feat: NMFS Agents 全量入库（org v2 + dashboard + 多队列 + 组织架构）

包含从项目创建至今的全部代码首次入库：核心框架： - TaskQueue（SQLite，6 种状态，原子 dequeue） - Executor（asyncio 并发，semaphore 限流，agent_role 路由） - Scheduler（APScheduler，三队列独立调度） - Watchdog（超时检测 + 优先级防饥饿） - Scanner（项目扫描，README/TODO/CLAUDE.md 提取） Agent 体系（20+ 角色）： - 项目交付组：architect/developer/tester/productizer - 基础技术组：base-architect/validator/hw/os/kernel/lowlevel/system-tester - 算法组：algo-antishake/position/nav - 洞察组：planner/vision-analyst/media-producer - 市场组：market-pm/sport/elder/safety - 组织层：boss/group-leader/senior-dev/ops - 平台扩展：nrf-dev/esp32-dev - 三专项调研：rtp-researcher/net-researcher/kernel-analyzer 工具层： - ProjectMemory（goals/facts/history SQLite） - VersionPipeline（并行版本流水线 + 反幻觉门控） - EventBus（Redis Streams，5 consumer groups） - DailyReport（append-only 日报 + Boss 决策视图） - AgentScorer（4 维评分：完成度/质量/自主性/协作） - DeviceAgent（SSH rsync + 板端执行） - EnvCollector（本机+SSH 环境采集） - ArchVersion（快照/优化/promote/rollback） Dashboard（React + Vite + Tailwind）： - 看板/项目/配置/洞察/日报/Visual 六标签页 - WebSocket 实时更新 - Markdown 渲染（react-markdown + remark-gfm）测试：314 passed Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-10 07:24:19 +08:00
parent 36985d0774
commit df59fc0c83
160 changed files with 31806 additions and 1707 deletions
@@ -0,0 +1,28 @@
+# Python
+__pycache__/
+*.pyc
+*.egg-info/
+venv/
+.venv/
+
+# Runtime data
+data/
+logs/
+tmp/
+
+# IDE
+.vscode/
+.idea/
+
+# OS
+.DS_Store
+
+# Dashboard build output
+dashboard/dist/
+dashboard/node_modules/
+
+# Playwright
+.playwright-mcp/
+
+# ArchVersion snapshots
+.arch_versions/
@@ -4,7 +4,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co

 ## 项目概述

-Rockchip Agents：基于 Claude CLI 的 AI 项目自动化维护与调研系统。支持代码修复、板端测试、成熟度评估、架构整合、市场洞察与基础技术调研，运行在 RK3588 NPU 开发平台上。
+NMFS Agents：基于 Claude CLI 的 AI 项目自动化维护与调研系统。支持代码修复、板端测试、成熟度评估、架构整合、市场洞察与基础技术调研，运行在 RK3588 NPU 开发平台上。

 ## 目录约定

@@ -70,7 +70,7 @@ sqlite3 data/tasks.db "SELECT id,project,agent_role,title,status FROM tasks ORDE
 ## 项目结构

 ```
-src/rockchip_agents/
+src/nmfs_agents/
 ├── config.py                  # 配置加载（AgentsConfig dataclass）
 ├── agents/
 │   ├── developer.py           # Developer Agent（claude --print 子进程，三模式）
@@ -1,4 +1,4 @@
-# Rockchip Agents 开发指南
+# NMFS Agents 开发指南

 ## 环境搭建

@@ -73,6 +73,35 @@ Scanner.scan_all()            # 扫描 README/TODO/CLAUDE.md
              └─> FeishuNotifier.send_task_result()
 ```

+## Web Dashboard
+
+Dashboard 是首选操作入口，访问地址：`http://localhost:9080`（Docker 环境同端口）。
+
+### 主要功能
+
+| 标签页 | 功能 |
+|--------|------|
+| 看板（Kanban） | 按项目查看任务状态（待处理/运行中/完成/失败），可拖拽调整优先级 |
+| 洞察（Insight） | 市场目标卡片（运动/适老/AI安全）+ 基础技术调研展示 + 历史记录 |
+| 配置（Config） | 查看/编辑全局和项目级 CLAUDE.md / README.md，触发 CLAUDE.md 刷新 |
+
+### 常见操作
+
+**提交开发需求**（点击右上角「提交需求」）：
+- 选择项目 → 选择类型（修复问题/新功能/代码审查等）→ 填写描述 → 提交
+- 提交后自动路由到 `project_delivery` 队列
+
+**发起技术攻关**：
+- 需求类型选「技术攻关」→ 选择角色（RTP/网络栈/内核调度/OS/硬件）→ 填写描述 → 提交
+- 自动路由到 `base_opt` 队列；调研结论写入 `__research__.*`，洞察页自动展示
+
+**触发市场洞察**（洞察页）：
+- 点击目标卡片的「触发洞察」按钮 → insight 队列入队 → market-pm 执行
+- 调研结论写入 `__market__.{goal}.*`，目标卡片下次刷新时自动更新
+
+**刷新项目指导**（配置页，选中项目后可见）：
+- 点击「🔄 刷新项目指导」→ 创建 `architect` 任务 → Agent 综合调研结论自动重写 CLAUDE.md
+
 ## 常用命令

 ```bash
@@ -91,6 +120,9 @@ python scripts/run.py scan
 # 执行队列中下一个任务（须在普通终端执行）
 python scripts/run.py run-next

+# 入队三专项调研任务（须在普通终端执行）
+python scripts/run.py research
+
 # 全量测试
 pytest tests/ -q
 ```
@@ -115,9 +147,9 @@ sqlite3 data/insight_tasks.db "SELECT id,project,agent_role,title,status FROM ta
 ```python
 source venv/bin/activate
 python - <<'EOF'
-from rockchip_agents.config import load_config
-from rockchip_agents.core.queue import Task, TaskQueue
-import asyncio, rockchip_agents.core.executor as ex
+from nmfs_agents.config import load_config
+from nmfs_agents.core.queue import Task, TaskQueue
+import asyncio, nmfs_agents.core.executor as ex

 cfg = load_config()
 q = TaskQueue()
@@ -174,7 +206,7 @@ projects:

 ## 添加新 Agent

-1. 在 `src/rockchip_agents/agents/` 新建 `my_agent.py`，继承 `DeveloperAgent` 或独立实现 `run(task) -> AgentResult`
+1. 在 `src/nmfs_agents/agents/` 新建 `my_agent.py`，继承 `DeveloperAgent` 或独立实现 `run(task) -> AgentResult`
 2. 在 `executor.py` 的 `_make_agent()` 添加路由条目
 3. 在 `configs/agents.yaml` 的对应队列 `roles:` 中添加新角色名
 4. 在 `tests/` 添加测试文件
@@ -242,6 +274,45 @@ devices:
 | `DB_PATH` | 任务数据库路径（默认 `data/tasks.db`） | 否 |
 | `DASHBOARD_PORT` | Dashboard 端口（默认 8080） | 否 |

+## API 端点参考
+
+Dashboard 后端 FastAPI，主要端点：
+
+| 方法 | 路径 | 说明 |
+|------|------|------|
+| GET | `/api/tasks` | 获取任务列表（支持 `?project=` 过滤） |
+| POST | `/api/tasks` | 创建任务（`queue` 字段路由到不同数据库） |
+| PATCH | `/api/tasks/{id}` | 更新优先级 |
+| POST | `/api/tasks/{id}/retry` | 重试失败任务 |
+| POST | `/api/tasks/{id}/cancel` | 取消任务 |
+| POST | `/api/tasks/{id}/approve` | 审批 waiting_approval 任务 |
+| POST | `/api/tasks/{id}/reject` | 拒绝任务（附原因） |
+| GET | `/api/projects` | 获取项目列表 |
+| POST | `/api/projects/import` | 导入新项目 |
+| DELETE | `/api/projects/{name}` | 删除项目 |
+| GET | `/api/projects/{name}/config` | 获取项目 CLAUDE.md / README.md |
+| PUT | `/api/projects/{name}/config/{type}` | 保存配置文件（`claude`/`readme`） |
+| POST | `/api/projects/{name}/refresh-claude-md` | 触发 architect 更新 CLAUDE.md |
+| GET | `/api/config/global` | 获取全局 CLAUDE.md |
+| PUT | `/api/config/global` | 保存全局 CLAUDE.md |
+| GET | `/api/market/facts` | 获取市场洞察 facts（`__market__.*`） |
+| POST | `/api/market/trigger/{goal}` | 触发市场洞察任务 |
+| GET | `/api/market/history` | 获取洞察任务历史 |
+| GET | `/api/research/facts` | 获取三专项调研 facts（`__research__.*`） |
+| GET | `/api/executor/status` | 执行器状态 |
+| POST | `/api/executor/start` | 启动执行器 |
+| POST | `/api/executor/stop` | 停止执行器 |
+
+### 队列路由规则
+
+`POST /api/tasks` 的 `queue` 字段：
+
+| queue 值 | 路由到 | 典型 agent_role |
+|----------|--------|-----------------|
+| `project_delivery`（默认）| `data/tasks.db` | developer / architect |
+| `base_opt` | `data/base_opt_tasks.db` | rtp-researcher / kernel-analyzer / hw-engineer |
+| `insight` | `data/insight_tasks.db` | market-pm / planner / productizer |
+
 ## 项目记忆系统

 `data/project_memory.db` 分三层：
@@ -252,6 +323,16 @@ devices:
 | facts（事实） | 任意 agent；architect 通过 `[记忆]` 标签写入 |
 | history（历史） | executor 每次任务完成/失败后自动写入 |

+### 特殊 facts 命名约定
+
+| key 前缀 | 写入者 | 用途 | Dashboard 展示位置 |
+|----------|--------|------|-------------------|
+| `__research__.rtp.*` | rtp-researcher | RTP/RTCP 调研结论 | 洞察页"基础技术调研" |
+| `__research__.net.*` | net-researcher | 网络栈调优结论 | 洞察页"基础技术调研" |
+| `__research__.kernel.*` | kernel-analyzer | 内核调度分析结论 | 洞察页"基础技术调研" |
+| `__market__.{goal}.*` | market-pm | 市场洞察结论（goal=运动/适老/AI安全） | 洞察页目标卡片 |
+| `__env__.*` | env-collector | 板端硬件/软件环境信息 | Agent 系统提示 |
+
 ```bash
 python scripts/run.py goals    # 查看项目长期目标
 python scripts/run.py facts    # 查看项目知识事实
@@ -0,0 +1,37 @@
+# ── Stage 1: 构建 React 前端 ─────────────────────────────────────────
+FROM node:20-alpine AS frontend-builder
+WORKDIR /frontend
+
+COPY dashboard/package*.json ./
+RUN npm ci --prefer-offline
+
+COPY dashboard/ ./
+RUN npm run build
+
+# ── Stage 2: Python FastAPI 运行时 ───────────────────────────────────
+FROM python:3.10-slim AS runtime
+WORKDIR /app
+
+# 安装系统工具（rsync/sshpass 用于板端同步测试，bash 用于 source 命令）
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    sshpass rsync openssh-client \
+    && rm -rf /var/lib/apt/lists/*
+
+# 安装 Python 依赖
+COPY requirements-dashboard.txt ./
+RUN pip install --no-cache-dir -r requirements-dashboard.txt
+
+# 复制源码并安装包
+COPY src/ ./src/
+COPY pyproject.toml ./
+RUN pip install --no-cache-dir -e . --no-deps
+COPY scripts/dashboard.py ./scripts/
+
+# 复制前端构建产物
+COPY --from=frontend-builder /frontend/dist ./dashboard/dist/
+
+EXPOSE 8080
+
+ENV PYTHONPATH=/app/src
+
+CMD ["python", "scripts/dashboard.py"]
@@ -0,0 +1,28 @@
+.PHONY: install test test-one clean scan run status start
+
+install:
+	python3.10 -m venv venv
+	venv/bin/pip install -e ".[dev]"
+
+test:
+	venv/bin/pytest tests/ -q
+
+test-one:
+	venv/bin/pytest $(FILE) -v
+
+clean:
+	rm -rf data/tasks.db __pycache__ src/rockchip_agents.egg-info
+	find . -name "*.pyc" -delete
+	find . -name "__pycache__" -type d -exec rm -rf {} + 2>/dev/null || true
+
+scan:
+	venv/bin/python scripts/run.py scan
+
+run:
+	venv/bin/python scripts/run.py run
+
+status:
+	venv/bin/python scripts/run.py status
+
+start:
+	venv/bin/python scripts/run.py start
@@ -1,4 +1,4 @@
-# Rockchip Agents
+# NMFS Agents

 基于 Claude CLI 的 AI 项目自动化维护与调研系统，运行在 RK3588 NPU 开发平台上。
 支持代码修复、板端测试、成熟度评估、架构整合、市场洞察与基础技术调研。
@@ -93,12 +93,33 @@ python scripts/run.py status        # 查看任务状态

 ## 工作原理

+系统有两条触发路径，均汇入同一个三队列执行引擎：
+
 ```
-APScheduler（多队列，1h/4h/6h）
-  └─> Scanner 扫描项目（README 已知问题 / TODO / CLAUDE.md 标记）
-        └─> TaskQueue 入队（SQLite，去重，优先级排序）
-              └─> Executor 并发分发（asyncio Semaphore 限流）
-                    └─> Agent.run(task) → claude --print 子进程
+┌─ Web Dashboard（主操作入口，http://localhost:9080）────────────────────────┐
+│  提交需求（修复/重构/文档）  →  project_delivery 队列  →  developer/architect│
+│  技术攻关（RTP/网络/内核）   →  base_opt 队列          →  专项研究 Agent     │
+│  触发市场洞察                →  insight 队列           →  market-pm/planner  │
+│  刷新项目指导（CLAUDE.md）   →  project_delivery 队列  →  architect 分析写入 │
+└────────────────────────────────────────────────────────────────────────────┘
+
+┌─ APScheduler（定时自动）───────────────────────────────────────────────────┐
+│  Scanner 扫描 README/TODO/CLAUDE.md → 自动入对应队列（1h/6h 间隔）          │
+└────────────────────────────────────────────────────────────────────────────┘
+
+                         ↓ 三队列并发执行
+
+project_delivery（tasks.db）  ─> developer / architect / tester
+base_opt（base_opt_tasks.db） ─> 算法专项 / 调研 / 硬件优化 Agent
+insight（insight_tasks.db）   ─> market-pm / planner / productizer
+
+                         ↓ 调研结论写入 ProjectMemory
+
+__research__.rtp.*   / __research__.net.*   / __research__.kernel.*
+  └─> Dashboard"基础技术调研"区块自动展示
+
+__market__.{goal}.*（运动 / 适老 / AI安全）
+  └─> Dashboard"市场洞察"目标卡片自动更新
 ```

 ## 执行模式
@@ -14,7 +14,7 @@
 - **Dashboard Web UI**：FastAPI + React 看板，WebSocket 实时推送，支持 KanbanBoard / InsightView / ConfigView / VisualView
 - **市场洞察（InsightView）**：planner/productizer/vision-analyst/media-producer 角色，前端触发
 - **三专项 Agent**：RtpResearcherAgent / NetResearcherAgent / KernelAnalyzerAgent
- **Dashboard Docker 部署**：`rockchip-agents-dashboard:latest`，端口 9080
+- **Dashboard Docker 部署**：`nmfs-agents-dashboard:latest`，端口 9080
 - **rk3566 设备接入**：算法任务默认用 rk3566（192.168.123.183）验证，rk3588 作为高性能备选

 ### 修复
@@ -39,6 +39,7 @@ queues:
      - dev-embedded
      - os-engineer
      - system-tester
+      - tester
      - architect
      - market-pm
      - rtp-researcher
@@ -97,28 +97,51 @@ roles:

  productizer:
    - sequential-thinking   # 成熟度评分推理、维度分析
+    - fetch                 # 行业标准、同类产品成熟度参考
    - sqlite-memory         # 查询项目目标和能力事实作为评分依据

  vision-analyst:
    - sequential-thinking   # 视觉质量分析推理
+    - fetch                 # 视觉效果参考、UI 设计规范
    - playwright            # 板端推理界面截图、视觉交互验证

  market-pm:
    - sequential-thinking   # 市场文档推理、方案书结构规划
    - fetch                 # 市场资料、竞品信息、行业数据抓取
    - context7              # 行业标准和规范文档
+    - sqlite-memory         # 查询本平台能力事实作为竞品对比基准

  media-producer:
    - sequential-thinking   # 媒体生产规划、效果图合成逻辑推理
+    - fetch                 # 行业内容参考、风格灵感

