-OPENAI_BASE_URL=http://openclaw:18789/v1
-OPENAI_API_KEY=<OPENCLAW_GATEWAY_TOKEN>
-
-# OpenClaw 函数路由
-OPENCLAW_ROUTING_ENABLED=true
-OPENCLAW_ROUTING_RULES_PATH=/app/backend/.openclaw/pmflow-routing-rules.json
-# OpenClaw agent 默认超时（秒）：用于注入 chat/completions -> agents.defaults.timeoutSeconds
-OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS=120
-
-# 关闭 PMFlow 内部 mixed routing，优先以 OpenClaw 规则为准
-LLM_MIXED_ROUTING_ENABLED=false
-LLM_ROUTING_DEFAULT_TIER=active
-LLM_LOW_TO_HIGH_ESCALATION_ENABLED=true
-
-# low/high/safety tier 默认模型（由网关解释）
-LLM_ROUTING_LOW_PROVIDER=openai
-LLM_ROUTING_LOW_MODEL=zai/glm-4-flash
-LLM_ROUTING_HIGH_PROVIDER=openai
-LLM_ROUTING_HIGH_MODEL=minimax/MiniMax-M2.5
-LLM_ROUTING_SAFETY_PROVIDER=openai
-LLM_ROUTING_SAFETY_MODEL=deepseek/deepseek-chat
-
-# 应急开关：仅在网关不可用时允许直连 provider（默认关闭）
-LLM_DIRECT_EMERGENCY_ENABLED=false
-LLM_FALLBACK_PROVIDER=ollama
-
-# Adaptive RAG 深入思考模式（前端开关=deep 时生效）
-RAG_DEEP_RETRIEVAL_STEPS=5
-RAG_DEEP_VECTOR_TOP_K=12
-RAG_DEEP_CONTEXT_BLOCKS=16
-RAG_DEEP_ANSWER_MAX_TOKENS=1400
-
-# --- Emergency Direct Provider Keys (optional) ---
-GLM_API_KEY=
-GLM_BASE_URL=https://open.bigmodel.cn/api/paas/v4
-GLM_MODEL=glm-4-flash
-
-DEEPSEEK_API_KEY=
-DEEPSEEK_BASE_URL=https://api.deepseek.com/v1
-
-# --- Ollama 本地配置（应急本地兜底） ---
-OLLAMA_HOST=http://127.0.0.1:11434
-OLLAMA_CHAT_MODEL=llama3:8b
-OLLAMA_EMBEDDING_MODEL=nomic-embed-text
-OLLAMA_BASE_URL=http://127.0.0.1:11434
-OLLAMA_MODEL=llama3:latest
-
-# ========================================
-
-# Multi-Model Voting Configuration (仅 Ollama 支持)
-# For high-impact signals (impact_score >= VOTING_THRESHOLD), multiple models will vote
-# Available models: llama3, llama3.2, qwen2.5, mistral, deepseek-coder, etc.
-VOTING_MODELS=llama3,qwen2.5,mistral
-VOTING_THRESHOLD=0.8
-
-# Business Context for AI Strategic Analysis
-# Customize this to match your actual product/team
-# This context is used by the AI to evaluate external signals
-BUSINESS_CONTEXT=Pixelle-Video: An AI-powered automated short video generation platform. Core features: topic-to-video automation (script generation → AI images/videos → TTS → BGM → final video). Tech stack: Python, Streamlit, FastAPI, ComfyUI workflows. AI models: FLUX/Qwen/SDXL for images, WAN 2.1/2.2 for videos, Edge TTS/Index TTS for voice. Supports multiple pipelines: standard generation, custom assets, digital human, image-to-video, action transfer.
-
-# Plane PM Tool Integration (optional, simulation mode if not set)
-PLANE_API_URL=https://api.plane.so/api/v1
-PLANE_API_KEY=
-PLANE_WORKSPACE_SLUG=
-PLANE_PROJECT_ID=
-
-# Lark/Feishu Bot Integration (optional)
-LARK_APP_ID=
-LARK_APP_SECRET=
-LARK_VERIFICATION_TOKEN=
-# 飞书长连接重试策略（指数退避 1s -> 2s -> 4s，最多 3 次）
-FEISHU_LONG_CONNECTION_RETRY_BASE_SECONDS=1
-FEISHU_LONG_CONNECTION_RETRY_MAX_ATTEMPTS=3
--- a/backend/.openclaw/pmflow-routing-rules.json
+++ b/backend/.openclaw/pmflow-routing-rules.json
@ -1,74 +0,0 @@
-{
-  "version": "v2",
-  "updated_at": "2026-03-05T00:00:00Z",
-  "source_of_truth": "openclaw",
-  "description": "PMFlow 三引擎统一路由：所有主链路统一经 OpenClaw OpenAI-Compatible 网关。",
-  "rules": [
-    {
-      "name": "engine_low_execute_glm_flash",
-      "enabled": true,
-      "priority": 300,
-      "tier": "low",
-      "provider": "glm",
-      "model": "GLM-4.7",
-      "match_any": [
-        "feishu_level_classification",
-        "feishu_signal_extraction",
-        "rag_intent_classification",
-        "classify_intent",
-        "feishu_humanized_reply",
-        "extract_context_insights"
-      ]
-    },
-    {
-      "name": "engine_high_reasoning_minimax",
-      "enabled": true,
-      "priority": 220,
-      "tier": "high",
-      "provider": "openai",
-      "model": "MiniMax-M2.5",
-      "match_any": [
-        "adaptive_rag_answer",
-        "feishu_project_prefill",
-        "feishu_task_orchestration",
-        "generate_task_breakdown",
-        "generate_charter_from_signal",
-        "decompose_tasks_from_charter",
-        "recommend_assignees_for_task",
-        "evaluate_model_requirement",
-        "calculate_alignment_score",
-        "detect_cognitive_conflicts",
-        "rag_db_project_semantic_retrieval",
-        "workload_optimize_prompt",
-        "analyze_cost_roi"
-      ]
-    },
-    {
-      "name": "engine_high_extended_reasoning",
-      "enabled": true,
-      "priority": 200,
-      "tier": "high",
-      "provider": "openai",
-      "model": "MiniMax-M2.5",
-      "match_any": [
-        "analyze_food_calories",
-        "feishu_new_project_brief",
-        "feishu_project_signal_classification",
-        "sstc_ai_breakdown",
-        "sstc_charter_localize_zh",
-        "sstc_charter_prefill",
-        "skill_score_update",
-        "test_engine_connection"
-      ]
-    },
-    {
-      "name": "engine_safety_fallback",
-      "enabled": true,
-      "priority": 50,
-      "tier": "safety",
-      "provider": "openai",
-      "model": "deepseek/deepseek-chat",
-      "match_all": true
-    }
-  ]
-}
--- a/backend/Dockerfile
+++ b/backend/Dockerfile
@ -13,6 +13,7 @@ COPY backend/requirements.txt /tmp/requirements.txt
 RUN pip install --no-cache-dir -r /tmp/requirements.txt