  # ── 兼容旧角色（保持向后兼容）────────────────────────────────────────────
-  tester:        [sequential-thinking, fetch, playwright]
+  tester:        [sequential-thinking, fetch, playwright, sqlite-memory]
  dev-embedded:  [sequential-thinking, fetch, context7]
  dev-backend:   [sequential-thinking, fetch, context7]
  dev-ml:        [sequential-thinking, fetch, context7]
-  dev-test:      [sequential-thinking, playwright]
-  dev-refactor:  [sequential-thinking]
+  dev-test:      [sequential-thinking, fetch, playwright]
+  dev-refactor:  [sequential-thinking, fetch]
  arch-system:   [sequential-thinking, fetch, context7]
  arch-security: [sequential-thinking, fetch]
-  arch-reviewer: [sequential-thinking]
-  pm:            [sequential-thinking, fetch]
+  arch-reviewer: [sequential-thinking, fetch]
+  pm:            [sequential-thinking, fetch, sqlite-memory]
+
+  # ── 专项调研（project_delivery 队列）────────────────────────────────────
+  rtp-researcher:
+    - sequential-thinking
+    - fetch                 # RTP/RTCP RFC 文档、媒体传输最佳实践
+    - context7              # 协议库文档
+    - sqlite-memory         # 读取已有调研结论
+
+  net-researcher:
+    - sequential-thinking
+    - fetch                 # 网络栈文档、内核网络子系统资料
+    - context7              # Linux 网络编程文档
+    - sqlite-memory
+
+  kernel-analyzer:
+    - sequential-thinking
+    - fetch                 # 内核文档、调度器源码参考
+    - context7              # Linux 内核 API 文档
+    - sqlite-memory
@@ -0,0 +1,77 @@
+# 扫描任务标题/上下文中的关键词，路由到对应 agent_role
+role_keywords:
+  dev-kernel:
+    - 驱动
+    - 内核
+    - kernel module
+    - linux driver
+    - platform_driver
+    - kernel
+    - dts
+    - device tree
+    - insmod
+    - modprobe
+    - kconfig
+  dev-lowlevel:
+    - bsp
+    - uboot
+    - bootloader
+    - 裸机
+    - rtos
+    - 引导
+    - 启动
+    - u-boot
+    - spl
+  dev-embedded:
+    - rknn
+    - npu
+    - 推理
+    - 板端
+    - rk3588
+    - lite2
+    - rknn_lite
+    - opencv
+    - mediapipe
+  dev-ml:
+    - 模型
+    - 精度
+    - 量化
+    - 训练
+    - onnx
+    - accuracy
+    - map
+    - calibration
+  dev-backend:
+    - api
+    - fastapi
+    - 数据库
+    - 服务
+    - http
+    - redis
+    - django
+    - flask
+  hw-engineer:
+    - 选型
+    - 硬件
+    - 芯片
+    - 外设
+    - 接口
+    - bom
+    - 功耗
+    - datasheet
+    - schematic
+  dev-test:
+    - 测试覆盖
+    - coverage
+    - 缺少测试
+    - test case
+    - pytest
+  dev-refactor:
+    - 重构
+    - 技术债
+    - 代码质量
+    - 冗余
+    - refactor
+
+# 未匹配时的默认角色
+default_role: dev-embedded
@@ -0,0 +1,3 @@
+node_modules
+dist
+.git
@@ -0,0 +1,24 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+pnpm-debug.log*
+lerna-debug.log*
+
+node_modules
+dist
+dist-ssr
+*.local
+
+# Editor directories and files
+.vscode/*
+!.vscode/extensions.json
+.idea
+.DS_Store
+*.suo
+*.ntvs*
+*.njsproj
+*.sln
+*.sw?
@@ -0,0 +1,73 @@
+# React + TypeScript + Vite
+
+This template provides a minimal setup to get React working in Vite with HMR and some ESLint rules.
+
+Currently, two official plugins are available:
+
+- [@vitejs/plugin-react](https://github.com/vitejs/vite-plugin-react/blob/main/packages/plugin-react) uses [Babel](https://babeljs.io/) (or [oxc](https://oxc.rs) when used in [rolldown-vite](https://vite.dev/guide/rolldown)) for Fast Refresh
+- [@vitejs/plugin-react-swc](https://github.com/vitejs/vite-plugin-react/blob/main/packages/plugin-react-swc) uses [SWC](https://swc.rs/) for Fast Refresh
+
+## React Compiler
+
+The React Compiler is not enabled on this template because of its impact on dev & build performances. To add it, see [this documentation](https://react.dev/learn/react-compiler/installation).
+
+## Expanding the ESLint configuration
+
+If you are developing a production application, we recommend updating the configuration to enable type-aware lint rules:
+
+```js
+export default defineConfig([
+  globalIgnores(['dist']),
+  {
+    files: ['**/*.{ts,tsx}'],
+    extends: [
+      // Other configs...
+
+      // Remove tseslint.configs.recommended and replace with this
+      tseslint.configs.recommendedTypeChecked,
+      // Alternatively, use this for stricter rules
+      tseslint.configs.strictTypeChecked,
+      // Optionally, add this for stylistic rules
+      tseslint.configs.stylisticTypeChecked,
+
+      // Other configs...
+    ],
+    languageOptions: {
+      parserOptions: {
+        project: ['./tsconfig.node.json', './tsconfig.app.json'],
+        tsconfigRootDir: import.meta.dirname,
+      },
+      // other options...
+    },
+  },
+])
+```
+
+You can also install [eslint-plugin-react-x](https://github.com/Rel1cx/eslint-react/tree/main/packages/plugins/eslint-plugin-react-x) and [eslint-plugin-react-dom](https://github.com/Rel1cx/eslint-react/tree/main/packages/plugins/eslint-plugin-react-dom) for React-specific lint rules:
+
+```js
+// eslint.config.js
+import reactX from 'eslint-plugin-react-x'
+import reactDom from 'eslint-plugin-react-dom'
+
+export default defineConfig([
+  globalIgnores(['dist']),
+  {
+    files: ['**/*.{ts,tsx}'],
+    extends: [
+      // Other configs...
+      // Enable lint rules for React
+      reactX.configs['recommended-typescript'],
+      // Enable lint rules for React DOM
+      reactDom.configs.recommended,
+    ],
+    languageOptions: {
+      parserOptions: {
+        project: ['./tsconfig.node.json', './tsconfig.app.json'],
+        tsconfigRootDir: import.meta.dirname,
+      },
+      // other options...
+    },
+  },
+])
+```
@@ -0,0 +1,23 @@
+import js from '@eslint/js'
+import globals from 'globals'
+import reactHooks from 'eslint-plugin-react-hooks'
+import reactRefresh from 'eslint-plugin-react-refresh'
+import tseslint from 'typescript-eslint'
+import { defineConfig, globalIgnores } from 'eslint/config'
+
+export default defineConfig([
+  globalIgnores(['dist']),
+  {
+    files: ['**/*.{ts,tsx}'],
+    extends: [
+      js.configs.recommended,
+      tseslint.configs.recommended,
+      reactHooks.configs.flat.recommended,
+      reactRefresh.configs.vite,
+    ],
+    languageOptions: {
+      ecmaVersion: 2020,
+      globals: globals.browser,
+    },
+  },
+])
@@ -0,0 +1,13 @@
+<!doctype html>
+<html lang="en" class="dark">
+  <head>
+    <meta charset="UTF-8" />
+    <link rel="icon" type="image/svg+xml" href="/vite.svg" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>dashboard</title>
+  </head>
+  <body>
+    <div id="root"></div>
+    <script type="module" src="/src/main.tsx"></script>
+  </body>
+</html>
@@ -0,0 +1,42 @@
+{
+  "name": "dashboard",
+  "private": true,
+  "version": "0.0.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "tsc -b && vite build",
+    "lint": "eslint .",
+    "preview": "vite preview"
+  },
+  "dependencies": {
+    "@xyflow/react": "^12.10.1",
+    "class-variance-authority": "^0.7.1",
+    "clsx": "^2.1.1",
+    "lucide-react": "^0.577.0",
+    "react": "^19.2.0",
+    "react-dom": "^19.2.0",
+    "react-markdown": "^10.1.0",
+    "remark-gfm": "^4.0.1",
+    "tailwind-merge": "^3.5.0",
+    "zustand": "^5.0.11"
+  },
+  "devDependencies": {
+    "@eslint/js": "^9.39.1",
+    "@tailwindcss/typography": "^0.5.19",
+    "@types/node": "^24.10.1",
+    "@types/react": "^19.2.7",
+    "@types/react-dom": "^19.2.3",
+    "@vitejs/plugin-react": "^5.1.1",
+    "autoprefixer": "^10.4.27",
+    "eslint": "^9.39.1",
+    "eslint-plugin-react-hooks": "^7.0.1",
+    "eslint-plugin-react-refresh": "^0.4.24",
+    "globals": "^16.5.0",
+    "postcss": "^8.5.8",
+    "tailwindcss": "^3.4.19",
+    "typescript": "~5.9.3",
+    "typescript-eslint": "^8.48.0",
+    "vite": "^7.3.1"
+  }
+}
@@ -0,0 +1,6 @@
+export default {
+  plugins: {
+    tailwindcss: {},
+    autoprefixer: {},
+  },
+}
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" class="iconify iconify--logos" width="31.88" height="32" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 257"><defs><linearGradient id="IconifyId1813088fe1fbc01fb466" x1="-.828%" x2="57.636%" y1="7.652%" y2="78.411%"><stop offset="0%" stop-color="#41D1FF"></stop><stop offset="100%" stop-color="#BD34FE"></stop></linearGradient><linearGradient id="IconifyId1813088fe1fbc01fb467" x1="43.376%" x2="50.316%" y1="2.242%" y2="89.03%"><stop offset="0%" stop-color="#FFEA83"></stop><stop offset="8.333%" stop-color="#FFDD35"></stop><stop offset="100%" stop-color="#FFA800"></stop></linearGradient></defs><path fill="url(#IconifyId1813088fe1fbc01fb466)" d="M255.153 37.938L134.897 252.976c-2.483 4.44-8.862 4.466-11.382.048L.875 37.958c-2.746-4.814 1.371-10.646 6.827-9.67l120.385 21.517a6.537 6.537 0 0 0 2.322-.004l117.867-21.483c5.438-.991 9.574 4.796 6.877 9.62Z"></path><path fill="url(#IconifyId1813088fe1fbc01fb467)" d="M185.432.063L96.44 17.501a3.268 3.268 0 0 0-2.634 3.014l-5.474 92.456a3.268 3.268 0 0 0 3.997 3.378l24.777-5.718c2.318-.535 4.413 1.507 3.936 3.838l-7.361 36.047c-.495 2.426 1.782 4.5 4.151 3.78l15.304-4.649c2.372-.72 4.652 1.36 4.15 3.788l-11.698 56.621c-.732 3.542 3.979 5.473 5.943 2.437l1.313-2.028l72.516-144.72c1.215-2.423-.88-5.186-3.54-4.672l-25.505 4.922c-2.396.462-4.435-1.77-3.759-4.114l16.646-57.705c.677-2.35-1.37-4.583-3.769-4.113Z"></path></svg>
@@ -0,0 +1,42 @@
+#root {
+  max-width: 1280px;
+  margin: 0 auto;
+  padding: 2rem;
+  text-align: center;
+}
+
+.logo {
+  height: 6em;
+  padding: 1.5em;
+  will-change: filter;
+  transition: filter 300ms;
+}
+.logo:hover {
+  filter: drop-shadow(0 0 2em #646cffaa);
+}
+.logo.react:hover {
+  filter: drop-shadow(0 0 2em #61dafbaa);
+}
+
+@keyframes logo-spin {
+  from {
+    transform: rotate(0deg);
+  }
+  to {
+    transform: rotate(360deg);
+  }
+}
+
+@media (prefers-reduced-motion: no-preference) {
+  a:nth-of-type(2) .logo {
+    animation: logo-spin infinite 20s linear;
+  }
+}
+
+.card {
+  padding: 2em;
+}
+
+.read-the-docs {
+  color: #888;
+}
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" class="iconify iconify--logos" width="35.93" height="32" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 228"><path fill="#00D8FF" d="M210.483 73.824a171.49 171.49 0 0 0-8.24-2.597c.465-1.9.893-3.777 1.273-5.621c6.238-30.281 2.16-54.676-11.769-62.708c-13.355-7.7-35.196.329-57.254 19.526a171.23 171.23 0 0 0-6.375 5.848a155.866 155.866 0 0 0-4.241-3.917C100.759 3.829 77.587-4.822 63.673 3.233C50.33 10.957 46.379 33.89 51.995 62.588a170.974 170.974 0 0 0 1.892 8.48c-3.28.932-6.445 1.924-9.474 2.98C17.309 83.498 0 98.307 0 113.668c0 15.865 18.582 31.778 46.812 41.427a145.52 145.52 0 0 0 6.921 2.165a167.467 167.467 0 0 0-2.01 9.138c-5.354 28.2-1.173 50.591 12.134 58.266c13.744 7.926 36.812-.22 59.273-19.855a145.567 145.567 0 0 0 5.342-4.923a168.064 168.064 0 0 0 6.92 6.314c21.758 18.722 43.246 26.282 56.54 18.586c13.731-7.949 18.194-32.003 12.4-61.268a145.016 145.016 0 0 0-1.535-6.842c1.62-.48 3.21-.974 4.76-1.488c29.348-9.723 48.443-25.443 48.443-41.52c0-15.417-17.868-30.326-45.517-39.844Zm-6.365 70.984c-1.4.463-2.836.91-4.3 1.345c-3.24-10.257-7.612-21.163-12.963-32.432c5.106-11 9.31-21.767 12.459-31.957c2.619.758 5.16 1.557 7.61 2.4c23.69 8.156 38.14 20.213 38.14 29.504c0 9.896-15.606 22.743-40.946 31.14Zm-10.514 20.834c2.562 12.94 2.927 24.64 1.23 33.787c-1.524 8.219-4.59 13.698-8.382 15.893c-8.067 4.67-25.32-1.4-43.927-17.412a156.726 156.726 0 0 1-6.437-5.87c7.214-7.889 14.423-17.06 21.459-27.246c12.376-1.098 24.068-2.894 34.671-5.345a134.17 134.17 0 0 1 1.386 6.193ZM87.276 214.515c-7.882 2.783-14.16 2.863-17.955.675c-8.075-4.657-11.432-22.636-6.853-46.752a156.923 156.923 0 0 1 1.869-8.499c10.486 2.32 22.093 3.988 34.498 4.994c7.084 9.967 14.501 19.128 21.976 27.15a134.668 134.668 0 0 1-4.877 4.492c-9.933 8.682-19.886 14.842-28.658 17.94ZM50.35 144.747c-12.483-4.267-22.792-9.812-29.858-15.863c-6.35-5.437-9.555-10.836-9.555-15.216c0-9.322 13.897-21.212 37.076-29.293c2.813-.98 5.757-1.905 8.812-2.773c3.204 10.42 7.406 21.315 12.477 32.332c-5.137 11.18-9.399 22.249-12.634 32.792a134.718 134.718 0 0 1-6.318-1.979Zm12.378-84.26c-4.811-24.587-1.616-43.134 6.425-47.789c8.564-4.958 27.502 2.111 47.463 19.835a144.318 144.318 0 0 1 3.841 3.545c-7.438 7.987-14.787 17.08-21.808 26.988c-12.04 1.116-23.565 2.908-34.161 5.309a160.342 160.342 0 0 1-1.76-7.887Zm110.427 27.268a347.8 347.8 0 0 0-7.785-12.803c8.168 1.033 15.994 2.404 23.343 4.08c-2.206 7.072-4.956 14.465-8.193 22.045a381.151 381.151 0 0 0-7.365-13.322Zm-45.032-43.861c5.044 5.465 10.096 11.566 15.065 18.186a322.04 322.04 0 0 0-30.257-.006c4.974-6.559 10.069-12.652 15.192-18.18ZM82.802 87.83a323.167 323.167 0 0 0-7.227 13.238c-3.184-7.553-5.909-14.98-8.134-22.152c7.304-1.634 15.093-2.97 23.209-3.984a321.524 321.524 0 0 0-7.848 12.897Zm8.081 65.352c-8.385-.936-16.291-2.203-23.593-3.793c2.26-7.3 5.045-14.885 8.298-22.6a321.187 321.187 0 0 0 7.257 13.246c2.594 4.48 5.28 8.868 8.038 13.147Zm37.542 31.03c-5.184-5.592-10.354-11.779-15.403-18.433c4.902.192 9.899.29 14.978.29c5.218 0 10.376-.117 15.453-.343c-4.985 6.774-10.018 12.97-15.028 18.486Zm52.198-57.817c3.422 7.8 6.306 15.345 8.596 22.52c-7.422 1.694-15.436 3.058-23.88 4.071a382.417 382.417 0 0 0 7.859-13.026a347.403 347.403 0 0 0 7.425-13.565Zm-16.898 8.101a358.557 358.557 0 0 1-12.281 19.815a329.4 329.4 0 0 1-23.444.823c-7.967 0-15.716-.248-23.178-.732a310.202 310.202 0 0 1-12.513-19.846h.001a307.41 307.41 0 0 1-10.923-20.627a310.278 310.278 0 0 1 10.89-20.637l-.001.001a307.318 307.318 0 0 1 12.413-19.761c7.613-.576 15.42-.876 23.31-.876H128c7.926 0 15.743.303 23.354.883a329.357 329.357 0 0 1 12.335 19.695a358.489 358.489 0 0 1 11.036 20.54a329.472 329.472 0 0 1-11 20.722Zm22.56-122.124c8.572 4.944 11.906 24.881 6.52 51.026c-.344 1.668-.73 3.367-1.15 5.09c-10.622-2.452-22.155-4.275-34.23-5.408c-7.034-10.017-14.323-19.124-21.64-27.008a160.789 160.789 0 0 1 5.888-5.4c18.9-16.447 36.564-22.941 44.612-18.3ZM128 90.808c12.625 0 22.86 10.235 22.86 22.86s-10.235 22.86-22.86 22.86s-22.86-10.235-22.86-22.86s10.235-22.86 22.86-22.86Z"></path></svg>
@@ -0,0 +1,196 @@
+import { useMemo, useState } from 'react'
+import { useTaskStore } from '../store/tasks'
+import type { Task } from '../types'
+
+const STATUS_ORDER: Record<string, number> = {
+  running: 0, waiting_approval: 1, pending: 2, failed: 3, done: 4, vetoed: 5,
+}
+
+const STATUS_CLS: Record<string, string> = {
+  running:          'bg-blue-100 text-blue-800 dark:bg-blue-900 dark:text-blue-200',
+  waiting_approval: 'bg-orange-100 text-orange-800 dark:bg-orange-900 dark:text-orange-200',
+  pending:          'bg-slate-100 text-slate-600 dark:bg-gray-800 dark:text-gray-400',
+  failed:           'bg-red-100 text-red-800 dark:bg-red-900 dark:text-red-200',
+  done:             'bg-green-100 text-green-800 dark:bg-green-900 dark:text-green-200',
+  vetoed:           'bg-purple-100 text-purple-800 dark:bg-purple-900 dark:text-purple-200',
+}
+
+const STATUS_LABEL: Record<string, string> = {
+  running: '运行中', waiting_approval: '待审批',
+  pending: '排队中', failed: '失败', done: '完成', vetoed: '已否决',
+}
+
+const ROLE_ABBR: Record<string, string> = {
+  developer: 'Dev', architect: 'Arch', tester: 'Test',
+  'market-pm': 'Mkt', planner: 'Plan', productizer: 'Prod',
+  'rtp-researcher': 'RTP', 'net-researcher': 'Net', 'kernel-analyzer': 'Kern',
+  'os-engineer': 'OS', 'hw-engineer': 'HW', 'dev-kernel': 'KDev',
+  'base-architect': 'BArch', 'base-validator': 'BVal',
+  'vision-analyst': 'Vision', 'media-producer': 'Media',
+}
+
+type FilterStatus = 'active' | 'all' | 'done'
+
+function elapsed(since: string | null): string {
+  if (!since) return ''
+  const secs = Math.floor((Date.now() - new Date(since).getTime()) / 1000)
+  if (secs < 60) return `${secs}s`
+  if (secs < 3600) return `${Math.floor(secs / 60)}m`
+  return `${Math.floor(secs / 3600)}h${Math.floor((secs % 3600) / 60)}m`
+}
+
+export function ActiveTasksView() {
+  const tasksMap = useTaskStore(s => s.tasks)
+  const setSelectedTask = useTaskStore(s => s.setSelectedTask)
+  const [filter, setFilter] = useState<FilterStatus>('active')
+  const [search, setSearch] = useState('')
+
+  const tasks = useMemo(() => {
+    let arr = Array.from(tasksMap.values())
+    if (filter === 'active') {
+      arr = arr.filter(t => ['running', 'waiting_approval', 'pending'].includes(t.status))
+    } else if (filter === 'done') {
+      arr = arr.filter(t => ['done', 'failed', 'vetoed'].includes(t.status))
+        .sort((a, b) => (b.completed_at ?? '').localeCompare(a.completed_at ?? ''))
+        .slice(0, 50)
+    }
+    if (search.trim()) {
+      const q = search.toLowerCase()
+      arr = arr.filter(t =>
+        t.title.toLowerCase().includes(q) ||
+        t.project.toLowerCase().includes(q) ||
+        t.agent_role.toLowerCase().includes(q)
+      )
+    }
+    return arr.sort((a, b) =>
+      (STATUS_ORDER[a.status] ?? 9) - (STATUS_ORDER[b.status] ?? 9) ||
+      a.priority - b.priority ||
+      b.id - a.id
+    )
+  }, [tasksMap, filter, search])
+
+  const counts = useMemo(() => {
+    const all = Array.from(tasksMap.values())
+    return {
+      running: all.filter(t => t.status === 'running').length,
+      waiting: all.filter(t => t.status === 'waiting_approval').length,
+      pending: all.filter(t => t.status === 'pending').length,
+    }
+  }, [tasksMap])
+
+  return (
+    <div className="p-4 max-w-5xl mx-auto">
+      {/* 概览计数条 */}
+      <div className="flex items-center gap-4 mb-4 text-xs">
+        <div className="flex items-center gap-1.5">
+          <span className="w-2 h-2 rounded-full bg-blue-400 animate-pulse inline-block" />
+          <span className="text-slate-600 dark:text-slate-400">运行中 <strong className="text-slate-900 dark:text-white">{counts.running}</strong></span>
+        </div>
+        {counts.waiting > 0 && (
+          <div className="flex items-center gap-1.5">
+            <span className="w-2 h-2 rounded-full bg-orange-400 inline-block" />
+            <span className="text-slate-600 dark:text-slate-400">待审批 <strong className="text-orange-500">{counts.waiting}</strong></span>
+          </div>
+        )}
+        <div className="flex items-center gap-1.5">
+          <span className="w-2 h-2 rounded-full bg-slate-400 inline-block" />
+          <span className="text-slate-600 dark:text-slate-400">排队中 <strong className="text-slate-900 dark:text-white">{counts.pending}</strong></span>
+        </div>
+      </div>
+
+      {/* 过滤栏 */}
+      <div className="flex items-center gap-2 mb-3">
+        <div className="flex gap-1 bg-slate-100 dark:bg-gray-800 rounded-lg p-0.5">
+          {(['active', 'all', 'done'] as FilterStatus[]).map(f => (
+            <button key={f}
+              onClick={() => setFilter(f)}
+              className={`px-3 py-1 text-xs rounded-md transition-colors
+                ${filter === f
+                  ? 'bg-white dark:bg-gray-700 text-slate-900 dark:text-white shadow-sm'
+                  : 'text-slate-500 dark:text-gray-500 hover:text-slate-800 dark:hover:text-gray-300'
+                }`}
+            >
+              {{ active: '进行中', all: '全部', done: '已完成' }[f]}
+            </button>
+          ))}
+        </div>
+        <input
+          value={search}
+          onChange={e => setSearch(e.target.value)}
+          placeholder="搜索任务/项目/角色..."
+          className="flex-1 max-w-xs bg-white dark:bg-gray-800 border border-slate-200 dark:border-gray-700
+                     rounded-lg px-3 py-1 text-xs text-slate-700 dark:text-gray-300
+                     focus:outline-none focus:ring-1 focus:ring-blue-500"
+        />
+        <span className="text-xs text-slate-400">{tasks.length} 条</span>
+      </div>
+
+      {/* 任务列表 */}
+      <div className="bg-white dark:bg-gray-900 rounded-xl border border-slate-200 dark:border-gray-700 overflow-hidden">
+        {tasks.length === 0 ? (
+          <div className="text-center py-12 text-sm text-slate-400 dark:text-gray-500">
+            {filter === 'active' ? '当前没有进行中的任务' : '没有匹配的任务'}
+          </div>
+        ) : (
+          tasks.map(task => <TaskRow key={task.id} task={task} onClick={() => setSelectedTask(task.id)} />)
+        )}
+      </div>
+    </div>
+  )
+}
+
+function TaskRow({ task, onClick }: { task: Task; onClick: () => void }) {
+  const isRunning = task.status === 'running'
+  return (
+    <button
+      onClick={onClick}
+      className="w-full px-4 py-2.5 border-b border-slate-100 dark:border-gray-800 last:border-0
+                 flex items-center gap-3 text-left hover:bg-slate-50 dark:hover:bg-gray-800/50
+                 transition-colors group"
+    >
+      {/* 状态 badge */}
+      <span className={`text-[10px] px-1.5 py-0.5 rounded-full shrink-0 font-medium
+                        ${STATUS_CLS[task.status] ?? STATUS_CLS.pending}`}>
+        {STATUS_LABEL[task.status] ?? task.status}
+      </span>
+
+      {/* 项目 badge */}
+      <span className="text-[10px] px-1.5 py-0.5 bg-slate-100 dark:bg-gray-800
+                       text-slate-500 dark:text-gray-400 rounded font-mono shrink-0">
+        {task.project}
+      </span>
+
+      {/* 标题 */}
+      <div className="flex-1 min-w-0">
+        <div className={`text-sm truncate ${
+          isRunning ? 'text-slate-900 dark:text-white' : 'text-slate-700 dark:text-slate-300'
+        }`}>
+          {isRunning && (
+            <span className="inline-block w-1.5 h-1.5 rounded-full bg-blue-400 animate-pulse mr-1.5 align-middle" />
+          )}
+          {task.title}
+        </div>
+      </div>
+
+      {/* 角色 */}
+      <span className="text-[10px] text-slate-400 dark:text-gray-500 shrink-0 font-mono">
+        {ROLE_ABBR[task.agent_role] ?? task.agent_role}
+      </span>
+
+      {/* P 级 */}
+      <span className="text-[10px] text-slate-400 dark:text-gray-500 shrink-0">
+        P{task.priority}
+      </span>
+
+      {/* 计时 */}
+      {(task.started_at || task.completed_at) && (
+        <span className="text-[10px] text-slate-400 dark:text-gray-500 shrink-0 w-10 text-right">
+          {task.status === 'running'
+            ? elapsed(task.started_at)
+            : task.completed_at?.slice(5, 16)
+          }
+        </span>
+      )}
+    </button>
+  )
+}
@@ -0,0 +1,241 @@
+import { useEffect, useState, useCallback } from 'react'
+import {
+  fetchGlobalConfig, updateGlobalConfig,
+  fetchProjectConfig, updateProjectConfig,
+  refreshClaudeMd,
+} from '../lib/api'
+
+interface Props {
+  activeProject: string | null
+}
+
+interface FileSection {
+  label: string
+  badge: string
+  badgeColor: string
+  content: string
+  path: string
+  exists: boolean
+  editable: boolean
+  warningOnEdit?: string
+  onSave?: (content: string) => Promise<void>
+}
+
+export function ConfigView({ activeProject }: Props) {
+  const [sections, setSections] = useState<FileSection[]>([])
+  const [loading, setLoading] = useState(true)
+  const [error, setError] = useState('')
+  const [refreshing, setRefreshing] = useState(false)
+  const [refreshMsg, setRefreshMsg] = useState('')
+
+  async function handleRefreshClaudeMd() {
+    if (!activeProject) return
+    setRefreshing(true)
+    setRefreshMsg('')
+    try {
+      const task = await refreshClaudeMd(activeProject)
+      setRefreshMsg(`✓ 已创建刷新任务 #${task.id}，AI 将根据最新调研结论更新 CLAUDE.md`)
+    } catch (e) {
+      setRefreshMsg(`✕ 失败：${e}`)
+    } finally {
+      setRefreshing(false)
+    }
+  }
+
+  const load = useCallback(async () => {
+    setLoading(true)
+    setError('')
+    try {
+      const result: FileSection[] = []
+
+      // 全局 CLAUDE.md
+      try {
+        const g = await fetchGlobalConfig()
+        result.push({
+          label: '全局 CLAUDE.md',
+          badge: 'Global',
+          badgeColor: 'bg-purple-700 text-purple-100',
+          content: g.content,
+          path: g.path,
+          exists: g.exists,
+          editable: true,
+          warningOnEdit: '⚠ 这是全局配置，修改将影响所有项目。确认要继续吗？',
+          onSave: updateGlobalConfig,
+        })
+      } catch {
+        result.push({
+          label: '全局 CLAUDE.md',
+          badge: 'Global',
+          badgeColor: 'bg-purple-700 text-purple-100',
+          content: '',
+          path: '未配置（容器未挂载 ~/.claude）',
+          exists: false,
+          editable: false,
+        })
+      }
+
+      // 项目级配置（仅在选中项目时显示）
+      if (activeProject) {
+        try {
+          const p = await fetchProjectConfig(activeProject)
+          result.push({
+            label: `${activeProject} / CLAUDE.md`,
+            badge: '项目',
+            badgeColor: 'bg-blue-700 text-blue-100',
+            content: p.claude.content,
+            path: p.claude.path,
+            exists: p.claude.exists,
+            editable: true,
+            onSave: (c) => updateProjectConfig(activeProject, 'claude', c),
+          })
+          result.push({
+            label: `${activeProject} / README.md`,
+            badge: '项目',
+            badgeColor: 'bg-green-700 text-green-100',
+            content: p.readme.content,
+            path: p.readme.path,
+            exists: p.readme.exists,
+            editable: true,
+            onSave: (c) => updateProjectConfig(activeProject, 'readme', c),
+          })
+        } catch (e) {
+          setError(String(e))
+        }
+      }
+
+      setSections(result)
+    } catch (e) {
+      setError(String(e))
+    } finally {
+      setLoading(false)
+    }
+  }, [activeProject])
+
+  useEffect(() => { load() }, [load])
+
+  if (loading) return <div className="p-6 text-slate-500 dark:text-gray-400 text-sm">加载中...</div>
+
+  return (
+    <div className="p-4 space-y-4 max-w-4xl mx-auto">
+      {!activeProject && (
+        <div className="text-xs text-slate-400 dark:text-gray-500 bg-slate-100/60 dark:bg-gray-800/60 rounded-lg px-3 py-2">
+          选择左上角项目标签后，可查看并编辑该项目的 CLAUDE.md 和 README.md。当前仅显示全局配置。
+        </div>
+      )}
+      {activeProject && (
+        <div className="flex items-center gap-3">
+          <button
+            onClick={handleRefreshClaudeMd}
+            disabled={refreshing}
+            className="px-4 py-1.5 text-xs bg-amber-700 hover:bg-amber-600 text-white rounded-lg disabled:opacity-50 transition-colors"
+          >
+            {refreshing ? '创建任务中...' : '🔄 刷新项目指导'}
+          </button>
+          {refreshMsg && (
+            <span className={`text-xs ${refreshMsg.startsWith('✓') ? 'text-green-400' : 'text-red-400'}`}>
+              {refreshMsg}
+            </span>
+          )}
+        </div>
+      )}
+      {error && <div className="text-xs text-red-400 bg-red-900/20 rounded px-3 py-2">{error}</div>}
+      {sections.map((s, i) => (
+        <SectionCard key={i} section={s} onReload={load} />
+      ))}
+    </div>
+  )
+}
+
+function SectionCard({ section, onReload }: { section: FileSection; onReload: () => void }) {
+  const [editing, setEditing] = useState(false)
+  const [draft, setDraft] = useState(section.content)
+  const [saving, setSaving] = useState(false)
+  const [saveError, setSaveError] = useState('')
+
+  // keep draft in sync when section reloads
+  useEffect(() => { setDraft(section.content); setEditing(false) }, [section.content])
+
+  async function handleEdit() {
+    if (section.warningOnEdit && !confirm(section.warningOnEdit)) return
+    setEditing(true)
+  }
+
+  async function handleSave() {
+    if (!section.onSave) return
+    setSaving(true)
+    setSaveError('')
+    try {
+      await section.onSave(draft)
+      setEditing(false)
+      onReload()
+    } catch (e) {
+      setSaveError(String(e))
+    } finally {
+      setSaving(false)
+    }
+  }
+
+  return (
+    <div className="bg-white dark:bg-gray-900 border border-slate-200 dark:border-gray-700 rounded-xl overflow-hidden">
+      {/* Header */}
+      <div className="flex items-center gap-2 px-4 py-2.5 border-b border-slate-200 dark:border-gray-700 bg-slate-100/50 dark:bg-gray-800/50">
+        <span className={`text-[10px] font-semibold px-2 py-0.5 rounded-full ${section.badgeColor}`}>
+          {section.badge}
+        </span>
+        <span className="text-sm font-medium text-slate-800 dark:text-gray-200">{section.label}</span>
+        <span className="text-xs text-slate-400 dark:text-gray-500 ml-1 truncate">{section.path}</span>
+        {!section.exists && (
+          <span className="text-xs text-yellow-600 ml-auto">（文件不存在）</span>
+        )}
+        <div className="ml-auto flex gap-2">
+          {editing ? (
+            <>
+              <button
+                onClick={() => { setEditing(false); setDraft(section.content) }}
+                className="text-xs text-slate-500 dark:text-gray-400 hover:text-slate-800 dark:hover:text-gray-200 px-2 py-1"
+              >
+                取消
+              </button>
+              <button
+                onClick={handleSave}
+                disabled={saving}
+                className="text-xs bg-blue-700 hover:bg-blue-600 text-white px-3 py-1 rounded disabled:opacity-50"
+              >
+                {saving ? '保存中...' : '保存'}
+              </button>
+            </>
+          ) : section.editable ? (
+            <button
+              onClick={handleEdit}
+              className="text-xs text-slate-500 dark:text-gray-400 hover:text-slate-800 dark:hover:text-gray-200 px-2 py-1 rounded hover:bg-slate-200 dark:hover:bg-gray-700"
+            >
+              编辑
+            </button>
+          ) : null}
+        </div>
+      </div>
+
+      {saveError && (
+        <div className="px-4 py-2 text-xs text-red-400 bg-red-900/20">{saveError}</div>
+      )}
+
+      {/* Content */}
+      <div className="relative">
+        {editing ? (
+          <textarea
+            value={draft}
+            onChange={e => setDraft(e.target.value)}
+            className="w-full bg-slate-50 dark:bg-gray-950 text-slate-800 dark:text-gray-200 text-xs font-mono p-4 min-h-[300px] resize-y
+                       focus:outline-none focus:ring-1 focus:ring-blue-600"
+            spellCheck={false}
+          />
+        ) : (
+          <pre className="text-xs text-slate-700 dark:text-gray-300 font-mono p-4 overflow-x-auto whitespace-pre-wrap
+                          max-h-[400px] overflow-y-auto">
+            {section.content || <span className="text-slate-300 dark:text-gray-600 italic">（空文件）</span>}
+          </pre>
+        )}
+      </div>
+    </div>
+  )
+}
@@ -0,0 +1,233 @@
+import { useState } from 'react'
+import { importProject, fetchProjects } from '../lib/api'
+import { useTaskStore } from '../store/tasks'
+
+interface Props { onClose: () => void }
+
+const MODES = ['auto', 'confirm', 'report']
+
+const TASK_TEMPLATE = [
+  { priority: 'P1', label: '全量代码审查', role: 'developer', desc: '架构设计与技术债务识别' },
+  { priority: 'P1', label: '测试覆盖率补全', role: 'tester', desc: '关键路径 + mock 无依赖 CI' },
+  { priority: 'P2', label: '模块重构', role: 'developer', desc: '接口规范化与解耦' },
+  { priority: 'P2', label: '性能分析', role: 'developer', desc: '推理瓶颈识别 + 向量化优化' },
+  { priority: 'P2', label: '板端集成测试', role: 'tester', desc: 'RKNN 部署验证（如含 RKNN）' },
+  { priority: 'P2', label: '跨项目数据对接', role: 'developer', desc: '关联项目数据流集成（若填写）' },
+  { priority: 'P3', label: '架构评估', role: 'architect', desc: '耦合度与扩展性路线图' },
+  { priority: 'P3', label: '文档完善', role: 'developer', desc: 'README/CLAUDE.md 更新' },
+]
+
+const ROLE_COLOR: Record<string, string> = {
+  developer: 'text-blue-400',
+  tester:    'text-green-400',
+  architect: 'text-purple-400',
+}
+
+export function ImportProjectModal({ onClose }: Props) {
+  const setProjects = useTaskStore(s => s.setProjects)
+  const [form, setForm] = useState({
+    path: '',
+    name: '',
+    mode: 'auto',
+    description: '',
+    scan_now: true,
+    linked_project: '',
+  })
+  const [loading, setLoading] = useState(false)
+  const [error, setError] = useState('')
+  const [result, setResult] = useState<{ tasks_created: number; tasks: { title: string }[] } | null>(null)
+  const [showTemplate, setShowTemplate] = useState(false)
+
+  function handlePathChange(val: string) {
+    const derived = val.replace(/\/$/, '').split('/').pop() ?? ''
+    setForm(f => ({ ...f, path: val, name: f.name || derived }))
+  }
+
+  async function submit(e: React.FormEvent) {
+    e.preventDefault()
+    if (!form.path.trim()) { setError('目录路径不能为空'); return }
+    setLoading(true)
+    setError('')
+    try {
+      const res = await importProject({
+        path: form.path.trim(),
+        name: form.name.trim() || undefined,
+        mode: form.mode,
+        description: form.description.trim(),
+        scan_now: form.scan_now,
+        linked_project: form.linked_project.trim() || undefined,
+      })
+      const projects = await fetchProjects()
+      setProjects(projects)
+      if (res.tasks_created > 0) {
+        setResult(res)
+      } else {
+        onClose()
+      }
+    } catch (e) {
+      setError(String(e))
+    } finally {
+      setLoading(false)
+    }
+  }
+
+  // 成功后显示任务创建结果
+  if (result) {
+    return (
+      <div className="fixed inset-0 bg-black/60 z-50 flex items-center justify-center">
+        <div className="bg-white dark:bg-gray-900 border border-slate-200 dark:border-gray-700 rounded-xl p-6 w-[520px] shadow-2xl">
+          <div className="flex items-center gap-2 mb-3">
+            <span className="text-green-400 text-lg">✓</span>
+            <h2 className="text-base font-semibold text-slate-900 dark:text-gray-100">
+              项目已导入，{result.tasks_created} 个任务已入队
+            </h2>
+          </div>
+          <div className="space-y-1 max-h-60 overflow-y-auto mb-4">
+            {result.tasks.map((t, i) => (
+              <div key={i} className="text-xs text-slate-700 dark:text-gray-300 bg-slate-100 dark:bg-gray-800 rounded px-3 py-1.5">
+                {t.title}
+              </div>
+            ))}
+          </div>
+          <p className="text-xs text-slate-400 dark:text-gray-500 mb-4">
+            点击 TopBar 的「开始执行」按钮启动自动执行，或在 Kanban 中查看任务状态。
+          </p>
+          <div className="flex justify-end">
+            <button onClick={onClose}
+              className="px-4 py-1.5 text-sm bg-blue-700 hover:bg-blue-600 text-white rounded">
+              完成
+            </button>
+          </div>
+        </div>
+      </div>
+    )
+  }
+
+  return (
+    <div className="fixed inset-0 bg-black/60 z-50 flex items-center justify-center">
+      <div className="bg-white dark:bg-gray-900 border border-slate-200 dark:border-gray-700 rounded-xl p-6 w-[540px] shadow-2xl">
+        <h2 className="text-base font-semibold text-slate-900 dark:text-gray-100 mb-1">导入项目目录</h2>
+        <p className="text-xs text-slate-400 dark:text-gray-500 mb-4">
+          自动检测项目特征，生成适合软硬件协同研发的默认任务集
+        </p>
+
+        <form onSubmit={submit} className="space-y-3">
+          <Row label="目录路径">
+            <input
+              value={form.path}
+              onChange={e => handlePathChange(e.target.value)}
+              className={inputCls}
+              placeholder="/data/rockchip/my-project"
+            />
+          </Row>
+          <Row label="项目名称">
+            <input
+              value={form.name}
+              onChange={e => setForm(f => ({ ...f, name: e.target.value }))}
+              className={inputCls}
+              placeholder="自动从路径推断"
+            />
+          </Row>
+          <Row label="模式">
+            <select
+              value={form.mode}
+              onChange={e => setForm(f => ({ ...f, mode: e.target.value }))}
+              className={inputCls}
+            >
+              {MODES.map(m => (
+                <option key={m} value={m}>
+                  {m === 'auto'    ? 'auto — 全自动持续开发（推荐）' :
+                   m === 'confirm' ? 'confirm — 每步需审批' :
+                   'report — 仅分析不修改'}
+                </option>
+              ))}
+            </select>
+          </Row>
+          <Row label="关联项目">
+            <input
+              value={form.linked_project}
+              onChange={e => setForm(f => ({ ...f, linked_project: e.target.value }))}
+              className={inputCls}
+              placeholder="/data/test-platform/logo-collect（可选，生成集成任务）"
+            />
+          </Row>
+          <Row label="描述">
+            <input
+              value={form.description}
+              onChange={e => setForm(f => ({ ...f, description: e.target.value }))}
+              className={inputCls}
+              placeholder="项目简短描述（可选）"
+            />
+          </Row>
+
+          {/* 任务模板开关 */}
+          <div className="flex items-center gap-2 pl-[108px]">
+            <input
+              type="checkbox"
+              id="scan_now"
+              checked={form.scan_now}
+              onChange={e => setForm(f => ({ ...f, scan_now: e.target.checked }))}
+              className="accent-blue-500"
+            />
+            <label htmlFor="scan_now" className="text-xs text-slate-500 dark:text-gray-400 cursor-pointer">
+              生成默认任务集（根据项目特征自动适配）
+            </label>
+            {form.scan_now && (
+              <button
+                type="button"
+                onClick={() => setShowTemplate(v => !v)}
+                className="text-xs text-blue-400 hover:text-blue-300 ml-1"
+              >
+                {showTemplate ? '收起' : '预览任务'}
+              </button>
+            )}
+          </div>
+
+          {/* 任务模板预览 */}
+          {form.scan_now && showTemplate && (
+            <div className="ml-[108px] bg-slate-100/60 dark:bg-gray-800/60 rounded-lg p-3 space-y-1.5">
+              {TASK_TEMPLATE.map((t, i) => (
+                <div key={i} className="flex items-start gap-2 text-xs">
+                  <span className={`text-[10px] font-mono px-1.5 py-0.5 rounded shrink-0
+                    ${t.priority === 'P1' ? 'bg-red-900/60 text-red-300' :
+                      t.priority === 'P2' ? 'bg-yellow-900/60 text-yellow-300' :
+                      'bg-slate-200 dark:bg-gray-700 text-slate-500 dark:text-gray-400'}`}>
+                    {t.priority}
+                  </span>
+                  <span className="text-slate-800 dark:text-gray-200 font-medium shrink-0">{t.label}</span>
+                  <span className={`shrink-0 ${ROLE_COLOR[t.role]}`}>[{t.role}]</span>
+                  <span className="text-slate-400 dark:text-gray-500">{t.desc}</span>
+                </div>
+              ))}
+            </div>
+          )}
+
+          {error && <p className="text-xs text-red-400">{error}</p>}
+
+          <div className="flex justify-end gap-2 pt-2">
+            <button type="button" onClick={onClose}
+              className="px-4 py-1.5 text-sm text-slate-500 dark:text-gray-400 hover:text-slate-800 dark:hover:text-gray-200">
+              取消
+            </button>
+            <button type="submit" disabled={loading}
+              className="px-4 py-1.5 text-sm bg-blue-700 hover:bg-blue-600 text-white
+                         rounded disabled:opacity-50">
+              {loading ? '导入中...' : '导入并生成任务'}
+            </button>
+          </div>
+        </form>
+      </div>
+    </div>
+  )
+}
+
+const inputCls = 'w-full bg-slate-100 dark:bg-gray-800 border border-slate-300 dark:border-gray-600 rounded px-2 py-1.5 text-xs text-slate-800 dark:text-gray-200'
+
+function Row({ label, children }: { label: string; children: React.ReactNode }) {
+  return (
+    <div className="grid grid-cols-[100px_1fr] items-start gap-2">
+      <label className="text-xs text-slate-500 dark:text-gray-400 pt-1.5">{label}</label>
+      {children}
+    </div>
+  )
+}
@@ -1,7 +1,10 @@
 import { useEffect, useState } from 'react'
+import ReactMarkdown from 'react-markdown'
+import remarkGfm from 'remark-gfm'
 import {
  fetchMarketFacts, fetchMarketHistory, triggerMarketInsight,
-  type MarketFacts, type MarketHistoryItem,
+  fetchResearchFacts, fetchTaskLog,
+  type MarketFacts, type MarketHistoryItem, type ResearchFacts,
 } from '../lib/api'

 const GOAL_META: Record<string, { icon: string; label: string; color: string }> = {
@@ -53,6 +56,12 @@ function GoalCard({ goal, facts, onTrigger }: {
              {facts.trends}
            </div>
          )}
+          {facts.opportunities && (
+            <div>
+              <span className="font-medium text-slate-500 dark:text-slate-400">机会  </span>
+              {facts.opportunities}
+            </div>
+          )}
          {facts.updated_at && (
            <div className="text-slate-400 dark:text-slate-500 text-right">
              更新于 {facts.updated_at.slice(0, 16)}
@@ -69,6 +78,23 @@ function GoalCard({ goal, facts, onTrigger }: {
 }

 function HistoryRow({ item }: { item: MarketHistoryItem }) {
+  const [expanded, setExpanded] = useState(false)
+  const [log, setLog] = useState<string[]>([])
+  const [logLoading, setLogLoading] = useState(false)
+
+  async function toggle() {
+    if (!expanded && log.length === 0) {
+      setLogLoading(true)
+      try {
+        const r = await fetchTaskLog(item.id)
+        setLog(r.lines)
+      } catch { /* ignore */ } finally {
+        setLogLoading(false)
+      }
+    }
+    setExpanded(e => !e)
+  }
+
  const statusCls: Record<string, string> = {
    done: 'bg-green-100 text-green-800 dark:bg-green-900 dark:text-green-200',
    failed: 'bg-red-100 text-red-800 dark:bg-red-900 dark:text-red-200',
@@ -76,34 +102,105 @@ function HistoryRow({ item }: { item: MarketHistoryItem }) {
    pending: 'bg-slate-100 text-slate-600',
  }
  return (
-    <div className="px-4 py-3 border-b border-slate-100 dark:border-gray-800 flex items-start gap-3">
-      <span className={`text-xs px-2 py-0.5 rounded-full shrink-0 ${statusCls[item.status] ?? statusCls.pending}`}>
-        {item.status}
-      </span>
-      <div className="flex-1 min-w-0">
-        <div className="text-sm text-slate-800 dark:text-slate-200 truncate">{item.title}</div>
-        {item.result_summary && (
-          <div className="text-xs text-slate-500 dark:text-slate-400 mt-0.5 line-clamp-2">
-            {item.result_summary}
-          </div>
-        )}
-      </div>
-      {item.completed_at && (
-        <span className="text-xs text-slate-400 shrink-0">{item.completed_at.slice(0, 10)}</span>
+    <div className="border-b border-slate-100 dark:border-gray-800">
+      <button
+        onClick={toggle}
+        className="w-full px-4 py-3 flex items-start gap-3 text-left hover:bg-slate-50 dark:hover:bg-gray-800/50 transition-colors"
+      >
+        <span className={`text-xs px-2 py-0.5 rounded-full shrink-0 mt-0.5 ${statusCls[item.status] ?? statusCls.pending}`}>
+          {item.status}
+        </span>
+        <div className="flex-1 min-w-0">
+          <div className="text-sm text-slate-800 dark:text-slate-200 truncate">{item.title}</div>
+          {item.result_summary && !expanded && (
+            <div className="text-xs text-slate-500 dark:text-slate-400 mt-0.5 line-clamp-2">
+              {item.result_summary}
+            </div>
+          )}
+        </div>
+        <div className="flex items-center gap-2 shrink-0">
+          {item.completed_at && (
+            <span className="text-xs text-slate-400">{item.completed_at.slice(0, 10)}</span>
+          )}
+          <span className="text-xs text-slate-400">{expanded ? '▲' : '▼'}</span>
+        </div>
+      </button>
+      {expanded && (
+        <div className="px-4 pb-4">
+          {logLoading ? (
+            <div className="text-xs text-slate-400 py-2">加载中...</div>
+          ) : (() => {
+            const content = log.length > 0 ? log.join('\n') : item.result_summary
+            return content ? (
+              <div className="max-h-[32rem] overflow-y-auto bg-slate-50 dark:bg-gray-950 rounded-lg p-4
+                              prose prose-sm dark:prose-invert max-w-none
+                              prose-table:text-xs prose-td:py-1 prose-th:py-1
+                              prose-headings:text-slate-800 dark:prose-headings:text-slate-100
+                              prose-p:text-slate-700 dark:prose-p:text-slate-300
+                              prose-code:text-xs prose-code:bg-slate-200 dark:prose-code:bg-gray-800
+                              prose-code:px-1 prose-code:rounded">
+                <ReactMarkdown remarkPlugins={[remarkGfm]}>{content}</ReactMarkdown>
+              </div>
+            ) : (
+              <div className="text-xs text-slate-400 py-2">暂无内容</div>
+            )
+          })()}
+        </div>
      )}
    </div>
  )
 }

+const DOMAIN_META: Record<string, { icon: string; label: string }> = {
+  rtp:    { icon: '📡', label: 'RTP/RTCP 实时传输' },
+  net:    { icon: '🌐', label: '网络栈调优' },
+  kernel: { icon: '⚙️', label: '内核调度分析' },
+}
+
+function ResearchSection({ research }: { research: ResearchFacts }) {
+  const domains = Object.entries(research.domains)
+  if (domains.length === 0) {
+    return (
+      <div className="text-xs text-slate-400 dark:text-gray-500 text-center py-6">
+        暂无调研数据，基础技术组完成调研后自动展示
+      </div>
+    )
+  }
+  return (
+    <div className="grid grid-cols-1 md:grid-cols-3 gap-4">
+      {domains.map(([domain, facts]) => {
+        const meta = DOMAIN_META[domain] ?? { icon: '🔬', label: domain }
+        return (
+          <div key={domain}
+            className="rounded-xl border border-slate-200 dark:border-gray-700 bg-white dark:bg-gray-900 p-4 space-y-2">
+            <div className="flex items-center gap-2 mb-3">
+              <span className="text-lg">{meta.icon}</span>
+              <span className="text-sm font-semibold text-slate-900 dark:text-white">{meta.label}</span>
+            </div>
+            {Object.entries(facts).slice(0, 6).map(([k, v]) => (
+              <div key={k} className="text-xs">
+                <span className="font-medium text-slate-500 dark:text-slate-400 capitalize">{k.replace(/_/g, ' ')}  </span>
+                <span className="text-slate-700 dark:text-slate-300 line-clamp-2">{v}</span>
+              </div>
+            ))}
+          </div>
+        )
+      })}
+    </div>
+  )
+}
+
 export function InsightView() {
  const [data, setData] = useState<MarketFacts>({ goals: ['运动', '适老', 'AI安全'], facts: {} })
  const [history, setHistory] = useState<MarketHistoryItem[]>([])
+  const [research, setResearch] = useState<ResearchFacts>({ domains: {} })
  const [_triggering, setTriggering] = useState<string | null>(null)
  const [toast, setToast] = useState('')

  useEffect(() => {
    fetchMarketFacts().then(setData).catch(console.warn)
    fetchMarketHistory().then(setHistory).catch(console.warn)
+    fetchResearchFacts().then(setResearch).catch(console.warn)
  }, [])

  async function handleTrigger(goal: string) {
@@ -142,6 +239,11 @@ export function InsightView() {
        ))}
      </div>

+      <div className="mb-8">
+        <h2 className="text-sm font-semibold text-slate-600 dark:text-slate-400 mb-3">基础技术调研</h2>
+        <ResearchSection research={research} />
+      </div>
+
      <div>
        <h2 className="text-sm font-semibold text-slate-600 dark:text-slate-400 mb-3">洞察任务历史</h2>
        <div className="bg-white dark:bg-gray-900 rounded-xl border border-slate-200 dark:border-gray-700 overflow-hidden">
@@ -0,0 +1,237 @@
+import { useMemo, useState, useCallback, useRef, useEffect } from 'react'
+import { useTaskStore } from '../store/tasks'
+import { clearProjectTasks, deleteProject } from '../lib/api'
+import { KanbanColumn } from './KanbanColumn'
+import { STATUS_COLUMNS } from '../types'
+import type { Task } from '../types'
+
+interface Props {
+  project?: string  // undefined = 全览（所有项目合并）
+  onImport?: () => void
+  onNewTask?: () => void
+  onBack?: () => void  // 返回概览
+}
+
+export function KanbanBoard({ project, onImport, onNewTask, onBack }: Props) {
+  const tasksMap = useTaskStore(s => s.tasks)
+  const projects = useTaskStore(s => s.projects)
+  const [collapsed, setCollapsed] = useState<Set<string>>(new Set())
+  const [search, setSearch] = useState('')
+  const [menuOpen, setMenuOpen] = useState<string | null>(null)  // which project's menu is open
+  const menuRef = useRef<HTMLDivElement>(null)
+
+  const tasksByProject = useMemo(() => {
+    const result = new Map<string, Task[]>()
+    for (const task of tasksMap.values()) {
+      const arr = result.get(task.project) ?? []
+      arr.push(task)
+      result.set(task.project, arr)
+    }
+    return result
+  }, [tasksMap])
+
+  const projectsToShow = project
+    ? [project]
+    : Array.from(tasksByProject.keys()).sort()
+
+  const getTasksByStatus = useCallback((proj: string | undefined): Map<string, Task[]> => {
+    const source = proj ? (tasksByProject.get(proj) ?? []) : Array.from(tasksMap.values())
+    const q = search.trim().toLowerCase()
+    const filtered = q ? source.filter(t => t.title.toLowerCase().includes(q)) : source
+    const result = new Map<string, Task[]>()
+    STATUS_COLUMNS.forEach(s => result.set(s, []))
+    filtered.forEach(t => {
+      const key = STATUS_COLUMNS.includes(t.status as any) ? t.status : 'pending'
+      result.get(key)?.push(t)
+    })
+    return result
+  }, [tasksByProject, tasksMap, search])
+
+  function toggleCollapse(proj: string) {
+    setCollapsed(prev => {
+      const next = new Set(prev)
+      if (next.has(proj)) next.delete(proj)
+      else next.add(proj)
+      return next
+    })
+  }
+
+  useEffect(() => {
+    function handleClick(e: MouseEvent) {
+      if (menuRef.current && !menuRef.current.contains(e.target as Node)) {
+        setMenuOpen(null)
+      }
+    }
+    document.addEventListener('mousedown', handleClick)
+    return () => document.removeEventListener('mousedown', handleClick)
+  }, [])
+
+  async function handleClearTasks(proj: string) {
+    setMenuOpen(null)
+    const n = [...useTaskStore.getState().tasks.values()].filter(t => t.project === proj).length
+    if (!confirm(`确认清空项目「${proj}」的全部 ${n} 条任务记录？\n（此操作不可恢复）`)) return
+    try {
+      await clearProjectTasks(proj)
+      useTaskStore.getState().removeProjectTasks(proj)
+    } catch (e) {
+      alert('清空失败：' + e)
+    }
+  }
+
+  async function handleDeleteProject(proj: string) {
+    setMenuOpen(null)
+    if (!confirm(`确认删除项目「${proj}」配置并清空所有任务？\n（此操作不可恢复）`)) return
+    try {
+      await deleteProject(proj)
+      useTaskStore.getState().removeProjectTasks(proj)
+      useTaskStore.getState().setProjects(
+        useTaskStore.getState().projects.filter(p => p.name !== proj)
+      )
+    } catch (e) {
+      alert('删除失败：' + e)
+    }
+  }
+
+  if (projectsToShow.length === 0) {
+    return (
+      <div className="flex flex-col items-center justify-center h-64 gap-4">
+        <p className="text-slate-400 dark:text-gray-600 text-sm">暂无任务数据</p>
+        <div className="flex gap-2">
+          {onImport && (
+            <button
+              onClick={onImport}
+              className="px-4 py-2 text-xs bg-slate-100 dark:bg-gray-800 hover:bg-slate-200 dark:hover:bg-gray-700 text-slate-700 dark:text-gray-300 rounded-lg"
+            >
+              导入项目
+            </button>
+          )}
+          {onNewTask && projects.length > 0 && (
+            <button
+              onClick={onNewTask}
+              className="px-4 py-2 text-xs bg-blue-800 hover:bg-blue-700 text-blue-100 rounded-lg"
+            >
+              提交需求
+            </button>
+          )}
+        </div>
+      </div>
+    )
+  }
+
+  return (
+    <div className="flex flex-col gap-6 p-4">
+      {/* 面包屑：在特定项目模式下显示 */}
+      {project && (
+        <div className="flex items-center gap-2">
+          <button
+            onClick={onBack}
+            className="text-xs text-slate-400 dark:text-gray-500 hover:text-slate-700 dark:hover:text-gray-300 transition-colors"
+          >
+            ← 所有项目
+          </button>
+          <span className="text-slate-500 dark:text-gray-700">/</span>
+          <span className="text-xs font-medium text-slate-800 dark:text-gray-200">{project}</span>
+        </div>
+      )}
+
+      {/* 搜索过滤 */}
+      <div className="flex items-center gap-2">
+        <input
+          value={search}
+          onChange={e => setSearch(e.target.value)}
+          placeholder="搜索任务..."
+          className="bg-slate-100 dark:bg-gray-800 border border-slate-200 dark:border-gray-700 rounded-lg px-3 py-1.5 text-xs
+                     text-slate-800 dark:text-gray-200 placeholder-slate-400 dark:placeholder-gray-600 w-56 focus:outline-none focus:border-slate-400 dark:focus:border-gray-500"
+        />
+        {search && (
+          <button
+            onClick={() => setSearch('')}
+            className="text-xs text-slate-400 dark:text-gray-500 hover:text-slate-700 dark:hover:text-gray-300"
+          >
+            清除
+          </button>
+        )}
+      </div>
+
+      {projectsToShow.map(proj => {
+        const byStatus = getTasksByStatus(proj)
+        const isCollapsed = collapsed.has(proj)
+        const runningCount = byStatus.get('running')?.length ?? 0
+        const waitingCount = byStatus.get('waiting_approval')?.length ?? 0
+        const failedCount  = byStatus.get('failed')?.length ?? 0
+
+        return (
+          <div key={proj}>
+            {/* 项目标题栏 */}
+            <div className="flex items-center gap-3 mb-3 group">
+              <div
+                className="flex items-center gap-3 cursor-pointer select-none flex-1"
+                onClick={() => toggleCollapse(proj)}
+              >
+                <span className="text-slate-500 dark:text-gray-400 text-xs">{isCollapsed ? '▶' : '▼'}</span>
+                <span className="text-slate-800 dark:text-gray-200 font-semibold text-sm group-hover:text-blue-700 dark:group-hover:text-white transition-colors">
+                  {proj}
+                </span>
+                <div className="flex items-center gap-2 text-xs">
+                  {runningCount > 0 && (
+                    <span className="flex items-center gap-1 text-blue-600 dark:text-blue-400">
+                      <span className="w-1.5 h-1.5 rounded-full bg-blue-500 dark:bg-blue-400 animate-pulse inline-block" />
+                      {runningCount} 运行中
+                    </span>
+                  )}
+                  {waitingCount > 0 && (
+                    <span className="text-orange-600 dark:text-orange-400">⚠ {waitingCount} 待确认</span>
+                  )}
+                  {failedCount > 0 && (
+                    <span className="text-red-600 dark:text-red-400">✕ {failedCount} 失败</span>
+                  )}
+                </div>
+              </div>
+
+              {/* 项目操作菜单 */}
+              <div className="relative" ref={menuOpen === proj ? menuRef : undefined}>
+                <button
+                  onClick={e => { e.stopPropagation(); setMenuOpen(menuOpen === proj ? null : proj) }}
+                  className="text-slate-400 dark:text-gray-600 hover:text-slate-700 dark:hover:text-gray-300 px-1.5 py-0.5 rounded text-xs opacity-0 group-hover:opacity-100 transition-opacity"
+                  title="项目操作"
+                >
+                  ⋮
+                </button>
+                {menuOpen === proj && (
+                  <div className="absolute right-0 top-6 z-50 bg-white dark:bg-gray-800 border border-slate-200 dark:border-gray-700 rounded-lg shadow-lg py-1 min-w-[140px]">
+                    <button
+                      onClick={() => handleClearTasks(proj)}
+                      className="w-full text-left px-3 py-2 text-xs text-orange-600 dark:text-orange-400 hover:bg-orange-50 dark:hover:bg-orange-900/20 transition-colors"
+                    >
+                      🗑 清空所有任务
+                    </button>
+                    <div className="border-t border-slate-100 dark:border-gray-700 my-1" />
+                    <button
+                      onClick={() => handleDeleteProject(proj)}
+                      className="w-full text-left px-3 py-2 text-xs text-red-600 dark:text-red-400 hover:bg-red-50 dark:hover:bg-red-900/20 transition-colors"
+                    >
+                      ✕ 删除项目
+                    </button>
+                  </div>
+                )}
+              </div>
+            </div>
+
+            {/* Kanban 列 */}
+            {!isCollapsed && (
+              <div className="flex gap-4 overflow-x-auto pb-2">
+                {STATUS_COLUMNS.map(status => (
+                  <KanbanColumn
+                    key={status}
+                    status={status}
+                    tasks={byStatus.get(status) ?? []}
+                  />
+                ))}
+              </div>
+            )}
+          </div>
+        )
+      })}
+    </div>
+  )
+}
@@ -0,0 +1,71 @@
+import { useState } from 'react'
+import type { Task } from '../types'
+import { STATUS_LABELS } from '../types'
+import { TaskCard } from './TaskCard'
+
+interface Props {
+  status: string
+  tasks: Task[]
+}
+
+const COLUMN_HEADER_COLORS: Record<string, string> = {
+  pending:          'text-slate-500 dark:text-gray-400 border-slate-300 dark:border-gray-600',
+  running:          'text-blue-400 border-blue-500',
+  done:             'text-green-400 border-green-600',
+  failed:           'text-red-400 border-red-600',
+  waiting_approval: 'text-orange-400 border-orange-500',
+  vetoed:           'text-slate-400 dark:text-gray-500 border-slate-300 dark:border-gray-600',
+}
+
+// 完成列默认只显示最近 5 条
+const DONE_PREVIEW_COUNT = 5
+
+export function KanbanColumn({ status, tasks }: Props) {
+  const [showAll, setShowAll] = useState(false)
+  const label = STATUS_LABELS[status] ?? status
+  const headerColor = COLUMN_HEADER_COLORS[status] ?? 'text-slate-500 dark:text-gray-400 border-slate-300 dark:border-gray-600'
+
+  // 完成列截断
+  const isDone = status === 'done'
+  const displayTasks = isDone && !showAll
+    ? tasks.slice(0, DONE_PREVIEW_COUNT)
+    : tasks
+  const hiddenCount = isDone ? tasks.length - DONE_PREVIEW_COUNT : 0
+
+  return (
+    <div className="flex flex-col min-w-[240px] w-[240px]">
+      {/* 列头 */}
+      <div className={`flex items-center justify-between mb-3 pb-2 border-b ${headerColor}`}>
+        <span className="text-sm font-semibold">{label}</span>
+        <span className="text-xs bg-slate-200/60 dark:bg-gray-700/60 rounded-full px-2 py-0.5 font-mono">{tasks.length}</span>
+      </div>
+
+      {/* 任务卡片 */}
+      <div className="flex flex-col gap-2 overflow-y-auto max-h-[calc(100vh-280px)]">
+        {displayTasks.length === 0 ? (
+          <p className="text-xs text-slate-500 dark:text-gray-700 text-center py-4">—</p>
+        ) : (
+          displayTasks.map(t => <TaskCard key={t.id} task={t} />)
+        )}
+
+        {/* 显示更多 / 收起 */}
+        {isDone && hiddenCount > 0 && !showAll && (
+          <button
+            onClick={() => setShowAll(true)}
+            className="text-xs text-slate-400 dark:text-gray-600 hover:text-slate-600 dark:hover:text-gray-400 text-center py-2 transition-colors"
+          >
+            查看更早的 {hiddenCount} 条
+          </button>
+        )}
+        {isDone && showAll && tasks.length > DONE_PREVIEW_COUNT && (
+          <button
+            onClick={() => setShowAll(false)}
+            className="text-xs text-slate-400 dark:text-gray-600 hover:text-slate-600 dark:hover:text-gray-400 text-center py-2 transition-colors"
+          >
+            收起
+          </button>
+        )}
+      </div>
+    </div>
+  )
+}
@@ -0,0 +1,232 @@
+import { useState } from 'react'
+import { createTask } from '../lib/api'
+import { useTaskStore } from '../store/tasks'
+
+interface Props { onClose: () => void }
+
+// 用户友好的需求类型（隐藏技术实现细节）
+const GOAL_TYPES = [
+  { value: 'fix_bug',      label: '修复一个问题', desc: '描述您遇到的 Bug 或异常行为' },
+  { value: 'new_feature',  label: '添加新功能',   desc: '描述您希望新增的能力' },
+  { value: 'improve_perf', label: '提升性能',     desc: '描述需要优化的性能瓶颈' },
+  { value: 'code_review',  label: '代码审查',     desc: '检查代码质量和潜在问题' },
+  { value: 'refactor',     label: '代码优化',     desc: '改善代码结构和可维护性' },
+  { value: 'documentation','label': '完善文档',   desc: '更新 README 或补充注释' },
+  { value: 'tech_support', label: '技术攻关',     desc: '向基础技术组请求专项调研或优化支持' },
+]
+
+const TECH_ROLES = [
+  { value: 'rtp-researcher',   label: 'RTP/RTCP 调研',  desc: '实时传输协议、低延迟音视频' },
+  { value: 'net-researcher',   label: '网络栈调优',     desc: 'QoS/拥塞控制/多链路' },
+  { value: 'kernel-analyzer',  label: '内核调度分析',   desc: 'CPUFreq/NPU/多核调度' },
+  { value: 'os-engineer',      label: 'OS 系统优化',    desc: '系统层配置与驱动调优' },
+  { value: 'hw-engineer',      label: '硬件工程',       desc: '硬件接口/外设/信号分析' },
+]
+
+export function NewTaskModal({ onClose }: Props) {
+  const projects = useTaskStore(s => s.projects)
+  const applyUpdate = useTaskStore(s => s.applyUpdate)
+  const [form, setForm] = useState({
+    project: projects[0]?.name ?? '',
+    type: 'fix_bug',
+    title: '',
+    context: '',
+    initiator: '',
+    discussion: '',
+    techRole: 'rtp-researcher',
+  })
+  const isTechSupport = form.type === 'tech_support'
+  const [loading, setLoading] = useState(false)
+  const [error, setError] = useState('')
+
+  const selectedType = GOAL_TYPES.find(t => t.value === form.type)
+
+  async function submit(e: React.FormEvent) {
+    e.preventDefault()
+    if (!form.title.trim()) { setError('请描述您的需求'); return }
+    if (!form.project) { setError('请选择项目'); return }
+    setLoading(true)
+    try {
+      const task = await createTask({
+        project: form.project,
+        type: form.type,
+        title: form.title.trim(),
+        context: form.context.trim(),
+        priority: 2,
+        mode: projects.find(p => p.name === form.project)?.mode ?? 'auto',
+        agent_role: isTechSupport ? form.techRole : 'developer',
+        initiator: form.initiator.trim() || 'user',
+        discussion: form.discussion.trim(),
+        queue: isTechSupport ? 'base_opt' : 'project_delivery',
+      })
+      applyUpdate(task)
+      onClose()
+    } catch (e) {
+      setError(String(e))
+    } finally {
+      setLoading(false)
+    }
+  }
+
+  if (projects.length === 0) {
+    return (
+      <div className="fixed inset-0 bg-black/60 z-50 flex items-center justify-center">
+        <div className="bg-white dark:bg-gray-900 border border-slate-200 dark:border-gray-700 rounded-xl p-6 w-[420px] shadow-2xl text-center">
+          <p className="text-slate-500 dark:text-gray-400 mb-4">请先导入一个项目</p>
+          <button onClick={onClose}
+            className="px-4 py-2 text-sm bg-slate-200 dark:bg-gray-700 hover:bg-slate-300 dark:hover:bg-gray-600 text-slate-800 dark:text-gray-200 rounded-lg">
+            关闭
+          </button>
+        </div>
+      </div>
+    )
+  }
+
+  return (
+    <div className="fixed inset-0 bg-black/60 z-50 flex items-center justify-center">
+      <div className="bg-white dark:bg-gray-900 border border-slate-200 dark:border-gray-700 rounded-xl p-6 w-[520px] shadow-2xl">
+        <h2 className="text-base font-semibold text-slate-900 dark:text-gray-100 mb-1">提交开发需求</h2>
+        <p className="text-xs text-slate-400 dark:text-gray-500 mb-5">
+          AI 将根据您的描述自动规划并完成开发工作
+        </p>
+
+        <form onSubmit={submit} className="space-y-4">
+          {/* 项目选择 */}
+          {projects.length > 1 && (
+            <div>
+              <label className="text-xs text-slate-500 dark:text-gray-400 mb-1.5 block">选择项目</label>
+              <select
+                value={form.project}
+                onChange={e => setForm(f => ({ ...f, project: e.target.value }))}
+                className={inputCls}
+              >
+                {projects.map(p => (
+                  <option key={p.name} value={p.name}>{p.name}</option>
+                ))}
+              </select>
+            </div>
+          )}
+
+          {/* 需求类型（图标化选择） */}
+          <div>
+            <label className="text-xs text-slate-500 dark:text-gray-400 mb-2 block">需求类型</label>
+            <div className="grid grid-cols-3 gap-2">
+              {GOAL_TYPES.map(t => (
+                <button
+                  key={t.value}
+                  type="button"
+                  onClick={() => setForm(f => ({ ...f, type: t.value }))}
+                  className={`px-3 py-2 text-xs rounded-lg text-left transition-colors border
+                    ${form.type === t.value
+                      ? 'bg-blue-900/50 border-blue-600 text-blue-200'
+                      : 'bg-slate-100 dark:bg-gray-800 border-slate-200 dark:border-gray-700 text-slate-500 dark:text-gray-400 hover:border-slate-300 dark:hover:border-gray-600'
+                    }`}
+                >
+                  {t.label}
+                </button>
+              ))}
+            </div>
+          </div>
+
+          {/* 技术角色（仅技术攻关） */}
+          {isTechSupport && (
+            <div>
+              <label className="text-xs text-slate-500 dark:text-gray-400 mb-2 block">指派给</label>
+              <div className="grid grid-cols-1 gap-1.5">
+                {TECH_ROLES.map(r => (
+                  <button
+                    key={r.value}
+                    type="button"
+                    onClick={() => setForm(f => ({ ...f, techRole: r.value }))}
+                    className={`px-3 py-2 text-xs rounded-lg text-left transition-colors border flex items-center gap-2
+                      ${form.techRole === r.value
+                        ? 'bg-amber-900/50 border-amber-600 text-amber-200'
+                        : 'bg-slate-100 dark:bg-gray-800 border-slate-200 dark:border-gray-700 text-slate-500 dark:text-gray-400 hover:border-slate-300 dark:hover:border-gray-600'
+                      }`}
+                  >
+                    <span className="font-medium">{r.label}</span>
+                    <span className="text-slate-400 dark:text-gray-600">{r.desc}</span>
+                  </button>
+                ))}
+              </div>
+            </div>
+          )}
+
+          {/* 需求描述 */}
+          <div>
+            <label className="text-xs text-slate-500 dark:text-gray-400 mb-1.5 block">
+              需求描述
+              {selectedType && (
+                <span className="text-slate-300 dark:text-gray-600 ml-1">— {selectedType.desc}</span>
+              )}
+            </label>
+            <input
+              value={form.title}
+              onChange={e => setForm(f => ({ ...f, title: e.target.value }))}
+              className={inputCls}
+              placeholder="用一句话描述您的需求..."
+              autoFocus
+            />
+          </div>
+
+          {/* 发起人 */}
+          <div>
+            <label className="text-xs text-slate-500 dark:text-gray-400 mb-1.5 block">
+              发起人 <span className="text-slate-300 dark:text-gray-600">（您的姓名）</span>
+            </label>
+            <input
+              value={form.initiator}
+              onChange={e => setForm(f => ({ ...f, initiator: e.target.value }))}
+              className={inputCls}
+              placeholder="请填写您的姓名..."
+            />
+          </div>
+
+          {/* 沟通纪要 */}
+          <div>
+            <label className="text-xs text-slate-500 dark:text-gray-400 mb-1.5 block">
+              沟通纪要 <span className="text-slate-300 dark:text-gray-600">（可选，任务背景与决策依据）</span>
+            </label>
+            <textarea
+              value={form.discussion}
+              onChange={e => setForm(f => ({ ...f, discussion: e.target.value }))}
+              rows={2}
+              className={inputCls + ' resize-none'}
+              placeholder="记录相关讨论、背景原因或决策依据..."
+            />
+          </div>
+
+          {/* 详细背景（可选） */}
+          <div>
+            <label className="text-xs text-slate-500 dark:text-gray-400 mb-1.5 block">
+              补充说明 <span className="text-slate-300 dark:text-gray-600">（可选）</span>
+            </label>
+            <textarea
+              value={form.context}
+              onChange={e => setForm(f => ({ ...f, context: e.target.value }))}
+              rows={3}
+              className={inputCls + ' resize-none'}
+              placeholder="如有需要，可补充复现步骤、期望结果、相关文件路径等..."
+            />
+          </div>
+
+          {error && <p className="text-xs text-red-400">{error}</p>}
+
+          <div className="flex justify-end gap-2 pt-1">
+            <button type="button" onClick={onClose}
+              className="px-4 py-2 text-sm text-slate-500 dark:text-gray-400 hover:text-slate-800 dark:hover:text-gray-200">
+              取消
+            </button>
+            <button type="submit" disabled={loading}
+              className="px-5 py-2 text-sm bg-blue-700 hover:bg-blue-600 text-white
+                         rounded-lg disabled:opacity-50 font-medium">
+              {loading ? '提交中...' : '提交需求'}
+            </button>
+          </div>
+        </form>
+      </div>
+    </div>
+  )
+}
+
+const inputCls = 'w-full bg-slate-100 dark:bg-gray-800 border border-slate-300 dark:border-gray-600 rounded-lg px-3 py-2 text-xs text-slate-800 dark:text-gray-200'
@@ -0,0 +1,312 @@
+import { useMemo } from 'react'
+import { useTaskStore } from '../store/tasks'
+import { approveTask, rejectTask, deleteProject, fetchProjects } from '../lib/api'
+import type { Task, Project } from '../types'
+
+interface Props {
+  onImport: () => void
+  onSelectProject: (name: string) => void
+}
+
+function timeAgo(iso: string | null): string {
+  if (!iso) return ''
+  const sec = Math.floor((Date.now() - new Date(iso + 'Z').getTime()) / 1000)
+  if (sec < 60) return '刚刚'
+  if (sec < 3600) return `${Math.floor(sec / 60)} 分钟前`
+  if (sec < 86400) return `${Math.floor(sec / 3600)} 小时前`
+  return `${Math.floor(sec / 86400)} 天前`
+}
+
+export function ProjectOverview({ onImport, onSelectProject }: Props) {
+  const tasksMap = useTaskStore(s => s.tasks)
+  const projects = useTaskStore(s => s.projects)
+  const applyUpdate = useTaskStore(s => s.applyUpdate)
+
+  const setProjects = useTaskStore(s => s.setProjects)
+
+  const { projectStats, approvalTasks } = useMemo(() => {
+    const all = Array.from(tasksMap.values())
+    const byProj = new Map<string, Task[]>()
+    for (const t of all) {
+      const arr = byProj.get(t.project) ?? []
+      arr.push(t)
+      byProj.set(t.project, arr)
+    }
+
+    const projectStats = projects.map(p => {
+      const tasks = byProj.get(p.name) ?? []
+      const done    = tasks.filter(t => t.status === 'done').length
+      const failed  = tasks.filter(t => t.status === 'failed').length
+      const running = tasks.filter(t => t.status === 'running').length
+      const pending = tasks.filter(t => t.status === 'pending').length
+      const waiting = tasks.filter(t => t.status === 'waiting_approval').length
+      const total   = tasks.length
+      const progress = total > 0 ? Math.round((done / total) * 100) : 0
+      // 最近活动时间
+      const lastTs = tasks
+        .map(t => t.completed_at || t.started_at || t.created_at)
+        .filter(Boolean)
+        .sort()
+        .pop() ?? null
+      return { project: p, tasks, done, failed, running, pending, waiting, total, progress, lastTs }
+    })
+
+    const approvalTasks = all.filter(t => t.status === 'waiting_approval')
+    return { projectStats, approvalTasks }
+  }, [tasksMap, projects])
+
+  // 空状态
+  if (projects.length === 0) {
+    return <EmptyState onImport={onImport} />
+  }
+
+  return (
+    <div className="p-6 max-w-6xl mx-auto space-y-6">
+      {/* 待审批横幅 */}
+      {approvalTasks.length > 0 && (
+        <ApprovalBanner
+          tasks={approvalTasks}
+          onApprove={async (id) => {
+            const t = await approveTask(id)
+            applyUpdate(t)
+          }}
+          onReject={async (id) => {
+            const t = await rejectTask(id, '用户拒绝')
+            applyUpdate(t)
+          }}
+        />
+      )}
+
+      {/* 统计栏 */}
+      <div className="grid grid-cols-4 gap-3">
+        <StatCard label="项目总数" value={projects.length} color="text-blue-400" />
+        <StatCard label="今日完成" value={projectStats.reduce((s, p) => s + p.done, 0)} color="text-green-400" />
+        <StatCard label="进行中" value={projectStats.reduce((s, p) => s + p.running, 0)} color="text-sky-400" />
+        <StatCard label="待处理" value={projectStats.reduce((s, p) => s + p.failed, 0)} color="text-red-400" />
+      </div>
+
+      {/* 项目卡片网格 */}
+      <div>
+        <h2 className="text-sm font-medium text-slate-500 dark:text-gray-400 mb-3">所有项目</h2>
+        <div className="grid grid-cols-1 sm:grid-cols-2 lg:grid-cols-3 gap-4">
+          {projectStats.map(s => (
+            <ProjectCard
+              key={s.project.name}
+              data={s}
+              onClick={() => onSelectProject(s.project.name)}
+              onDelete={async () => {
+                if (!confirm(`确认删除项目「${s.project.name}」？\n（任务记录不会被删除）`)) return
+                await deleteProject(s.project.name)
+                const updated = await fetchProjects()
+                setProjects(updated)
+              }}
+            />
+          ))}
+          {/* 添加项目 */}
+          <button
+            onClick={onImport}
+            className="h-40 rounded-xl border-2 border-dashed border-slate-200 dark:border-gray-700 hover:border-slate-400 dark:hover:border-gray-500
+                       flex flex-col items-center justify-center gap-2 text-slate-400 dark:text-gray-600
+                       hover:text-slate-600 dark:hover:text-gray-400 transition-colors"
+          >
+            <span className="text-2xl">+</span>
+            <span className="text-sm">添加项目</span>
+          </button>
+        </div>
+      </div>
+    </div>
+  )
+}
+
+// ── 空状态引导页 ──────────────────────────────────────────────────────
+
+function EmptyState({ onImport }: { onImport: () => void }) {
+  return (
+    <div className="flex flex-col items-center justify-center h-[calc(100vh-120px)] gap-6 px-4">
+      <div className="text-6xl opacity-30">🤖</div>
+      <div className="text-center">
+        <h2 className="text-xl font-semibold text-slate-800 dark:text-gray-200 mb-2">还没有项目</h2>
+        <p className="text-sm text-slate-400 dark:text-gray-500 max-w-sm">
+          导入一个本地项目目录，AI 将自动分析代码、提出改进方案并持续优化
+        </p>
+      </div>
+      <button
+        onClick={onImport}
+        className="px-6 py-2.5 bg-blue-700 hover:bg-blue-600 text-white rounded-xl
+                   text-sm font-medium transition-colors shadow-lg shadow-blue-900/30"
+      >
+        导入第一个项目
+      </button>
+      <div className="flex gap-6 text-xs text-slate-400 dark:text-gray-600 mt-2">
+        <span>✓ 自动代码审查</span>
+        <span>✓ 持续优化改进</span>
+        <span>✓ 实时进度跟踪</span>
+      </div>
+    </div>
+  )
+}
+
+// ── 项目卡片 ──────────────────────────────────────────────────────────
+
+interface ProjectStat {
+  project: Project
+  done: number; failed: number; running: number; pending: number; waiting: number
+  total: number; progress: number; lastTs: string | null
+}
+
+function ProjectCard({ data, onClick, onDelete }: {
+  data: ProjectStat; onClick: () => void; onDelete: () => void
+}) {
+  const { project, done, failed, running, pending, waiting, progress, lastTs } = data
+  const isActive = running > 0 || pending > 0
+
+  return (
+    <button
+      onClick={onClick}
+      className="bg-white dark:bg-gray-900 border border-slate-200/60 dark:border-gray-700/60 hover:border-slate-400 dark:hover:border-gray-500 rounded-xl p-4
+                 text-left transition-all hover:shadow-lg hover:shadow-slate-300/50 dark:shadow-black/30 group"
+    >
+      {/* 头部：名称 + 状态 + 删除 */}
+      <div className="flex items-start justify-between mb-3">
+        <div className="min-w-0 flex-1">
+          <h3 className="text-sm font-semibold text-slate-900 dark:text-gray-100 group-hover:text-blue-700 dark:group-hover:text-white truncate">
+            {project.name}
+          </h3>
+          {project.description && (
+            <p className="text-xs text-slate-400 dark:text-gray-500 mt-0.5 line-clamp-1">{project.description}</p>
+          )}
+        </div>
+        <div className="flex items-center gap-1.5 shrink-0 ml-2">
+          <StatusPill running={running} waiting={waiting} isActive={isActive} />
+          <button
+            onClick={e => { e.stopPropagation(); onDelete() }}
+            className="text-slate-300 dark:text-gray-600 hover:text-red-400 transition-colors text-sm leading-none opacity-0 group-hover:opacity-100"
+            title="删除项目"
+          >
+            ×
+          </button>
+        </div>
+      </div>
+
+      {/* 进度条 */}
+      <div className="mb-3">
+        <div className="flex justify-between text-xs text-slate-400 dark:text-gray-500 mb-1">
+          <span>完成进度</span>
+          <span>{progress}%</span>
+        </div>
+        <div className="h-1.5 bg-slate-200 dark:bg-gray-700 rounded-full overflow-hidden">
+          <div
+            className="h-full bg-blue-500 rounded-full transition-all duration-500"
+            style={{ width: `${progress}%` }}
+          />
+        </div>