 COPY backend /app/backend
+COPY .openclaw /app/.openclaw

 EXPOSE 8000

--- a/backend/app/api/init.py
+++ b/backend/app/api/init.py
@ -22,6 +22,7 @@ from .budget import router as budget_router
 from .scheduler import router as scheduler_router
 from .notifications import router as notifications_router
 from .llm_engines import router as llm_engines_router
+from .llm_providers import router as llm_providers_router
 from .prompt_profiles import router as prompt_profiles_router
 from .knowledge_graph import router as knowledge_graph_router
 from .operation_catalog import router as operation_catalog_router
@ -46,6 +47,7 @@ __all__ = [
    "scheduler_router",
    "notifications_router",
    "llm_engines_router",
+    "llm_providers_router",
    "prompt_profiles_router",
    "knowledge_graph_router",
    "operation_catalog_router",
--- a/backend/app/api/llm_providers.py
+++ b/backend/app/api/llm_providers.py
@ -0,0 +1,58 @@
+"""LLM provider preset API."""
+
+from __future__ import annotations
+
+import os
+from typing import Any
+
+from fastapi import APIRouter, Depends
+
+from ..core.response import success_response
+from ..dependencies.auth import get_current_user
+from ..models import User
+
+router = APIRouter(prefix="/ai-steward/llm", tags=["AI Steward"])
+
+
+def _build_provider_payload() -> dict[str, Any]:
+    low_model = str(os.getenv("LLM_ROUTING_LOW_MODEL", "GLM-4.7") or "GLM-4.7").strip()
+    high_model = str(os.getenv("LLM_ROUTING_HIGH_MODEL", "MiniMax-M2.5") or "MiniMax-M2.5").strip()
+    base_url = str(os.getenv("OPENAI_BASE_URL", "https://api.edgefn.net/v1") or "https://api.edgefn.net/v1").strip()
+    ollama_base_url = str(
+        os.getenv("OLLAMA_BASE_URL", os.getenv("OLLAMA_HOST", "http://127.0.0.1:11434")) or "http://127.0.0.1:11434"
+    ).strip()
+    return {
+        "low": low_model,
+        "high": high_model,
+        "base_url": base_url,
+        "providers": {
+            "glm": {
+                "model": str(os.getenv("GLM_MODEL", low_model) or low_model).strip(),
+                "base_url": str(os.getenv("GLM_BASE_URL", base_url) or base_url).strip(),
+                "api_env": "GLM_API_KEY",
+            },
+            "openai": {
+                "model": str(os.getenv("OPENAI_MODEL", high_model) or high_model).strip(),
+                "base_url": base_url,
+                "api_env": "OPENAI_API_KEY",
+            },
+            "deepseek": {
+                "model": str(os.getenv("DEEPSEEK_MODEL", "deepseek-chat") or "deepseek-chat").strip(),
+                "base_url": str(os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com/v1") or "").strip(),
+                "api_env": "DEEPSEEK_API_KEY",
+            },
+            "ollama": {
+                "model": str(os.getenv("OLLAMA_MODEL", "llama3:latest") or "llama3:latest").strip(),
+                "base_url": ollama_base_url,
+                "api_env": "",
+            },
+        },
+    }
+
+
+@router.get("/providers")
+async def get_llm_provider_presets(
+    current_user: User = Depends(get_current_user),
+):
+    _ = current_user
+    return success_response(_build_provider_payload())
--- a/backend/app/api/router.py
+++ b/backend/app/api/router.py
@ -25,6 +25,7 @@ from .feishu import router as feishu_router
 from .wecom import router as wecom_router
 from .notifications import router as notifications_router
 from .llm_engines import router as llm_engines_router
+from .llm_providers import router as llm_providers_router
 from .ai_logs import router as ai_logs_router
 from .prompt_profiles import router as prompt_profiles_router
 from .rag import router as rag_router
@ -57,6 +58,7 @@ api_router.include_router(feishu_router)
 api_router.include_router(wecom_router)
 api_router.include_router(notifications_router)
 api_router.include_router(llm_engines_router)
+api_router.include_router(llm_providers_router)
 api_router.include_router(ai_logs_router)
 api_router.include_router(prompt_profiles_router)
 api_router.include_router(rag_router)
--- a/backend/app/config/ai_prompts.py
+++ b/backend/app/config/ai_prompts.py
@ -525,7 +525,8 @@ PROMPT_AI_PMO_INTENT_CLASSIFICATION = """
 3. project_id 无法确定时返回 null，不要猜测无依据项目。
 4. 噪音/寒暄/无关消息返回 ignore。
 5. 禁止输出 thinking、analysis、解释句、Markdown、代码块。
-6. 只输出 JSON，不要额外文本。
+6. 若消息表达“新增/增加/添加/新建/创建/建立 + 项目”（如“给我新增个项目”），优先判定 project.create；仅当明确是“信号源/来源/webhook/rss/github 接入或创建信号”时才判 signal_source.onboard 或 signal.create。
+7. 只输出 JSON，不要额外文本。
 """


--- a/backend/app/core/config.py
+++ b/backend/app/core/config.py
@ -9,10 +9,9 @@ from pathlib import Path
 from typing import Optional
 from dotenv import load_dotenv

-# 统一从 backend/.env 加载配置，避免因启动目录不同读到错误环境
+# 仅从项目根目录 .env 加载，作为唯一配置源
 BACKEND_DIR = Path(__file__).resolve().parents[2]
 PROJECT_ROOT = BACKEND_DIR.parent
-load_dotenv(BACKEND_DIR / ".env", override=False)
 load_dotenv(PROJECT_ROOT / ".env", override=False)

 # 默认数据库固定到 backend/neural_sync.db，避免相对路径导致多库混用
--- a/backend/app/services/ai_pmo_kernel_service.py
+++ b/backend/app/services/ai_pmo_kernel_service.py
@ -139,7 +139,24 @@ FOLLOWUP_CONFIRM_HINTS = ("确认归属", "确认项目", "确认到项目", "
 FOLLOWUP_CONTINUE_HINTS = ("继续处理", "继续执行", "接着处理", "往下做", "继续推进")
 FOLLOWUP_WORK_ITEM_ID_RE = re.compile(r"(?:work[_ -]?item|工作项|队列项)\s*(?:id|ID)?\s*[:#：]?\s*(\d+)", re.IGNORECASE)
 SIGNAL_CREATE_HINTS = ("新增一个信号", "新增一条信号", "增加一个信号", "录入一个信号", "关注", "跟踪")
-PROJECT_CREATE_HINTS = ("创建项目", "新建项目", "建立项目", "做个项目", "立项", "开个项目")
+PROJECT_CREATE_HINTS = (
+    "创建项目",
+    "新建项目",
+    "建立项目",
+    "做个项目",
+    "立项",
+    "开个项目",
+    "新增项目",
+    "新增个项目",
+    "新增一个项目",
+    "增加项目",
+    "增加个项目",
+    "增加一个项目",
+    "添加项目",
+    "添加一个项目",
+    "加个项目",
+    "加一个项目",
+)
 POLICY_GATE_RULES: tuple[tuple[str, tuple[str, ...]], ...] = (
    ("delete_or_archive", ("删除项目", "删除信号", "归档项目", "归档", "archive", "清空项目")),
    ("budget_overrun", ("超预算", "突破预算", "额外预算", "加预算")),
@ -268,7 +285,10 @@ class IntentResolver:
                pretty[:300],
                {"source_request": True},
            )
-        if any(token in pretty for token in PROJECT_CREATE_HINTS) or OpenClawCommanderService._extract_project_create_title(pretty):
+        if (
+            IntentResolver._looks_like_project_create_request(text=pretty)
+            or OpenClawCommanderService._extract_project_create_title(pretty)
+        ):
            return CanonicalIntent(
                INTENT_PROJECT_CREATE,
                0.9,
@ -321,6 +341,21 @@ class IntentResolver:
        has_operation = any(token in normalized for token in IntentResolver.OPERATION_VERBS)
        return not has_operation

+    @staticmethod
+    def _looks_like_project_create_request(*, text: str) -> bool:
+        normalized = re.sub(r"\s+", "", str(text or "")).lower()
+        if not normalized:
+            return False
+        if any(token in normalized for token in PROJECT_CREATE_HINTS):
+            return True
+        return bool(
+            re.search(
+                r"(?:新增|增加|添加|新建|创建|建立|开个|开一个|建个|建一个|做个|做一个|加个|加一个|立项)(?:一个|个)?项目",
+                normalized,
+                re.IGNORECASE,
+            )
+        )
+
    @staticmethod
    def _build_signal_title(text: str, *, default: str) -> str:
        cleaned = re.sub(r"\s+", " ", str(text or "")).strip(" ，,:：")
--- a/backend/app/services/ai_service.py
+++ b/backend/app/services/ai_service.py
@ -2,7 +2,7 @@
 PMFlow AI 核心能力封装