+      </div>
+
+      {/* 统计数字 */}
+      <div className="flex gap-3 text-xs">
+        <span className="text-green-400">{done} 完成</span>
+        {failed > 0 && <span className="text-red-400">{failed} 失败</span>}
+        {pending > 0 && <span className="text-slate-400 dark:text-gray-500">{pending} 待执行</span>}
+        {waiting > 0 && <span className="text-orange-400 font-medium">⚠ {waiting} 待审批</span>}
+      </div>
+
+      {/* 最近活动 */}
+      {lastTs && (
+        <p className="text-xs text-slate-400 dark:text-gray-600 mt-2">{timeAgo(lastTs)}</p>
+      )}
+    </button>
+  )
+}
+
+function StatusPill({ running, waiting, isActive }: {
+  running: number; waiting: number; isActive: boolean
+}) {
+  if (waiting > 0) {
+    return (
+      <span className="text-[10px] px-2 py-0.5 rounded-full bg-orange-100 text-orange-700 border border-orange-300 dark:bg-orange-900/50 dark:text-orange-300 dark:border-orange-700/50 shrink-0">
+        待审批
+      </span>
+    )
+  }
+  if (running > 0) {
+    return (
+      <span className="text-[10px] px-2 py-0.5 rounded-full bg-blue-100 text-blue-700 border border-blue-300 dark:bg-blue-900/50 dark:text-blue-300 dark:border-blue-700/50 shrink-0 flex items-center gap-1">
+        <span className="w-1.5 h-1.5 rounded-full bg-blue-500 dark:bg-blue-400 animate-pulse inline-block" />
+        运行中
+      </span>
+    )
+  }
+  if (isActive) {
+    return (
+      <span className="text-[10px] px-2 py-0.5 rounded-full bg-slate-200 dark:bg-gray-700 text-slate-600 dark:text-gray-400 shrink-0">
+        排队中
+      </span>
+    )
+  }
+  return (
+    <span className="text-[10px] px-2 py-0.5 rounded-full bg-slate-100 dark:bg-gray-800 text-slate-400 dark:text-gray-600 shrink-0">
+      待机
+    </span>
+  )
+}
+
+// ── 待审批横幅 ─────────────────────────────────────────────────────────
+
+function ApprovalBanner({ tasks, onApprove, onReject }: {
+  tasks: Task[]
+  onApprove: (id: number) => Promise<void>
+  onReject: (id: number) => Promise<void>
+}) {
+  const first = tasks[0]
+  return (
+    <div className="bg-orange-50 border border-orange-200 dark:bg-orange-950/40 dark:border-orange-700/50 rounded-xl p-4">
+      <div className="flex items-center justify-between mb-2">
+        <div className="flex items-center gap-2">
+          <span className="text-orange-500 dark:text-orange-400 text-base">⚠</span>
+          <span className="text-sm font-medium text-orange-700 dark:text-orange-200">
+            {tasks.length} 个任务需要您的确认
+          </span>
+        </div>
+        <span className="text-xs text-slate-400 dark:text-gray-500">{first.project}</span>
+      </div>
+      <p className="text-sm text-slate-700 dark:text-gray-300 mb-3 line-clamp-2">{first.title}</p>
+      {first.result_summary && (
+        <p className="text-xs text-slate-500 dark:text-gray-400 bg-slate-100/60 dark:bg-gray-900/60 rounded p-2 mb-3 line-clamp-3">
+          {first.result_summary}
+        </p>
+      )}
+      <div className="flex gap-2">
+        <button
+          onClick={() => onApprove(first.id)}
+          className="px-4 py-1.5 text-sm bg-green-800 hover:bg-green-700 text-green-100
+                     rounded-lg transition-colors"
+        >
+          允许执行
+        </button>
+        <button
+          onClick={() => onReject(first.id)}
+          className="px-4 py-1.5 text-sm bg-slate-100 dark:bg-gray-800 hover:bg-slate-200 dark:hover:bg-gray-700 text-slate-700 dark:text-gray-300
+                     rounded-lg transition-colors"
+        >
+          拒绝
+        </button>
+        {tasks.length > 1 && (
+          <span className="self-center text-xs text-slate-400 dark:text-gray-500 ml-1">
+            还有 {tasks.length - 1} 个待确认
+          </span>
+        )}
+      </div>
+    </div>
+  )
+}
+
+// ── 统计卡片 ──────────────────────────────────────────────────────────
+
+function StatCard({ label, value, color }: { label: string; value: number; color: string }) {
+  return (
+    <div className="bg-white dark:bg-gray-900 border border-slate-200/60 dark:border-gray-700/60 rounded-xl px-4 py-3">
+      <div className={`text-2xl font-bold ${color}`}>{value}</div>
+      <div className="text-xs text-slate-400 dark:text-gray-500 mt-0.5">{label}</div>
+    </div>
+  )
+}
@@ -0,0 +1,110 @@
+import type { Task } from '../types'
+import { STATUS_COLORS, ROLE_COLORS } from '../types'
+import { useTaskStore } from '../store/tasks'
+
+interface Props { task: Task }
+
+function elapsed(startedAt: string | null): string {
+  if (!startedAt) return ''
+  const sec = Math.floor((Date.now() - new Date(startedAt + 'Z').getTime()) / 1000)
+  if (sec < 60) return `${sec}s`
+  if (sec < 3600) return `${Math.floor(sec / 60)}m`
+  return `${Math.floor(sec / 3600)}h ${Math.floor((sec % 3600) / 60)}m`
+}
+
+const TYPE_LABELS: Record<string, string> = {
+  fix_bug: '修复问题',
+  improve_perf: '性能优化',
+  code_review: '代码审查',
+  new_feature: '新功能',
+  arch_optimize: '架构优化',
+  refactor: '代码重构',
+  test_improvement: '测试完善',
+  documentation: '文档更新',
+  performance: '性能分析',
+  integration: '项目集成',
+  architecture: '架构评估',
+  device_test: '设备测试',
+  architect: '架构分析',
+}
+
+const PRIORITY_BADGE: Record<number, { label: string; cls: string }> = {
+  1: { label: 'P1', cls: 'bg-red-500/20 text-red-400 border border-red-500/30' },
+  2: { label: 'P2', cls: 'bg-orange-500/20 text-orange-400 border border-orange-500/30' },
+  3: { label: 'P3', cls: 'bg-yellow-500/20 text-yellow-500 border border-yellow-500/30' },
+  4: { label: 'P4', cls: 'bg-blue-500/20 text-blue-400 border border-blue-500/30' },
+  5: { label: 'P5', cls: 'bg-slate-500/20 text-slate-400 border border-slate-500/30' },
+}
+
+export function TaskCard({ task }: Props) {
+  const setSelected = useTaskStore(s => s.setSelectedTask)
+  const colorClass = STATUS_COLORS[task.status] ?? 'border-slate-300 dark:border-gray-600 bg-slate-300/10 dark:bg-gray-600/10'
+
+  const typeLabel = TYPE_LABELS[task.type] ?? task.type
+  const isRunning = task.status === 'running'
+  const needsApproval = task.status === 'waiting_approval'
+  const pri = PRIORITY_BADGE[task.priority] ?? PRIORITY_BADGE[3]
+  const roleCls = ROLE_COLORS[task.agent_role] ?? 'bg-slate-500'
+  const initiatorShort = (task.initiator || 'system').split(':').pop()?.slice(0, 8) ?? 'system'
+
+  return (
+    <div
+      className={`rounded-lg border p-3 cursor-pointer hover:opacity-90 transition-opacity ${colorClass}`}
+      onClick={() => setSelected(task.id)}
+    >
+      {/* 行1: 优先级 + 类型 + 待确认标记 */}
+      <div className="flex items-center gap-1.5 mb-2">
+        <span className={`text-[10px] font-bold px-1.5 py-0.5 rounded font-mono ${pri.cls}`}>
+          {pri.label}
+        </span>
+        <span className="text-[10px] text-slate-400 dark:text-gray-500 bg-slate-200/60 dark:bg-gray-700/60 px-1.5 py-0.5 rounded">
+          {typeLabel}
+        </span>
+        {needsApproval && (
+          <span className="text-[10px] text-orange-700 bg-orange-100 dark:text-orange-300 dark:bg-orange-900/40 px-1.5 py-0.5 rounded font-medium ml-auto">
+            待确认
+          </span>
+        )}
+        {task.retry_count > 0 && (
+          <span className="text-[10px] text-slate-400 dark:text-gray-500 ml-auto">×{task.retry_count}</span>
+        )}
+      </div>
+
+      {/* 行2: 标题 */}
+      <p className="text-sm text-slate-800 dark:text-gray-200 font-medium line-clamp-2 mb-2 leading-snug">
+        {task.title}
+      </p>
+
+      {/* 行3: 沟通纪要摘要（如有） */}
+      {task.discussion && (
+        <p className="text-[10px] text-slate-400 dark:text-gray-600 italic line-clamp-1 mb-2 border-l-2 border-slate-300 dark:border-gray-700 pl-1.5">
+          {task.discussion}
+        </p>
+      )}
+
+      {/* 结果预览（done 状态） */}
+      {task.status === 'done' && task.result_summary && (
+        <p className="text-xs text-slate-400 dark:text-gray-500 line-clamp-1 mb-2">
+          {task.result_summary}
+        </p>
+      )}
+
+      {/* 行4: 发起人 → 执行者 + 计时 */}
+      <div className="flex items-center justify-between mt-1">
+        <div className="flex items-center gap-1 text-[10px] text-slate-400 dark:text-gray-600">
+          <span className="font-mono">{initiatorShort}</span>
+          <span>→</span>
+          <span className={`px-1.5 py-0.5 rounded text-white text-[10px] font-mono ${roleCls}`}>
+            {task.agent_role}
+          </span>
+        </div>
+        {isRunning && task.started_at && (
+          <div className="flex items-center gap-1 text-[10px] text-blue-400">
+            <span className="w-1.5 h-1.5 rounded-full bg-blue-400 animate-pulse inline-block" />
+            <span>{elapsed(task.started_at)}</span>
+          </div>
+        )}
+      </div>
+    </div>
+  )
+}
@@ -0,0 +1,53 @@
+import { useEffect } from 'react'
+import { useTaskStore } from '../store/tasks'
+import type { Toast as ToastItem } from '../store/tasks'
+
+const TOAST_DURATION = 5000
+
+function ToastBubble({ toast }: { toast: ToastItem }) {
+  const removeToast = useTaskStore(s => s.removeToast)
+  const setSelected = useTaskStore(s => s.setSelectedTask)
+
+  useEffect(() => {
+    const timer = setTimeout(() => removeToast(toast.id), TOAST_DURATION)
+    return () => clearTimeout(timer)
+  }, [toast.id, removeToast])
+
+  const colorCls =
+    toast.type === 'success' ? 'bg-green-900/90 border-green-700 text-green-100' :
+    toast.type === 'error'   ? 'bg-red-900/90 border-red-700 text-red-100' :
+    toast.type === 'warning' ? 'bg-orange-900/90 border-orange-700 text-orange-100' :
+    'bg-slate-100/90 dark:bg-gray-800/90 border-slate-300 dark:border-gray-600 text-slate-900 dark:text-gray-100'
+
+  function handleClick() {
+    if (toast.taskId != null) setSelected(toast.taskId)
+    removeToast(toast.id)
+  }
+
+  return (
+    <div
+      className={`flex items-center justify-between gap-3 px-4 py-2.5 rounded-lg border
+                  text-xs shadow-lg cursor-pointer animate-in slide-in-from-right ${colorCls}`}
+      onClick={handleClick}
+    >
+      <span className="leading-snug">{toast.message}</span>
+      <button
+        onClick={e => { e.stopPropagation(); removeToast(toast.id) }}
+        className="opacity-60 hover:opacity-100 text-base leading-none shrink-0"
+      >
+        ×
+      </button>
+    </div>
+  )
+}
+
+export function ToastContainer() {
+  const toasts = useTaskStore(s => s.toasts)
+  if (toasts.length === 0) return null
+
+  return (
+    <div className="fixed bottom-4 right-4 z-[100] flex flex-col gap-2 max-w-[360px]">
+      {toasts.map(t => <ToastBubble key={t.id} toast={t} />)}
+    </div>
+  )
+}
@@ -0,0 +1,195 @@
+import { useEffect, useState } from 'react'
+import {
+  fetchVisualRuns, fetchVisualCases, approveVisualCase, rejectVisualCase,
+  type VisualRun, type VisualCase,
+} from '../lib/api'
+
+const STATUS_BADGE: Record<string, string> = {
+  passed: 'bg-green-100 text-green-800 dark:bg-green-900 dark:text-green-200',
+  failed: 'bg-red-100 text-red-800 dark:bg-red-900 dark:text-red-200',
+  needs_review: 'bg-amber-100 text-amber-800 dark:bg-amber-900 dark:text-amber-200',
+  running: 'bg-blue-100 text-blue-800 dark:bg-blue-900 dark:text-blue-200',
+  pending: 'bg-slate-100 text-slate-600 dark:bg-gray-800 dark:text-slate-400',
+  skipped: 'bg-gray-100 text-gray-500 dark:bg-gray-800 dark:text-gray-400',
+  done: 'bg-green-100 text-green-800 dark:bg-green-900 dark:text-green-200',
+  interrupted: 'bg-red-100 text-red-800 dark:bg-red-900 dark:text-red-200',
+}
+
+function RunRow({ run, selected, onClick }: {
+  run: VisualRun
+  selected: boolean
+  onClick: () => void
+}) {
+  const pct = run.total > 0 ? Math.round(run.passed / run.total * 100) : 0
+  return (
+    <div
+      onClick={onClick}
+      className={`cursor-pointer px-4 py-3 border-b border-slate-200 dark:border-gray-700
+        hover:bg-slate-50 dark:hover:bg-gray-800 ${selected ? 'bg-blue-50 dark:bg-blue-900/20' : ''}`}
+    >
+      <div className="flex items-center justify-between">
+        <span className="font-medium text-sm">{run.project}</span>
+        <span className={`text-xs px-2 py-0.5 rounded-full ${STATUS_BADGE[run.status] ?? ''}`}>
+          {run.status}
+        </span>
+      </div>
+      <div className="mt-1 flex items-center gap-3 text-xs text-slate-500 dark:text-slate-400">
+        <span>{run.started_at.slice(0, 16)}</span>
+        <span>{run.passed}/{run.total} passed</span>
+        {run.pending_review > 0 && (
+          <span className="text-amber-600 dark:text-amber-400">⚠ {run.pending_review} 待审批</span>
+        )}
+      </div>
+      {run.total > 0 && (
+        <div className="mt-1.5 h-1 bg-slate-200 dark:bg-gray-700 rounded-full overflow-hidden">
+          <div className="h-full bg-green-500" style={{ width: `${pct}%` }} />
+        </div>
+      )}
+    </div>
+  )
+}
+
+type ImageKey = 'baseline_path' | 'screenshot_path' | 'diff_path'
+
+function CaseRow({ c, onApprove, onReject }: {
+  c: VisualCase
+  onApprove: (id: number) => void
+  onReject: (id: number) => void
+}) {
+  const imageKeys: ImageKey[] = ['baseline_path', 'screenshot_path', 'diff_path']
+  const labels = ['基线', '当前截图', '差异']
+
+  return (
+    <div className="border border-slate-200 dark:border-gray-700 rounded-lg p-4">
+      <div className="flex items-center justify-between mb-3">
+        <div>
+          <span className="text-sm font-medium">{c.name}</span>
+          <span className="text-xs text-slate-500 dark:text-slate-400 ml-2">{c.url}</span>
+        </div>
+        <div className="flex items-center gap-2">
+          {c.diff_pct != null && (
+            <span className="text-xs text-slate-500">diff {c.diff_pct.toFixed(2)}%</span>
+          )}
+          <span className={`text-xs px-2 py-0.5 rounded-full ${STATUS_BADGE[c.status] ?? ''}`}>
+            {c.status}
+          </span>
+        </div>
+      </div>
+      <div className="grid grid-cols-3 gap-3">
+        {imageKeys.map((key, i) => {
+          const path = c[key]
+          return (
+            <div key={key} className="text-center">
+              <div className="text-xs text-slate-500 dark:text-slate-400 mb-1">{labels[i]}</div>
+              {path ? (
+                <img
+                  src={`/api/visual/image?path=${encodeURIComponent(path)}`}
+                  alt={labels[i]}
+                  className="w-full h-32 object-cover rounded border border-slate-200 dark:border-gray-700"
+                  onError={e => { (e.target as HTMLImageElement).style.display = 'none' }}
+                />
+              ) : (
+                <div className="w-full h-32 bg-slate-100 dark:bg-gray-800 rounded flex items-center justify-center text-xs text-slate-400">
+                  无
+                </div>
+              )}
+            </div>
+          )
+        })}
+      </div>
+      {c.status === 'needs_review' && (
+        <div className="flex gap-2 mt-3">
+          <button
+            onClick={() => onApprove(c.id)}
+            className="flex-1 py-1.5 text-xs font-medium bg-green-600 hover:bg-green-700 text-white rounded transition-colors"
+          >
+            ✓ 批准（更新基线）
+          </button>
+          <button
+            onClick={() => onReject(c.id)}
+            className="flex-1 py-1.5 text-xs font-medium bg-red-600 hover:bg-red-700 text-white rounded transition-colors"
+          >
+            ✗ 拒绝（保留旧基线）
+          </button>
+        </div>
+      )}
+    </div>
+  )
+}
+
+export function VisualView() {
+  const [runs, setRuns] = useState<VisualRun[]>([])
+  const [selectedRun, setSelectedRun] = useState<string | null>(null)
+  const [cases, setCases] = useState<VisualCase[]>([])
+
+  const pendingReviewTotal = runs.reduce((s, r) => s + r.pending_review, 0)
+
+  const loadRuns = () => fetchVisualRuns().then(setRuns).catch(console.warn)
+
+  useEffect(() => {
+    loadRuns()
+  }, [])
+
+  async function loadCases(runId: string) {
+    setSelectedRun(runId)
+    const cs = await fetchVisualCases(runId)
+    setCases(cs)
+  }
+
+  async function handleApprove(caseId: number) {
+    await approveVisualCase(caseId)
+    if (selectedRun) await loadCases(selectedRun)
+    loadRuns()
+  }
+
+  async function handleReject(caseId: number) {
+    await rejectVisualCase(caseId)
+    if (selectedRun) await loadCases(selectedRun)
+    loadRuns()
+  }
+
+  return (
+    <div className="flex h-full">
+      {/* 左侧：运行列表 */}
+      <div className="w-72 flex-shrink-0 border-r border-slate-200 dark:border-gray-700 overflow-y-auto">
+        <div className="sticky top-0 bg-white dark:bg-gray-950 px-4 py-3 border-b border-slate-200 dark:border-gray-700">
+          <h2 className="text-sm font-semibold">视觉测试运行</h2>
+          {pendingReviewTotal > 0 && (
+            <p className="text-xs text-amber-600 dark:text-amber-400 mt-0.5">
+              ⚠ {pendingReviewTotal} 个截图待审批
+            </p>
+          )}
+        </div>
+        {runs.length === 0 ? (
+          <div className="p-4 text-sm text-slate-400">暂无运行记录</div>
+        ) : (
+          runs.map(r => (
+            <RunRow
+              key={r.id}
+              run={r}
+              selected={selectedRun === r.id}
+              onClick={() => loadCases(r.id)}
+            />
+          ))
+        )}
+      </div>
+
+      {/* 右侧：case 列表 */}
+      <div className="flex-1 overflow-y-auto p-4">
+        {!selectedRun ? (
+          <div className="flex items-center justify-center h-full text-slate-400 text-sm">
+            选择左侧运行记录查看测试详情
+          </div>
+        ) : cases.length === 0 ? (
+          <div className="p-4 text-sm text-slate-400">该次运行暂无测试用例</div>
+        ) : (
+          <div className="grid gap-4">
+            {cases.map(c => (
+              <CaseRow key={c.id} c={c} onApprove={handleApprove} onReject={handleReject} />
+            ))}
+          </div>
+        )}
+      </div>
+    </div>
+  )
+}
@@ -0,0 +1,23 @@
+@tailwind base;
+@tailwind components;
+@tailwind utilities;
+
+body {
+  background-color: #f8fafc;
+  color: #1e293b;
+  font-family: 'Inter', system-ui, sans-serif;
+}
+
+.dark body {
+  background-color: #0f172a;
+  color: #e2e8f0;
+}
+
+@keyframes pulse-ring {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0.5; }
+}
+
+.running-pulse {
+  animation: pulse-ring 1.5s ease-in-out infinite;
+}
@@ -16,7 +16,7 @@ export async function fetchProjects() {
 export async function createTask(body: {
  project: string; type: string; title: string;
  priority?: number; mode?: string; agent_role?: string; context?: string;
-  initiator?: string; discussion?: string
+  initiator?: string; discussion?: string; queue?: string
 }) {
  const res = await fetch(`${BASE}/tasks`, {
    method: 'POST',
@@ -243,3 +243,22 @@ export async function fetchMarketHistory(): Promise<MarketHistoryItem[]> {
  if (!res.ok) throw new Error(`fetchMarketHistory failed: ${res.status}`)
  return res.json()
 }
+
+// ── Research Facts API ─────────────────────────────────────────────
+export interface ResearchFacts {
+  domains: Record<string, Record<string, string>>  // domain -> {sub_key -> value}
+}
+
+export async function fetchResearchFacts(): Promise<ResearchFacts> {
+  const res = await fetch(`${BASE}/research/facts`)
+  if (!res.ok) throw new Error(`fetchResearchFacts failed: ${res.status}`)
+  return res.json()
+}
+
+export async function refreshClaudeMd(project: string): Promise<{ id: number; title: string }> {
+  const res = await fetch(`${BASE}/projects/${encodeURIComponent(project)}/refresh-claude-md`, {
+    method: 'POST',
+  })
+  if (!res.ok) throw new Error(await res.text())
+  return res.json()
+}
@@ -0,0 +1,44 @@
+import { useTaskStore } from '../store/tasks'
+
+const WS_URL = `${location.protocol === 'https:' ? 'wss' : 'ws'}://${location.host}/ws`
+
+let ws: WebSocket | null = null
+let retryDelay = 1000
+
+export function connectWS() {
+  const store = useTaskStore.getState()
+
+  ws = new WebSocket(WS_URL)
+  store.setWsStatus('connecting')
+
+  ws.onopen = () => {
+    store.setWsStatus('connected')
+    retryDelay = 1000
+  }
+
+  ws.onmessage = (ev) => {
+    try {
+      const msg = JSON.parse(ev.data)
+      if (msg.type === 'snapshot') {
+        store.applySnapshot(msg.tasks)
+      } else if (msg.type === 'task_update' || msg.type === 'task_create') {
+        store.applyUpdate(msg.task)
+      }
+    } catch (e) {
+      console.warn('WS parse error', e)
+    }
+  }
+
+  ws.onclose = () => {
+    store.setWsStatus('reconnecting')
+    setTimeout(() => {
+      retryDelay = Math.min(retryDelay * 2, 30000)
+      connectWS()
+    }, retryDelay)
+  }
+
+  ws.onerror = () => {
+    store.setWsStatus('offline')
+    ws?.close()
+  }
+}
@@ -0,0 +1,10 @@
+import { StrictMode } from 'react'
+import { createRoot } from 'react-dom/client'
+import './index.css'
+import App from './App.tsx'
+
+createRoot(document.getElementById('root')!).render(
+  <StrictMode>
+    <App />
+  </StrictMode>,
+)
@@ -0,0 +1,94 @@
+import { create } from 'zustand'
+import type { Task, Project } from '../types'
+
+export interface Toast {
+  id: string
+  message: string
+  type: 'success' | 'error' | 'warning' | 'info'
+  taskId?: number
+}
+
+interface TaskStore {
+  tasks: Map<number, Task>
+  projects: Project[]
+  wsStatus: 'connecting' | 'connected' | 'reconnecting' | 'offline'
+  selectedTaskId: number | null
+  activeProject: string | null  // null = 全览
+  toasts: Toast[]
+  theme: 'dark' | 'light'
+
+  // actions
+  applySnapshot: (tasks: Task[]) => void
+  applyUpdate: (task: Task) => void
+  setProjects: (projects: Project[]) => void
+  setWsStatus: (s: TaskStore['wsStatus']) => void
+  setSelectedTask: (id: number | null) => void
+  setActiveProject: (p: string | null) => void
+  removeProjectTasks: (project: string) => void
+  removeToast: (id: string) => void
+  toggleTheme: () => void
+}
+
+const _savedTheme = (localStorage.getItem('theme') as 'dark' | 'light') || 'dark'
+document.documentElement.classList.toggle('dark', _savedTheme === 'dark')
+
+export const useTaskStore = create<TaskStore>((set) => ({
+  tasks: new Map(),
+  projects: [],
+  wsStatus: 'connecting',
+  selectedTaskId: null,
+  activeProject: null,
+  toasts: [],
+  theme: _savedTheme,
+
+  applySnapshot: (tasks) => {
+    const m = new Map<number, Task>()
+    tasks.forEach(t => m.set(t.id, t))
+    set({ tasks: m })
+  },
+
+  applyUpdate: (task) => {
+    set(state => {
+      const prev = state.tasks.get(task.id)
+      const m = new Map(state.tasks)
+      m.set(task.id, task)
+
+      // 检测状态转换，生成通知
+      const toasts = [...state.toasts]
+      if (prev && prev.status !== task.status) {
+        const id = `${task.id}-${task.status}-${Date.now()}`
+        const name = task.title.length > 30
+          ? task.title.slice(0, 30) + '…'
+          : task.title
+        if (task.status === 'done') {
+          toasts.push({ id, type: 'success', message: `✓ 完成：${name}`, taskId: task.id })
+        } else if (task.status === 'failed') {
+          toasts.push({ id, type: 'error', message: `✕ 失败：${name}`, taskId: task.id })
+        } else if (task.status === 'waiting_approval') {
+          toasts.push({ id, type: 'warning', message: `⚠ 待审批：${name}`, taskId: task.id })
+        }
+      }
+
+      return { tasks: m, toasts }
+    })
+  },
+
+  setProjects: (projects) => set({ projects }),
+  setWsStatus: (wsStatus) => set({ wsStatus }),
+  setSelectedTask: (selectedTaskId) => set({ selectedTaskId }),
+  setActiveProject: (activeProject) => set({ activeProject }),
+  removeProjectTasks: (project) => set(state => {
+    const m = new Map(state.tasks)
+    for (const [id, task] of m) {
+      if (task.project === project) m.delete(id)
+    }
+    return { tasks: m, selectedTaskId: null }
+  }),
+  removeToast: (id) => set(s => ({ toasts: s.toasts.filter(t => t.id !== id) })),
+  toggleTheme: () => set(s => {
+    const next = s.theme === 'dark' ? 'light' : 'dark'
+    document.documentElement.classList.toggle('dark', next === 'dark')
+    localStorage.setItem('theme', next)
+    return { theme: next }
+  }),
+}))
@@ -0,0 +1,74 @@
+export interface Task {
+  id: number
+  project: string
+  type: string
+  title: string
+  context: string
+  priority: number
+  mode: string
+  agent_role: string
+  status: 'pending' | 'running' | 'done' | 'failed' | 'waiting_approval' |
+          'vetoed' | 'adversarial_pending' | 'waiting_confirm'
+  created_at: string
+  started_at: string | null
+  completed_at: string | null
+  result_summary: string
+  retry_count: number
+  parent_task_id: number
+  constraint_type: string
+  awaiting_role: string
+  initiator: string
+  discussion: string
+}
+
+export interface Project {
+  name: string
+  mode: string
+  kind: string
+  description: string
+}
+
+export const STATUS_COLUMNS = [
+  'pending', 'running', 'done', 'failed', 'waiting_approval', 'vetoed'
+] as const
+
+export const STATUS_LABELS: Record<string, string> = {
+  pending: '待执行',
+  running: '运行中',
+  done: '完成',
+  failed: '失败',
+  waiting_approval: '待审批',
+  vetoed: '已否决',
+  adversarial_pending: '对抗中',
+  waiting_confirm: '等确认',
+}
+
+export const STATUS_COLORS: Record<string, string> = {
+  pending: 'border-gray-500 bg-gray-500/10',
+  running: 'border-blue-500 bg-blue-500/10 running-pulse',
+  done: 'border-green-500/50 bg-green-500/5 opacity-70',
+  failed: 'border-red-500 bg-red-500/10',
+  waiting_approval: 'border-orange-500 bg-orange-500/10',
+  vetoed: 'border-purple-500 bg-purple-500/10',
+  adversarial_pending: 'border-yellow-500 bg-yellow-500/10',
+  waiting_confirm: 'border-cyan-500 bg-cyan-500/10',
+}
+
+export const ROLE_COLORS: Record<string, string> = {
+  'developer':    'bg-blue-600',
+  'dev-embedded': 'bg-blue-500',
+  'dev-kernel':   'bg-indigo-600',
+  'dev-lowlevel': 'bg-indigo-500',
+  'dev-ml':       'bg-violet-600',
+  'dev-backend':  'bg-cyan-600',
+  'dev-test':     'bg-teal-600',
+  'dev-refactor': 'bg-slate-600',
+  'hw-engineer':  'bg-amber-600',
+  'tester':       'bg-teal-500',
+  'arch-system':  'bg-rose-600',
+  'arch-security':'bg-red-600',
+  'arch-reviewer':'bg-pink-600',
+  'pm':           'bg-fuchsia-600',
+  'productizer':  'bg-emerald-600',
+  'architect':    'bg-rose-500',
+}
@@ -0,0 +1,18 @@
+/** @type {import('tailwindcss').Config} */
+export default {
+  content: ['./index.html', './src/**/*.{ts,tsx}'],
+  darkMode: 'class',
+  theme: {
+    extend: {
+      colors: {
+        pending:  '#6b7280',
+        running:  '#3b82f6',
+        done:     '#22c55e',
+        failed:   '#ef4444',
+        waiting:  '#f97316',
+        vetoed:   '#a855f7',
+      },
+    },
+  },
+  plugins: [require('@tailwindcss/typography')],
+}
@@ -0,0 +1,28 @@
+{
+  "compilerOptions": {
+    "tsBuildInfoFile": "./node_modules/.tmp/tsconfig.app.tsbuildinfo",
+    "target": "ES2022",
+    "useDefineForClassFields": true,
+    "lib": ["ES2022", "DOM", "DOM.Iterable"],
+    "module": "ESNext",
+    "types": ["vite/client"],
+    "skipLibCheck": true,
+
+    /* Bundler mode */
+    "moduleResolution": "bundler",
+    "allowImportingTsExtensions": true,
+    "verbatimModuleSyntax": true,
+    "moduleDetection": "force",
+    "noEmit": true,
+    "jsx": "react-jsx",
+
+    /* Linting */
+    "strict": true,
+    "noUnusedLocals": true,
+    "noUnusedParameters": true,
+    "erasableSyntaxOnly": true,
+    "noFallthroughCasesInSwitch": true,
+    "noUncheckedSideEffectImports": true
+  },
+  "include": ["src"]
+}
@@ -0,0 +1,7 @@
+{
+  "files": [],
+  "references": [
+    { "path": "./tsconfig.app.json" },
+    { "path": "./tsconfig.node.json" }
+  ]
+}
@@ -0,0 +1,26 @@
+{
+  "compilerOptions": {
+    "tsBuildInfoFile": "./node_modules/.tmp/tsconfig.node.tsbuildinfo",
+    "target": "ES2023",
+    "lib": ["ES2023"],
+    "module": "ESNext",
+    "types": ["node"],
+    "skipLibCheck": true,
+
+    /* Bundler mode */
+    "moduleResolution": "bundler",
+    "allowImportingTsExtensions": true,
+    "verbatimModuleSyntax": true,
+    "moduleDetection": "force",
+    "noEmit": true,
+
+    /* Linting */
+    "strict": true,
+    "noUnusedLocals": true,
+    "noUnusedParameters": true,
+    "erasableSyntaxOnly": true,
+    "noFallthroughCasesInSwitch": true,
+    "noUncheckedSideEffectImports": true
+  },
+  "include": ["vite.config.ts"]
+}
@@ -0,0 +1,15 @@
+import { defineConfig } from 'vite'
+import react from '@vitejs/plugin-react'
+
+export default defineConfig({
+  plugins: [react()],
+  build: {
+    outDir: 'dist',
+  },
+  server: {
+    proxy: {
+      '/api': 'http://localhost:8080',
+      '/ws': { target: 'ws://localhost:8080', ws: true },
+    },
+  },
+})
@@ -0,0 +1,40 @@
+version: "3.9"
+
+services:
+  dashboard:
+    build:
+      context: .
+      dockerfile: Dockerfile.dashboard
+    image: nmfs-agents-dashboard:latest
+    container_name: agents-dashboard
+    ports:
+      - "9080:8080"
+    volumes:
+      - ./data:/app/data        # 共享 tasks.db（SQLite WAL 模式并发安全）
+      - ./configs:/app/configs  # 读取项目配置
+      - /data:/data             # 宿主机项目目录（用于读写项目 CLAUDE.md/README.md）
+      - /home/qiurui/.claude:/home/qiurui/.claude  # 全局 CLAUDE.md + OAuth 认证
+      - /home/qiurui/.claude.json:/home/qiurui/.claude.json  # Claude CLI 配置
+      - /home/qiurui/.local:/home/qiurui/.local    # claude CLI 二进制
+      - /data/company/visual-tester:/app/visual-tester
+    environment:
+      - DB_PATH=/app/data/tasks.db
+      - CONFIG_DIR=/app/configs
+      - DASHBOARD_PORT=8080
+      - CLAUDE_GLOBAL_DIR=/home/qiurui/.claude
+      - HOME=/home/qiurui
+      - PATH=/home/qiurui/.local/bin:/usr/local/bin:/usr/local/sbin:/usr/bin:/usr/sbin:/bin
+      - http_proxy=http://172.17.0.1:8889/
+      - https_proxy=http://172.17.0.1:8889/
+      - HTTP_PROXY=http://172.17.0.1:8889/
+      - HTTPS_PROXY=http://172.17.0.1:8889/
+      - ALL_PROXY=socks://172.17.0.1:1089/
+      - all_proxy=socks://172.17.0.1:1089/
+      - NO_PROXY=localhost,127.0.0.1,172.17.0.0/16
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "python", "-c",
+             "import urllib.request; urllib.request.urlopen('http://localhost:8080/api/tasks')"]
+      interval: 30s
+      timeout: 5s
+      retries: 3
@@ -0,0 +1,123 @@