 技术栈：LangChain 0.3+、OpenAI SDK、Redis
- 封装 GPT-4o mini 调用（流式/非流式）
+- 封装统一 LLM 路由入口调用（流式/非流式）
 - Prompt 集中在 config/ai_prompts
 - AI 调用日志写入 ai_call_logs
 - Redis 结果缓存（1 小时 TTL）
@ -39,6 +39,9 @@ INPUT_PRICE_PER_1K = 0.00015
 OUTPUT_PRICE_PER_1K = 0.0006
 CACHE_TTL_SECONDS = 3600  # 1 小时
 PROMPT_PROFILE_CACHE_TTL_SECONDS = 60
+DEFAULT_OPENAI_BASE_URL = "https://api.edgefn.net/v1"
+DEFAULT_OPENAI_MODEL = "MiniMax-M2.5"
+DEFAULT_GLM_MODEL = "GLM-4.7"

 from ..config.ai_prompts import build_default_role_prompt_bundle
 from ..database import AsyncSessionLocal
@ -73,27 +76,95 @@ class AIService:
    def __init__(
        self,
        api_key: Optional[str] = None,
-        model: str = "gpt-4o-mini",
+        model: Optional[str] = None,
+        base_url: Optional[str] = None,
        redis_url: Optional[str] = None,
        max_retries: int = 3,
    ):
-        self.api_key = api_key or os.getenv("OPENAI_API_KEY", "")
-        self.model = model
+        runtime = self._resolve_llm_runtime(
+            api_key=api_key,
+            model=model,
+            base_url=base_url,
+        )
+        self.api_key = runtime["api_key"]
+        self.model = runtime["model"]
+        self.base_url = runtime["base_url"]
        self.redis_url = redis_url or os.getenv("REDIS_URL")
        self.max_retries = max_retries
        self._llm: Optional[BaseChatModel] = None
        self._redis: Optional[Any] = None
        self._prompt_profile_cache: dict[int, tuple[float, tuple[str, str]]] = {}

+    @staticmethod
+    def _is_env_true(name: str, default: bool) -> bool:
+        raw = str(os.getenv(name, str(default)) or "").strip().lower()
+        if not raw:
+            return default
+        return raw in {"1", "true", "yes", "y", "on", "是"}
+
+    def _resolve_llm_runtime(
+        self,
+        *,
+        api_key: Optional[str],
+        model: Optional[str],
+        base_url: Optional[str],
+    ) -> dict[str, str]:
+        routing_enabled = self._is_env_true("OPENCLAW_ROUTING_ENABLED", True)
+        direct_emergency_enabled = self._is_env_true("LLM_DIRECT_EMERGENCY_ENABLED", False)
+        if routing_enabled and not direct_emergency_enabled:
+            return {
+                "api_key": str(api_key or os.getenv("OPENAI_API_KEY", "") or "").strip(),
+                "model": str(
+                    model
+                    or os.getenv("LLM_ROUTING_HIGH_MODEL")
+                    or os.getenv("OPENAI_MODEL")
+                    or DEFAULT_OPENAI_MODEL
+                    or ""
+                ).strip(),
+                "base_url": str(base_url or os.getenv("OPENAI_BASE_URL", DEFAULT_OPENAI_BASE_URL) or "").strip(),
+            }
+
+        provider = str(os.getenv("LLM_PROVIDER", "openai") or "openai").strip().lower()
+        if provider == "glm":
+            return {
+                "api_key": str(api_key or os.getenv("GLM_API_KEY", "") or "").strip(),
+                "model": str(model or os.getenv("GLM_MODEL", DEFAULT_GLM_MODEL) or "").strip(),
+                "base_url": str(base_url or os.getenv("GLM_BASE_URL", DEFAULT_OPENAI_BASE_URL) or "").strip(),
+            }
+        if provider == "deepseek":
+            return {
+                "api_key": str(api_key or os.getenv("DEEPSEEK_API_KEY", "") or "").strip(),
+                "model": str(model or os.getenv("DEEPSEEK_MODEL", "deepseek-chat") or "").strip(),
+                "base_url": str(base_url or os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com/v1") or "").strip(),
+            }
+        if provider == "ollama":
+            return {
+                "api_key": str(api_key or "").strip(),
+                "model": str(model or os.getenv("OLLAMA_MODEL", "llama3:latest") or "").strip(),
+                "base_url": str(
+                    base_url
+                    or os.getenv("OLLAMA_BASE_URL")
+                    or os.getenv("OLLAMA_HOST", "http://127.0.0.1:11434")
+                    or ""
+                ).strip(),
+            }
+
+        return {
+            "api_key": str(api_key or os.getenv("OPENAI_API_KEY", "") or "").strip(),
+            "model": str(model or os.getenv("OPENAI_MODEL", DEFAULT_OPENAI_MODEL) or "").strip(),
+            "base_url": str(base_url or os.getenv("OPENAI_BASE_URL", DEFAULT_OPENAI_BASE_URL) or "").strip(),
+        }
+
    def _get_llm(self) -> BaseChatModel:
        if not LANGCHAIN_AVAILABLE:
            raise RuntimeError("请安装 langchain-openai: pip install langchain-openai langchain-core")
        if not self.api_key:
-            raise RuntimeError("请设置 OPENAI_API_KEY 环境变量")
+            raise RuntimeError("请设置统一路由所需 API Key 环境变量")
        if self._llm is None:
            self._llm = ChatOpenAI(
                model=self.model,
                api_key=self.api_key,
+                base_url=self.base_url,
                temperature=0.3,
            )
        return self._llm
--- a/backend/app/services/llm.py
+++ b/backend/app/services/llm.py
@ -43,9 +43,9 @@ logger = logging.getLogger(__name__)

 SUPPORTED_PROVIDERS = {"deepseek", "glm", "openai", "ollama"}
 OPENCLAW_PRIMARY_GATEWAY_URL = "http://openclaw:18789/v1"
-OPENCLAW_DEFAULT_LOW_MODEL = "zai/glm-4-flash"
-OPENCLAW_DEFAULT_HIGH_MODEL = "minimax/MiniMax-M2.5"
-OPENCLAW_DEFAULT_SAFETY_MODEL = "deepseek/deepseek-chat"
+OPENCLAW_DEFAULT_LOW_MODEL = "GLM-4.7"
+OPENCLAW_DEFAULT_HIGH_MODEL = "MiniMax-M2.5"
+OPENCLAW_DEFAULT_SAFETY_MODEL = "deepseek-chat"
 PROVIDER_DEFAULTS: dict[str, dict[str, str]] = {
    "deepseek": {
        "name": "DeepSeek",
@ -56,7 +56,7 @@ PROVIDER_DEFAULTS: dict[str, dict[str, str]] = {
    "glm": {
        "name": "GLM",
        "base_url": "https://open.bigmodel.cn/api/paas/v4",
-        "model": "glm-4-flash",
+        "model": "GLM-4.7",
        "api_env": "GLM_API_KEY",
    },
    "openai": {
@ -81,8 +81,7 @@ PROVIDER_ALIASES = {
 PROMPT_PROFILE_CACHE_TTL_SECONDS = 60.0
 BOOL_TRUE_VALUES = {"1", "true", "yes", "y", "on", "是"}
 OPENCLAW_ROUTING_CACHE_TTL_SECONDS = 30.0
-OPENCLAW_BACKEND_RULES_PATH = Path(__file__).resolve().parents[2] / ".openclaw" / "pmflow-routing-rules.json"
-OPENCLAW_PROJECT_RULES_PATH = Path(__file__).resolve().parents[3] / ".openclaw" / "pmflow-routing-rules.json"
+OPENCLAW_DEFAULT_RULES_RELATIVE_PATH = "../.openclaw/pmflow-routing-rules.json"
 OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS = 120
 OPENCLAW_AGENTS_TIMEOUT_SECONDS_MAX = 120
 DEFAULT_TIER_MODEL_OVERRIDES = {
@ -1456,14 +1455,18 @@ class LLMService:
        return runtime

    def _openclaw_rules_path(self) -> str:
-        custom = str(os.getenv("OPENCLAW_ROUTING_RULES_PATH", "") or "").strip()
-        if custom:
-            return custom
-        if OPENCLAW_BACKEND_RULES_PATH.exists():
-            return str(OPENCLAW_BACKEND_RULES_PATH)
-        if OPENCLAW_PROJECT_RULES_PATH.exists():
-            return str(OPENCLAW_PROJECT_RULES_PATH)
-        return str(OPENCLAW_BACKEND_RULES_PATH)
+        custom = str(
+            os.getenv(
+                "OPENCLAW_ROUTING_RULES_PATH",
+                OPENCLAW_DEFAULT_RULES_RELATIVE_PATH,
+            )
+            or ""
+        ).strip()
+        raw_path = custom or OPENCLAW_DEFAULT_RULES_RELATIVE_PATH
+        candidate = Path(raw_path)
+        if not candidate.is_absolute():
+            candidate = (Path(__file__).resolve().parents[2] / candidate).resolve()
+        return str(candidate)

    def _load_openclaw_routing_bundle(self) -> dict[str, Any]:
        now = time.time()
@ -1580,6 +1583,7 @@ class LLMService:
        engine_id = str(rule.get("engine_id") or "").strip()
        model_override = str(rule.get("model") or "").strip()
        base_url_override = str(rule.get("base_url") or "").strip().rstrip("/")
+        api_env_override = str(rule.get("api_env") or "").strip()

        routed: Optional[dict[str, Any]] = None
        if engine_id:
@ -1618,6 +1622,9 @@ class LLMService:
            routed["model"] = model_override
        if base_url_override:
            routed["base_url"] = base_url_override
+        if api_env_override:
+            routed["api_key"] = str(os.getenv(api_env_override, "") or "").strip()
+            routed["api_env"] = api_env_override
        return routed

    def _resolve_function_tier(self, function_name: str) -> str:
--- a/backend/app/services/llm_client.py
+++ b/backend/app/services/llm_client.py
@ -34,8 +34,8 @@ class LLMClient:
        elif self.provider == "openai":
            return OpenAICompatibleClient(
                api_key=os.getenv("OPENAI_API_KEY"),
-                base_url=os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1"),
-                model=os.getenv("OPENAI_MODEL", "gpt-4o-mini")
+                base_url=os.getenv("OPENAI_BASE_URL", "https://api.edgefn.net/v1"),
+                model=os.getenv("OPENAI_MODEL", "MiniMax-M2.5")
            )
        elif self.provider == "qwen":
            return OpenAICompatibleClient(
--- a/backend/app/services/openclaw_commander_service.py
+++ b/backend/app/services/openclaw_commander_service.py
@ -819,8 +819,8 @@ class OpenClawCommanderService:
        if quoted_match:
            return OpenClawCommanderService._normalize_project_title_candidate(quoted_match.group(1))
        for pattern in (
-            r"(?:帮我|请|麻烦)?(?:创建|新建|建立|开个|开一个|建个|建一个|做个|做一个)(?:一个|个)?项目(?:叫|名为|名称是|名称为|是|为)?[：:，,\s]*(.+)",
-            r"(?:创建|新建|建立)项目[：:，,\s]*(.+)",
+            r"(?:帮我|请|麻烦|给我)?(?:新增|增加|添加|创建|新建|建立|开个|开一个|建个|建一个|做个|做一个|加个|加一个)(?:一个|个)?项目(?:叫|名为|名称是|名称为|是|为)?[：:，,\s]*(.+)",
+            r"(?:新增|增加|添加|创建|新建|建立)项目[：:，,\s]*(.+)",
            r"项目(?:叫|名为|名称是|名称为)[：:，,\s]*(.+)",
        ):
            match = re.search(pattern, cleaned, re.IGNORECASE)
@ -840,7 +840,7 @@ class OpenClawCommanderService:
        if title:
            cleaned = cleaned.replace(title, "", 1).strip(" ，,:：")
        cleaned = re.sub(
-            r"^(?:帮我|请|麻烦)?(?:创建|新建|建立|开个|开一个|建个|建一个|做个|做一个)(?:一个|个)?项目(?:叫|名为|名称是|名称为|是|为)?",
+            r"^(?:帮我|请|麻烦|给我)?(?:新增|增加|添加|创建|新建|建立|开个|开一个|建个|建一个|做个|做一个|加个|加一个)(?:一个|个)?项目(?:叫|名为|名称是|名称为|是|为)?",
            "",
            cleaned,
            flags=re.IGNORECASE,
--- a/backend/tests/test_ai_pmo_kernel_service.py
+++ b/backend/tests/test_ai_pmo_kernel_service.py
@ -25,6 +25,7 @@ from app.models import (
 from app.services.ai_pmo_cadence_service import AIPMOCadenceService
 from app.services.ai_pmo_kernel_service import (
    AIPMOKernelService,
+    INTENT_PROJECT_CREATE,
    INTENT_SIGNAL_SOURCE_ONBOARD,
    IntentResolver,
 )
@ -183,6 +184,13 @@ class AIPMOKernelServiceTest(unittest.IsolatedAsyncioTestCase):
        )
        self.assertEqual(intent.name, INTENT_SIGNAL_SOURCE_ONBOARD)

+    def test_intent_resolver_detects_project_create_for_add_project_phrase(self) -> None:
+        intent = IntentResolver.resolve(
+            text="给我新增个项目，openclaw自动化部署。",
+            project_id=None,
+        )
+        self.assertEqual(intent.name, INTENT_PROJECT_CREATE)
+
    async def test_build_world_model_includes_budget_capacity_and_skill_graph(self) -> None:
        async with self.session_factory() as db:
            service = AIPMOKernelService(db, tenant_id=1)
@ -375,7 +383,7 @@ class AIPMOKernelServiceTest(unittest.IsolatedAsyncioTestCase):
                            },
                            source_field="content",
                            provider="openai",
-                            model="zai/glm-4.7",
+                            model="GLM-4.7",
                            raw_text='{"intent":"signal_source.onboard"}',
                            validation="ok",
                        )
@ -422,7 +430,7 @@ class AIPMOKernelServiceTest(unittest.IsolatedAsyncioTestCase):
                            payload=None,
                            source_field="",
                            provider="openai",
-                            model="zai/glm-4.7",
+                            model="GLM-4.7",
                            error_code="model_not_allowed",
                            raw_text="403 ModelNotAllowed",
                            validation="invalid",
@ -509,6 +517,43 @@ class AIPMOKernelServiceTest(unittest.IsolatedAsyncioTestCase):
        self.assertTrue(result["direct_reply_required"])
        self.assertEqual(work_item.project_id, 9)

+    async def test_process_add_project_phrase_routes_to_project_create(self) -> None:
+        async with self.session_factory() as db:
+            service = AIPMOKernelService(db, tenant_id=1)
+            with patch(
+                "app.services.ai_pmo_kernel_service.AIStewardService.create_project_from_natural_language",
+                AsyncMock(
+                    return_value={
+                        "project": {"id": 12, "title": "OpenClaw 自动化部署", "name": "OpenClaw 自动化部署"},
+                        "signal": {"id": 31},
+                        "tasks": [],
+                    }
+                ),
+            ):
+                result = await service.process_ingress(
+                    source_system="pmflow_manual",
+                    source_message_id="project-phrase-1",
+                    message_text="给我新增个项目，openclaw自动化部署。",
+                    sender="老板",
+                    channel_ref="pmflow",
+                )
+            work_item = (
+                await db.execute(
+                    select(WorkItem).where(
+                        and_(
+                            WorkItem.tenant_id == 1,
+                            WorkItem.source_message_id == "project-phrase-1",
+                        )
+                    )
+                )
+            ).scalar_one()
+
+        self.assertEqual(result["status"], "captured")
+        self.assertEqual(result["created_project_id"], 12)
+        self.assertEqual(result["linked_project_ids"], [12])
+        self.assertEqual(work_item.intent, INTENT_PROJECT_CREATE)
+        self.assertEqual(work_item.project_id, 12)
+
    async def test_signal_source_request_hits_simulation_policy_gate_when_skill_gap_exists(self) -> None:
        async with self.session_factory() as db:
            service = AIPMOKernelService(db, tenant_id=1)
--- a/backend/tests/test_llm_providers_api.py
+++ b/backend/tests/test_llm_providers_api.py
@ -0,0 +1,40 @@
+import asyncio
+
+from app.api.llm_providers import _build_provider_payload, get_llm_provider_presets
+
+
+def test_build_provider_payload_reads_env(monkeypatch) -> None:
+    monkeypatch.setenv("LLM_ROUTING_LOW_MODEL", "GLM-4.7")
+    monkeypatch.setenv("LLM_ROUTING_HIGH_MODEL", "MiniMax-M2.5")
+    monkeypatch.setenv("OPENAI_BASE_URL", "https://api.edgefn.net/v1")
+    monkeypatch.setenv("GLM_BASE_URL", "https://api.edgefn.net/v1")
+    monkeypatch.setenv("GLM_MODEL", "GLM-4.7")
+    monkeypatch.setenv("OPENAI_MODEL", "MiniMax-M2.5")
+    monkeypatch.setenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com/v1")
+    monkeypatch.setenv("DEEPSEEK_MODEL", "deepseek-chat")
+    monkeypatch.setenv("OLLAMA_BASE_URL", "http://host.docker.internal:11434")
+    monkeypatch.setenv("OLLAMA_MODEL", "llama3:latest")
+
+    payload = _build_provider_payload()
+
+    assert payload["low"] == "GLM-4.7"
+    assert payload["high"] == "MiniMax-M2.5"
+    assert payload["base_url"] == "https://api.edgefn.net/v1"
+    assert payload["providers"]["glm"]["api_env"] == "GLM_API_KEY"
+    assert payload["providers"]["openai"]["api_env"] == "OPENAI_API_KEY"
+    assert payload["providers"]["glm"]["model"] == "GLM-4.7"
+    assert payload["providers"]["openai"]["model"] == "MiniMax-M2.5"
+
+
+def test_get_llm_provider_presets_returns_envelope(monkeypatch) -> None:
+    monkeypatch.setenv("LLM_ROUTING_LOW_MODEL", "GLM-4.7")
+    monkeypatch.setenv("LLM_ROUTING_HIGH_MODEL", "MiniMax-M2.5")
+    monkeypatch.setenv("OPENAI_BASE_URL", "https://api.edgefn.net/v1")
+
+    result = asyncio.run(get_llm_provider_presets(current_user=object()))
+
+    assert result["code"] == 0
+    assert result["message"] == "success"
+    assert isinstance(result["data"], dict)
+    assert result["data"]["low"] == "GLM-4.7"
+    assert result["data"]["high"] == "MiniMax-M2.5"
--- a/backend/tests/test_openclaw_commander_service.py
+++ b/backend/tests/test_openclaw_commander_service.py
@ -373,6 +373,23 @@ class OpenClawCommanderServiceTest(unittest.TestCase):
        self.assertEqual(plan["operations"][0]["operation_code"], "project.create")
        self.assertEqual(plan["operations"][0]["payload"]["title"], "财务冲账协同")