+# AI安全领域市场深度分析报告
+
+> 更新时间：2026-03-10 | 分析人：Market-PM Agent | 平台：RK3588 NPU
+
+---
+
+## 一、市场规模
+
+| 维度 | 数据 |
+|------|------|
+| 全球 AI 视频监控市场（2025） | USD 59.8 亿（另一口径：$65.1亿） |
+| 全球 AI 视频监控市场（2030 预测） | USD 117.6–287.6 亿，CAGR 14.46%–30.6% |
+| 全球视频监控总市场（2025） | USD 387.4 亿，2026 年 $425.4 亿 |
+| 全球视频监控总市场（2035 预测） | USD 2,673.9 亿，CAGR 12.26% |
+| 全球边缘计算市场（2025） | USD 2,650 亿，2029 年接近翻倍 |
+| 中国智能安防市场（2028 预测） | 突破 2,000 亿元，AI 安防占比>50% |
+| 中国安防市场（海康+大华合计） | 国内市占超 70% |
+| 亚太地区 | 增速最高，中国/日本/韩国领跑 |
+
+---
+
+## 二、竞品全景（≥5 家）
+
+### 硬件/平台层
+
+| # | 厂商 | 产品 | 定价 | 核心技术 | 市场地位 |
+|---|------|------|------|---------|---------|
+| 1 | **海康威视** | DeepinView AI 摄像头/边缘盒子 | 消费摄像头¥250–1,000；AI 深眸¥5,000+ | 全球市占 23%，海思 Hi3516DV300 NPU 集成，20+ 风险行为实时预警 | 全球第一 |
+| 2 | **大华股份** | WizMind/WizSense 系列 | 商用定价 | 全球市占 10.5%，AI 算法+前端智能摄像机 | 全球第二 |
+| 3 | **宇视科技** | AI Smart 系列 | 商用定价 | 全栈安防产品，与阿里云深度合作 | 国内第三 |
+| 4 | **萤石（海康子公司）** | 家用 AI 摄像头 | ¥149–299 | 消费级人形识别、跌倒检测、一键呼叫 | 国内家用第一 |
+| 5 | **小米生态链** | 米家摄像头 | ¥199–499 | 人脸识别、人形预警、AI 降噪 | 中低端市占领先 |
+
+### 算法/AI 层
+
+| # | 厂商 | 定位 | 核心能力 |
+|---|------|------|---------|
+| 1 | **商汤科技** | 视觉 AI 算法 | SenseVideo 城市级视频分析、大模型支持 |
+| 2 | **旷视科技** | 人脸/视频 AI | 人脸识别全球领先，Brain++平台 |
+| 3 | **云从科技** | 金融/安防算法 | 行为分析、轨迹追踪，与海康深度合作 |
+| 4 | **华为云（Ascend）** | AI 算力+安防生态 | 商汤/旷视/依图均是华为安防生态伙伴 |
+
+### 边缘计算芯片/盒子层
+
+| 方案 | 算力 | 定价 | 场景 |
+|------|------|------|------|
+| 海思 Hi3516DV300 | ~0.5 TOPS | 模组¥50–200 | 端侧摄像头 NPU |
+| NVIDIA Jetson Nano | ~0.5 TOPS | $149 | 开发者/高端摄像头 |
+| NVIDIA Jetson TX1 | ~1 TOPS | $299+ | 高端 AI 摄像头（海康深眸¥5000+） |
+| **RK3588（本平台）** | **6 TOPS INT8** | 模块 ¥800–1,200 | NPU 算力是海思方案 10x+ |
+
+---
+
+## 三、近期动态（2025年9月–2026年3月）
+
+| 时间 | 事件 | 影响 |
+|------|------|------|
+| 2026.05（预告） | 杭州安博会发布《2026-2028 全球智能安防产业发展白皮书》 | 行业趋势指向大模型+边缘 AI 融合 |
+| 2026.01 | 中国智能安防市场预测 2028 年突破 2000 亿，AI 占比>50% | AI 渗透安防进入加速拐点 |
+| 2025 | 边缘 AI 渗透率仍不足 1%，巨头加速投入 | 蓝海窗口期，先发优势明显 |
+| 2025 | 全球边缘计算 $2650 亿，AI 驱动近翻倍增长（IDC） | 边缘算力需求量级提升 |
+| 2023 | 安博会：大模型应用扎堆，AI 摄像头满场 | 行业已完成 AI 化初步认知教育 |
+
+---
+
+## 四、用户痛点
+
+| 痛点 | 类型 | 描述 |
+|------|------|------|
+| **算法授权费高昂** | B/G 端 | 商汤/旷视算法 SaaS 按通道数收费，中小客户难以承受 |
+| **私有化部署难** | B 端 | 云端方案数据安全顾虑；工厂/机构不允许数据出厂 |
+| **边缘算力不足** | 技术 | 主流边缘摄像头 NPU ≤1 TOPS，复杂场景精度差 |
+| **多厂商不兼容** | 运维 | 海康/大华/宇视协议碎片化，统一管理成本高 |
+| **实施周期长** | B/G 端 | 传统安防项目动辄 3–6 个月，快速交付需求未满足 |
+| **误报率高** | 终端用户 | AI 行为分析误报导致报警疲劳，运维人力浪费 |
+
+---
+
+## 五、RK3588 NPU 差异化机会
+
+### 平台算力优势
+| 对比维度 | RK3588 NPU | 海思 Hi3516DV300 | Jetson TX1 |
+|---------|-----------|----------------|-----------|
+| NPU 算力 | **6 TOPS INT8** | ~0.5 TOPS | ~1 TOPS |
+| 成本（模块） | ¥800–1,200 | ¥50–200 | $299+ |
+| 算力/成本比 | ★★★★★ | ★★★ | ★★★ |
+| 已有模型 | YOLO+MediaPipe+Embedding | - | - |
+
+### 3 大切入机会
+
+#### 机会 1：中小企业私有化 AI 安防盒子（最高优先级）
+- **痛点**：海康/大华私有化 AI 盒子 ¥5,000–50,000，算法 SaaS 按通道收费贵
+- **方案**：RK3588 + 已有 YOLO + 行为分析算法，一次性硬件 ¥3,000–8,000 包含 5 路分析
+- **竞争优势**：算力 10x（vs 海思方案），无月费，私有化本地部署
+- **目标客户**：中小工厂、仓储、商超、物业
+
+#### 机会 2：适老场景 AI 安全监控（与适老领域协同）
+- **方案**：养老院/社区 AI 跌倒检测+异常行为分析，合规隐私保护（深度图而非 RGB）
+- **差异化**：专项适老算法（骨骼+行为），与适老领域产品协同，一盒多用
+
+#### 机会 3：运动场馆安防+分析一体（与运动领域协同）
+- **方案**：运动场馆 AI 摄像头，安防+动作分析双功能，单台设备实现双价值
+- **差异化**：安防厂商只做安全，本方案额外输出训练数据，提升客户价值
+
+### 技术路径
+```
+摄像头 RTSP → RK3588 NPU 推理 → YOLO（多目标检测）
+                               → 行为异常分析（本地规则引擎）
+                               → MediaPipe（姿态检测，适老/运动）
+                               → Embedding（人员 Re-ID）
+                               → 本地告警 + HTTP API → 监控平台
+```
+
+---
+
+## 六、切入建议
+
+1. **算力降维打击**：以 6 TOPS vs 竞品 0.5 TOPS 作为核心卖点
+2. **私有化本地**：无月费、无云依赖、一次性交付，打破算法 SaaS 订阅壁垒
+3. **横向协同**：运动+适老+安全三场景共用一套推理平台，降低边际成本
+4. **先做 B 端**：中小工厂/物业/养老院，避开海康/大华的渠道体系
+
+*参考来源：Fortune Business Insights、Grand View Research、Mordor Intelligence、佰泰盛世科技、智东西*
@@ -0,0 +1,112 @@
+# 适老领域市场深度分析报告
+
+> 更新时间：2026-03-10 | 分析人：Market-PM Agent | 平台：RK3588 NPU
+
+---
+
+## 一、市场规模
+
+| 维度 | 数据 |
+|------|------|
+| 全球 AI 养老市场规模（2025） | USD 56.78 亿 |
+| 全球 AI 养老市场规模（2035 预测） | USD 387.52 亿，CAGR 21.3% |
+| 中国智能养老设备市场（2024） | 1,560 亿元 |
+| 中国智能养老设备市场（2029 预测） | 3,420 亿元，CAGR 17%（2024-2029） |
+| 中国养老机器人市场（2024） | 79 亿元，预计 2025 年 91 亿元 |
+| 智能护理机器人（2025 预测） | 突破 300 亿元，年均增长>20%；2035 年突破万亿元 |
+| 中国 60 岁以上人口 | >3 亿（占总人口 22%），65岁以上跌倒率 28~42% |
+| 中国独居老人数量 | 约 3,000 万 |
+| 全球跌倒检测设备市场（2025） | USD 14.1 亿，2034 年 USD 41.8 亿 |
+
+---
+
+## 二、竞品全景（≥5 家）
+
+### 国内竞品
+
+| # | 厂商 | 产品/方案 | 定价区间 | 技术特点 | 目标用户 |
+|---|------|----------|---------|---------|---------|
+| 1 | **华为** | WATCH D2/全屋 AI 康养传感器 | 手表¥1,099–3,299；传感器未公开 | ECG+血压+鸿蒙闭环；全屋传感器卫生间跌倒/卧室坠床检测；与优必选/智元联合推康养人形机器人（2025.05） | 银发+全年龄，24.5% 市场份额 |
+| 2 | **小米** | 小米手环 H1（适老款） | ¥399–799 | 一键 SOS、跌倒检测、血压/心率异常预警 | 60 岁以上独居老人 |
+| 3 | **爱牵挂** | 老人定位手环/手表 | ¥199–599 设备 + 年服务费 | GPS 定位+SOS+健康监测，SaaS 平台 | 居家/机构老人 |
+| 4 | **德心小宝** | 5D 毫米波 + AI 养老机器人 | ¥2 万+ | 非接触式，5D 毫米波+AI，申报 CFDA；目标 2025 年营收 1 亿，2026 年冲刺港股 | 养老机构/社区 |
+| 5 | **猎户星空（豹小秘）** | 豹小秘 Mini 养老机器人 | ¥2–5 万 | 视觉+4D 雷达融合，3秒跌倒预警，FHIR 医疗数据链路 | 医疗机构/高端养老 |
+| 6 | **麦迪科技** | 康养护理机器人（联合优必选） | 未公开（面向机构） | 7 大功能：情感陪伴/苏醒刺激/健康监测等；傅利叶/优必选机器人+麦迪临床数据 | 医院/康复/养老院 |
+| 7 | **傅利叶智能** | GR-1（科研）/ GR-3 Care-bot（陪伴） | ¥3–14 万 | 通用人形机器人，康复机器人+外骨骼助行；占养老机器人市场 55% 康复方向 | 康复机构/养老机构 |
+
+### 国际竞品
+
+| # | 厂商 | 产品 | 定价 | 特点 |
+|---|------|------|------|------|
+| 1 | **Apple Watch S10** | 适老跌倒检测 | USD 399+ | 碰撞+跌倒检测，AI 升级 S10 芯片，误报率降低 |
+| 2 | **Teton.ai（美国）** | AI 预测性护理平台 | SaaS 机构定价 | 2025.09 Series A $2000 万 |
+| 3 | **Panasonic** | 居家 AI 跌倒检测系统 | 商用定价 | 2026.01 发布，日本先发 |
+
+---
+
+## 三、近期动态（2025年9月–2026年3月）
+
+| 时间 | 事件 | 影响 |
+|------|------|------|
+| 2026.01 | Panasonic 推出居家 AI 跌倒检测系统（日本） | 国际巨头入场，证明场景可行性 |
+| 2025.11 | 中国适老化智能手环销售额同比增速超 200%；智能手表超 350% | 银发适老硬件已进入爆发增长期 |
+| 2025.05 | 华为+优必选+智元机器人+中坚科技合作康养人形机器人 | 头部企业联合布局，行业加速整合 |
+| 2025.09 | Teton.ai Series A $2000 万（AI 预测性老年护理） | 海外资本验证智能看护 SaaS 模式 |
+| 2025.10 | 阿里《2025 银发+AI 应用趋势报告》：老年人深度参与 AI | 老年用户对 AI 接受度大幅提升 |
+| 2025.07 | 田园之家（居家智慧养老）完成 A 轮融资 | 居家养老 SaaS 赛道融资活跃 |
+
+---
+
+## 四、用户痛点
+
+| 痛点 | 影响人群 | 现有方案局限 |
+|------|---------|------------|
+| **跌倒安全风险** | 65岁以上跌倒率 28~42% | 可穿戴接受度低（老人不愿戴）；摄像头侵犯隐私 |
+| **独居无人陪伴** | 3000 万独居老人 | 定时呼叫成本高；子女无法实时感知 |
+| **穿戴设备难用** | 高龄老人 | 操作复杂、字体小、充电遗忘 |
+| **指纹门锁失效** | 皮肤干燥老人 | 指纹磨损误识率高，掌静脉/人脸成刚需 |
+| **隐私数据泄露** | 全量老人 | 云端上传健康数据监管风险 |
+| **慢病管理断层** | 高血压患者（2.45亿，60岁以上占50%+） | 数据孤岛，医疗-家庭无法联通 |
+| **设备成本过高** | 机构采购方 | 毫米波方案 ¥2万+，高端机器人 ¥3–20万，性价比不足 |
+
+---
+
+## 五、RK3588 NPU 差异化机会
+
+### 平台能力基准
+| 能力 | 参数 |
+|------|------|
+| NPU 算力 | 6 TOPS INT8，三核并行最高 1.9x |
+| 姿态估计（mediapipe） | 33 关键点骨架，板端实时推理 |
+| 视觉推理 | YOLO 多人检测 30+ FPS |
+| 成本优势 | NanoPC-T6（RK3588）模块 ¥800–1,200 |
+
+### 差异化机会矩阵
+
+| 场景 | 痛点 | RK3588 方案 | 优势 |
+|------|------|------------|------|
+| **非接触跌倒检测** | 摄像头泄露隐私，毫米波成本高（¥2万+） | 深度摄像头+mediapipe 骨骼估计，边缘推理，数据不出设备 | 隐私保护 + 成本 ¥2,000–5,000 |
+| **卧室/卫生间监护** | 传统摄像头无法合规部署 | 深度图+骨骼检测（无 RGB 画面传输） | 合规可部署私密场所 |
+| **睡眠监测** | 毫米波精度差，商用设备贵 | 毫米波+视觉融合，RK3588 并行推理 | 多模态精度，成本可控 |
+| **摔倒后自动呼叫** | 需要手动 SOS | NPU 实时姿态异常→本地拨号 | 端侧自动化，无网络依赖 |
+| **异常行为检测** | 需 7×24 人工看护 | 长时间不动/异常行为时序分析，本地告警 | 替代人工看护 |
+
+### 产品 BOM 估算
+```
+[居家适老 AI 盒子] = RK3588 + 深度摄像头 + 本地推理
+- RK3588 核心板（NanoPC-T6 类）：¥800–1,200
+- 深度摄像头（OAK-D Lite 类）：¥600–900
+- 外壳 + 电源：¥100–200
+- 合计 BOM：¥1,500–2,300（量产可降至 ¥1,000 以内）
+- 竞品对比：德心小宝（毫米波方案）¥2万+，猎户豹小秘 ¥2–5万
+```
+
+---
+
+## 六、切入建议
+
+1. **主打"隐私保护+边缘推理"**，避开华为/小米可穿戴红海
+2. **B 端先行**（养老机构/社区），积累数据与 CFDA 资质 → C 端扩展
+3. **技术 MVP**：基于现有 mediapipe + RK3588，快速验证跌倒检测准确率（目标 >95%）
+
+*参考来源：AgeClub、新浪财经、东方财富、OFweek 可穿戴网、Precedence Research*
@@ -0,0 +1,134 @@
+# 适老领域 AI 市场深度分析报告
+
+> 生成日期：2026-03-09 | 作者：Market-PM Agent | 版本：v1.0
+
+---
+
+## 1. 市场概况
+
+| 维度 | 数据 |
+|------|------|
+| 全球 AI 养老市场规模（2025） | USD 56.78 亿 |
+| 全球 AI 养老市场规模（2035 预测） | USD 387.52 亿 |
+| CAGR（2026-2035） | 21.30% |
+| 中国 60 岁以上人口 | >3 亿（占总人口 22%） |
+| 中国独居老人数量 | 约 3000 万 |
+| 全球跌倒检测设备市场（2025） | USD 14.1 亿 |
+| 全球跌倒检测设备市场（2034 预测） | USD 41.8 亿 |
+| 中国智能腕戴市场（2025） | 7100 万台，同比 +13.1% |
+
+---
+
+## 2. 竞品全景（≥5 家）
+
+### 2.1 国内竞品
+
+| # | 厂商 | 产品/方案 | 定价区间 | 技术特点 | 目标用户 |
+|---|------|----------|---------|---------|---------|
+| 1 | **华为** | WATCH D2 / Watch GT 系列 | ¥1,099~3,299 | ECG+血压+鸿蒙生态闭环，"腕戴-手机-家居-医疗"全场景 | 银发 + 全年龄段，24.5% 市场份额 |
+| 2 | **小米** | 小米手环 H1（适老款） | ¥399~799 | 一键 SOS，跌倒检测，血压/心率异常预警 | 60 岁以上独居老人 |
+| 3 | **爱牵挂** | 老人定位手环/手表 | ¥199~599/年服务费 | GPS 定位 + SOS + 健康监测，SaaS 平台 | 居家/机构老人 |
+| 4 | **德心小宝** | 5D 毫米波 + AI 养老机器人 | 约 ¥2 万+ | 5D 毫米波雷达+AI 算法，非接触式，2025 年申报 CFDA | 养老机构 / 社区 |
+| 5 | **云米科技** | 心肺监测雷达 | ¥1,500~3,000 | 24GHz 毫米波，无感呼吸/心率监测，无需穿戴 | 家庭安装 |
+| 6 | **猎户星空（豹小秘）** | 豹小秘 Mini 养老机器人 | ¥2~5 万 | 视觉+4D雷达融合，3秒跌倒预警，FHIR 医疗数据链路 | 医疗机构/高端养老 |
+| 7 | **国脉科技** | AI 居家养老智能体 | 政府采购定价 | 大模型+居家服务调度，主打 G 端 | 政府/街道养老站 |
+
+### 2.2 国际竞品
+
+| # | 厂商 | 产品 | 定价（美元） | 特点 |
+|---|------|------|------------|------|
+| 1 | **Teton.ai** | AI 预测性护理平台 | SaaS 机构定价 | 2025.09 获 Series A $2000万 |
+| 2 | **Panasonic** | AI 跌倒检测系统 | 商用定价 | 2026.01 发布，家用版本 |
+| 3 | **Cera Care（英国）** | 家庭医疗 AI | 订阅制 | 看护人员调度+AI 监测 |
+| 4 | **Apple Watch** | Series 10 血糖监测 | USD 399+ | 无创血糖（误差<5%），AI 跌倒检测 |
+
+---
+
+## 3. 近 6 个月重要动态（2025.09 ~ 2026.03）
+
+### 3.1 技术发布
+- **2026.01** Panasonic 推出居家 AI 跌倒检测系统（日本市场先发）
+- **2025.11** 中国智能腕戴全年预计 7100 万台，银发适老化产品增速 200~350%
+- **2025.10** 阿里巴巴发布《2025 银发+AI 应用趋势报告》，老年人 AI 使用深度参与者比例上升
+- **2025.09** 掌静脉门锁因适老优势销售额同比增长 150%，指纹识别被替代趋势明显
+
+### 3.2 融资动态
+- **2025.09** Teton.ai（美国）Series A 融资 $2000 万，专注 AI 预测性老年护理
+- **2025年** 欧盟拨款 €5000 万用于养老机构 AI 整合项目
+- **2025.07** 田园之家（居家智慧养老）完成 A 轮，金沙江联合资本领投
+- **德心小宝** 目标 2025 年营收 1 亿，2026 年冲刺港股 Pre-IPO
+
+---
+
+## 4. 用户痛点调研
+
+### 4.1 核心痛点（重要度排序）
+
+| 痛点 | 影响人群 | 现有方案局限 |
+|------|---------|------------|
+| **跌倒安全风险** | 65岁以上跌倒率 28~42% | 可穿戴接受度低（老人不愿戴）；摄像头侵犯隐私 |
+| **独居无人陪伴** | 3000 万独居老人 | 定时呼叫人力成本高；子女无法实时感知 |
+| **穿戴设备难用** | 认知能力下降的高龄老人 | 操作复杂、字体小、充电难（遗忘） |
+| **指纹门锁失效** | 皮肤干燥、指纹磨损老人 | 指纹锁误识率高 → 掌静脉/人脸成刚需 |
+| **隐私数据泄露** | 全量老人 | 云端上传健康数据存在风险（监管加强） |
+| **慢病管理断层** | 高血压患者（2.45亿，60岁以上占50%+） | 数据孤岛，医疗-家庭无法联通 |
+| **认知障碍走失** | 阿尔茨海默症老人 | GPS 定位精度不足室内使用 |
+
+### 4.2 社区/论坛调研发现
+- 子女最关注：**跌倒实时报警** > 健康数据查看 > 位置确认
+- 老人本人最关注：**不想被监视感**（隐私）、操作简单、不额外充电
+- 机构采购关注：数据合规、CFDA 资质、与现有护理系统集成
+
+---
+
+## 5. RK3588 NPU 差异化机会分析
+
+### 5.1 平台能力基准（基于已有项目事实）
+
+| 能力 | 参数 |
+|------|------|
+| NPU 算力 | 6 TOPS（INT8），三核并行最高 1.9x 加速 |
+| 视觉推理 | YOLO v8n 30+ FPS，姿态估计（mediapipe）实时 |
+| 向量检索 | embedding 本地化，<10ms 延迟 |
+| 内存/算力比 | 远优于树莓派 4B，接近工作站级嵌入式 |
+
+### 5.2 差异化机会矩阵
+
+| 场景 | 现有痛点 | RK3588 方案 | 竞争优势 |
+|------|---------|------------|---------|
+| **非接触跌倒检测** | 摄像头泄露隐私，毫米波成本高 | 深度摄像头+骨骼估计（mediapipe）边缘推理，数据不出设备 | 隐私保护 + 成本低于毫米波方案 |
+| **卧室/卫生间监护** | 传统摄像头无法部署 | 深度图+骨骼检测（无 RGB 画面传输） | 合规可部署私密场所 |
+| **掌静脉门禁** | 老人指纹识别失败率高 | 红外+静脉纹路 CNN，NPU 加速推理<50ms | 替代方案，无需云端验证 |
+| **睡眠监测** | 单纯毫米波精度差，商用设备贵 | 毫米波+视觉融合，RK3588 并行推理 | 多模态精度高，成本可控 |
+| **摔倒后自动呼叫** | 需要手动 SOS | NPU 实时姿态分析→异常触发→本地拨号 | 端侧自动化，无网络依赖 |
+| **独居老人异常行为** | 需要 7×24 人工看护 | 行为时序分析（长时间不动/异常）本地触发告警 | 替代人工看护，成本极低 |
+
+### 5.3 可行产品形态
+
+```
+[居家适老 AI 盒子] = RK3588 + 深度摄像头 + 本地推理
+├── 跌倒检测（mediapipe 骨骼估计）→ 误报率 <5%
+├── 睡眠呼吸监测（可选毫米波模块）
+├── 异常行为检测（长时间不动 / 徘徊）
+├── 隐私保护：所有推理本地，不上传视频
+└── 告警通道：本地 WiFi → 子女 App / 护理平台
+
+BOM 估算（初版）：
+- RK3588 核心板（NanoPC-T6 类）：¥800~1200
+- 深度摄像头（OAK-D Lite 类）：¥600~900
+- 外壳 + 电源：¥100~200
+- 合计 BOM：¥1500~2300（量产可降至 ¥1000 以内）
+```
+
+---
+
+## 6. 结论与建议
+
+1. **市场时机成熟**：银发经济政策加持 + 人口结构红利，适老 AI 设备进入快速增长期
+2. **差异化定位**：主打"隐私保护+边缘推理"，避开华为/小米的可穿戴红海
+3. **切入路径**：B 端先行（养老机构/社区）→ 积累数据与 CFDA 资质 → C 端扩展
+4. **技术 MVP**：基于现有 mediapipe + RK3588，快速验证跌倒检测准确率（目标 >95%）
+
+---
+
+*参考来源：AgeClub、OFweek 可穿戴网、前瞻研究院、InsightAce Analytics、Precedence Research*
@@ -0,0 +1,96 @@
+# 运动领域市场深度分析报告
+
+> 更新时间：2026-03-10 | 分析人：Market-PM Agent | 平台：RK3588 NPU
+
+---
+
+## 一、竞品全景（7 家核心玩家）
+
+| 品牌/产品 | 定价区间 | 用户规模 | 核心技术特点 | 盈利模式 |
+|---------|--------|--------|------------|--------|
+| **Apple Watch S10/Ultra2** | $399–$799（国内¥2999起），Ultra2 $799 | 全球可穿戴市占 23% | S10/S9 芯片、ECG/血氧/摔倒检测、AI Siri 整合、跑步功率 | 硬件+Fitness+ 订阅 |
+| **Garmin Fenix 8** | $999–$1,199（国内¥6800–8700） | 2026年Q2运动线收入$6.05亿（YoY +41%） | 专业运动模式 50+、TopoActive 地图、VO2 Max/体力预测、Elevate Gen5 传感器 | 纯硬件高毛利 |
+| **华为 Watch GT5 Pro** | ¥2,699–3,999 | 全球可穿戴市占 17%（Q1 2025），出货同比+36% | 连续血氧+ECG+盘古大模型健康问答、鸿蒙生态闭环，心率追踪接近 Garmin Fenix 8 水准 | 硬件（生态锁定） |
+| **Amazfit T-Rex Ultra 2** | $299–$399 | 快速增长，TikTok 营销覆盖 800万次曝光 | Garmin 专业功能约一半价格，耐用性强，运动姿态分析 | 纯硬件 |
+| **WHOOP 5.0** | 硬件免费 + $30/月订阅 | 约 500 万付费会员 | 持续 HRV/睡眠/恢复评分；Advanced Labs 血液生物标志物（65项）整合；无屏幕极简 | 纯 SaaS 订阅 |
+| **Oura Ring 4** | 硬件$349 + $5.99/月 | 累计 550 万枚；2025 年营收$10亿（YoY +100%） | 睡眠/经期/准备度评分；AI 顾问聊天机器人；Dexcom 血糖整合 | 硬件+轻订阅 |
+| **Keep（Kinetic.ai）** | 会员¥25/月；硬件¥599–2,999 | 月活 3000 万+；2024 年营收¥20.66亿，H1 2025 扭亏为盈 | 垂直运动大模型；AI 教练"卡卡"姿态识别+个性化计划；线上课 | 订阅+硬件+广告 |
+
+### 补充玩家
+- **小米 Redmi Band/Watch**：Q1 2025 出货 870 万台（全球第一），¥199–599，主打下沉市场
+- **COROS/Suunto/Polar**：专业耐力运动（铁三/越野跑），$300–700，GPS 精度和续航领先
+- **RingConn**：智能戒指，70% 品牌知名度（2025.01），舒适佩戴，专注健康监测
+
+---
+
+## 二、近期重要动态（2025年9月–2026年3月）
+
+| 时间 | 事件 | 影响 |
+|------|------|------|
+| 2026.02 | Garmin 运动线 Q2 收入 $6.05亿，同比 +41% | 专业运动市场快速扩张，抢夺 Apple 高端份额 |
+| 2025.11 | 中国智能腕戴全年达 7100 万台（历史新高） | 国内运动健康穿戴持续大规模增长 |
+| 2025.09 | WHOOP 上线 Advanced Labs（65项血液生物标志物+AI整合） | 运动健康进入医疗级生物标志物赛道 |
+| 2025.08 | Keep H1 扭亏为盈，调整净利润¥1035万，毛利率 52.2% | AI 驱动运动 SaaS 模式已跑通盈利 |
+| 2025.03 | Keep 上线 AI 教练"卡卡"，基于 Kinetic.ai 大模型 | 国内运动 AI 教练首发，3000万月活覆盖 |
+| 2025 | Oura 累计融资超$5亿，估值约$50亿 | 轻订阅可穿戴已是十亿美元级产品 |
+
+## 市场规模
+- 全球运动手表市场（2024）：**$97.55 亿**，预计 2031 年 $173 亿（CAGR 8.6%）
+- 中国智能健身市场（2025 预测）：**¥820 亿**
+- 全球智能手表市场（2025）：**$320 亿**，预计 2033 年 $795 亿
+
+---
+
+## 三、用户痛点
+
+| 痛点 | 严重程度 | 描述 |
+|------|---------|------|
+| **续航 vs 功能取舍** | ★★★★★ | 开 GPS 后 Garmin 续航仅 15h，Apple Watch 不足 24h |
+| **AI/传感器精准度不足** | ★★★★☆ | 心率较准，血氧/血压仅供参考，ECG 无法诊断 |
+| **专项运动识别偏差** | ★★★★☆ | 骑行距离偏差大，球类运动（羽毛球/篮球）识别率低 |
+| **个性化训练不足** | ★★★☆☆ | 建议通用化，不考虑疲劳状态和历史水平 |
+| **订阅费用叠加** | ★★★☆☆ | 设备本身价高，完整功能还需额外订阅 |
+| **专业分析系统价格门槛高** | ★★★★★ | 商用 AI 运动分析系统（商汤/海康）¥10万+，中小场馆无法承受 |
+| **云端隐私顾虑** | ★★★☆☆ | 云端视频分析让用户担忧运动数据安全 |
+
+---
+
+## 四、RK3588 NPU 差异化机会分析
+
+### 平台核心能力（实测数据）
+| 能力 | 规格 |
+|------|------|
+| NPU 算力 | 6 TOPS INT8（三核并行 1.3~1.9x 加速） |
+| MediaPipe Pose | 已完成板端部署（/data/company/mediapipe），33 关键点实时骨架 |
+| YOLO 目标检测 | 已完成板端部署（/data/company/yolo），多人实时检测 |
+| Embedding 向量检索 | 已完成（/data/company/embedding），动作序列比对 |
+| 视频处理 | 硬解码 8K@60fps，H.265/H.264 |
+
+### 4 大差异化机会
+
+#### 机会 1：中小场馆智能运动分析终端（最高优先级）
+- **痛点**：商用方案 ¥10万+，中小场馆负担不起；云端有隐私风险
+- **方案**：RK3588 本地推理（MediaPipe 关键点 + YOLO 多人检测），设备成本 ¥3,000–8,000/台
+- **竞争壁垒**：全本地、低延迟（<100ms）、无订阅、无隐私风险
+- **目标客户**：社区球场、中学体育馆、健身房、球类俱乐部
+
+#### 机会 2：视觉辅助纠姿系统（高优先级）
+- **痛点**：手环传感器无法捕捉动作质量，实时反馈延迟大
+- **方案**：固定摄像头 + RK3588 实时姿态分析，与穿戴数据融合（心率+骨架联合分析）
+- **差异化**：弥补纯手环无视觉盲区，提供"教练级"实时动作反馈
+
+#### 机会 3：本地 AI 运动教练盒子（中优先级）
+- **痛点**：Keep 等 App 需联网、订阅费高
+- **方案**：RK3588 离线推理盒，含标准化动作库（150+ 动作），家庭/企业健身房使用
+- **商业模式**：硬件一次性购买 + 动作包扩展
+
+#### 机会 4：运动视频自动精彩剪辑（中优先级）
+- **方案**：RK3588 本地视频事件检测（YOLO + 事件识别），自动导出精彩片段
+- **对标**：影石 AI 高光助手（依赖手机），本方案完全本地化
+
+---
+
+## 五、可复用模块
+`摄像头RTSP` → `YOLO 多人检测` → `MediaPipe 骨架提取` → `Embedding 动作分析` → `本地展示/API`
+
+*参考来源：洛图科技、Tom's Guide、TechRadar、新浪财经、Keep 财报*
@@ -0,0 +1,222 @@
+# Rockchip Agents 系统设计文档
+
+**日期**: 2026-02-27
+**状态**: 已批准，待实施
+
+---
+
+## 1. 目标
+
+构建一个 AI Agent 系统，通过管理 `/data/rockchip/` 下的模型相关项目（yolo / mediapipe / embedding / info-privacy / logodetect-x86），使用 Claude API 持续进行综合优化（bug 修复、性能提升、功能迭代、代码质量），并支持飞书通知与交互。
+
+---
+
+## 2. 整体架构
+
+### 2.1 核心循环
+
+```
+APScheduler（每小时）
+    └─→ Scanner 扫描所有项目（CLAUDE.md / 测试结果 / TODO）
+            └─→ 生成任务 → 写入 SQLite 任务队列
+                    └─→ Manager Agent 调度
+                            ├─→ asyncio 并发分发（最多 N 个并发）
+                            └─→ 专门 Agent 执行（Developer / Tester / ...）
+                                    └─→ 飞书通知结果
+```
+
+### 2.2 Agent 角色体系
+
+```
+Manager Agent（项目经理）
+  │  协调工作、分配任务、跟踪进度、飞书汇报
+  │
+  ├── Developer Agent（开发者）
+  │     修复 bug、实现功能、重构代码
+  │     工具：read_file, write_file, run_tests
+  │
+  ├── Tester Agent（测试者）
+  │     验证修复、运行基准、板端部署测试
+  │     工具：run_local_tests, ssh_device, collect_benchmark
+  │
+  ├── Productizer Agent（产品化）
+  │     评估成熟度、完善文档/API/配置、打包发布
+  │     工具：read_project, write_docs, check_release_readiness
+  │
+  └── Architect Agent（架构师）
+        跨项目分析 → 识别整合机会 → 定义统一产品架构
+        安排预研任务（如：yolo + embedding → 视觉语义搜索）
+        工具：analyze_all_projects, create_research_task, write_architecture_doc
+```
+
+每个角色 = 不同的 System Prompt + 工具集 + 决策权限。
+
+---
+
+## 3. 组件设计
+
+### 3.1 目录结构
+
+```
+src/rockchip_agents/
+├── agents/
+│   ├── manager.py       # Manager Agent（调度协调）
+│   ├── developer.py     # Developer Agent（代码修改）
+│   ├── tester.py        # Tester Agent（测试验证）
+│   ├── productizer.py   # Productizer Agent（产品化）
+│   └── architect.py     # Architect Agent（架构整合）
+├── core/
+│   ├── scheduler.py     # APScheduler 定时入口
+│   ├── scanner.py       # 项目扫描器
+│   ├── queue.py         # 任务队列（SQLite）
+│   └── executor.py      # 并发执行协调（asyncio）
+├── tools/
+│   ├── file_tools.py    # read_file / write_file
+│   ├── test_tools.py    # run_tests / collect_results
+│   ├── device_agent.py  # SSH 设备管理（部署/板端测试）
+│   └── feishu.py        # 飞书通知 + 交互卡片
+└── config.py            # 配置加载
+configs/
+├── projects.yaml        # 被管理项目列表及设置
+├── agents.yaml          # 全局配置（API Key、飞书、并发数）
+└── devices.yaml         # 设备 SSH 配置
+tests/
+scripts/
+└── run.py               # 启动入口
+```
+
+### 3.2 任务队列（SQLite）
+
+```sql
+CREATE TABLE tasks (
+    id          INTEGER PRIMARY KEY AUTOINCREMENT,
+    project     TEXT NOT NULL,
+    type        TEXT NOT NULL,  -- fix_bug | improve_perf | add_feature | code_review | productize | architect
+    title       TEXT NOT NULL,
+    context     TEXT,           -- 相关代码/日志片段
+    priority    INTEGER,        -- 1（严重）→ 5（低）
+    mode        TEXT,           -- auto | confirm | report
+    agent_role  TEXT,           -- developer | tester | productizer | architect
+    status      TEXT DEFAULT 'pending',  -- pending | running | done | failed | waiting_confirm
+    created_at  DATETIME,
+    started_at  DATETIME,
+    completed_at DATETIME,
+    result_summary TEXT
+);
+```
+
+### 3.3 三种执行模式
+
+| 模式 | Agent 行为 | 飞书消息 |
+|------|-----------|---------|
+| `auto` | 直接写代码 → 跑测试 → 完成 | 结果通知 |
+| `confirm` | 生成 diff 草案 → 发飞书卡片等待审批 → 执行 | 交互卡片（批准/拒绝） |
+| `report` | 只输出分析报告，不修改代码 | 报告摘要 |
+
+### 3.4 Device Agent（双侧同步）
+
+```python
+# 同步工作流
+x86: Developer Agent 修复 → 本地测试通过
+                         ↓
+               DeviceAgent.sync(project, device)  # rsync
+                         ↓
+               DeviceAgent.run(device, "pytest tests/ -v")
+                         ↓
+               收集板端 benchmark 结果 → 回传 → 飞书通知
+```
+
+支持多设备配置，当前默认 `pi@192.168.0.127`（RK3588），可扩展 RV1106 等。
+
+---
+
+## 4. 飞书集成
+
+### 4.1 通知消息（单向 Webhook）
+
+- 任务完成摘要
+- 板端验证结果（FPS / 精度对比）
+- 错误告警
+- 架构师定期跨项目分析报告
+
+### 4.2 交互卡片（`confirm` 模式，需飞书 Bot）
+
+```
+┌──────────────────────────────────────┐
+│ [yolo] 修复 yolo12n 零检出问题        │
+│                                      │
+│ 修改文件: inference/postprocess.py   │
+│ 变更摘要: 补充 sigmoid 后处理...       │
+│                                      │
+│  [批准执行]    [查看 Diff]   [拒绝]   │
+└──────────────────────────────────────┘
+```
+
+需要简单 HTTP callback server（`configs/agents.yaml` 中配置 `callback_port`）。
+
+---
+
+## 5. 项目配置示例
+
+```yaml
+# configs/projects.yaml
+projects:
+  yolo:
+    path: /data/rockchip/yolo
+    device_workspace: /home/pi/Desktop/yolo
+    mode: confirm
+  mediapipe:
+    path: /data/rockchip/mediapipe
+    device_workspace: /home/pi/Desktop/mediapipe
+    mode: auto
+  embedding:
+    path: /data/rockchip/embedding
+    device_workspace: /home/pi/Desktop/embed-db
+    mode: report
+  info-privacy:
+    path: /data/rockchip/info-privacy
+    device_workspace: /home/pi/Desktop/info-privacy
+    mode: confirm
+  logodetect-x86:
+    path: /data/rockchip/logodetect-x86
+    mode: report   # 纯 x86 训练项目，无设备同步
+```
+
+---
+
+## 6. Architect Agent 整合场景
+
+| 当前独立项目 | 整合机会识别 | 目标产品方向 |
+|---|---|---|
+| yolo + embedding | 图像语义检索 | 「看图找图」本地搜索引擎 |
+| mediapipe + info-privacy | 人脸检测 + 隐私遮罩 | 会议/监控隐私保护系统 |
+| logodetect + embedding | 图标检测 + 向量匹配 | 桌面 UI 自动化测试 |
+
+Architect Agent 定期（每周）对所有项目做一次全局分析，输出：
+- 各项目成熟度评估
+- 整合机会及优先级
+- 预研任务建议（提交至任务队列，`type=architect`）
+
+---
+
+## 7. 测试策略
+
+```
+tests/
+├── test_scanner.py      # Scanner 正确识别已知问题（Mock 项目文件）
+├── test_queue.py        # 任务入队/出队/优先级
+├── test_worker.py       # Agent 模式逻辑（Mock Claude API）
+├── test_device_agent.py # SSH 命令生成（Mock SSH）
+└── test_feishu.py       # Webhook 消息格式（Mock HTTP）
+```
+
+---
+
+## 8. 实施阶段规划
+
+| 阶段 | 内容 | 产出 |
+|------|------|------|
+| Phase 1 | 基础框架：Scanner + Queue + Developer Agent + 飞书通知 | 能扫描项目、生成任务、执行修复、通知 |
+| Phase 2 | Device Agent + Tester Agent + 三模式完整支持 | 双侧同步、板端验证、confirm 交互 |
+| Phase 3 | Productizer + Architect Agent | 成熟度评估、跨项目整合分析 |
+| Phase 4 | Manager Agent 全面协调 + 并发执行 | 完整多角色团队运行 |
@@ -0,0 +1,969 @@
+# Project Generalization & Memory Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** 消除 agent 提示词中的平台硬编码，建立分层项目记忆（SQLite facts/history + 向量后端接口），使 agents 每次处理同一项目时能读取并利用历史记忆。
+
+**Architecture:** ProjectMemory 扩展两张新表（project_facts / project_history）；MemoryBackend Protocol 预留向量后端接口；DeveloperAgent 和 ArchitectAgent 移除 Rockchip 硬编码，改为动态读取记忆；Executor 在任务完成后自动写入 history；Architect 通过 `[记忆]` 标签写入 facts。
+
+**Tech Stack:** Python 3.10, SQLite, dataclasses, Protocol (typing), pytest
+
+---
+
+### Task 1: 扩展 ProjectMemory — facts + history 表
+
+**Files:**
+- Modify: `src/rockchip_agents/tools/project_memory.py`
+- Test: `tests/test_project_memory.py`
+
+**背景：** 当前 ProjectMemory 只有 `project_goals` 一张表。需要新增：
+- `project_facts`：key-value 事实知识库，任意 agent 可写（无角色限制）
+- `project_history`：任务完成记录，executor 自动写入
+
+**Step 1: 写失败测试**
+
+在 `tests/test_project_memory.py` 末尾追加：
+
+```python
+def test_set_and_get_fact(tmp_path):
+    mem = _mem(tmp_path)
+    mem.set_fact("yolo", "test_cmd", "pytest tests/", role="developer")
+    facts = mem.get_facts("yolo")
+    assert facts["test_cmd"] == "pytest tests/"
+
+
+def test_set_fact_overwrites_existing(tmp_path):
+    mem = _mem(tmp_path)
+    mem.set_fact("yolo", "lang", "Python", role="developer")
+    mem.set_fact("yolo", "lang", "Python/C++", role="architect")
+    assert mem.get_facts("yolo")["lang"] == "Python/C++"
+
+
+def test_get_facts_returns_empty_when_absent(tmp_path):
+    mem = _mem(tmp_path)
+    assert mem.get_facts("nonexistent") == {}
+
+
+def test_add_and_get_history(tmp_path):
+    mem = _mem(tmp_path)
+    mem.add_history("yolo", "修复零检出", "fix_bug", "done", "已修复推理配置")
+    history = mem.get_recent_history("yolo")
+    assert len(history) == 1
+    assert history[0]["task_title"] == "修复零检出"
+    assert history[0]["status"] == "done"
+
+
+def test_history_limit(tmp_path):
+    mem = _mem(tmp_path)
+    for i in range(8):
+        mem.add_history("yolo", f"任务{i}", "code_review", "done", f"摘要{i}")
+    history = mem.get_recent_history("yolo", limit=5)
+    assert len(history) == 5
+    # 最新的在前
+    assert history[0]["task_title"] == "任务7"
+
+
+def test_context_includes_facts_and_history(tmp_path):
+    mem = _mem(tmp_path)
+    mem.set_goal("yolo", "实时目标检测", role="productizer")
+    mem.set_fact("yolo", "lang", "Python", role="developer")
+    mem.add_history("yolo", "修复零检出", "fix_bug", "done", "摘要")
+    ctx = mem.get_context_for_prompt("yolo")
+    assert "项目长期目标" in ctx
+    assert "lang" in ctx
+    assert "Python" in ctx
+    assert "近期任务" in ctx
+    assert "修复零检出" in ctx
+```
+
+**Step 2: 运行验证测试失败**
+
+```bash
+cd /data/rockchip/agents && source venv/bin/activate
+pytest tests/test_project_memory.py::test_set_and_get_fact -v
+```
+预期: FAIL — `AttributeError: 'ProjectMemory' object has no attribute 'set_fact'`
+
+**Step 3: 实现新方法**
+
+完整替换 `src/rockchip_agents/tools/project_memory.py`：
+
+```python
+from __future__ import annotations
+
+import sqlite3
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Optional, Protocol, runtime_checkable
+
+DEFAULT_DB_PATH = Path(__file__).parent.parent.parent.parent / "data" / "project_memory.db"
+
+WRITE_ROLES: frozenset[str] = frozenset({"productizer", "architect"})
+
+
+@runtime_checkable
+class MemoryBackend(Protocol):
+    """Layer-2 vector backend interface. Implement to enable semantic search."""
+
+    def search(self, query: str, project: str, top_k: int = 5) -> list[str]: ...
+    def upsert(self, project: str, key: str, text: str) -> None: ...
+
+
+@dataclass
+class ProjectGoal:
+    project: str
+    goal_text: str
+    updated_by: str
+    updated_at: str
+
+
+class ProjectMemory:
+    """Persistent project knowledge store.
+
+    Layer 1 — SQLite: goals, facts (key-value), task history.
+    Layer 2 — Optional MemoryBackend: vector semantic search (future).
+
+    Facts can be written by any agent; goals only by WRITE_ROLES.
+    """
+
+    def __init__(
+        self,
+        db_path: Path | None = None,
+        vector_backend: Optional[MemoryBackend] = None,
+    ) -> None:
+        self._db_path = db_path or DEFAULT_DB_PATH
+        self._db_path.parent.mkdir(parents=True, exist_ok=True)
+        self._vector = vector_backend
+        self._init_db()
+
+    # ── goals (write-restricted) ──────────────────────────────────────────────
+
+    def get_goal(self, project: str) -> Optional[ProjectGoal]:
+        conn = self._conn()
+        try:
+            row = conn.execute(
+                "SELECT project, goal_text, updated_by, updated_at "
+                "FROM project_goals WHERE project = ?",
+                (project,),
+            ).fetchone()
+            return ProjectGoal(**dict(row)) if row else None
+        finally:
+            conn.close()
+
+    def set_goal(self, project: str, goal_text: str, role: str) -> None:
+        if role not in WRITE_ROLES:
+            raise PermissionError(f"角色 {role!r} 无权修改项目目标（需要 {WRITE_ROLES}）")
+        conn = self._conn()
+        try:
+            conn.execute(
+                "INSERT OR REPLACE INTO project_goals "
+                "(project, goal_text, updated_by, updated_at) VALUES (?, ?, ?, ?)",
+                (project, goal_text, role, datetime.now().isoformat()),
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+    def list_goals(self) -> dict[str, ProjectGoal]:
+        conn = self._conn()
+        try:
+            rows = conn.execute(
+                "SELECT project, goal_text, updated_by, updated_at FROM project_goals"
+            ).fetchall()
+            return {r["project"]: ProjectGoal(**dict(r)) for r in rows}
+        finally:
+            conn.close()
+
+    # ── facts (open write) ────────────────────────────────────────────────────
+
+    def set_fact(self, project: str, key: str, value: str, role: str) -> None:
+        """Store a project fact. Any agent role may write."""
+        conn = self._conn()
+        try:
+            conn.execute(
+                "INSERT OR REPLACE INTO project_facts "
+                "(project, key, value, updated_by, updated_at) VALUES (?, ?, ?, ?, ?)",
+                (project, key, value, role, datetime.now().isoformat()),
+            )
+            conn.commit()
+            if self._vector is not None:
+                self._vector.upsert(project, key, f"{key}: {value}")
+        finally:
+            conn.close()
+
+    def get_facts(self, project: str) -> dict[str, str]:
+        conn = self._conn()
+        try:
+            rows = conn.execute(
+                "SELECT key, value FROM project_facts WHERE project = ? ORDER BY key",
+                (project,),
+            ).fetchall()
+            return {r["key"]: r["value"] for r in rows}
+        finally:
+            conn.close()
+
+    # ── history (auto-written by executor) ───────────────────────────────────
+
+    def add_history(
+        self,
+        project: str,
+        task_title: str,
+        task_type: str,
+        status: str,
+        summary: str,
+    ) -> None:
+        conn = self._conn()
+        try:
+            conn.execute(
+                "INSERT INTO project_history "
+                "(project, task_title, task_type, status, summary, created_at) "
+                "VALUES (?, ?, ?, ?, ?, ?)",
+                (
+                    project,
+                    task_title[:200],
+                    task_type,
+                    status,
+                    summary[:500],
+                    datetime.now().isoformat(),
+                ),
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+    def get_recent_history(self, project: str, limit: int = 5) -> list[dict]:
+        conn = self._conn()
+        try:
+            rows = conn.execute(
+                "SELECT task_title, task_type, status, summary, created_at "
+                "FROM project_history WHERE project = ? "
+                "ORDER BY id DESC LIMIT ?",
+                (project, limit),
+            ).fetchall()
+            return [dict(r) for r in rows]
+        finally:
+            conn.close()
+
+    # ── composite context builder ─────────────────────────────────────────────
+
+    def get_context_for_prompt(self, project: str) -> str:
+        """Build a context block to inject into agent system prompts.
+
+        Includes goal, facts, and recent task history. If a vector backend
+        is configured, appends semantic search results too.
+        """
+        parts: list[str] = []
+
+        goal = self.get_goal(project)
+        if goal:
+            parts.append(
+                f"【项目长期目标（由 {goal.updated_by} 设定，{goal.updated_at[:10]}）】\n"
+                f"{goal.goal_text}"
+            )
+
+        facts = self.get_facts(project)
+        if facts:
+            facts_lines = "\n".join(f"  {k}: {v}" for k, v in facts.items())
+            parts.append(f"【项目知识库】\n{facts_lines}")
+
+        history = self.get_recent_history(project)
+        if history:
+            hist_lines = "\n".join(
+                f"  [{h['status']}] {h['task_title'][:60]}" for h in history
+            )
+            parts.append(f"【近期任务记录（最近 {len(history)} 条）】\n{hist_lines}")
+
+        if self._vector is not None:
+            extra = self._vector.search(project, project, top_k=3)
+            if extra:
+                parts.append("【语义检索相关记忆】\n" + "\n".join(f"  {e}" for e in extra))
+
+        return "\n\n".join(parts) + "\n" if parts else ""
+