+    def test_project_create_fallback_supports_add_project_phrase(self) -> None:
+        commander = _ProjectFallbackCommander()
+
+        plan = self.async_run(
+            commander._build_fallback_plan(
+                tenant_id=1,
+                project_id=None,
+                instruction="给我新增个项目，openclaw自动化部署。",
+                max_operations=4,
+                error="structured_http_error",
+                project_directory=None,
+            )
+        )
+
+        self.assertEqual(plan["operations"][0]["operation_code"], "project.create")
+        self.assertEqual(plan["operations"][0]["payload"]["title"], "openclaw自动化部署")
+
    def test_project_status_fallback_resolves_project_id_from_directory(self) -> None:
        commander = _ProjectFallbackCommander()

--- a/deploy.sh
+++ b/deploy.sh
@ -300,7 +300,7 @@ User=root
 WorkingDirectory=$PROJECT_DIR/backend
 Environment="PATH=$PROJECT_DIR/backend/venv/bin"
 Environment="PYTHONPATH=$PROJECT_DIR/backend"
-EnvironmentFile=$PROJECT_DIR/backend/.env
+EnvironmentFile=$PROJECT_DIR/.env
 ExecStart=$PROJECT_DIR/backend/venv/bin/uvicorn app.main:app --host 127.0.0.1 --port 8000
 Restart=always
 RestartSec=10
--- a/docker-compose.dev.yml
+++ b/docker-compose.dev.yml
@ -55,41 +55,45 @@ services:
      redis:
        condition: service_healthy
    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-pmflow}:${POSTGRES_PASSWORD:-pmflow_pass}@db:5432/${POSTGRES_DB:-pmflow}
-      REDIS_URL: redis://:${REDIS_PASSWORD:-pmflow_redis_pass}@redis:6379/0
-      CORS_ORIGINS: ${CORS_ORIGINS:-http://localhost:5174,http://127.0.0.1:5174}
-      LLM_PROVIDER: ${LLM_PROVIDER:-openai}
-      LLM_FALLBACK_PROVIDER: ${LLM_FALLBACK_PROVIDER:-ollama}
-      OPENAI_BASE_URL: ${OPENAI_BASE_URL:-http://openclaw:18789/v1}
-      OPENCLAW_ROUTING_ENABLED: ${OPENCLAW_ROUTING_ENABLED:-true}
-      OPENCLAW_ROUTING_RULES_PATH: ${OPENCLAW_ROUTING_RULES_PATH:-/app/backend/.openclaw/pmflow-routing-rules.json}
-      LLM_MIXED_ROUTING_ENABLED: ${LLM_MIXED_ROUTING_ENABLED:-false}
-      LLM_LOW_TO_HIGH_ESCALATION_ENABLED: ${LLM_LOW_TO_HIGH_ESCALATION_ENABLED:-true}
-      LLM_ROUTING_LOW_PROVIDER: ${LLM_ROUTING_LOW_PROVIDER:-glm}
-      LLM_ROUTING_LOW_MODEL: ${LLM_ROUTING_LOW_MODEL:-GLM-4.7}
-      LLM_ROUTING_HIGH_PROVIDER: ${LLM_ROUTING_HIGH_PROVIDER:-openai}
-      LLM_ROUTING_HIGH_MODEL: ${LLM_ROUTING_HIGH_MODEL:-MiniMax-M2.5}
-      LLM_ROUTING_SAFETY_PROVIDER: ${LLM_ROUTING_SAFETY_PROVIDER:-glm}
-      LLM_ROUTING_SAFETY_MODEL: ${LLM_ROUTING_SAFETY_MODEL:-GLM-4.7}
-      LLM_DIRECT_EMERGENCY_ENABLED: ${LLM_DIRECT_EMERGENCY_ENABLED:-false}
-      LLM_REQUEST_TIMEOUT: ${LLM_REQUEST_TIMEOUT:-90}
+      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:?missing_POSTGRES_USER}:${POSTGRES_PASSWORD:?missing_POSTGRES_PASSWORD}@db:5432/${POSTGRES_DB:?missing_POSTGRES_DB}
+      REDIS_URL: redis://:${REDIS_PASSWORD:?missing_REDIS_PASSWORD}@redis:6379/0
+      PMFLOW_EXPECTED_DB_DIALECT: ${PMFLOW_EXPECTED_DB_DIALECT:?missing_PMFLOW_EXPECTED_DB_DIALECT}
+      CORS_ORIGINS: ${CORS_ORIGINS:?missing_CORS_ORIGINS}
+      LLM_PROVIDER: ${LLM_PROVIDER:?missing_LLM_PROVIDER}
+      LLM_FALLBACK_PROVIDER: ${LLM_FALLBACK_PROVIDER:?missing_LLM_FALLBACK_PROVIDER}
+      OPENAI_BASE_URL: ${OPENAI_BASE_URL:?missing_OPENAI_BASE_URL}
+      OPENCLAW_ROUTING_ENABLED: ${OPENCLAW_ROUTING_ENABLED:?missing_OPENCLAW_ROUTING_ENABLED}
+      OPENCLAW_ROUTING_RULES_PATH: ${OPENCLAW_ROUTING_RULES_PATH:?missing_OPENCLAW_ROUTING_RULES_PATH}
+      OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS: ${OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS:?missing_OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS}
+      LLM_MIXED_ROUTING_ENABLED: ${LLM_MIXED_ROUTING_ENABLED:?missing_LLM_MIXED_ROUTING_ENABLED}
+      LLM_ROUTING_DEFAULT_TIER: ${LLM_ROUTING_DEFAULT_TIER:?missing_LLM_ROUTING_DEFAULT_TIER}
+      LLM_LOW_TO_HIGH_ESCALATION_ENABLED: ${LLM_LOW_TO_HIGH_ESCALATION_ENABLED:?missing_LLM_LOW_TO_HIGH_ESCALATION_ENABLED}
+      LLM_ROUTING_LOW_PROVIDER: ${LLM_ROUTING_LOW_PROVIDER:?missing_LLM_ROUTING_LOW_PROVIDER}
+      LLM_ROUTING_LOW_MODEL: ${LLM_ROUTING_LOW_MODEL:?missing_LLM_ROUTING_LOW_MODEL}
+      LLM_ROUTING_HIGH_PROVIDER: ${LLM_ROUTING_HIGH_PROVIDER:?missing_LLM_ROUTING_HIGH_PROVIDER}
+      LLM_ROUTING_HIGH_MODEL: ${LLM_ROUTING_HIGH_MODEL:?missing_LLM_ROUTING_HIGH_MODEL}
+      LLM_ROUTING_SAFETY_PROVIDER: ${LLM_ROUTING_SAFETY_PROVIDER:?missing_LLM_ROUTING_SAFETY_PROVIDER}
+      LLM_ROUTING_SAFETY_MODEL: ${LLM_ROUTING_SAFETY_MODEL:?missing_LLM_ROUTING_SAFETY_MODEL}
+      LLM_DIRECT_EMERGENCY_ENABLED: ${LLM_DIRECT_EMERGENCY_ENABLED:?missing_LLM_DIRECT_EMERGENCY_ENABLED}
+      LLM_REQUEST_TIMEOUT: ${LLM_REQUEST_TIMEOUT:?missing_LLM_REQUEST_TIMEOUT}
      AUTO_CREATE_TABLES: "true"
-      GLM_API_KEY: ${GLM_API_KEY:-}
-      GLM_BASE_URL: ${GLM_BASE_URL:-https://open.bigmodel.cn/api/paas/v4}
-      GLM_MODEL: ${GLM_MODEL:-GLM-4.7}
-      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
-      OPENAI_MODEL: ${OPENAI_MODEL:-MiniMax-M2.5}
-      DEEPSEEK_API_KEY: ${DEEPSEEK_API_KEY:-}
-      OLLAMA_HOST: ${OLLAMA_HOST:-http://host.docker.internal:11434}
-      OLLAMA_BASE_URL: ${OLLAMA_BASE_URL:-http://host.docker.internal:11434}
-      OLLAMA_MODEL: ${OLLAMA_MODEL:-llama3:latest}
+      GLM_API_KEY: ${GLM_API_KEY:?missing_GLM_API_KEY}
+      GLM_BASE_URL: ${GLM_BASE_URL:?missing_GLM_BASE_URL}
+      GLM_MODEL: ${GLM_MODEL:?missing_GLM_MODEL}
+      OPENAI_API_KEY: ${OPENAI_API_KEY:?missing_OPENAI_API_KEY}
+      OPENAI_MODEL: ${OPENAI_MODEL:?missing_OPENAI_MODEL}
+      DEEPSEEK_API_KEY: ${DEEPSEEK_API_KEY}
+      OLLAMA_HOST: ${OLLAMA_HOST:?missing_OLLAMA_HOST}