+    # ── db helpers ────────────────────────────────────────────────────────────
+
+    def _init_db(self) -> None:
+        conn = self._conn()
+        try:
+            conn.executescript("""
+                CREATE TABLE IF NOT EXISTS project_goals (
+                    project     TEXT PRIMARY KEY,
+                    goal_text   TEXT NOT NULL,
+                    updated_by  TEXT NOT NULL,
+                    updated_at  TEXT NOT NULL
+                );
+                CREATE TABLE IF NOT EXISTS project_facts (
+                    id          INTEGER PRIMARY KEY AUTOINCREMENT,
+                    project     TEXT NOT NULL,
+                    key         TEXT NOT NULL,
+                    value       TEXT NOT NULL,
+                    updated_by  TEXT NOT NULL,
+                    updated_at  TEXT NOT NULL,
+                    UNIQUE(project, key)
+                );
+                CREATE TABLE IF NOT EXISTS project_history (
+                    id          INTEGER PRIMARY KEY AUTOINCREMENT,
+                    project     TEXT NOT NULL,
+                    task_title  TEXT NOT NULL,
+                    task_type   TEXT NOT NULL DEFAULT '',
+                    status      TEXT NOT NULL,
+                    summary     TEXT NOT NULL DEFAULT '',
+                    created_at  TEXT NOT NULL
+                );
+            """)
+            conn.commit()
+        finally:
+            conn.close()
+
+    def _conn(self) -> sqlite3.Connection:
+        conn = sqlite3.connect(str(self._db_path))
+        conn.row_factory = sqlite3.Row
+        return conn
+```
+
+**Step 4: 运行所有新测试**
+
+```bash
+pytest tests/test_project_memory.py -v
+```
+预期: 全部 PASS（原有 10 个 + 新增 6 个 = 16 个）
+
+**Step 5: 运行全量测试确认无回归**
+
+```bash
+pytest tests/ -q
+```
+预期: 73 passed（数量与之前一致）
+
+---
+
+### Task 2: DeveloperAgent 去硬编码，注入记忆
+
+**Files:**
+- Modify: `src/rockchip_agents/agents/developer.py`
+- Test: `tests/test_developer.py`
+
+**背景：** `_build_context()` 在 line 95 硬编码了 "Rockchip RKNN NPU 推理项目"；且 DeveloperAgent 没有读取 ProjectMemory。修改后，context 从项目的 description 字段和 ProjectMemory 动态构建。
+
+**Step 1: 写失败测试**
+
+读取 `tests/test_developer.py` 找到现有测试的结构，然后追加：
+
+```python
+def test_developer_context_uses_description_not_hardcoded(tmp_path):
+    """_build_context must not mention 'Rockchip RKNN' when description differs."""
+    from rockchip_agents.config import AgentsConfig, ProjectConfig, SchedulerConfig
+    from rockchip_agents.config import ClaudeConfig, FeishuConfig
+    from rockchip_agents.agents.developer import DeveloperAgent
+    from rockchip_agents.core.queue import Task
+
+    cfg = AgentsConfig(
+        projects={"myproj": ProjectConfig(
+            path=tmp_path, mode="report",
+            description="通用 Web 后端服务"
+        )},
+        scheduler=SchedulerConfig(),
+        claude=ClaudeConfig(model="claude-sonnet-4-6"),
+        feishu=FeishuConfig(),
+        devices={},
+    )
+    agent = DeveloperAgent(cfg)
+    task = Task(project="myproj", type="fix_bug", title="修复登录",
+                context="", priority=2, mode="report")
+    ctx = agent._build_context(task)
+    assert "通用 Web 后端服务" in ctx
+    assert "Rockchip RKNN" not in ctx
+    assert "myproj" in ctx
+
+
+def test_developer_context_includes_memory(tmp_path):
+    """_build_context includes ProjectMemory context when available."""
+    from rockchip_agents.config import AgentsConfig, ProjectConfig, SchedulerConfig
+    from rockchip_agents.config import ClaudeConfig, FeishuConfig
+    from rockchip_agents.agents.developer import DeveloperAgent
+    from rockchip_agents.tools.project_memory import ProjectMemory
+    from rockchip_agents.core.queue import Task
+
+    mem = ProjectMemory(db_path=tmp_path / "mem.db")
+    mem.set_goal("myproj", "稳定高性能后端", role="productizer")
+    mem.set_fact("myproj", "test_cmd", "pytest tests/", role="developer")
+
+    cfg = AgentsConfig(
+        projects={"myproj": ProjectConfig(path=tmp_path, mode="auto",
+                                          description="通用 Web 后端服务")},
+        scheduler=SchedulerConfig(),
+        claude=ClaudeConfig(model="claude-sonnet-4-6"),
+        feishu=FeishuConfig(),
+        devices={},
+    )
+    agent = DeveloperAgent(cfg, memory=mem)
+    task = Task(project="myproj", type="fix_bug", title="修复登录",
+                context="", priority=2, mode="auto")
+    ctx = agent._build_context(task)
+    assert "稳定高性能后端" in ctx
+    assert "test_cmd" in ctx
+```
+
+**Step 2: 验证测试失败**
+
+```bash
+pytest tests/test_developer.py::test_developer_context_uses_description_not_hardcoded -v
+```
+预期: FAIL — `assert "Rockchip RKNN" not in ctx` 失败（因为还没改）
+
+**Step 3: 修改 DeveloperAgent**
+
+完整替换 `src/rockchip_agents/agents/developer.py`：
+
+```python
+from __future__ import annotations
+
+import json
+import logging
+import os
+import subprocess
+from dataclasses import dataclass
+from typing import Optional
+
+from rockchip_agents.config import AgentsConfig
+from rockchip_agents.core.queue import Task
+from rockchip_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+TIMEOUT = 300
+
+
+@dataclass
+class AgentResult:
+    status: str   # done | failed
+    summary: str
+
+
+def _parse_cli_result(r: subprocess.CompletedProcess) -> AgentResult:
+    """解析 claude --output-format json 的输出。"""
+    if r.returncode != 0:
+        return AgentResult(status="failed", summary=r.stderr or r.stdout or "未知错误")
+    try:
+        data = json.loads(r.stdout)
+        summary = data.get("result", r.stdout)
+        is_error = data.get("is_error", False)
+        return AgentResult(status="failed" if is_error else "done", summary=summary)
+    except (json.JSONDecodeError, KeyError):
+        return AgentResult(status="done", summary=r.stdout)
+
+
+class DeveloperAgent:
+    def __init__(
+        self,
+        config: AgentsConfig,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        self._config = config
+        self._memory = memory or ProjectMemory()
+
+    def run(self, task: Task) -> AgentResult:
+        proj_cfg = self._config.projects.get(task.project)
+        proj_path = str(proj_cfg.path) if proj_cfg else None
+        cmd = self._build_cmd(task)
+        env = self._build_env()
+        try:
+            r = subprocess.run(
+                cmd,
+                capture_output=True, text=True,
+                timeout=TIMEOUT,
+                cwd=proj_path,
+                env=env,
+            )
+            return _parse_cli_result(r)
+        except subprocess.TimeoutExpired:
+            return AgentResult(status="failed", summary=f"执行超时（{TIMEOUT}s）")
+        except Exception as e:
+            logger.error("DeveloperAgent 执行失败: %s", e)
+            return AgentResult(status="failed", summary=str(e))
+
+    def _build_cmd(self, task: Task) -> list[str]:
+        cmd = [
+            "claude", "--print",
+            "--output-format", "json",
+            "--no-session-persistence",
+            "--model", self._config.claude.model,
+            "--append-system-prompt", self._build_context(task),
+        ]
+        if task.mode == "report":
+            cmd += ["--allowedTools", "Read,Glob,Grep"]
+        elif task.mode == "confirm":
+            cmd += ["--allowedTools", "Read,Glob,Grep,Bash"]
+        else:  # auto
+            cmd.append("--dangerously-skip-permissions")
+        cmd.append(self._build_prompt(task))
+        return cmd
+
+    def _build_env(self) -> dict:
+        env = os.environ.copy()
+        if self._config.claude.api_key:
+            env["ANTHROPIC_API_KEY"] = self._config.claude.api_key
+        env.pop("CLAUDECODE", None)
+        env.pop("CLAUDE_CODE_ENTRYPOINT", None)
+        return env
+
+    def _build_context(self, task: Task) -> str:
+        proj_cfg = self._config.projects.get(task.project)
+        proj_path = str(proj_cfg.path) if proj_cfg else ""
+        proj_desc = (proj_cfg.description if proj_cfg and proj_cfg.description
+                     else task.project)
+        mode_desc = {
+            "auto":    "直接修复文件并运行测试验证。",
+            "confirm": "只分析并给出修改方案摘要，不实际写文件。",
+            "report":  "只读取和分析，输出详细报告，不做任何修改。",
+        }.get(task.mode, "")
+        memory_ctx = self._memory.get_context_for_prompt(task.project)
+        return (
+            f"你是 {task.project} 项目的 Developer Agent。\n"
+            f"项目路径: {proj_path}。\n"
+            f"项目描述: {proj_desc}。\n"
+            f"当前模式: {task.mode} — {mode_desc}\n"
+            f"{memory_ctx}"
+            f"请用中文回复，保持简洁，聚焦问题。"
+        )
+
+    def _build_prompt(self, task: Task) -> str:
+        return (
+            f"任务类型: {task.type}\n"
+            f"任务标题: {task.title}\n"
+            f"上下文:\n{task.context}\n\n"
+            f"请分析并处理这个任务。"
+        )
+```
+
+**Step 4: 运行新测试**
+
+```bash
+pytest tests/test_developer.py -v
+```
+预期: 全部 PASS
+
+**Step 5: 全量测试**
+
+```bash
+pytest tests/ -q
+```
+预期: 所有测试通过
+
+---
+
+### Task 3: ArchitectAgent 去硬编码 + [记忆] 标签支持
+
+**Files:**
+- Modify: `src/rockchip_agents/agents/architect.py`
+- Test: `tests/test_architect.py`
+
+**背景：** `ARCHITECT_SYSTEM` 在 line 21-32 硬编码了 "Rockchip AI 平台"、"RKNN NPU 推理项目" 以及 YOLO/embedding/MediaPipe 等具体集成示例。需要改成通用架构师角色。同时新增 `[记忆]` 标签解析，让 architect 可以写入项目 facts。
+
+**Step 1: 写失败测试**
+
+在 `tests/test_architect.py` 末尾追加：
+
+```python
+def test_architect_parses_memory_tag_and_writes_fact(tmp_path):
+    """[记忆] tag in report should call memory.set_fact for each key=value pair."""
+    from unittest.mock import MagicMock, patch
+    from rockchip_agents.agents.architect import ArchitectAgent
+    from rockchip_agents.tools.project_memory import ProjectMemory
+    from rockchip_agents.core.queue import TaskQueue
+
+    cfg = _make_cfg(tmp_path)
+    mem = ProjectMemory(db_path=tmp_path / "mem.db")
+    queue = TaskQueue(db_path=tmp_path / "q.db")
+    agent = ArchitectAgent(cfg, queue=queue, memory=mem)
+
+    report = (
+        "架构分析完成。\n"
+        "[记忆] yolo: test_cmd=pytest tests/, main_lang=Python\n"
+        "[优化] yolo: 优化推理管道\n"
+    )
+    agent._extract_and_enqueue(report, requester="test")
+
+    facts = mem.get_facts("yolo")
+    assert facts.get("test_cmd") == "pytest tests/"
+    assert facts.get("main_lang") == "Python"
+
+
+def test_architect_system_prompt_is_generic():
+    """ARCHITECT_SYSTEM must not contain hardcoded project names."""
+    from rockchip_agents.agents.architect import ARCHITECT_SYSTEM
+    for hardcoded in ["Rockchip", "RKNN", "YOLO", "MediaPipe", "embedding"]:
+        assert hardcoded not in ARCHITECT_SYSTEM, (
+            f"ARCHITECT_SYSTEM 不应包含硬编码词语: {hardcoded!r}"
+        )
+```
+
+**Step 2: 验证测试失败**
+
+```bash
+pytest tests/test_architect.py::test_architect_system_prompt_is_generic -v
+```
+预期: FAIL — `assert "Rockchip" not in ARCHITECT_SYSTEM` 失败
+
+**Step 3: 修改 architect.py**
+
+替换 `ARCHITECT_SYSTEM` 常量（line 21-32）：
+
+```python
+ARCHITECT_SYSTEM = """你是一位技术架构师，负责分析多个软件项目，识别整合机会，提出架构改进建议。
+
+分析维度：
+- 功能互补性：哪些项目的能力可以组合，产生新价值
+- 技术共性：是否有可复用的基础设施或数据管道
+- 产品可能性：整合后能解决什么实际问题
+
+请用中文输出结构化分析报告，末尾按格式列出任务（每行一条）：
+- [预研] 任务名称  ← 跨项目探索、新功能研究
+- [优化] <项目名>: 优化描述  ← 对特定项目的架构改进（项目名须在上面列出的项目中）
+- [目标] <项目名>: 目标描述  ← 更新项目长期目标（仅当需要修订时）
+- [记忆] <项目名>: key=value, key2=value2  ← 记录项目关键技术事实
+"""
+```
+
+在 `_extract_and_enqueue()` 方法末尾（`[目标]` 解析块之后）追加 `[记忆]` 解析：
+
+```python
+        # [记忆] <project>: key=value, key2=value2 → update project facts
+        for m in re.finditer(r"\[记忆\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            kv_text = m.group(2).strip()
+            if project not in self._config.projects:
+                logger.warning("记忆更新项目 %r 不在配置中，跳过", project)
+                continue
+            for kv in kv_text.split(","):
+                kv = kv.strip()
+                if "=" in kv:
+                    key, _, value = kv.partition("=")
+                    key = key.strip()[:50]
+                    value = value.strip()[:200]
+                    if key:
+                        try:
+                            self._memory.set_fact(project, key, value, "architect")
+                            logger.info("项目事实已记录 [%s] %s=%s", project, key, value)
+                        except Exception as e:
+                            logger.warning("[%s] 写入项目事实失败: %s", project, e)
+```
+
+**Step 4: 运行新测试**
+
+```bash
+pytest tests/test_architect.py -v
+```
+预期: 全部 PASS
+
+**Step 5: 全量测试**
+
+```bash
+pytest tests/ -q
+```
+预期: 所有测试通过
+
+---
+
+### Task 4: Executor 自动写任务历史 + 传递 ProjectMemory 给 agents
+
+**Files:**
+- Modify: `src/rockchip_agents/core/executor.py`
+- Test: `tests/test_executor.py`
+
+**背景：** 每次任务完成（done 或 failed）后，应把结果写入 `project_history`，让下一轮 agent 处理同一项目时看到历史记录。同时 `_make_agent()` 需要接受 `memory` 参数，将同一 ProjectMemory 实例传递给所有 agents。
+
+**Step 1: 写失败测试**
+
+在 `tests/test_executor.py` 末尾追加（或找到该文件后追加）：
+
+```python
+def test_executor_writes_history_after_task(tmp_path):
+    """Executor must write task result to ProjectMemory history after completion."""
+    import asyncio
+    from unittest.mock import patch, MagicMock
+    from rockchip_agents.core.executor import Executor
+    from rockchip_agents.core.queue import TaskQueue, Task
+    from rockchip_agents.tools.project_memory import ProjectMemory
+    from rockchip_agents.agents.developer import AgentResult
+
+    mem = ProjectMemory(db_path=tmp_path / "mem.db")
+    queue = TaskQueue(db_path=tmp_path / "q.db")
+    queue.enqueue(Task(project="yolo", type="fix_bug", title="修复零检出",
+                       context="", priority=2, mode="report", agent_role="developer"))
+
+    cfg = _make_cfg(tmp_path)   # reuse helper from existing tests
+    executor = Executor(cfg, queue=queue, memory=mem)
+
+    mock_result = AgentResult(status="done", summary="已修复")
+    with patch("rockchip_agents.agents.developer.DeveloperAgent.run",
+               return_value=mock_result):
+        asyncio.run(executor.run_next())
+
+    history = mem.get_recent_history("yolo")
+    assert len(history) == 1
+    assert history[0]["status"] == "done"
+    assert history[0]["task_title"] == "修复零检出"
+```
+
+**Step 2: 验证测试失败**
+
+```bash
+pytest tests/test_executor.py::test_executor_writes_history_after_task -v
+```
+预期: FAIL — `Executor.__init__` 不接受 `memory` 参数，或 history 为空
+
+**Step 3: 修改 executor.py**
+
+完整替换 `src/rockchip_agents/core/executor.py`：
+
+```python
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Optional
+
+from rockchip_agents.agents.developer import DeveloperAgent
+from rockchip_agents.config import AgentsConfig
+from rockchip_agents.core.queue import TaskQueue, Task
+from rockchip_agents.tools.feishu import FeishuNotifier
+from rockchip_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+
+def _make_agent(config: AgentsConfig, task: Task, memory: ProjectMemory):
+    """根据 agent_role 路由到对应 Agent，传入共享 memory 实例。"""
+    role = task.agent_role
+    if role == "tester":
+        from rockchip_agents.agents.tester import TesterAgent
+        return TesterAgent(config)
+    if role == "productizer":
+        from rockchip_agents.agents.productizer import ProductizerAgent
+        return ProductizerAgent(config, memory=memory)
+    if role == "architect":
+        from rockchip_agents.agents.architect import ArchitectAgent
+        return ArchitectAgent(config, memory=memory)
+    return DeveloperAgent(config, memory=memory)
+
+
+class Executor:
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        self._config = config
+        self._queue = queue or TaskQueue()
+        self._memory = memory or ProjectMemory()
+        self._feishu = FeishuNotifier(
+            webhook_url=config.feishu.webhook_url,
+            bot_token=config.feishu.bot_token,
+        )
+        self._semaphore = asyncio.Semaphore(config.scheduler.max_concurrent)
+
+    async def run_next(self) -> bool:
+        task = self._queue.dequeue()
+        if not task:
+            logger.debug("任务队列为空")
+            return False
+        async with self._semaphore:
+            await self._run_task(task)
+        return True
+
+    async def run_all_pending(self) -> int:
+        tasks_to_run: list[Task] = []
+        while True:
+            task = self._queue.dequeue()
+            if not task:
+                break
+            tasks_to_run.append(task)
+
+        async def _run_with_sem(t: Task) -> None:
+            async with self._semaphore:
+                await self._run_task(t)
+
+        if tasks_to_run:
+            await asyncio.gather(
+                *[_run_with_sem(t) for t in tasks_to_run],
+                return_exceptions=True,
+            )
+        return len(tasks_to_run)
+
+    async def _run_task(self, task: Task) -> None:
+        logger.info(
+            "开始执行任务 [%s] %s (role=%s, mode=%s)",
+            task.project, task.title, task.agent_role, task.mode,
+        )
+        try:
+            agent = _make_agent(self._config, task, self._memory)
+            result = await asyncio.get_running_loop().run_in_executor(
+                None, agent.run, task
+            )
+            if task.mode == "confirm" and result.status == "done":
+                callback_url = f"http://localhost:{self._config.feishu.callback_port}"
+                self._feishu.send_confirm_card(
+                    task_id=task.id,
+                    project=task.project,
+                    task_title=task.title,
+                    diff_summary=result.summary,
+                    callback_url=callback_url,
+                )
+                self._queue.mark_waiting_confirm(task.id, diff=result.summary)
+                logger.info("任务等待确认 [%s] %s", task.project, task.title)
+            else:
+                self._queue.mark_done(task.id, result.summary)
+                self._feishu.send_task_result(
+                    project=task.project, task_title=task.title,
+                    status=result.status, summary=result.summary,
+                )
+                logger.info("任务完成 [%s] %s", task.project, task.title)
+
+            # Write history so future agents can learn from this run
+            try:
+                self._memory.add_history(
+                    project=task.project,
+                    task_title=task.title,
+                    task_type=task.type,
+                    status=result.status,
+                    summary=result.summary[:300],
+                )
+            except Exception as hist_err:
+                logger.warning("写入任务历史失败: %s", hist_err)
+
+        except Exception as e:
+            logger.error("任务失败 [%s] %s: %s", task.project, task.title, e)
+            self._queue.mark_failed(task.id, str(e))
+            try:
+                self._memory.add_history(
+                    project=task.project,
+                    task_title=task.title,
+                    task_type=task.type,
+                    status="failed",
+                    summary=str(e)[:300],
+                )
+            except Exception:
+                pass
+```
+
+**Step 4: 运行新测试**
+
+```bash
+pytest tests/test_executor.py -v
+```
+预期: 全部 PASS
+
+**Step 5: 全量测试**
+
+```bash
+pytest tests/ -q
+```
+预期: 所有测试通过（原 73 个 + 新增若干 = N passed）
+
+---
+
+### Task 5: 新增 CLI 命令 + 更新文档
+
+**Files:**
+- Modify: `scripts/run.py`
+- Modify: `DEVELOP.md`
+
+**背景：** 新增 `facts` 命令，方便查看各项目的知识库事实。同时更新 DEVELOP.md 里的记忆架构说明。
+
+**Step 1: 在 scripts/run.py 中新增 `facts` 子命令**
+
+在 `sub.add_parser("goals", ...)` 行之后追加：
+
+```python
+    sub.add_parser("facts",    help="查看所有项目的知识库事实（由 agents 自动记录）")
+    p_sf = sub.add_parser("set-fact",  help="手动写入项目事实")
+    p_sf.add_argument("project", help="项目名")
+    p_sf.add_argument("key",     help="事实键名，如 test_cmd")
+    p_sf.add_argument("value",   help="事实值")
+```
+
+在 `elif args.cmd == "goals":` 块之后追加：
+
+```python
+    elif args.cmd == "facts":
+        from rockchip_agents.tools.project_memory import ProjectMemory
+        mem = ProjectMemory()
+        from rockchip_agents.config import load_config
+        cfg = load_config()
+        any_found = False
+        for proj_name in sorted(cfg.projects):
+            facts = mem.get_facts(proj_name)
+            if facts:
+                any_found = True
+                print(f"\n[{proj_name}]")
+                for k, v in facts.items():
+                    print(f"  {k}: {v}")
+        if not any_found:
+            print("暂无项目事实记录（运行 architect 任务后自动写入）")
+
+    elif args.cmd == "set-fact":
+        from rockchip_agents.tools.project_memory import ProjectMemory
+        mem = ProjectMemory()
+        mem.set_fact(args.project, args.key, args.value, role="developer")
+        print(f"已写入: [{args.project}] {args.key} = {args.value}")
+```
+
+**Step 2: 手动验证命令可用**
+
+```bash
+source venv/bin/activate
+python scripts/run.py facts
+python scripts/run.py set-fact memory-analyze test_cmd "cd test_cases && python test_runner.py"
+python scripts/run.py facts
+```
+预期: 正确打印 memory-analyze 的 test_cmd 事实
+
+**Step 3: 更新 DEVELOP.md 记忆章节**
+
+在 DEVELOP.md 中找到"项目记忆"相关章节（或在末尾追加），写入以下内容：
+
+```markdown
+## 项目记忆系统
+
+ProjectMemory 采用分层设计：
+
+| 层 | 存储 | 读写规则 |
+|----|------|---------|
+| Layer 1a — goals | SQLite project_goals | 只有 productizer/architect 可写 |
+| Layer 1b — facts | SQLite project_facts | 任意 agent 可写；architect 通过 [记忆] 标签写入 |
+| Layer 1c — history | SQLite project_history | executor 自动写入，每次任务完成后记录 |
+| Layer 2 — vector | MemoryBackend Protocol | 可选；实现接口后自动启用语义检索 |
+
+### Agent 写入记忆的方式
+
+- **Developer**: 任务执行后由 executor 自动写 history
+- **Architect**: 分析报告中添加 `[记忆] <项目名>: key=value` 标签写 facts
+- **Productizer**: 评估后自动调用 `set_goal()` 更新项目目标
+
+### 查看项目记忆
+
+```bash
+python scripts/run.py goals    # 项目长期目标
+python scripts/run.py facts    # 项目知识库事实
+```
+
+### 接入向量后端（Layer 2）
+
+实现 `MemoryBackend` Protocol，传入 `ProjectMemory(vector_backend=your_backend)`：
+
+```python
+class MyVectorBackend:
+    def search(self, query: str, project: str, top_k: int = 5) -> list[str]: ...
+    def upsert(self, project: str, key: str, text: str) -> None: ...
+```
+```
+
+**Step 4: 全量测试最终确认**
+
+```bash
+pytest tests/ -q
+```
+预期: 所有测试通过，0 warnings
@@ -0,0 +1,217 @@
+# Agent 全面升级设计文档
+
+**日期**: 2026-03-06
+**状态**: 已批准，待实施
+**约束**: 仅使用 `claude --print` CLI，零额外 token，MCP 工具在单次调用内生效
+
+---
+
+## 一、Agent 角色扩展（4 → 13 个）
+
+### 开发者层（5 个，并行执行）
+
+| 角色 | 职责 |
+|------|------|
+| `dev-embedded` | RKNN/NPU/板端推理、AI 应用层 |
+| `dev-backend` | Python 服务、API、数据管道（x86）|
+| `dev-ml` | 模型量化、精度优化、训练脚本 |
+| `dev-test` | 专门生成/补全测试代码，TDD 驱动 |
+| `dev-refactor` | 重构、技术债、代码质量 |
+
+### 内核与底层层（2 个）
+
+| 角色 | 职责 |
+|------|------|
+| `dev-kernel` | Linux 内核模块、驱动开发、内核配置（Kconfig/Device Tree）|
+| `dev-lowlevel` | BSP、Bootloader（U-Boot）、裸机/RTOS、硬件 bring-up |
+
+### 硬件工程师（1 个）
+
+| 角色 | 职责 |
+|------|------|
+| `hw-engineer` | 芯片选型（RK3588/RV1106/等）、外设接口设计、BOM 评审、功耗分析、Datasheet 查阅 |
+
+### 增强测试工程师（1 个，多维度）
+
+| 角色 | 测试维度 |
+|------|---------|
+| `tester` | ① 本地 pytest + 覆盖率报告 → ② 自动补全缺失测试 → ③ rsync 同步嵌入式设备 → ④ 板端实际运行（dmesg/崩溃日志/内存占用）→ ⑤ Playwright 验证 Web UI |
+
+### 架构师层（3 个，相互制约）
+
+| 角色 | 职责 | 制约关系 |
+|------|------|---------|
+| `arch-system` | 系统设计、性能、可扩展性 | 被 arch-security 和 pm 制约 |
+| `arch-security` | 安全审查、权限、数据合规 | 被 arch-system 和 pm 制约 |
+| `arch-reviewer` | 专门 review 其他 arch 提案 | 持有否决权 |
+
+### 管理层（1 个）
+
+| 角色 | 职责 |
+|------|------|
+| `pm` | 产品经理：用户价值、需求对齐、否决架构决策 |
+
+---
+
+## 二、MCP 工具矩阵
+
+### 工具注册（`configs/mcp_tools.yaml`，支持热更新）
+
+```yaml
+registry:
+  sequential-thinking: "npx -y @modelcontextprotocol/server-sequential-thinking"
+  fetch:               "npx -y @modelcontextprotocol/server-fetch"
+  playwright:          "npx -y @playwright/mcp@latest"
+  context7:            "npx -y @upstash/context7-mcp@latest"
+
+roles:
+  dev-embedded:  [sequential-thinking, fetch, context7]
+  dev-backend:   [sequential-thinking, fetch, context7]
+  dev-ml:        [sequential-thinking, fetch, context7]
+  dev-test:      [sequential-thinking, playwright]
+  dev-refactor:  [sequential-thinking]
+  dev-kernel:    [sequential-thinking, fetch]
+  dev-lowlevel:  [sequential-thinking, fetch]
+  hw-engineer:   [sequential-thinking, fetch, playwright]
+  tester:        [sequential-thinking, fetch, playwright]
+  arch-system:   [sequential-thinking, fetch, context7]
+  arch-security: [sequential-thinking, fetch]
+  arch-reviewer: [sequential-thinking]
+  pm:            [sequential-thinking, fetch]
+```
+
+### Plan-First 系统提示（所有角色统一前缀）
+
+```
+在执行任何写文件/命令操作前，必须先使用 sequentialthinking 工具完成：
+1. 问题拆解（≥3 步）
+2. 风险识别
+3. 方案选择（列出 2 种，说明选择理由）
+只有 sequentialthinking 调用完成后才能调用 Write/Edit/Bash。
+```
+
+---
+
+## 三、制约工作流（D 方案混合）
+
+### 触发路由
+
+| 任务类型 | 制约模式 |
+|---------|---------|
+| `fix_bug` / `improve_perf` / `code_review` | C 异步否决 |
+| `arch_optimize` / `security_review` / `kernel_change` / `hw_selection` | A 串行审批链 |
+| `new_feature` / `brainstorm` | B 对抗式提案 |
+
+### 模式 C：异步否决（日常改动）
+
+```
+dev-* 完成任务
+  └─ [SPAWN:arch-reviewer] 自动注入
+        ├─ [VETO] 原因  →  父任务 vetoed，重入队（retry_count+1，上限 2）
+        └─ [APPROVE]   →  保持 done
+  无响应 48h  →  自动通过
+```
+
+### 模式 A：串行审批链（重大变更）
+
+```
+arch-system / dev-kernel 提案
+  └─ [AWAIT:arch-security]  →  waiting_approval
+        ├─ [REJECT] 原因  →  返回提案者重试
+        └─ [APPROVE]
+              └─ [AWAIT:pm]  →  waiting_approval
+                    ├─ [REJECT]  →  返回提案者重试
+                    └─ [APPROVE]  →  dev-* 实施
+```
+
+### 模式 B：对抗式提案（新功能）
+
+```
+pm 接收 new_feature
+  ├─ [SPAWN:arch-system]    →  [PROPOSAL:arch-system] 方案 A
+  └─ [SPAWN:arch-security]  →  [PROPOSAL:arch-security] 方案 B
+        └─ arch-reviewer 汇总
+              └─ [SELECT:arch-system 或 arch-security]  →  dev-* 实施
+```
+
+### 结构化输出标签
+
+| 标签 | 效果 |
+|------|------|
+| `[SPAWN:role] 描述` | 入队新任务，指定 agent_role |
+| `[AWAIT:role]` | 当前任务进入 waiting_approval，入队审查任务 |
+| `[APPROVE]` | 解除 waiting_approval，链路推进 |
+| `[REJECT] 原因` | 任务重入队，原因注入上下文 |
+| `[VETO] 原因` | 已完成任务回退为 vetoed，重入队 |
+| `[PROPOSAL:role] 内容` | 对抗式提案，等待 SELECT |
+| `[SELECT:role]` | 选中提案，触发 dev-* 实施 |
+
+---
+
+## 四、Scanner 升级 + Executor 改造
+
+### Scanner 智能角色路由（`configs/role_routing.yaml`）
+
+```yaml
+role_keywords:
+  dev-kernel:   [驱动, 内核, driver, kernel, dts, device tree, insmod]
+  dev-lowlevel: [bsp, uboot, bootloader, 裸机, rtos, 引导, 启动]
+  dev-embedded: [rknn, npu, 推理, 板端, rk3588, lite2, rknn_lite]
+  dev-ml:       [模型, 精度, 量化, 训练, onnx, accuracy, mAP]
+  dev-backend:  [api, fastapi, 数据库, 服务, http, redis]
+  hw-engineer:  [选型, 硬件, 芯片, 外设, 接口, bom, 功耗, datasheet]
+  dev-test:     [测试覆盖, coverage, 缺少测试, test case]
+  dev-refactor: [重构, 技术债, 代码质量, 冗余]
+  # 未匹配 → dev-embedded（嵌入式项目默认）
+```
+
+### 新增 Scanner 信号源
+
+| 来源 | 提取内容 | 任务类型 |
+|------|---------|---------|
+| `CLAUDE.md` 的 `## 规划功能` 章节 | 新功能描述 | `new_feature` |
+| 设备 dmesg（SSH 拉取） | kernel panic / oops | `kernel_change` |
+| `README.md` 的 `## 硬件` 章节 | 选型待定标记 | `hw_selection` |
+| `git log --since=7d` 失败 commit | revert 记录 | `fix_bug` + 失败上下文 |
+
+### 任务状态机扩展
+
+```
+pending → running → done ✅
+                 → failed ❌        (retry ≤2，携带失败上下文)
+                 → waiting_approval ⏳ (等待串行链审批，不占 Semaphore)
+                 → vetoed 🚫       (被否决，重入队 retry)
+                 → adversarial_pending ⚔️ (等待 SELECT)
+```
+
+### Task 新增字段
+
+```python
+retry_count:     int = 0   # 最多重试 2 次
+parent_task_id:  int = 0   # 追踪任务血缘
+constraint_type: str = ""  # "veto_pending"|"approval_pending"|"adversarial"
+awaiting_role:   str = ""  # 串行链中等待哪个角色
+```
+
+### Retry with Context 继承
+
+失败/否决时重入队，前缀注入原因：
+```
+⚠️ 上次执行失败（第 N 次重试）
+原因：<reason>
+
+--- 原始任务上下文 ---
+<original context>
+```
+
+---
+
+## 五、实施顺序（建议）
+
+1. **基础设施**：Task 新字段 + 新状态 + `mcp_tools.yaml` + `role_routing.yaml`
+2. **MCP 注入**：`_build_cmd()` 按角色加载工具，Plan-First system prompt
+3. **新角色实现**：`dev-kernel`, `dev-lowlevel`, `hw-engineer` agent 文件
+4. **标签解析引擎**：Executor `_parse_agent_tags()` + `_handle_tags()`
+5. **制约状态机**：`waiting_approval`, `vetoed`, `adversarial_pending` 状态处理
+6. **Scanner 升级**：智能路由 + 新信号源
+7. **测试补全**：每个新 agent 和新状态的 pytest 覆盖
@@ -0,0 +1,218 @@
+# Rockchip Agents Dashboard 设计文档
+
+**日期**: 2026-03-06
+**状态**: 已批准，待实施
+
+---
+
+## 一、目标
+
+构建一个独立的实时可视化看板，展示所有 Agent 任务的状态流转，支持多项目并行展示，并提供完整的操作控制台（手动入队、审批、重试、取消）。
+
+---
+
+## 二、架构总览
+
+```
+scripts/dashboard.py              ← 独立启动入口
+    └─ FastAPI (port 8080)
+         ├─ GET  /api/tasks                  所有任务（可按 project 过滤）
+         ├─ POST /api/tasks                  手动入队新任务
+         ├─ PATCH /api/tasks/{id}            修改优先级
+         ├─ POST /api/tasks/{id}/retry       重试 failed/vetoed 任务
+         ├─ POST /api/tasks/{id}/cancel      取消 running 任务（→ failed）
+         ├─ POST /api/tasks/{id}/approve     APPROVE waiting_approval
+         ├─ POST /api/tasks/{id}/reject      REJECT waiting_approval（附 reason）
+         ├─ GET  /api/projects               已配置项目列表
+         ├─ WS   /ws                         实时推送（WebSocket）
+         └─ GET  /                           静态托管 dashboard/dist/
+
+dashboard/                        ← React + Vite + Tailwind + shadcn/ui
+    src/
+      components/
+        KanbanBoard.tsx            多项目 Kanban（按 project 分组）
+        TaskCard.tsx               任务卡片（角色徽章、状态颜色、操作按钮）
+        TaskDetail.tsx             右侧滑出详情面板
+        DagView.tsx                DAG 关系图（React Flow）
+        NewTaskModal.tsx           手动入队对话框
+      store/
+        tasks.ts                   Zustand store
+      lib/
+        ws.ts                      WebSocket 管理（断线重连）
+        api.ts                     REST API 封装
+```
+
+---
+
+## 三、UI 布局
+
+### 主布局
+
+```
+┌──────────────────────────────────────────────────────────────────────┐
+│  🤖 Rockchip Agents Dashboard                          [+ 新任务]    │
+│  ● WS connected   3 运行中 | 2 待审批 | 1 失败                       │
+├──────────────────────────────────────────────────────────────────────┤
+│  [全览] [yolo] [claude-wx] [memory-analyze]  …（项目标签页）          │
+├──────────────────────────────────────────────────────────────────────┤
+│  [Kanban 看板] [DAG 关系图]                                           │
+├──────────────────────────────────────────────────────────────────────┤
+│                                                                      │
+│  ▼ yolo (auto)  ─────────────────────────────────────────────────    │
+│  ┌─────────┬─────────┬──────┬──────┬────────────────┬──────────┐   │
+│  │ PENDING │ RUNNING │ DONE │ FAIL │ WAIT_APPROVAL  │ VETOED   │   │