+      OLLAMA_BASE_URL: ${OLLAMA_BASE_URL:?missing_OLLAMA_BASE_URL}
+      OLLAMA_MODEL: ${OLLAMA_MODEL:?missing_OLLAMA_MODEL}
      PYTHONUNBUFFERED: "1"
    command: ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000", "--reload", "--reload-dir", "app"]
    volumes:
      - ./backend:/app/backend
+      - ./.openclaw:/app/.openclaw:ro
      - ./logs:/app/logs
    ports:
-      - "${BACKEND_PORT:-8000}:8000"
+      - "${BACKEND_PORT:?missing_BACKEND_PORT}:8000"
    healthcheck:
      test: ["CMD-SHELL", "curl -fsS http://localhost:8000/healthz >/dev/null || exit 1"]
      interval: 10s
--- a/docker-compose.yml
+++ b/docker-compose.yml
@ -55,37 +55,41 @@ services:
      redis:
        condition: service_healthy
    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-pmflow}:${POSTGRES_PASSWORD:-pmflow_pass}@db:5432/${POSTGRES_DB:-pmflow}
-      REDIS_URL: redis://:${REDIS_PASSWORD:-pmflow_redis_pass}@redis:6379/0
-      CORS_ORIGINS: ${CORS_ORIGINS:-http://localhost:5174,http://127.0.0.1:5174}
-      LLM_PROVIDER: ${LLM_PROVIDER:-openai}
-      LLM_FALLBACK_PROVIDER: ${LLM_FALLBACK_PROVIDER:-ollama}
-      OPENAI_BASE_URL: ${OPENAI_BASE_URL:-http://openclaw:18789/v1}
-      OPENCLAW_ROUTING_ENABLED: ${OPENCLAW_ROUTING_ENABLED:-true}
-      OPENCLAW_ROUTING_RULES_PATH: ${OPENCLAW_ROUTING_RULES_PATH:-/app/backend/.openclaw/pmflow-routing-rules.json}
-      LLM_MIXED_ROUTING_ENABLED: ${LLM_MIXED_ROUTING_ENABLED:-false}
-      LLM_LOW_TO_HIGH_ESCALATION_ENABLED: ${LLM_LOW_TO_HIGH_ESCALATION_ENABLED:-true}
-      LLM_ROUTING_LOW_PROVIDER: ${LLM_ROUTING_LOW_PROVIDER:-glm}
-      LLM_ROUTING_LOW_MODEL: ${LLM_ROUTING_LOW_MODEL:-GLM-4.7}
-      LLM_ROUTING_HIGH_PROVIDER: ${LLM_ROUTING_HIGH_PROVIDER:-openai}
-      LLM_ROUTING_HIGH_MODEL: ${LLM_ROUTING_HIGH_MODEL:-MiniMax-M2.5}
-      LLM_ROUTING_SAFETY_PROVIDER: ${LLM_ROUTING_SAFETY_PROVIDER:-glm}
-      LLM_ROUTING_SAFETY_MODEL: ${LLM_ROUTING_SAFETY_MODEL:-GLM-4.7}
-      LLM_DIRECT_EMERGENCY_ENABLED: ${LLM_DIRECT_EMERGENCY_ENABLED:-false}
-      LLM_REQUEST_TIMEOUT: ${LLM_REQUEST_TIMEOUT:-90}
+      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:?missing_POSTGRES_USER}:${POSTGRES_PASSWORD:?missing_POSTGRES_PASSWORD}@db:5432/${POSTGRES_DB:?missing_POSTGRES_DB}
+      REDIS_URL: redis://:${REDIS_PASSWORD:?missing_REDIS_PASSWORD}@redis:6379/0
+      PMFLOW_EXPECTED_DB_DIALECT: ${PMFLOW_EXPECTED_DB_DIALECT:?missing_PMFLOW_EXPECTED_DB_DIALECT}
+      CORS_ORIGINS: ${CORS_ORIGINS:?missing_CORS_ORIGINS}
+      LLM_PROVIDER: ${LLM_PROVIDER:?missing_LLM_PROVIDER}
+      LLM_FALLBACK_PROVIDER: ${LLM_FALLBACK_PROVIDER:?missing_LLM_FALLBACK_PROVIDER}
+      OPENAI_BASE_URL: ${OPENAI_BASE_URL:?missing_OPENAI_BASE_URL}
+      OPENCLAW_ROUTING_ENABLED: ${OPENCLAW_ROUTING_ENABLED:?missing_OPENCLAW_ROUTING_ENABLED}
+      OPENCLAW_ROUTING_RULES_PATH: ${OPENCLAW_ROUTING_RULES_PATH:?missing_OPENCLAW_ROUTING_RULES_PATH}
+      OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS: ${OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS:?missing_OPENCLAW_AGENTS_DEFAULT_TIMEOUT_SECONDS}
+      LLM_MIXED_ROUTING_ENABLED: ${LLM_MIXED_ROUTING_ENABLED:?missing_LLM_MIXED_ROUTING_ENABLED}
+      LLM_ROUTING_DEFAULT_TIER: ${LLM_ROUTING_DEFAULT_TIER:?missing_LLM_ROUTING_DEFAULT_TIER}
+      LLM_LOW_TO_HIGH_ESCALATION_ENABLED: ${LLM_LOW_TO_HIGH_ESCALATION_ENABLED:?missing_LLM_LOW_TO_HIGH_ESCALATION_ENABLED}
+      LLM_ROUTING_LOW_PROVIDER: ${LLM_ROUTING_LOW_PROVIDER:?missing_LLM_ROUTING_LOW_PROVIDER}
+      LLM_ROUTING_LOW_MODEL: ${LLM_ROUTING_LOW_MODEL:?missing_LLM_ROUTING_LOW_MODEL}
+      LLM_ROUTING_HIGH_PROVIDER: ${LLM_ROUTING_HIGH_PROVIDER:?missing_LLM_ROUTING_HIGH_PROVIDER}
+      LLM_ROUTING_HIGH_MODEL: ${LLM_ROUTING_HIGH_MODEL:?missing_LLM_ROUTING_HIGH_MODEL}
+      LLM_ROUTING_SAFETY_PROVIDER: ${LLM_ROUTING_SAFETY_PROVIDER:?missing_LLM_ROUTING_SAFETY_PROVIDER}
+      LLM_ROUTING_SAFETY_MODEL: ${LLM_ROUTING_SAFETY_MODEL:?missing_LLM_ROUTING_SAFETY_MODEL}
+      LLM_DIRECT_EMERGENCY_ENABLED: ${LLM_DIRECT_EMERGENCY_ENABLED:?missing_LLM_DIRECT_EMERGENCY_ENABLED}
+      LLM_REQUEST_TIMEOUT: ${LLM_REQUEST_TIMEOUT:?missing_LLM_REQUEST_TIMEOUT}
      AUTO_CREATE_TABLES: "true"
-      GLM_API_KEY: ${GLM_API_KEY:-}
-      GLM_BASE_URL: ${GLM_BASE_URL:-https://open.bigmodel.cn/api/paas/v4}
-      GLM_MODEL: ${GLM_MODEL:-GLM-4.7}
-      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
-      OPENAI_MODEL: ${OPENAI_MODEL:-MiniMax-M2.5}
-      DEEPSEEK_API_KEY: ${DEEPSEEK_API_KEY:-}
-      OLLAMA_HOST: ${OLLAMA_HOST:-http://host.docker.internal:11434}
-      OLLAMA_BASE_URL: ${OLLAMA_BASE_URL:-http://host.docker.internal:11434}
-      OLLAMA_MODEL: ${OLLAMA_MODEL:-llama3:latest}
+      GLM_API_KEY: ${GLM_API_KEY:?missing_GLM_API_KEY}
+      GLM_BASE_URL: ${GLM_BASE_URL:?missing_GLM_BASE_URL}
+      GLM_MODEL: ${GLM_MODEL:?missing_GLM_MODEL}
+      OPENAI_API_KEY: ${OPENAI_API_KEY:?missing_OPENAI_API_KEY}
+      OPENAI_MODEL: ${OPENAI_MODEL:?missing_OPENAI_MODEL}
+      DEEPSEEK_API_KEY: ${DEEPSEEK_API_KEY}
+      OLLAMA_HOST: ${OLLAMA_HOST:?missing_OLLAMA_HOST}
+      OLLAMA_BASE_URL: ${OLLAMA_BASE_URL:?missing_OLLAMA_BASE_URL}
+      OLLAMA_MODEL: ${OLLAMA_MODEL:?missing_OLLAMA_MODEL}
    ports:
-      - "${BACKEND_PORT:-8000}:8000"
+      - "${BACKEND_PORT:?missing_BACKEND_PORT}:8000"
    volumes:
+      - ./.openclaw:/app/.openclaw:ro
      - ./logs:/app/logs
    healthcheck:
      test: ["CMD-SHELL", "curl -fsS http://localhost:8000/healthz >/dev/null || exit 1"]
--- a/frontend/src/api/aiSteward.ts
+++ b/frontend/src/api/aiSteward.ts
@ -63,6 +63,21 @@ export interface OpenClawCommandDispatchResponse {
  };
 }