+│  │   (2)   │   (1)   │ (8)  │ (1)  │    (1) ⚠️      │   (0)    │   │
+│  │ [卡片]  │ [卡片]  │[卡片]│[卡片]│ [卡片]         │          │   │
+│  │         │         │      │[重试]│ [APPROVE][REJ] │          │   │
+│  └─────────┴─────────┴──────┴──────┴────────────────┴──────────┘   │
+│                                                                      │
+│  ▼ claude-wx (auto)  ──────────────────────────────────────────────  │
+│  ┌─────────┬─────────┬──────┬──────┬────────────────┬──────────┐   │
+│  │   (0)   │   (1)   │ (5)  │ (0)  │    (0)         │   (0)    │   │
+│  └─────────┴─────────┴──────┴──────┴────────────────┴──────────┘   │
+└──────────────────────────────────────────────────────────────────────┘
+```
+
+### 任务卡片
+
+```
+┌──────────────────────────────┐
+│ 🔵 dev-embedded   auto  P:1  │  ← 角色徽章 / 模式 / 优先级
+│ 修复零检出率（模型精度问题）  │  ← 标题（截断 60 字）
+│ ⏱ 4m 32s  🔄 retry: 0       │  ← 执行时长 / 重试次数
+│ parent: #12                  │  ← 父任务链路（可点击跳转）
+└──────────────────────────────┘
+```
+
+**状态颜色规则：**
+- `running`：蓝色边框 + 脉冲动画
+- `waiting_approval`：橙色边框 + ⚠️ 高亮
+- `failed`：红色边框
+- `vetoed`：紫色边框
+- `done`：绿色（半透明降调）
+- `pending`：灰色边框
+
+### 右侧详情面板（点击卡片打开）
+
+- 完整字段：title / type / context / result_summary / created_at / started_at / completed_at
+- parent_task_id 链接 → 点击跳转到父任务卡片
+- SPAWN 子任务列表（当前任务派生的所有子任务）
+- 操作按钮：按当前状态动态显示（Retry / Cancel / Approve / Reject）
+- 优先级滑块（1-5）
+
+---
+
+## 四、数据流 + 实时更新
+
+### WebSocket 事件格式
+
+```json
+// 初次连接：全量快照
+{ "type": "snapshot", "tasks": [...] }
+
+// 任务状态变化
+{ "type": "task_update", "task": { "id": 42, "project": "yolo", "status": "running", ... } }
+
+// 新任务入队
+{ "type": "task_create", "task": { ... } }
+```
+
+### 后端 Watcher 线程
+
+```
+后台线程（每 1 秒）：
+  SELECT * FROM tasks ORDER BY id DESC LIMIT 500
+  → 对比上次结果的哈希值（dict{id: hash(row)}）
+  → 检测到 UPDATE/INSERT → broadcast 给所有 WS 连接
+  → 无变化 → 跳过
+```
+
+### 前端状态管理（Zustand）
+
+```typescript
+interface TaskStore {
+  tasks: Map<number, Task>
+  wsStatus: 'connected' | 'reconnecting' | 'offline'
+  // 初次 snapshot 填充；后续 task_update/task_create 合并
+}
+```
+
+断线自动重连（指数退避：1s → 2s → 4s → 8s），重连后服务端自动发送 snapshot。
+
+### REST 操作（乐观更新）
+
+点击 Retry/Approve 等 → 立即更新本地 UI → POST 到后端 → SQLite 写入 → Watcher 检测 → WS 广播确认。
+
+---
+
+## 五、Docker 部署
+
+### Dockerfile.dashboard
+
+```dockerfile
+# Stage 1: 构建 React 前端
+FROM node:20-alpine AS frontend-builder
+WORKDIR /frontend
+COPY dashboard/package*.json ./
+RUN npm ci
+COPY dashboard/ ./
+RUN npm run build
+
+# Stage 2: Python FastAPI 运行时
+FROM python:3.10-slim AS runtime
+WORKDIR /app
+COPY requirements-dashboard.txt ./
+RUN pip install --no-cache-dir -r requirements-dashboard.txt
+COPY src/ ./src/
+COPY scripts/dashboard.py ./scripts/
+COPY --from=frontend-builder /frontend/dist ./dashboard/dist/
+EXPOSE 8080
+CMD ["python", "scripts/dashboard.py"]
+```
+
+### docker-compose.yml 新增服务
+
+```yaml
+services:
+  dashboard:
+    build:
+      context: .
+      dockerfile: Dockerfile.dashboard
+    ports:
+      - "8080:8080"
+    volumes:
+      - ./data:/app/data        # 共享 tasks.db（SQLite WAL 模式）
+      - ./configs:/app/configs  # 读取项目配置
+    environment:
+      - DB_PATH=/app/data/tasks.db
+      - CONFIG_DIR=/app/configs
+    restart: unless-stopped
+```
+
+---
+
+## 六、技术栈
+
+| 层 | 技术 |
+|---|---|
+| 后端 | FastAPI + uvicorn + websockets |
+| 前端框架 | React 18 + TypeScript + Vite |
+| UI 组件 | shadcn/ui + Tailwind CSS v3 |
+| 状态管理 | Zustand |
+| DAG 图 | React Flow |
+| 部署 | Docker 多阶段构建 + docker-compose |
+
+---
+
+## 七、实施顺序
+
+1. 后端 FastAPI 服务（`scripts/dashboard.py` + `src/rockchip_agents/dashboard/api.py`）
+2. WebSocket Watcher 线程
+3. React 前端脚手架（Vite + Tailwind + shadcn/ui）
+4. KanbanBoard + TaskCard 组件
+5. WebSocket 连接 + Zustand store
+6. TaskDetail 滑出面板 + REST 操作
+7. NewTaskModal 入队对话框
+8. DagView（React Flow）
+9. Dockerfile.dashboard + docker-compose 集成
+10. 构建验证 + 端到端测试
@@ -0,0 +1,385 @@
+# Framework V2 设计文档
+
+**日期**: 2026-03-08
+**状态**: 已确认
+
+## 背景
+
+当前框架（3 队列 × 15 角色）存在六个核心缺口：边缘验证无结构化、视觉测试单帧无基线、选题无人工门控、算法单线程、OS 覆盖过宽、设备接入仅 SSH。本文档描述 V2 框架的完整升级方案。
+
+---
+
+## §1 设备抽象层
+
+### 1.1 问题
+
+现有 `DeviceAgent` 仅封装 SSH，无法支持 ESP32/Nordic 等 MCU 设备（串口通信）及混合拓扑（部分 MCU 接开发机，部分接 RK3588）。
+
+### 1.2 设备类型与连接方式
+
+| 设备类型 | 连接方式 | 说明 |
+|---------|---------|------|
+| linux | ssh | RK3588、全志等 Linux 设备 |
+| esp32 | serial | 直连开发机 `/dev/ttyUSB*` |
+| nordic | serial | 直连开发机 `/dev/ttyACM*` |
+| esp32 | gateway | 经 RK3588 中转，RK3588 上的串口 |
+
+### 1.3 `devices.yaml` 扩展
+
+```yaml
+devices:
+  rk3588:
+    type: linux
+    connect: ssh
+    host: 192.168.123.181
+    user: pi
+    password: "123123"
+    workspace: /home/pi/Desktop
+    pre_cmd: "clashon && clashproxy on"
+
+  esp32-direct:           # 直连开发机
+    type: esp32
+    connect: serial
+    port: /dev/ttyUSB0
+    baud: 115200
+    workspace: /tmp/esp32
+
+  esp32-via-rk3588:       # 接在 RK3588 上
+    type: esp32
+    connect: gateway
+    via: rk3588
+    port: /dev/ttyUSB1    # RK3588 上的串口号
+    baud: 115200
+
+  nordic-dev:
+    type: nordic
+    connect: serial
+    port: /dev/ttyACM0
+    baud: 115200
+```
+
+### 1.4 `DeviceManager` 统一接口
+
+新建 `tools/device_manager.py`，替代 `tools/device_agent.py`：
+
+```python
+class DeviceAdapter(Protocol):
+    def run_test(self, cmd: str, cwd: str, timeout: int = 120) -> TestResult: ...
+    def get_metrics(self) -> DeviceMetrics: ...
+    def send_file(self, local: Path, remote: str) -> bool: ...
+    def read_logs(self, n: int = 50) -> list[str]: ...
+    def flash(self, firmware: Path) -> bool: ...  # MCU 专用
+
+class TestResult:
+    rc: int
+    stdout: str
+    stderr: str
+    duration_s: float
+    pass_: bool           # 解析自输出（PASS/FAIL/ERROR）
+
+class DeviceMetrics:
+    device: str
+    timestamp: str
+    # Linux 设备
+    cpu_pct: float | None
+    mem_used_mb: float | None
+    npu_util_pct: float | None  # RKNN 专用
+    fps: float | None           # 从推理日志解析
+    # MCU 设备
+    inference_latency_ms: float | None   # 从 UART 输出解析
+    ram_used_kb: int | None              # 从编译报告或 UART 解析
+    flash_used_kb: int | None
+```
+
+**适配器实现：**
+
+- `SSHAdapter`：复用现有 paramiko 逻辑，封装为统一接口
+- `SerialAdapter`：`pyserial` + 命令/响应解析（行尾 `\r\n`）
+  - ESP32：`esptool.py` 烧录，`AT` 命令或自定义协议测试
+  - Nordic：`nrfjprog` / `west flash` 烧录
+- `GatewayAdapter`：SSH 到中转设备，再通过其串口操作目标 MCU
+
+**`DeviceConfig` 扩展（`config.py`）：**
+
+```python
+@dataclass
+class DeviceConfig:
+    host: str = ""
+    user: str = ""
+    password: str = ""
+    workspace: str = ""
+    pre_cmd: str = ""
+    type: str = "linux"       # linux | esp32 | nordic
+    connect: str = "ssh"      # ssh | serial | gateway
+    via: str = ""             # gateway 时的中转设备名
+    port: str = ""            # 串口路径
+    baud: int = 115200
+```
+
+---
+
+## §2 EdgeValidatorService
+
+### 2.1 定位
+
+部署在 RK3588 上的轻量 FastAPI 服务（`services/edge_validator/`），负责：
+- 向开发机 agents 暴露 HTTP API（无需 agents 直接 SSH 操作）
+- 管理 RK3588 本机测试执行
+- 管理接在 RK3588 上的 MCU 设备（串口中转）
+
+### 2.2 API 端点
+
+```
+GET  /health
+GET  /devices                          # 列出所有已识别的串口设备
+GET  /metrics/{device_name}            # 实时指标（linux: /proc; mcu: UART 解析）
+POST /run-test/{device_name}           # 执行测试命令
+     body: {cmd, cwd, timeout}
+     resp: {rc, stdout, stderr, duration_s, pass_}
+POST /flash/{device_name}              # 烧录固件（MCU 专用）
+     body: {firmware_b64, offset}
+POST /screenshot                       # 截取 RK3588 显示内容（需 X11 或 framebuffer）
+     resp: {image_b64, width, height}
+GET  /logs/{device_name}/{n}           # 最近 N 行推理日志
+```
+
+### 2.3 部署
+
+```bash
+# RK3588 上
+cd /home/pi/Desktop/edge_validator
+pip install fastapi uvicorn pyserial
+uvicorn main:app --host 0.0.0.0 --port 8899
+# systemd 自启
+```
+
+### 2.4 全员边端感知
+
+`developer.py` 的 `_build_device_context()` 升级，所有 agents 系统提示包含：
+
+```
+【边端设备访问】
+本公司产品目标为边端落地，请优先验证边端实施效果。
+
+设备列表：
+- rk3588 [linux/ssh]：pi@192.168.123.181
+  EdgeValidatorService: http://192.168.123.181:8899
+  fetch GET  http://192.168.123.181:8899/metrics/rk3588
+  fetch POST http://192.168.123.181:8899/run-test/rk3588
+
+- esp32-direct [esp32/serial]：/dev/ttyUSB0 @115200
+  pyserial 直接操作，烧录：esptool.py --port /dev/ttyUSB0 ...
+
+- esp32-via-rk3588 [esp32/gateway]：
+  fetch POST http://192.168.123.181:8899/run-test/esp32-via-rk3588
+```
+
+---
+
+## §3 选题门控机制
+
+### 3.1 标签分类与审批规则
+
+| 标签 | 适用场景 | 处理方式 |
+|------|---------|---------|
+| `[需求] project: desc` | 现有项目的优化/修复 | 直接 `pending`，无需审批 |
+| `[集成] desc` | 现有能力整合 | 直接 `pending`，无需审批 |
+| `[选题] desc` | **新方向/新课题** | `topic_pending`，Dashboard 审批 |
+| `[市场情报] key=value` | 竞品/趋势信息 | 写入 `__market__` facts，无需审批 |
+
+**[选题] 使用条件（系统提示中明确）：**
+- ✓ 当前项目列表中不存在的全新技术领域
+- ✓ 需要新建项目的方向
+- ✓ 超出平台目标范围的探索性方向
+- ✗ 现有项目的功能/性能优化 → 用 `[需求]`
+- ✗ 现有能力的组合整合 → 用 `[集成]`
+
+### 3.2 Task 状态扩展
+
+```python
+# queue.py 新增状态
+# pending → running → done/failed（现有）
+# topic_pending → (用户确认) → pending → running → done
+#              → (用户否决) → vetoed
+```
+
+```sql
+-- topic_pending 任务不被 dequeue() 取出，直到用户手动确认
+-- 新增 status CHECK 约束
+status TEXT CHECK(status IN (
+    'pending', 'running', 'done', 'failed',
+    'vetoed', 'waiting_approval',
+    'topic_pending'   -- 新增
+)) DEFAULT 'pending'
+```
+
+### 3.3 Dashboard 操作界面
+
+在现有任务列表中新增"待确认选题"分区：
+
+```
+┌─ 待确认选题 (topic_pending) ──────────────────────────────────┐
+│  #47 [选题] 面向工业质检的边端视觉方案   planner  2026-03-08   │
+│       价值分析：当前平台 YOLO 精度满足工业质检需求，市场空白...  │
+│       [✅ 确认执行]  [❌ 否决]  [💬 修改意见...]               │
+└────────────────────────────────────────────────────────────────┘
+```
+
+Dashboard API 新增：
+```
+POST /api/tasks/{id}/approve-topic    → status: pending
+POST /api/tasks/{id}/reject-topic     body: {reason}  → status: vetoed
+```
+
+### 3.4 飞书通知（可选，Dashboard 设置开关）
+
+```
+飞书通知内容（单向 webhook POST，无回调）：
+  📌 新选题待确认 #47
+  方向：面向工业质检的边端视觉方案
+  提出人：planner
+  → 在 Dashboard 中审批：http://<dev-machine>:8080/topics
+```
+
+Dashboard 设置页新增：`飞书选题通知：[ 开启 / 关闭 ]`（默认关闭）
+
+### 3.5 Market-PM 长期情报积累
+
+- `scheduler.py` 新增 12h 定时任务触发 market-pm 情报收集
+- 收集输出：`[市场情报] __market__.competitor.rk3588=xxx`
+- 写入 ProjectMemory `__market__.*` facts（WRITE_ROLES 新增 `market-pm`）
+- 积累 ≥5 条情报后，market-pm 可生成 `[选题]` 供人工确认
+
+---
+
+## §4 算法角色三专项化
+
+### 4.1 角色分工
+
+| 角色 | 专项领域 | 边端验证设备 |
+|------|---------|------------|
+| `algo-antishake` | EIS/OIS、陀螺仪融合、光流防抖 | RK3588 (RKNN) + ESP32 (IMU) |
+| `algo-position` | UWB/BLE RSSI、ToF、多点定位 | Nordic + ESP32 |
+| `algo-nav` | GPS+IMU 融合、Kalman 滤波、姿态解算 | ESP32 (GPS+IMU模组) |
+
+### 4.2 实现
+
+三者均继承 `AlgoResearcherAgent`，追加领域专项系统提示，均注入：
+- `__hardware__` facts（硬件约束）
+- EdgeValidatorService URL（边端验证）
+- 对应设备串口/SSH 信息
+
+### 4.3 base_opt 队列并发调整
+
+```yaml
+base_opt:
+  max_concurrent: 5   # 3 (原) → 5
+  roles: [algo-antishake, algo-position, algo-nav, base-validator, base-architect]
+  # 移除 os-engineer（os-engineer 移至 project_delivery 队列，按需触发）
+```
+
+---
+
+## §5 OS Engineer 性能专项化
+
+### 5.1 新定位
+
+`os-engineer` 不再负责 RTOS/Linux 基础搭建（由 `dev-kernel`/`dev-lowlevel` 承接），专注**系统性能瓶颈识别与优化建议**。
+
+### 5.2 专项领域
+
+| 领域 | 工具 | 输出 |
+|------|------|------|
+| 内存 | `smem`, `valgrind --leak-check`, `/proc/meminfo`, OOM 溯源 | 内存泄漏报告、OOM 触发链 |
+| 网络 | `iperf3`, `ss`, `tc qdisc`, `tcpdump` | 带宽/延迟瓶颈报告 |
+| 视觉库 | OpenCV NEON/SIMD 利用率、Mat 内存开销时序 | 帧处理优化建议 |
+| AI 算子 | RKNN profiler、NPU 算子耗时、算子融合分析 | 推理优化路径 |
+
+输出：
+- `docs/perf_reports/YYYY-MM-DD-{project}.md` 火焰图 + 瓶颈分析
+- `[记忆] __hardware__: perf.{category}.bottleneck=xxx`
+- `[需求] project: 基于剖析结果的具体优化任务`
+
+### 5.3 队列归属
+
+移入 `project_delivery` 队列（按需触发，不占 base_opt 常驻资源）：
+```yaml
+project_delivery:
+  roles: [developer, dev-kernel, dev-lowlevel, system-tester, architect, os-engineer]
+```
+
+---
+
+## §6 边缘验证结构化
+
+### 6.1 base-validator 增强
+
+```
+输出标签：
+  [验证结果] project=yolo fps=15.3 precision=0.82 pass=true
+  [验证退化] project=yolo metric=fps current=12.1 baseline=15.3 drop_pct=20.9
+
+  退化阈值：fps drop >10% 或 precision drop >5% → 自动生成 [需求]
+
+ProjectMemory facts 写入（WRITE_ROLES 含 base-validator）：
+  __hardware__: yolo.fps_baseline=15.3
+  __hardware__: yolo.precision_baseline=0.82
+  __hardware__: yolo.last_validated=2026-03-08T10:00:00
+```
+
+### 6.2 vision-analyst 增强
+
+```
+多帧分析流程（default 5帧，间隔 2s）：
+  1. 调用 EdgeValidatorService /screenshot × 5
+  2. 逐帧 Claude Vision 分析：bbox 合理性、FPS 读取、漏检/误检
+  3. 跨帧分析：检测稳定性、bbox 抖动程度
+  4. 存档：docs/visual_tests/YYYY-MM-DD-{project}-{seq}.png
+
+输出标签：
+  [视觉结果] fps=15.3 stability=HIGH detected=3 false_pos=0 pass=true
+  [视觉退化] fps_trend=降 bbox_jitter=高 → [需求] yolo: 检测框抖动需平滑
+```
+
+---
+
+## §7 完整变更清单（按优先级）
+
+| P | 模块 | 变更说明 |
+|---|------|---------|
+| P0 | `configs/devices.yaml` | 扩展 type/connect/via/port/baud 字段 |
+| P0 | `src/.../config.py` | `DeviceConfig` 扩展对应字段 |
+| P0 | `src/.../tools/device_manager.py` | 新建，SSHAdapter/SerialAdapter/GatewayAdapter |
+| P0 | `services/edge_validator/` | 新建 FastAPI 服务 + systemd 配置 |
+| P1 | `src/.../core/queue.py` | 新增 `topic_pending` status；`execution_host` 字段 |
+| P1 | `src/.../agents/planner.py` | `[选题]` 标签解析；写入 `topic_pending` 任务 |
+| P1 | `src/.../agents/market_pm.py` | `[市场情报]` 标签；12h 定时情报收集 |
+| P1 | `src/.../core/scheduler.py` | market-pm 12h job；os-engineer 移入 project_delivery |
+| P1 | `src/.../dashboard/api.py` | approve-topic / reject-topic 端点；topic 列表页 |
+| P1 | `src/.../agents/developer.py` | `_build_device_context()` 多设备+EdgeValidatorService |
+| P2 | `src/.../agents/algo_antishake.py` | 新建，继承 AlgoResearcherAgent |
+| P2 | `src/.../agents/algo_position.py` | 新建 |
+| P2 | `src/.../agents/algo_nav.py` | 新建 |
+| P2 | `src/.../agents/os_engineer.py` | 系统提示重写，聚焦性能剖析 |
+| P2 | `src/.../agents/base_validator.py` | 结构化验证结果 + 退化检测 |
+| P2 | `src/.../agents/vision_analyst.py` | 多帧分析 + EdgeValidatorService 截图 |
+| P2 | `src/.../core/executor.py` | `_make_agent()` 新增三 algo 角色路由 |
+| P2 | `configs/agents.yaml` | base_opt max_concurrent 3→5；project_delivery 加 os-engineer |
+| P2 | `configs/mcp_tools.yaml` | 三 algo 角色工具配置 |
+| P3 | `src/.../tools/project_memory.py` | WRITE_ROLES 加 `base-validator`, `market-pm` |
+| P3 | 飞书通知设置 | Dashboard 设置页新增"选题飞书通知"开关 |
+
+---
+
+## 决策汇总
+
+| 维度 | 决策 |
+|------|------|
+| 多端协同 | EdgeValidatorService (FastAPI) + 全员 HTTP 调用 |
+| 设备协议 | SSHAdapter / SerialAdapter / GatewayAdapter 统一抽象 |
+| 拓扑 | 混合：MCU 可接开发机或 RK3588 |
+| 选题审批 | Dashboard 操作 + 可选飞书单向通知（无回调） |
+| 选题范围 | 仅新方向/新课题 [选题]；日常优化 [需求] 直通 |
+| 算法分工 | 三专项角色替换通用 algo-researcher |
+| OS 定位 | 性能瓶颈剖析专项，移入 project_delivery |
+| 边端感知 | 所有 agents 系统提示含边端设备 + EdgeValidatorService 信息 |
@@ -0,0 +1,184 @@
+# PlannerAgent 设计文档
+
+**日期**: 2026-03-08
+**状态**: 已确认
+
+## 背景
+
+rockchip-agents 目前缺少"需求侧"驱动力：Architect Agent 做跨项目分析，Developer Agent 执行任务，但没有角色持续评估平台差距、自主生成高价值需求。PlannerAgent 填补这一空白。
+
+核心理念：agents 是一个 **RK3588 生态协作团队**，不是传统项目组。部分角色（planner/architect）需要全局视野，其他角色（developer/tester）聚焦项目但具备平台方向感。
+
+---
+
+## §1 架构 & 触发机制
+
+### 角色定位
+
+PlannerAgent 扮演"产品经理"：
+
+- 持续读取三层目标（平台→项目→任务历史）
+- 评估现状与目标的差距
+- 生成高价值需求入队
+
+### 双触发机制
+
+| 触发点 | 配置 | 说明 |
+|--------|------|------|
+| Scheduler 定时 | `interval_hours: 6`（独立于主调度的 1h） | 低优先级后台扫描 |
+| Executor 队列空 | `on_queue_empty: true` | 无任务时主动补充需求 |
+
+### 并发控制
+
+- `agent_role = "planner"`，不占用 developer/tester 并发槽
+- `max_concurrent_planner: 1`（同时只有一个 PlannerAgent 运行）
+
+---
+
+## §2 输入上下文 & 系统提示
+
+### 三层上下文结构
+
+```
+[平台目标]
+构建 RK3588 NPU 平台上的模块化 AI 推理生态……
+
+[项目概况]
+yolo: 目标检测（当前目标：修复板端零检出）
+mediapipe: 关键点检测（当前目标：提升 Pose FPS）
+embedding: 向量检索（目标：完善 API）
+……
+
+[近期任务历史 - 最近 20 条]
+done: yolo - 修复 INT8 量化精度损失
+failed: ui-tars - RKNN 模型转换超时
+……
+
+[当前待处理任务数: N]
+```
+
+### 系统提示词（PLANNER_SYSTEM）
+
+```
+你是 RK3588 AI 生态团队的产品经理。
+
+你的职责：
+1. 分析平台目标与当前项目状态的差距
+2. 识别跨模块整合机会
+3. 提出高价值的下一步需求
+
+输出格式（报告末尾，每行一条）：
+- [需求] <项目名>: 具体需求描述
+- [集成] 跨模块整合需求描述
+- [目标] <项目名>: 更新后的项目目标
+- [平台目标] 更新后的平台整体目标（仅当需要修订时）
+
+去重原则：若相似需求已在队列中，请勿重复生成。
+```
+
+### 输出标签解析
+
+| 标签 | 目标 | 任务类型 | 优先级 |
+|------|------|---------|--------|
+| `[需求]` | 指定项目 | `feature` | 3 |
+| `[集成]` | `research`（跨项目） | `architect` | 3 |
+| `[目标]` | 指定项目 | — | 写入 ProjectMemory |
+| `[平台目标]` | `__platform__` | — | 写入 ProjectMemory |
+
+---
+
+## §3 三层目标体系
+
+### 存储
+
+`ProjectMemory.project_goals` 表，`project` 字段新增特殊值 `__platform__`：
+
+| project 值 | 含义 |
+|-----------|------|
+| `__platform__` | 平台级目标（整个 RK3588 生态方向） |
+| `yolo` / `embedding` / … | 项目级目标 |
+
+### 写入权限
+
+`WRITE_ROLES = {"productizer", "architect", "planner"}`（新增 `planner`）
+
+### 角色上下文注入分层
+
+```
+planner / architect（全局视野）
+  ├─ platform_goal（完整）
+  ├─ 所有 project_goal（完整）
+  └─ 所有 project_facts（精简）
+
+developer / tester（项目聚焦）
+  ├─ platform_goal（首句摘要，锚定方向感）
+  ├─ 当前 project_goal（完整）
+  └─ 当前 project_facts（完整）
+
+productizer（跨项目评分）
+  ├─ platform_goal（首句摘要）
+  └─ 所有 project_goal（摘要）
+```
+
+### API 变更
+
+```python
+# 新增 role 参数
+def get_context_for_prompt(
+    self,
+    project: str,
+    role: str = "developer",
+) -> str
+```
+
+### 平台目标初始化
+
+PlannerAgent 首次运行时，若 `__platform__` 无记录，自动写入：
+
+```
+构建 RK3588 NPU 平台上的模块化 AI 推理生态，
+持续提升各推理模块成熟度，探索跨模块整合产品价值。
+```
+
+---
+
+## §4 错误处理 & 测试
+
+### 错误处理
+
+| 场景 | 处理方式 |
+|------|---------|
+| Claude CLI 不可达 | 返回 failed，不阻塞 executor |
+| 输出无有效标签 | 正常返回（已达目标是合理结果） |
+| 重复需求 | 比对 pending+running 任务 title 前 40 字，存在则跳过 |
+| `__platform__` 未初始化 | 首次运行自动写入默认值 |
+| platform_goal 写入失败 | 降级：只注入项目级上下文，不中断 |
+
+### 测试
+
+```
+tests/test_planner.py
+├── test_planner_dedup          # 重复标题不重复入队
+├── test_planner_platform_init  # 首次运行自动初始化 platform_goal
+├── test_planner_parse_tags     # 四种标签解析正确
+├── test_planner_context_layers # role="planner" 包含三层上下文
+└── test_planner_empty_output   # 无输出时不崩溃
+
+tests/test_project_memory.py（新增）
+├── test_platform_goal_write    # set_goal("__platform__") 存储正确
+└── test_context_role_filter    # developer role 只获得平台摘要+项目完整
+```
+
+---
+
+## 决策汇总
+
+| 维度 | 决策 |
+|------|------|
+| 触发 | Scheduler 6h + executor 队列空双触发 |
+| 去重 | title 前 40 字 + pending 任务扫描混合 |
+| 输出标签 | `[需求]` / `[集成]` / `[目标]` / `[平台目标]` |
+| 存储 | ProjectMemory，`__platform__` 特殊 project |
+| 角色视野 | planner/architect 全三层；developer/tester 平台摘要+项目完整 |
+| 写权限 | `WRITE_ROLES` 新增 `planner`，可写 platform_goal |
+| 实现模式 | 复用 `ArchitectAgent._analyze` 模式（claude --print + report 工具集） |
@@ -0,0 +1,734 @@
+# PlannerAgent Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** 实现 PlannerAgent（产品经理角色），基于三层目标体系（platform/project/task）持续评估差距并自主生成高价值需求入队。
+
+**Architecture:**
+1. `ProjectMemory.get_context_for_prompt` 新增 `role` 参数，实现分层上下文注入
+2. `WRITE_ROLES` 新增 `"planner"`，支持写入 `__platform__` 级目标
+3. 新建 `agents/planner.py`，复用 `ArchitectAgent._analyze` 模式（`claude --print --output-format json`）
+4. Scheduler 新增独立 6h planner 定时任务；Executor 队列空时触发 planner
+
+**Tech Stack:** Python 3.10, SQLite, `claude --print` subprocess, APScheduler, pytest
+
+---
+
+### Task 1: ProjectMemory — 分层上下文 + planner 写权限
+
+**Files:**
+- Modify: `src/rockchip_agents/tools/project_memory.py`
+- Test: `tests/test_project_memory.py`
+
+**Step 1: 写失败测试**
+
+在 `tests/test_project_memory.py` 末尾追加：
+
+```python
+def test_platform_goal_write_and_read(tmp_path):
+    mem = ProjectMemory(db_path=tmp_path / "mem.db")
+    mem.set_goal("__platform__", "RK3588 生态目标", role="planner")
+    g = mem.get_goal("__platform__")
+    assert g is not None
+    assert g.goal_text == "RK3588 生态目标"
+    assert g.updated_by == "planner"
+
+
+def test_context_role_planner_includes_platform(tmp_path):
+    mem = ProjectMemory(db_path=tmp_path / "mem.db")
+    mem.set_goal("__platform__", "平台总目标", role="planner")
+    mem.set_goal("yolo", "yolo 项目目标", role="architect")
+    ctx = mem.get_context_for_prompt("yolo", role="planner")
+    assert "平台总目标" in ctx
+    assert "yolo 项目目标" in ctx
+
+
+def test_context_role_developer_gets_platform_summary(tmp_path):
+    mem = ProjectMemory(db_path=tmp_path / "mem.db")
+    mem.set_goal("__platform__", "平台总目标，后面很长很长的内容不应出现", role="planner")
+    mem.set_goal("yolo", "yolo 项目详细目标", role="architect")
+    ctx = mem.get_context_for_prompt("yolo", role="developer")
+    assert "平台总目标" in ctx           # 首句出现
+    assert "yolo 项目详细目标" in ctx    # 项目完整
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+cd /data/rockchip/agents && source venv/bin/activate
+pytest tests/test_project_memory.py::test_platform_goal_write_and_read \
+       tests/test_project_memory.py::test_context_role_planner_includes_platform \
+       tests/test_project_memory.py::test_context_role_developer_gets_platform_summary -v
+```
+
+预期：3 个 FAIL（`set_goal` 拒绝 planner / `get_context_for_prompt` 无 role 参数）
+
+**Step 3: 修改 project_memory.py**
+
+修改 `WRITE_ROLES`（第 14 行）：
+
+```python
+WRITE_ROLES: frozenset[str] = frozenset({"productizer", "architect", "planner"})
+```
+
+修改 `get_context_for_prompt` 签名（第 165 行）：
+
+```python
+def get_context_for_prompt(
+    self,
+    project: str,
+    role: str = "developer",
+    query: str | None = None,
+) -> str:
+```
+
+在方法体开头，`parts: list[str] = []` 之后，在读取 goal 之前插入平台目标逻辑：
+
+```python
+# 平台目标注入（role 分层）
+_GLOBAL_ROLES = frozenset({"planner", "architect"})
+platform_goal = self.get_goal("__platform__")
+if platform_goal:
+    if role in _GLOBAL_ROLES:
+        parts.append(
+            f"【平台长期目标（{platform_goal.updated_at[:10]}）】\n"
+            f"{platform_goal.goal_text}"
+        )
+    else:
+        # developer/tester/productizer 只获得首句摘要，锚定方向感
+        first_sentence = platform_goal.goal_text.split("，")[0].split("。")[0][:60]
+        parts.append(f"【平台方向】{first_sentence}")
+
+# planner/architect 还需要汇总所有项目目标
+if role in _GLOBAL_ROLES and project != "__platform__":
+    all_goals = self.list_goals()
+    proj_summaries = [
+        f"  {name}: {g.goal_text[:80]}"
+        for name, g in all_goals.items()
+        if name != "__platform__"
+    ]
+    if proj_summaries:
+        parts.append("【各项目目标概览】\n" + "\n".join(proj_summaries))
+```
+
+**Step 4: 运行测试**
+
+```bash
+pytest tests/test_project_memory.py -v
+```
+
+预期：所有测试 PASS
+
+**Step 5: Commit**
+
+```bash
+git add src/rockchip_agents/tools/project_memory.py tests/test_project_memory.py
+git commit -m "feat(memory): add role-based layered context + planner write permission"
+```
+
+---
+
+### Task 2: PlannerAgent — 核心实现
+
+**Files:**
+- Create: `src/rockchip_agents/agents/planner.py`
+- Test: `tests/test_planner.py`
+
+**Step 1: 写失败测试**
+
+创建 `tests/test_planner.py`：
+
+```python
+from __future__ import annotations
+
+import pytest
+from unittest.mock import MagicMock, patch
+
+from rockchip_agents.agents.planner import PlannerAgent, PLANNER_SYSTEM
+from rockchip_agents.agents.developer import AgentResult
+from rockchip_agents.config import AgentsConfig, ClaudeConfig, SchedulerConfig
+from rockchip_agents.core.queue import Task, TaskQueue
+from rockchip_agents.tools.project_memory import ProjectMemory
+
+
+def _make_config():
+    cfg = MagicMock(spec=AgentsConfig)
+    cfg.projects = {"yolo": MagicMock(description="目标检测", mode="confirm")}
+    cfg.claude = ClaudeConfig(model="claude-sonnet-4-6", api_key=None)
+    return cfg
+
+
+def _make_task():
+    return Task(
+        project="research", type="planner", title="PlannerAgent 周期分析",
+        context="", priority=5, mode="report", agent_role="planner",
+    )
+
+
+def test_planner_parse_tags(tmp_path):
+    """[需求]/[集成]/[目标]/[平台目标] 均被正确解析"""
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    q = TaskQueue(db_path=tmp_path / "t.db")
+    agent = PlannerAgent(_make_config(), queue=q, memory=mem)
+
+    report = (
+        "分析报告内容...\n"
+        "- [需求] yolo: 支持多目标追踪\n"
+        "- [集成] embedding+yolo 联合检索\n"
+        "- [目标] yolo: 提升板端 FPS 至 30\n"
+        "- [平台目标] 成为 RK3588 推理生态标准方案\n"
+    )
+    agent._extract_and_enqueue(report)
+
+    tasks = [dict(r) for r in q._conn().execute(
+        "SELECT project, type, title FROM tasks"
+    ).fetchall()]
+    titles = [t["title"] for t in tasks]
+    assert any("多目标追踪" in t for t in titles), "需求任务未入队"
+    assert any("embedding+yolo" in t for t in titles), "集成任务未入队"
+
+    yolo_goal = mem.get_goal("yolo")
+    assert yolo_goal is not None
+    assert "FPS" in yolo_goal.goal_text
+
+    platform_goal = mem.get_goal("__platform__")
+    assert platform_goal is not None
+    assert "标准方案" in platform_goal.goal_text
+
+
+def test_planner_dedup(tmp_path):
+    """相似标题（前 40 字匹配）不重复入队"""
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    q = TaskQueue(db_path=tmp_path / "t.db")
+    agent = PlannerAgent(_make_config(), queue=q, memory=mem)
+
+    report1 = "- [需求] yolo: 支持多目标追踪以提升产品能力\n"
+    report2 = "- [需求] yolo: 支持多目标追踪以提升产品能力（重复）\n"
+    agent._extract_and_enqueue(report1)
+    agent._extract_and_enqueue(report2)
+
+    count = q._conn().execute("SELECT count(*) FROM tasks").fetchone()[0]
+    assert count == 1, "重复需求被重复入队"
+
+
+def test_planner_platform_init(tmp_path):
+    """首次运行自动初始化 platform_goal"""
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    q = TaskQueue(db_path=tmp_path / "t.db")
+    agent = PlannerAgent(_make_config(), queue=q, memory=mem)
+    assert mem.get_goal("__platform__") is None
+
+    agent._ensure_platform_goal()
+
+    g = mem.get_goal("__platform__")
+    assert g is not None
+    assert len(g.goal_text) > 10
+
+
+def test_planner_empty_output(tmp_path):
+    """Claude 无有效输出时不崩溃"""
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    q = TaskQueue(db_path=tmp_path / "t.db")
+    agent = PlannerAgent(_make_config(), queue=q, memory=mem)
+    # 不应抛异常
+    agent._extract_and_enqueue("")
+    agent._extract_and_enqueue("这是一段没有标签的分析报告。")
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+pytest tests/test_planner.py -v
+```
+
+预期：ImportError（planner.py 不存在）
+
+**Step 3: 创建 planner.py**
+
+创建 `src/rockchip_agents/agents/planner.py`：
+
+```python
+from __future__ import annotations
+
+import logging
+import os
+import re
+from typing import Optional
+
+from rockchip_agents.agents.developer import (
+    AgentResult, _parse_cli_result, _make_drop_root_preexec,
+    _run_with_log, _task_log_path,
+)
+from rockchip_agents.config import AgentsConfig
+from rockchip_agents.core.queue import Task, TaskQueue
+from rockchip_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+TIMEOUT = 600  # planner 是 report 模式，只读分析
+
+_DEFAULT_PLATFORM_GOAL = (
+    "构建 RK3588 NPU 平台上的模块化 AI 推理生态，"
+    "持续提升各推理模块成熟度，探索跨模块整合产品价值。"
+)
+
+PLANNER_SYSTEM = """你是 RK3588 AI 生态团队的产品经理。
+
+你的职责：
+1. 分析平台目标与当前项目状态的差距
+2. 识别跨模块整合机会，提出高价值的下一步需求
+3. 不重复已经在执行或等待中的任务
+
+输出格式（报告末尾，每行一条）：
+- [需求] <项目名>: 具体需求描述
+- [集成] 跨模块整合需求描述
+- [目标] <项目名>: 更新后的项目目标（仅当需要修订时）
+- [平台目标] 更新后的平台整体目标（仅当需要修订时）
+
+去重原则：若相似需求已在队列中，请勿重复生成。
+请用中文输出，保持简洁。
+"""
+
+
+class PlannerAgent:
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        self._config = config
+        self._queue = queue or TaskQueue()
+        self._memory = memory or ProjectMemory()
+
+    def run(self, task: Task) -> AgentResult:
+        self._ensure_platform_goal()
+        prompt = self._build_prompt()
+        cmd = [
+            "claude", "--print",
+            "--output-format", "json",
+            "--no-session-persistence",
+            "--model", self._config.claude.model,
+            "--allowedTools", "Read,Glob,Grep",
+            "--append-system-prompt", PLANNER_SYSTEM,
+            "--", prompt,
+        ]
+        try:
+            env = self._build_env()
+            result = _run_with_log(
+                cmd, None, env, TIMEOUT, _task_log_path(task.id),
+                preexec_fn=_make_drop_root_preexec(env),
+            )
+            if result.status == "done":
+                self._extract_and_enqueue(result.summary)
+                logger.info("PlannerAgent 分析完成，报告 %d 字", len(result.summary))
+            return result
+        except Exception as e:
+            logger.error("PlannerAgent 运行失败: %s", e)
+            return AgentResult(status="failed", summary=str(e))
+
+    # ── helpers ───────────────────────────────────────────────────────────────
+
+    def _ensure_platform_goal(self) -> None:
+        """首次运行时初始化平台目标。"""
+        if not self._memory.get_goal("__platform__"):
+            try:
+                self._memory.set_goal("__platform__", _DEFAULT_PLATFORM_GOAL, role="planner")
+                logger.info("平台目标已初始化")
+            except Exception as e:
+                logger.warning("平台目标初始化失败: %s", e)
+
+    def _build_prompt(self) -> str:
+        platform_ctx = self._memory.get_context_for_prompt("__platform__", role="planner")
+        proj_names = "、".join(self._config.projects.keys())
+        lines = [
+            f"以下是当前 RK3588 AI 生态团队的状态（可用项目：{proj_names}）：\n",
+            platform_ctx.strip(),
+            "",
+        ]
+        for name, proj_cfg in self._config.projects.items():
+            proj_ctx = self._memory.get_context_for_prompt(name, role="planner")
+            desc = proj_cfg.description if hasattr(proj_cfg, "description") else ""
+            lines.append(f"**{name}**: {desc}")
+            if proj_ctx.strip():
+                lines.append(proj_ctx.strip())
+        # 附上当前队列摘要（防重复生成）
+        pending_titles = self._get_pending_titles()
+        if pending_titles:
+            lines.append(f"\n当前待处理任务（{len(pending_titles)} 条，避免重复）：")
+            for t in pending_titles[:15]:
+                lines.append(f"  - {t}")
+        lines.append(
+            "\n请分析平台目标与当前项目状态的差距，提出 3~5 条最高价值的需求或整合机会。"
+        )
+        return "\n".join(lines)
+
+    def _get_pending_titles(self) -> list[str]:
+        try:
+            conn = self._queue._conn()
+            rows = conn.execute(
+                "SELECT title FROM tasks WHERE status IN ('pending','running') ORDER BY id DESC LIMIT 30"
+            ).fetchall()
+            conn.close()
+            return [r["title"] for r in rows]
+        except Exception:
+            return []
+
+    def _is_duplicate(self, title: str) -> bool:
+        """模糊去重：与任何 pending/running 任务 title 前 40 字相同则视为重复。"""
+        key = title[:40]
+        for existing in self._get_pending_titles():
+            if existing[:40] == key:
+                return True
+        return False
+
+    def _extract_and_enqueue(self, report_text: str) -> None:
+        # [需求] <project>: <desc>
+        for m in re.finditer(r"\[需求\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            desc = m.group(2).strip()[:120]
+            if project not in self._config.projects:
+                logger.warning("需求任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            title = f"[需求] {desc}"