+export type LLMProvider = "deepseek" | "glm" | "openai" | "ollama";
+
+export interface LLMProviderPreset {
+  model: string;
+  base_url: string;
+  api_env: string;
+}
+
+export interface LLMProviderPresetResponse {
+  low: string;
+  high: string;
+  base_url: string;
+  providers: Record<LLMProvider, LLMProviderPreset>;
+}
+
 export const aiStewardApi = {
  async dispatchCommand(body: OpenClawCommandDispatchRequest): Promise<OpenClawCommandDispatchResponse> {
    const raw = await apiRequest<ApiEnvelope<OpenClawCommandDispatchResponse>>("/ai-steward/command/dispatch", {
@ -71,4 +86,8 @@ export const aiStewardApi = {
    });
    return unwrap(raw);
  },
+  async getLlmProviderPresets(): Promise<LLMProviderPresetResponse> {
+    const raw = await apiRequest<ApiEnvelope<LLMProviderPresetResponse>>("/ai-steward/llm/providers");
+    return unwrap(raw);
+  },
 };
--- a/frontend/src/pages/AIStewardPage.tsx
+++ b/frontend/src/pages/AIStewardPage.tsx
@ -20,6 +20,7 @@ import { formatProjectNameWithId, toZhPriority, toZhSeverity, toZhStatus } from
 import { formatBeijingDate } from "@/lib/datetime";
 import { QAHistoryPanel } from "@/components/ai-steward/QAHistoryPanel";
 import { OpenClawCommanderPanel } from "@/components/ai-steward/OpenClawCommanderPanel";
+import { aiStewardApi } from "@/api/aiSteward";

 // ==================== Types ====================

@ -83,7 +84,7 @@ const PROVIDER_ZH: Record<LLMEngine["provider"], string> = {
  ollama: "Ollama(本地)",
 };

-const ENGINE_PROVIDER_PRESETS: Record<LLMEngine["provider"], { model: string; base_url: string }> = {
+const ENGINE_PROVIDER_FALLBACK_PRESETS: Record<LLMEngine["provider"], { model: string; base_url: string }> = {
  deepseek: {
    model: "deepseek-chat",
    base_url: "https://api.deepseek.com/v1",
@ -119,12 +120,15 @@ export default function AIStewardPage() {
  const [briefing, setBriefing] = useState<DailyBriefing | null>(null);
  const [featureLoading, setFeatureLoading] = useState<string | null>(null);
  const [engineData, setEngineData] = useState<LLMEngineListResponse>({ active_engine_id: null, engines: [] });
+  const [engineProviderPresets, setEngineProviderPresets] = useState<Record<LLMEngine["provider"], { model: string; base_url: string }>>(
+    ENGINE_PROVIDER_FALLBACK_PRESETS
+  );
  const [engineForm, setEngineForm] = useState({
    name: "",
    provider: "glm" as "deepseek" | "glm" | "openai" | "ollama",
    api_key: "",
-    model: ENGINE_PROVIDER_PRESETS.glm.model,
-    base_url: ENGINE_PROVIDER_PRESETS.glm.base_url,
+    model: ENGINE_PROVIDER_FALLBACK_PRESETS.glm.model,
+    base_url: ENGINE_PROVIDER_FALLBACK_PRESETS.glm.base_url,
  });
  const [engineEditingId, setEngineEditingId] = useState<string | null>(null);
  const [engineMessage, setEngineMessage] = useState<{ type: "success" | "error"; text: string } | null>(null);
@ -276,14 +280,51 @@ export default function AIStewardPage() {
    }
  };

+  const loadEngineProviderPresets = async () => {
+    try {
+      const data = await aiStewardApi.getLlmProviderPresets();
+      const providers = data?.providers ?? {};
+      const nextPresets: Record<LLMEngine["provider"], { model: string; base_url: string }> = {
+        deepseek: {
+          model: providers.deepseek?.model || ENGINE_PROVIDER_FALLBACK_PRESETS.deepseek.model,
+          base_url: providers.deepseek?.base_url || ENGINE_PROVIDER_FALLBACK_PRESETS.deepseek.base_url,
+        },
+        glm: {
+          model: providers.glm?.model || data?.low || ENGINE_PROVIDER_FALLBACK_PRESETS.glm.model,
+          base_url: providers.glm?.base_url || data?.base_url || ENGINE_PROVIDER_FALLBACK_PRESETS.glm.base_url,
+        },
+        openai: {
+          model: providers.openai?.model || data?.high || ENGINE_PROVIDER_FALLBACK_PRESETS.openai.model,
+          base_url: providers.openai?.base_url || data?.base_url || ENGINE_PROVIDER_FALLBACK_PRESETS.openai.base_url,
+        },
+        ollama: {
+          model: providers.ollama?.model || ENGINE_PROVIDER_FALLBACK_PRESETS.ollama.model,
+          base_url: providers.ollama?.base_url || ENGINE_PROVIDER_FALLBACK_PRESETS.ollama.base_url,
+        },
+      };
+      setEngineProviderPresets(nextPresets);
+      setEngineForm((prev) => {
+        if (engineEditingId) return prev;
+        const preset = nextPresets[prev.provider];
+        return {
+          ...prev,
+          model: preset.model,
+          base_url: preset.base_url,
+        };
+      });
+    } catch (err) {
+      console.error("Failed to load provider presets:", err);
+    }
+  };
+
  const resetEngineForm = () => {
    setEngineEditingId(null);
    setEngineForm({
      name: "",
      provider: "glm",
      api_key: "",
-      model: ENGINE_PROVIDER_PRESETS.glm.model,
-      base_url: ENGINE_PROVIDER_PRESETS.glm.base_url,
+      model: engineProviderPresets.glm.model,
+      base_url: engineProviderPresets.glm.base_url,
    });
  };

@ -479,7 +520,10 @@ export default function AIStewardPage() {
    if (activeTab === "conflicts") loadConflicts();
    if (activeTab === "blockers") loadBlockers();
    if (activeTab === "briefing") loadBriefing();
-    if (activeTab === "engine") loadEngineConfigs();
+    if (activeTab === "engine") {
+      loadEngineProviderPresets();
+      loadEngineConfigs();
+    }
  }, [activeTab]);

  // ==================== Render ====================
@ -1559,7 +1603,7 @@ export default function AIStewardPage() {
                            onChange={(e) =>
                              setEngineForm((prev) => {
                                const provider = e.target.value as "deepseek" | "glm" | "openai" | "ollama";
-                                const preset = ENGINE_PROVIDER_PRESETS[provider];
+                                const preset = engineProviderPresets[provider];
                                return {
                                  ...prev,
                                  provider,
--- a/mdproject/docker-redis-context-plan-en.md
+++ b/mdproject/docker-redis-context-plan-en.md
@ -142,7 +142,7 @@ Return answer
 - [x] Added DB/Redis health checks and `depends_on` health gating.
 - [x] Added backend container build/runtime file: `backend/Dockerfile`.
 - [x] Added backend dependency lock baseline: `backend/requirements.txt`.
- [x] Added docker env template: `.env.example` and updated `backend/.env.example`.
+- [x] Added docker env template: `.env.example` as the single environment template source.
 - [x] Redis policy configured to `maxmemory 512mb` + `allkeys-lru`.

 ### Part 2 — Database Migration Steps
--- a/scripts/openclaw_routing_lint.py
+++ b/scripts/openclaw_routing_lint.py
@ -5,11 +5,15 @@ from __future__ import annotations

 import argparse
 import json
+import os
 from pathlib import Path
 from typing import Any, Optional

 VALID_PROVIDERS = {"glm", "openai", "deepseek", "ollama"}
 VALID_TIERS = {"low", "high", "safety"}
+DEFAULT_RULES_RELATIVE_PATH = "../.openclaw/pmflow-routing-rules.json"
+SCRIPT_ROOT = Path(__file__).resolve().parents[1]
+BACKEND_DIR = SCRIPT_ROOT / "backend"

 LOW_REQUIRED = {
    "feishu_level_classification",
@ -76,6 +80,14 @@ def _resolve_primary_tier(function_name: str, rules: list[dict[str, Any]]) -> Op
    return scored[0][2]


+def _resolve_rules_path(raw: str) -> Path:
+    token = str(raw or "").strip() or DEFAULT_RULES_RELATIVE_PATH
+    path = Path(token)
+    if path.is_absolute():
+        return path
+    return (BACKEND_DIR / path).resolve()
+
+
 def _lint(path: Path) -> tuple[bool, dict[str, Any]]:
    if not path.exists():
        return False, {"error": f"rules file not found: {path}"}
@ -135,9 +147,14 @@ def _lint(path: Path) -> tuple[bool, dict[str, Any]]:
                break

    model_hints = {
-        "low": "zai/glm-4-flash",
-        "high": "minimax/MiniMax-M2.5",
-        "safety": "deepseek/deepseek-chat",
+        "low": "GLM-4.7",
+        "high": "MiniMax-M2.5",
+        "safety": "deepseek-chat",
+    }
+    api_env_hints = {
+        "low": "GLM_API_KEY",
+        "high": "OPENAI_API_KEY",
+        "safety": "OPENAI_API_KEY",
    }
    for tier, expected_model in model_hints.items():
        tier_rows = [
@ -146,6 +163,9 @@ def _lint(path: Path) -> tuple[bool, dict[str, Any]]:
        ]
        if tier_rows and not any(str(row.get("model") or "").strip() == expected_model for row in tier_rows):
            problems.append(f"tier={tier} missing expected model hint: {expected_model}")
+        expected_api_env = api_env_hints[tier]
+        if tier_rows and not any(str(row.get("api_env") or "").strip() == expected_api_env for row in tier_rows):
+            problems.append(f"tier={tier} missing expected api_env hint: {expected_api_env}")

    return len(problems) == 0, {
        "path": str(path),
@ -159,15 +179,22 @@ def _lint(path: Path) -> tuple[bool, dict[str, Any]]:


 def main() -> None:
+    default_input = str(
+        os.getenv(
+            "OPENCLAW_ROUTING_RULES_PATH",
+            DEFAULT_RULES_RELATIVE_PATH,
+        )
+        or DEFAULT_RULES_RELATIVE_PATH
+    ).strip() or DEFAULT_RULES_RELATIVE_PATH
    parser = argparse.ArgumentParser(description="OpenClaw routing lint")
    parser.add_argument(
        "--input",
-        default=".openclaw/pmflow-routing-rules.json",
+        default=default_input,
        help="Rules file path",
    )
    args = parser.parse_args()

-    ok, report = _lint(Path(args.input))
+    ok, report = _lint(_resolve_rules_path(args.input))
    print(json.dumps(report, ensure_ascii=False, indent=2))
    raise SystemExit(0 if ok else 2)

--- a/tests/test_llm_fail_fast.py
+++ b/tests/test_llm_fail_fast.py
@ -30,7 +30,7 @@ def test_provider_circuit_breaker_opens_after_threshold():
    circuit_key = service._provider_circuit_key(
        provider="openai",
        base_url="https://example.com/v1",
-        model="gpt-4o-mini",
+        model="MiniMax-M2.5",
    )
    LLMService._provider_circuit_state.pop(circuit_key, None)

--- a/用户使用手册.md
+++ b/用户使用手册.md
@ -60,7 +60,7 @@

 **选项 A：使用云端 API（推荐，电脑配置低）**

-编辑 `backend/.env` 文件：
+编辑根目录 `.env` 文件：

 ```bash
 # 使用 DeepSeek（便宜，¥1/百万 tokens）
@ -96,7 +96,7 @@ ollama pull llama3:8b
 ollama pull nomic-embed-text
 ```

-编辑 `backend/.env` 文件：
+编辑根目录 `.env` 文件：

 ```bash
 LLM_PROVIDER=ollama
@ -116,10 +116,10 @@ source venv/bin/activate  # Windows: venv\Scripts\activate
 pip install -r requirements.txt

 # 配置环境变量 (可选，使用默认配置)
-cp .env.example .env
+cp ../.env.example ../.env

-# 编辑 .env 设置你的 LLM 提供商和 API 密钥
-vim .env
+# 编辑项目根目录 .env 设置你的 LLM 提供商和 API 密钥
+vim ../.env

 # 运行数据库迁移
 alembic upgrade head
@ -302,7 +302,7 @@ curl -X POST http://localhost:8000/signals/ingest \

 ### 5.1 环境变量配置

-编辑 `backend/.env` 文件：
+编辑根目录 `.env` 文件：

 ```bash
 # 数据库配置
@ -399,7 +399,7 @@ pip install -r requirements.txt

 **解决方案 A - 使用 SQLite（开发环境）**：

-编辑 `backend/.env`：
+编辑根目录 `.env`：
 ```bash
 DATABASE_URL=sqlite+aiosqlite:///./neural_sync.db
 ```
@ -604,8 +604,7 @@ Neural Sync/
 │   │   ├── database.py   # 数据库连接
 │   │   └── services/     # 业务逻辑
 │   ├── alembic/          # 数据库迁移
-│   ├── requirements.txt
-│   └── .env.example     # 环境变量模板
+│   └── requirements.txt
 │
 ├── frontend/              # Vue 3 前端
 │   ├── src/
@ -615,6 +614,7 @@ Neural Sync/
 │   ├── package.json
 │   └── vite.config.ts
 │
+├── .env.example          # 环境变量模板
 ├── CLAUDE.md             # Claude Code 指导
 └── README.md             # 项目说明
 ```