+            if self._is_duplicate(title):
+                logger.info("需求重复跳过: %s", title)
+                continue
+            self._queue.enqueue(Task(
+                project=project, type="feature", title=title,
+                context=f"由 PlannerAgent 从差距分析中提取\n{desc}",
+                priority=3, mode=proj_cfg.mode,
+                agent_role="developer", initiator="planner",
+                discussion="PlannerAgent 基于平台目标-项目现状差距分析生成",
+            ))
+            logger.info("需求任务入队: [%s] %s", project, title)
+
+        # [集成] <desc>  （跨项目，无 project 前缀）
+        for m in re.finditer(r"\[集成\]\s+(.+)", report_text):
+            desc = m.group(1).strip()[:120]
+            title = f"[集成] {desc}"
+            if self._is_duplicate(title):
+                logger.info("集成需求重复跳过: %s", title)
+                continue
+            self._queue.enqueue(Task(
+                project="research", type="architect", title=title,
+                context=f"由 PlannerAgent 提出的跨模块整合机会\n{desc}",
+                priority=3, mode="report",
+                agent_role="architect", initiator="planner",
+                discussion="PlannerAgent 跨项目整合分析",
+            ))
+            logger.info("集成任务入队: %s", title)
+
+        # [目标] <project>: <goal>
+        for m in re.finditer(r"\[目标\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            goal_text = m.group(2).strip()[:300]
+            if project not in self._config.projects:
+                logger.warning("目标更新项目 %r 不在配置中，跳过", project)
+                continue
+            try:
+                self._memory.set_goal(project, goal_text, role="planner")
+                logger.info("项目目标已更新 by planner: [%s]", project)
+            except Exception as e:
+                logger.warning("[%s] 写入项目目标失败: %s", project, e)
+
+        # [平台目标] <goal>
+        for m in re.finditer(r"\[平台目标\]\s+(.+)", report_text):
+            goal_text = m.group(1).strip()[:300]
+            try:
+                self._memory.set_goal("__platform__", goal_text, role="planner")
+                logger.info("平台目标已更新 by planner")
+            except Exception as e:
+                logger.warning("写入平台目标失败: %s", e)
+
+    def _build_env(self) -> dict:
+        env = os.environ.copy()
+        if self._config.claude.api_key:
+            env["ANTHROPIC_API_KEY"] = self._config.claude.api_key
+        env.pop("CLAUDECODE", None)
+        env.pop("CLAUDE_CODE_ENTRYPOINT", None)
+        return env
+```
+
+**Step 4: 运行测试**
+
+```bash
+pytest tests/test_planner.py -v
+```
+
+预期：4 个 PASS
+
+**Step 5: Commit**
+
+```bash
+git add src/rockchip_agents/agents/planner.py tests/test_planner.py
+git commit -m "feat(agents): add PlannerAgent with 3-layer goal context and dedup"
+```
+
+---
+
+### Task 3: Executor — 注册 planner 路由
+
+**Files:**
+- Modify: `src/rockchip_agents/core/executor.py:42-63`
+- Test: `tests/test_executor.py`（已有，检查路由）
+
+**Step 1: 写失败测试**
+
+在 `tests/test_executor.py` 末尾追加（若文件存在）：
+
+```python
+def test_executor_routes_planner(mock_config, tmp_path):
+    from rockchip_agents.core.executor import _make_agent
+    from rockchip_agents.agents.planner import PlannerAgent
+    from rockchip_agents.core.queue import Task
+    from rockchip_agents.tools.project_memory import ProjectMemory
+
+    task = Task(
+        project="research", type="planner", title="test",
+        context="", priority=5, mode="report", agent_role="planner",
+    )
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    agent = _make_agent(mock_config, task, mem)
+    assert isinstance(agent, PlannerAgent)
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+pytest tests/test_executor.py::test_executor_routes_planner -v
+```
+
+预期：FAIL（planner 路由不存在，返回 DeveloperAgent）
+
+**Step 3: 修改 executor.py**
+
+在 `_make_agent` 函数（第 42 行）的 `if role == "architect":` 块之后添加：
+
+```python
+if role == "planner":
+    from rockchip_agents.agents.planner import PlannerAgent
+    return PlannerAgent(config, queue=TaskQueue(), memory=memory)
+```
+
+注意：需要在文件顶部确认 `TaskQueue` 已导入（已有）。
+
+**Step 4: 运行测试**
+
+```bash
+pytest tests/test_executor.py -v
+```
+
+预期：全部 PASS
+
+**Step 5: Commit**
+
+```bash
+git add src/rockchip_agents/core/executor.py tests/test_executor.py
+git commit -m "feat(executor): route planner agent_role to PlannerAgent"
+```
+
+---
+
+### Task 4: Scheduler — 6h 定时触发 + 队列空触发
+
+**Files:**
+- Modify: `src/rockchip_agents/core/scheduler.py`
+- Test: `tests/test_scheduler.py`（新增简单验证）
+
+**Step 1: 写失败测试**
+
+创建 `tests/test_scheduler.py`（或追加）：
+
+```python
+def test_planner_task_enqueue(tmp_path, monkeypatch):
+    """_enqueue_planner_task 创建 agent_role=planner 的任务"""
+    from rockchip_agents.core.queue import TaskQueue
+    from rockchip_agents.core import scheduler as sched_module
+
+    q = TaskQueue(db_path=tmp_path / "t.db")
+    monkeypatch.setattr(sched_module, "_get_queue", lambda: q)
+
+    sched_module._enqueue_planner_task()
+
+    rows = q._conn().execute(
+        "SELECT agent_role, type FROM tasks WHERE agent_role='planner'"
+    ).fetchall()
+    assert len(rows) == 1
+    assert rows[0]["type"] == "planner"
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+pytest tests/test_scheduler.py::test_planner_task_enqueue -v
+```
+
+预期：FAIL（`_enqueue_planner_task` 不存在）
+
+**Step 3: 修改 scheduler.py**
+
+在现有 `_scan_and_enqueue` 函数之后，`_run_pending` 之前，添加：
+
+```python
+def _get_queue() -> "TaskQueue":
+    """可被测试 monkeypatch 替换的工厂函数。"""
+    return TaskQueue()
+
+
+def _enqueue_planner_task() -> None:
+    """将 planner 分析任务入队（同标题去重，不重复触发）。"""
+    from rockchip_agents.core.queue import Task
+    q = _get_queue()
+    tid = q.enqueue(Task(
+        project="research",
+        type="planner",
+        title="PlannerAgent 周期差距分析",
+        context="由 Scheduler 定时触发，分析平台目标与项目现状差距",
+        priority=5,
+        mode="report",
+        agent_role="planner",
+        initiator="scheduler",
+        discussion="6h 定时 planner 分析",
+    ))
+    if tid > 0:
+        logger.info("PlannerAgent 任务入队 #%d", tid)
+    else:
+        logger.debug("PlannerAgent 任务已在队列中，跳过")
+```
+
+在 `start()` 函数的 `scheduler.add_job` 部分，追加 planner 定时任务：
+
+```python
+scheduler.add_job(
+    _enqueue_planner_task, "interval", hours=6, id="planner",
+    next_run_time=datetime.now() + timedelta(minutes=10),
+)
+```
+
+**Step 4: 修改 `_run_pending` 支持队列空触发 planner**
+
+将 `_run_pending` 修改为：
+
+```python
+def _run_pending() -> None:
+    cfg = load_config()
+    executor = Executor(cfg)
+    count = asyncio.run(executor.run_all_pending())
+    logger.info("本轮执行任务数: %d", count)
+    if count == 0:
+        # 队列空时触发 planner 补充需求
+        _enqueue_planner_task()
+        logger.info("队列为空，已触发 PlannerAgent 差距分析")
+```
+
+**Step 5: 运行测试**
+
+```bash
+pytest tests/test_scheduler.py -v
+pytest tests/ -q  # 全量回归
+```
+
+预期：全部 PASS
+
+**Step 6: Commit**
+
+```bash
+git add src/rockchip_agents/core/scheduler.py tests/test_scheduler.py
+git commit -m "feat(scheduler): add 6h planner job + queue-empty trigger"
+```
+
+---
+
+### Task 5: get_context_for_prompt 调用方适配
+
+**Files:**
+- Modify: `src/rockchip_agents/agents/developer.py:252`
+- Modify: `src/rockchip_agents/agents/architect.py:190`
+
+确保传入 `role` 参数，让不同 agent 获得对应层级上下文。
+
+**Step 1: 修改 developer.py**
+
+第 252 行 `_build_context` 中：
+
+```python
+# 旧
+memory_ctx = self._memory.get_context_for_prompt(task.project)
+
+# 新
+memory_ctx = self._memory.get_context_for_prompt(task.project, role=task.agent_role)
+```
+
+**Step 2: 修改 architect.py**
+
+第 190 行 `_build_prompt` 中：
+
+```python
+# 旧
+goal_ctx = self._memory.get_context_for_prompt(name)
+
+# 新
+goal_ctx = self._memory.get_context_for_prompt(name, role="architect")
+```
+
+**Step 3: 全量测试**
+
+```bash
+pytest tests/ -q
+```
+
+预期：137+ PASS，0 FAIL
+
+**Step 4: Commit**
+
+```bash
+git add src/rockchip_agents/agents/developer.py src/rockchip_agents/agents/architect.py
+git commit -m "feat(agents): pass role to get_context_for_prompt for layered context"
+```
+
+---
+
+### Task 6: Dashboard — planner 任务在 UI 展示
+
+**Files:**
+- Modify: `dashboard/src/components/TaskList.tsx`（或相关列表组件）
+
+PlannerAgent 的任务 `project="research"、agent_role="planner"`，确认 Dashboard 任务列表能正常展示（无需特殊处理，只需 UI 显示 `[planner]` 标签即可）。
+
+**Step 1: 检查 TaskList 是否有 agent_role 展示**
+
+```bash
+grep -n "agent_role\|planner" dashboard/src/components/TaskList.tsx 2>/dev/null | head -20
+```
+
+若已有角色标签展示，无需修改。若无，在任务标题旁添加 `{task.agent_role}` badge。
+
+**Step 2: 全量测试（前端构建验证）**
+
+```bash
+cd dashboard && npm run build 2>&1 | tail -5
+```
+
+**Step 3: Commit（如有修改）**
+
+```bash
+git add dashboard/src/
+git commit -m "feat(dashboard): show planner role badge in task list"
+```
+
+---
+
+## 验收标准
+
+1. `pytest tests/ -q` 全部通过（包括新增的 `test_planner.py` 4 条 + `test_project_memory.py` 3 条新测试）
+2. 手动验证：`python scripts/run.py scan` 后队列中出现 `PlannerAgent 周期差距分析` 任务
+3. `python scripts/run.py goals` 展示 `__platform__` 目标
+4. Dashboard 任务列表能看到 `agent_role=planner` 的任务
+
+## 注意事项
+
+- PlannerAgent 只用 `claude --print --allowedTools Read,Glob,Grep`（report 模式），不修改任何代码文件
+- `enqueue()` 本身已有 `project+type+title` 精确去重，PlannerAgent 内部额外做 title 前 40 字模糊去重，两层配合防止语义重复任务
+- `__platform__` 不在 `self._config.projects` 中，写目标时直接调用 `memory.set_goal`，不做项目名校验
@@ -0,0 +1,275 @@
+# 三组 15 角色团队架构设计文档
+
+**日期**: 2026-03-08
+**状态**: 已确认
+
+## 背景
+
+当前系统为单队列 8 角色，缺乏明确的组织分工。随着业务扩展至多平台（RK3588/ESP32/全志/Nordic）、多产品线（方案级+整机出货），需要引入三组架构，覆盖基础技术研发、项目研发交付、产品洞察与市场落地三个维度。
+
+---
+
+## 业务上下文
+
+- **产品形态**：嵌入式 AI 方案（软硬全包含）为主 + 部分整机出货
+- **平台**：Rockchip（RK3588/RV1106/RK3562）、ESP32/Nordic（TinyML）、全志
+- **AI 能力**：RKNN 推理（板端）+ TinyML（微控制器）
+- **OS 栈**：Linux + RTOS 双栈
+
+---
+
+## 三组架构总览
+
+```
+┌─────────────────────────────────────────────────────────┐
+│ 🔬 基础技术研发组   base_opt_tasks.db   并发:3  周期:4h │
+│    hw-engineer / os-engineer / algo-researcher          │
+│    base-validator / base-architect                      │
+├─────────────────────────────────────────────────────────┤
+│ 🚀 项目研发与交付组  tasks.db（现有）   并发:3  周期:1h │
+│    developer / dev-kernel / dev-lowlevel                │
+│    system-tester / architect                            │
+├─────────────────────────────────────────────────────────┤
+│ 💡 产品洞察与市场组  insight_tasks.db   并发:2  事件触发│
+│    planner / productizer / vision-analyst               │
+│    market-pm / media-producer                           │
+└─────────────────────────────────────────────────────────┘
+```
+
+---
+
+## §1 基础技术研发组
+
+**队列**: `data/base_opt_tasks.db`
+**并发**: 3
+**扫描周期**: 4h（慢周期，长时间基线任务）
+**目标**: 建立平台级能力基线，输出能力事实供项目组消费
+
+### 角色定义
+
+#### `hw-engineer`（现有，增强）
+- 芯片选型（算力/功耗/接口/成本/供货）
+- BOM 评审与替代方案
+- 接口信号完整性、ESD 保护设计
+- **写入 `__hardware__` facts**：硬件规格约束，供全组读取
+- 模式：confirm（硬件决策需人工审批）
+- Skills：`fetch`、`sequential-thinking`
+
+#### `os-engineer`（新建，继承 LowlevelDev + KernelDev 能力）
+- RTOS 任务调度、中断优先级、驱动框架
+- Linux 内核基础层：Kconfig/DTS 模板、驱动框架
+- U-Boot / Bootloader 基础配置
+- 模式：auto/confirm
+- Skills：`ssh-device`、`sequential-thinking`
+
+#### `algo-researcher`（新建）
+- 防抖算法研究与基线实现
+- 短距离定位/测距算法
+- GPS + 惯导（IMU 融合）算法
+- **受硬件约束**：读取 `__hardware__` facts，确认算法在目标平台可行
+- 产出：算法能力事实写入 ProjectMemory
+- 模式：auto/confirm
+- Skills：`ssh-device`、`sequential-thinking`、`fetch`
+
+#### `base-validator`（新建，增强 TesterAgent）
+测试能力覆盖四层：
+
+| 层次 | 内容 |
+|------|------|
+| **精度测试** | 检出率、误检率、漏检率，基线对比 |
+| **性能测试** | FPS、推理延迟、NPU 利用率、CPU/内存占用 |
+| **稳定性测试** | 长时间压力运行、内存泄漏检测、异常恢复验证 |
+| **硬件约束验证** | 读取 `__hardware__` facts，确认算法在目标平台满足约束 |
+
+- 基线退化超阈值 → 自动触发 `[需求]` 任务入队
+- 测试结果写入 ProjectMemory facts
+- Skills：`ssh-device`、`sequential-thinking`
+
+#### `base-architect`（新建，继承 ArchitectAgent）
+- 分析基础技术栈（OS/算法/硬件）技术债和瓶颈
+- 跨方向整合建议（如"防抖算法与 GPS 惯导融合"）
+- 输出 `[优化]` 任务给 `algo-researcher` / `os-engineer`
+- **区别于项目组 architect**：聚焦平台能力本身，不关注项目交付
+- 模式：report
+
+---
+
+## §2 项目研发与交付组
+
+**队列**: `data/tasks.db`（现有，保持兼容）
+**并发**: 3
+**扫描周期**: 1h（快周期）
+**目标**: 消费基础组能力事实，完成面向客户的项目交付
+
+### 角色定义
+
+#### `developer`（现有）
+- 项目功能开发、Bug 修复
+- 读取基础组能力事实作为实现依据
+- 模式：auto/confirm
+
+#### `dev-kernel`（现有）
+- 项目级 Linux 驱动适配、DTS、Kconfig
+- 模式：auto/confirm
+
+#### `dev-lowlevel`（现有）
+- 项目级 BSP、Bootloader 适配、硬件 bring-up
+- 模式：auto/confirm
+
+#### `architect`（现有）
+- 跨项目架构分析 + arch_optimize
+- 消费基础组能力事实，制定项目架构方向
+- 模式：report + auto（optimize 时）
+
+#### `system-tester`（替换现有 tester，大幅增强）
+测试能力矩阵：
+
+| 能力 | 实现 |
+|------|------|
+| **功能测试** | pytest 全套，SSH 板端自动执行 |
+| **视觉测试** | Playwright 截图 → 传 vision-analyst 分析判断 |
+| **性能测试** | FPS/延迟/内存自动测量，写入 facts |
+| **跨平台** | RK3588/全志/ESP32 多平台 SSH 调度 |
+| **报告生成** | 结构化报告 → `docs/test_reports/` |
+| **回归闸门** | 测试失败自动触发 `[需求]` 入队 |
+
+- Skills：`playwright`、`ssh-device`、`sequential-thinking`、`fetch`
+
+---
+
+## §3 产品洞察与市场落地组
+
+**队列**: `data/insight_tasks.db`（新建）
+**并发**: 2
+**触发**: 事件驱动（planner 6h 定时 + 队列空触发）
+**目标**: 产品方向把控 + 市场文档 + 视觉质量把关
+
+### 角色定义
+
+#### `planner`（现有）
+- 平台级差距分析，跨三组需求生成
+- 三层目标体系（`__platform__` → project → task）
+- 模式：report
+
+#### `productizer`（现有）
+- 成熟度评分（100分制规则引擎）
+- 模式：report
+
+#### `vision-analyst`（新建）
+- **Playwright 截图** 板端推理界面/监控大屏
+- **Claude Vision（LLM多模态）** 分析：检测框合理性、漏检、FPS 数值读取
+- 定量补充：对接推理日志（JSON/txt）
+- 输出：视觉分析报告 + `[需求]` 任务入项目组队列
+- 触发：system-tester 完成后自动触发 / planner 主动下发
+- Skills：`playwright`、`sequential-thinking`
+- 模式：report
+
+#### `market-pm`（新建）
+- 读取项目目标、技术规格、能力事实
+- 生成：方案书、技术白皮书、BOM 估算（Markdown → `docs/market/`）
+- 协助 planner 提出新项目需求
+- 模式：report
+- Skills：`fetch`、`sequential-thinking`
+
+#### `media-producer`（新建）
+- **随时触发**（不依赖测试结果）
+- 技术侧：推理结果图标注（检测框/置信度/平台水印）
+- 市场侧：产品展示效果图（硬件渲染+算法效果合成）
+- 工具：OpenCV、ffmpeg、Pillow
+- 输出：`assets/` 目录
+- 模式：auto
+- Skills：`sequential-thinking`
+
+---
+
+## §4 信息流架构
+
+```
+hw-engineer
+    └─写入──▶ __hardware__ facts（硬件规格约束）
+                   │
+                   ▼
+algo-researcher / os-engineer
+    └─读取约束──▶ 实现算法/OS
+    └─写入──▶ 能力事实（algo.antishake.fps=60, os.rtos.latency=1ms）
+                   │
+base-validator──▶ 精度/性能/稳定性验证──▶ 写入验证通过事实
+base-architect──▶ 跨方向整合建议──▶ [优化] 任务
+
+                   │（能力事实传递）
+                   ▼
+developer / dev-kernel / dev-lowlevel
+    └─消费能力事实──▶ 项目实现
+system-tester
+    └─Playwright截图──▶ vision-analyst──▶ 视觉通过/失败
+    └─性能测试──▶ 写入项目 facts
+architect──▶ 架构优化建议
+
+                   │（项目状态传递）
+                   ▼
+planner──▶ 差距分析──▶ 跨组需求入队
+productizer──▶ 成熟度评分
+vision-analyst──▶ 视觉质量报告
+market-pm──▶ 方案书/白皮书
+media-producer──▶ 演示图/效果图（随时触发）
+```
+
+---
+
+## §5 Skills 授权配置（mcp_tools.yaml）
+
+```yaml
+roles:
+  hw-engineer:      [fetch, sequential-thinking]
+  os-engineer:      [ssh-device, sequential-thinking]
+  algo-researcher:  [ssh-device, sequential-thinking, fetch]
+  base-validator:   [ssh-device, sequential-thinking]
+  base-architect:   []
+  developer:        [sequential-thinking, ssh-device]
+  dev-kernel:       [ssh-device, sequential-thinking]
+  dev-lowlevel:     [ssh-device, sequential-thinking]
+  system-tester:    [playwright, ssh-device, sequential-thinking, fetch]
+  architect:        []
+  planner:          []
+  productizer:      []
+  vision-analyst:   [playwright, sequential-thinking]
+  market-pm:        [fetch, sequential-thinking]
+  media-producer:   [sequential-thinking]
+```
+
+---
+
+## §6 队列映射与调度配置
+
+```yaml
+queues:
+  base_opt:
+    db: data/base_opt_tasks.db
+    max_concurrent: 3
+    scan_interval_hours: 4
+    roles: [hw-engineer, os-engineer, algo-researcher, base-validator, base-architect]
+
+  project_delivery:
+    db: data/tasks.db
+    max_concurrent: 3
+    scan_interval_hours: 1
+    roles: [developer, dev-kernel, dev-lowlevel, system-tester, architect]
+
+  insight:
+    db: data/insight_tasks.db
+    max_concurrent: 2
+    trigger: event_driven  # planner 6h + 队列空
+    roles: [planner, productizer, vision-analyst, market-pm, media-producer]
+```
+
+---
+
+## 实现优先级
+
+| 优先级 | 内容 |
+|--------|------|
+| P0 | 三队列调度架构（Scheduler + Executor 扩展） |
+| P1 | `base-validator`、`system-tester` 增强（测试能力核心） |
+| P2 | `vision-analyst`（Playwright + LLM 视觉分析） |
+| P3 | `os-engineer`、`algo-researcher`、`base-architect`（新角色） |
+| P4 | `market-pm`、`media-producer`（文档和媒体产出） |
@@ -0,0 +1,138 @@
+# EnvCollector 设计文档
+
+## 概述
+
+自动收集本机环境变量中的模型路径、SSH 远端设备的模型文件 + 磁盘空间 + 关键包版本，
+持久化存储到 ProjectMemory，供所有 Agent 查询，避免重复下载模型。
+
+---
+
+## 1. 数据模型 & 存储结构
+
+结果全部写入 `ProjectMemory` 的 `__env__` 虚拟项目命名空间（类似现有 `__hardware__`）。
+
+### Key 规范
+
+```
+# 本机环境变量（含 MODEL/RKNN/ONNX/WEIGHT/PATH 关键字的变量）
+__env__.local.envvar.<VAR_NAME>  = "/path/to/model.rknn"
+
+# 本机文件系统扫描（/data/ 下递归，max_depth=5）
+__env__.local.models.rknn        = "/data/rockchip/yolo/model.rknn,/data/rockchip/embedding/emb.rknn"
+__env__.local.models.onnx        = "/data/rockchip/yolo/model.onnx"
+__env__.local.models.pt          = ""
+
+# SSH 设备模型文件
+__env__.{device_name}.models.rknn = "/home/pi/Desktop/yolo/model.rknn,/home/pi/models/det.rknn"
+__env__.{device_name}.models.onnx = ""
+
+# SSH 设备磁盘空间
+__env__.{device_name}.disk        = "Filesystem /dev/root Size=29G Used=8.2G Avail=19G"
+
+# SSH 设备已安装关键包
+__env__.{device_name}.packages    = "rknn-toolkit-lite2==2.3.2,opencv-python==4.13.0,numpy==2.4.2"
+
+# 设备离线时错误记录
+__env__.{device_name}.error       = "offline at 2026-03-09T10:30:00"
+
+# 收集时间戳
+__env__.collected_at              = "2026-03-09T10:30:00"
+```
+
+多值用逗号分隔，兼容现有 `get_facts()` 返回 dict 的结构，无需改表结构。
+
+---
+
+## 2. 收集逻辑 & 触发机制
+
+### EnvCollector 核心类（tools/env_collector.py）
+
+```python
+class EnvCollector:
+    def collect_local(self, scan_dirs: list[str] = None) -> dict[str, str]:
+        # 1. 扫描 os.environ，过滤含 MODEL/RKNN/ONNX/WEIGHT/PATH 的 key
+        # 2. 文件系统：在 scan_dirs（默认 ["/data/"]）下 find *.rknn/*.onnx/*.pt/*.pb
+        #    跳过 venv/__pycache__/.git，max_depth=5
+
+    def collect_device(self, name: str, device: DeviceConfig) -> dict[str, str]:
+        # 复用 DeviceAgent.run_command() 执行：
+        # 1. find $workspace -maxdepth 4 -name "*.rknn" -o -name "*.onnx" ...
+        # 2. df -h | grep -v tmpfs | head -5
+        # 3. pip list 2>/dev/null | grep -E "rknn|torch|opencv|numpy|pillow"
+        # 设备离线时捕获异常，写 error key
+
+    def save(self, facts: dict[str, str], memory: ProjectMemory) -> None:
+        # set_fact("__env__", key, value, role="env-collector")
+        # 最后写 collected_at
+```
+
+### 触发机制
+
+| 触发方式 | 实现位置 | 说明 |
+|---------|---------|------|
+| **手动** | `scripts/run.py collect-env` | 立即执行，输出收集摘要到 stdout |
+| **启动时** | `scheduler.start()` 末尾 daemon thread | 不阻塞调度器启动 |
+| **每日定时** | APScheduler `interval hours=24` | 后台安静更新，失败不影响主流程 |
+
+SSH 扫描失败（设备离线）时：静默跳过，记录 `__env__.{name}.error` 到 ProjectMemory。
+
+---
+
+## 3. Agent 使用方式
+
+### system prompt 注入（_build_device_context 扩展）
+
+`DeveloperAgent._build_device_context()` 末尾追加"已知模型"摘要块：
+
+```
+【已知模型文件（避免重复下载）】
+本机:
+  *.rknn: /data/rockchip/yolo/model.rknn, /data/rockchip/embedding/emb.rknn
+  RKNN_MODEL_PATH=/data/rockchip/models（来自环境变量）
+
+rk3588 (192.168.123.181):
+  *.rknn: /home/pi/Desktop/yolo/model.rknn
+  磁盘: /dev/root 29G 已用8.2G 剩余19G
+  已装: rknn-toolkit-lite2==2.3.2, opencv-python==4.13.0
+
+（更新于 2026-03-09 10:30，执行 `python scripts/run.py collect-env` 可刷新）
+```
+
+只在 memory 不为 None 且 `__env__` 有数据时注入（无数据时静默跳过）。
+
+### ProjectMemory 精确查询
+
+其他 Agent（base-validator、vision-analyst）可直接：
+
+```python
+facts = self._memory.get_facts("__env__")
+rk3588_models = facts.get("rk3588.models.rknn", "").split(",")
+disk_info = facts.get("rk3588.disk", "unknown")
+```
+
+### CLI 输出示例
+
+```bash
+$ python scripts/run.py collect-env
+
+[本机]
+  ENV: RKNN_MODEL_PATH=/data/rockchip/models (1 条)
+  *.rknn: 3 个文件
+  *.onnx: 1 个文件
+
+[rk3588 @ 192.168.123.181]
+  *.rknn: 2 个文件
+  磁盘: 剩余 19G / 29G
+  关键包: rknn-toolkit-lite2==2.3.2, opencv-python==4.13.0
+
+已保存到 ProjectMemory (__env__)
+```
+
+---
+
+## 实施原则
+
+- **TDD**：先写失败测试，再实现
+- **YAGNI**：不做 AI 分析模型版本、不做模型哈希校验
+- **轻依赖**：只用 paramiko（已有）+ pathlib + os.environ，无额外依赖
+- **容错优先**：任意设备离线或扫描失败不影响主流程
@@ -0,0 +1,721 @@
+# EnvCollector Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** 自动收集本机环境变量模型路径 + SSH 设备模型文件/磁盘/包信息，持久化至 ProjectMemory `__env__` 命名空间，注入 DeveloperAgent 上下文，避免重复下载模型。
+
+**Architecture:** 独立 `EnvCollector` 类（`tools/env_collector.py`），直接复用现有 `DeviceAgent._ssh()` + `ProjectMemory.set_fact()`；CLI / scheduler / developer 三处集成均为薄封装，不引入新依赖。
+
+**Tech Stack:** pathlib + os.environ + paramiko（已有）+ APScheduler（已有）+ ProjectMemory（已有）
+
+---
+
+### Task 1: tools/env_collector.py — 核心收集类
+
+**Files:**
+- Create: `src/rockchip_agents/tools/env_collector.py`
+- Modify: `src/rockchip_agents/tools/project_memory.py`（WRITE_ROLES 添加 "env-collector"）
+- Test: `tests/test_env_collector.py`
+
+**Step 1: 写失败测试**
+
+```python
+# tests/test_env_collector.py
+from __future__ import annotations
+import os
+import pytest
+from unittest.mock import MagicMock, patch
+from rockchip_agents.tools.env_collector import EnvCollector
+from rockchip_agents.tools.project_memory import ProjectMemory
+
+
+def test_collect_local_env_vars(tmp_path):
+    """collect_local 应收集含 MODEL/RKNN/ONNX 关键词的环境变量。"""
+    env = {
+        "RKNN_MODEL_PATH": "/data/rockchip/models",
+        "HOME": "/home/pi",           # 不含关键词，不应收集
+        "ONNX_WEIGHT_DIR": "/data/weights",
+    }
+    with patch.dict(os.environ, env, clear=True):
+        col = EnvCollector()
+        facts = col.collect_local(scan_dirs=[])
+    assert facts.get("local.envvar.RKNN_MODEL_PATH") == "/data/rockchip/models"
+    assert facts.get("local.envvar.ONNX_WEIGHT_DIR") == "/data/weights"
+    assert "local.envvar.HOME" not in facts
+
+
+def test_collect_local_file_scan(tmp_path):
+    """collect_local 文件扫描应找到 *.rknn / *.onnx / *.pt 文件。"""
+    (tmp_path / "model.rknn").touch()
+    (tmp_path / "sub").mkdir()
+    (tmp_path / "sub" / "det.onnx").touch()
+    col = EnvCollector()
+    facts = col.collect_local(scan_dirs=[str(tmp_path)])
+    assert "/model.rknn" in facts["local.models.rknn"]
+    assert "/det.onnx" in facts["local.models.onnx"]
+    assert facts.get("local.models.pt", "") == ""
+
+
+def test_collect_local_skips_venv(tmp_path):
+    """collect_local 应跳过 venv / __pycache__ 目录。"""
+    venv = tmp_path / "venv" / "lib"
+    venv.mkdir(parents=True)
+    (venv / "ignore.rknn").touch()
+    col = EnvCollector()
+    facts = col.collect_local(scan_dirs=[str(tmp_path)])
+    assert facts.get("local.models.rknn", "") == ""
+
+
+def test_collect_device_success():
+    """collect_device 成功时应返回 models/disk/packages facts。"""
+    from rockchip_agents.config import DeviceConfig
+    dev = DeviceConfig(host="1.2.3.4", user="pi", password="pi",
+                       workspace="/home/pi", type="linux", connect="ssh")
+
+    def fake_run(cmd, remote_cwd="", timeout=60):
+        if "find" in cmd:
+            return "[rc=0]\n/home/pi/model.rknn\n"
+        if "df" in cmd:
+            return "[rc=0]\n/dev/root  29G  8.2G  19G\n"
+        if "pip" in cmd:
+            return "[rc=0]\nrknn-toolkit-lite2 2.3.2\nopencv-python 4.13.0\n"
+        return "[rc=0]\n"
+
+    col = EnvCollector()
+    with patch.object(col, "_run_on_device", side_effect=fake_run):
+        facts = col.collect_device("rk3588", dev)
+    assert "/home/pi/model.rknn" in facts["rk3588.models.rknn"]
+    assert "29G" in facts["rk3588.disk"]
+    assert "rknn-toolkit-lite2" in facts["rk3588.packages"]
+
+
+def test_collect_device_offline():
+    """设备离线时应写入 error key，不抛异常。"""
+    from rockchip_agents.config import DeviceConfig
+    dev = DeviceConfig(host="1.2.3.4", user="pi", password="pi",
+                       workspace="/home/pi", type="linux", connect="ssh")
+    col = EnvCollector()
+    with patch.object(col, "_run_on_device", side_effect=Exception("Connection refused")):
+        facts = col.collect_device("rk3588", dev)
+    assert "offline" in facts.get("rk3588.error", "")
+
+
+def test_save_writes_to_memory(tmp_path):
+    """save() 应将 facts 写入 ProjectMemory __env__ 命名空间。"""
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    col = EnvCollector()
+    col.save({"local.models.rknn": "/data/a.rknn"}, mem)
+    stored = mem.get_facts("__env__")
+    assert stored["local.models.rknn"] == "/data/a.rknn"
+    assert "collected_at" in stored
+
+
+def test_save_requires_env_collector_role(tmp_path):
+    """只有 env-collector role 才能写 __env__ facts（需先确认 WRITE_ROLES 已扩展）。"""
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    # 不应抛 PermissionError
+    mem.set_fact("__env__", "test.key", "val", role="env-collector")
+    assert mem.get_facts("__env__")["test.key"] == "val"
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+source venv/bin/activate
+pytest tests/test_env_collector.py -v
+# Expected: ImportError / AttributeError — 模块不存在
+```
+
+**Step 3: 更新 WRITE_ROLES（project_memory.py）**
+
+在 `src/rockchip_agents/tools/project_memory.py` 找到 `WRITE_ROLES`，添加 `"env-collector"`：
+
+```python
+# 修改前
+WRITE_ROLES: frozenset[str] = frozenset({"productizer", "architect"})
+
+# 修改后
+WRITE_ROLES: frozenset[str] = frozenset({"productizer", "architect", "env-collector"})
+```
+
+**Step 4: 实现 env_collector.py**
+
+```python
+# src/rockchip_agents/tools/env_collector.py
+from __future__ import annotations
+
+import logging
+import os
+from datetime import datetime, timezone
+from pathlib import Path
+
+from rockchip_agents.config import DeviceConfig
+from rockchip_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+# 文件扫描跳过这些目录
+_SKIP_DIRS = {"venv", "__pycache__", ".git", "node_modules", ".tox"}
+# 环境变量关键词过滤
+_ENV_KEYWORDS = {"MODEL", "RKNN", "ONNX", "WEIGHT", "PATH"}
+# 扫描的模型扩展名
+_EXTENSIONS = {".rknn", ".onnx", ".pt", ".pb"}
+
+
+class EnvCollector:
+    """收集本机环境变量 + SSH 设备模型文件信息，写入 ProjectMemory __env__。"""
+
+    def collect_local(self, scan_dirs: list[str] | None = None) -> dict[str, str]:
+        """收集本机环境变量（过滤关键词）+ 文件系统模型文件。"""
+        facts: dict[str, str] = {}
+
+        # 1. 环境变量扫描
+        for key, val in os.environ.items():
+            if any(kw in key.upper() for kw in _ENV_KEYWORDS):
+                facts[f"local.envvar.{key}"] = val
+
+        # 2. 文件系统扫描
+        dirs = scan_dirs if scan_dirs is not None else ["/data/"]
+        found: dict[str, list[str]] = {ext: [] for ext in _EXTENSIONS}
+        for base in dirs:
+            base_path = Path(base)
+            if not base_path.exists():
+                continue
+            self._scan_dir(base_path, found, depth=0)
+
+        for ext, paths in found.items():
+            key = f"local.models.{ext.lstrip('.')}"
+            facts[key] = ",".join(sorted(set(paths)))
+
+        return facts
+
+    def _scan_dir(self, path: Path, found: dict[str, list[str]], depth: int) -> None:
+        if depth > 5:
+            return
+        try:
+            for entry in path.iterdir():
+                if entry.is_dir():
+                    if entry.name not in _SKIP_DIRS:
+                        self._scan_dir(entry, found, depth + 1)
+                elif entry.is_file() and entry.suffix in _EXTENSIONS:
+                    found[entry.suffix].append(str(entry))
+        except PermissionError:
+            pass
+
+    def collect_device(self, name: str, device: DeviceConfig) -> dict[str, str]:
+        """连接 SSH 设备，收集模型文件/磁盘/包信息。"""
+        facts: dict[str, str] = {}
+        try:
+            ws = device.workspace
+
+            # 模型文件
+            find_cmd = (
+                f'find {ws} -maxdepth 4 \\( '
+                f'-name "*.rknn" -o -name "*.onnx" -o -name "*.pt" -o -name "*.pb" '
+                f'\\) 2>/dev/null'
+            )
+            out = self._run_on_device(find_cmd, ws)
+            for ext in (".rknn", ".onnx", ".pt", ".pb"):
+                key_name = ext.lstrip(".")
+                paths = [l for l in out.splitlines() if l.endswith(ext) and not l.startswith("[")]
+                facts[f"{name}.models.{key_name}"] = ",".join(paths)
+
+            # 磁盘空间
+            disk_out = self._run_on_device("df -h | grep -v tmpfs | head -5", ws)
+            lines = [l for l in disk_out.splitlines() if not l.startswith("[")]
+            facts[f"{name}.disk"] = " | ".join(lines).strip()
+
+            # 已安装关键包
+            pkg_out = self._run_on_device(
+                'pip list 2>/dev/null | grep -iE "rknn|torch|opencv|numpy|pillow"', ws
+            )
+            pkgs = []
+            for line in pkg_out.splitlines():
+                if line.startswith("["):
+                    continue
+                parts = line.split()
+                if len(parts) >= 2:
+                    pkgs.append(f"{parts[0]}=={parts[1]}")
+            facts[f"{name}.packages"] = ",".join(pkgs)
+
+        except Exception as e:
+            logger.warning("设备 %s 收集失败: %s", name, e)
+            ts = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%S")
+            facts[f"{name}.error"] = f"offline at {ts}"
+
+        return facts
+
+    def _run_on_device(self, cmd: str, remote_cwd: str, timeout: int = 60) -> str:
+        """内部 SSH 执行（可被测试 monkeypatch）。"""
+        # 不在这里 import DeviceAgent，避免循环依赖
+        raise NotImplementedError("需在运行时传入 DeviceAgent 或 override")
+
+    def save(self, facts: dict[str, str], memory: ProjectMemory) -> None:
+        """将 facts 写入 ProjectMemory __env__ 命名空间。"""
+        for key, val in facts.items():
+            memory.set_fact("__env__", key, val, role="env-collector")
+        ts = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%S")
+        memory.set_fact("__env__", "collected_at", ts, role="env-collector")
+```
+
+> **注意：** `collect_device` 不直接依赖 `DeviceAgent`——由上层 `run_collection()` 函数组合，避免耦合：
+
+```python
+def run_collection(cfg, memory: ProjectMemory) -> dict[str, str]:
+    """顶层编排：收集本机 + 所有 SSH 设备，保存到 memory。"""
+    from rockchip_agents.tools.device_agent import DeviceAgent
+
+    col = EnvCollector()
+    all_facts: dict[str, str] = {}
+
+    # 本机收集
+    local = col.collect_local()
+    all_facts.update(local)
+
+    # SSH 设备收集
+    for name, dev in (cfg.devices or {}).items():
+        if dev.connect != "ssh":
+            continue
+        agent = DeviceAgent(dev)
+
+        def _run(cmd, remote_cwd="", timeout=60, _agent=agent):
+            return _agent.run_command(cmd, remote_cwd or dev.workspace, timeout)
+
+        col._run_on_device = _run
+        device_facts = col.collect_device(name, dev)
+        all_facts.update(device_facts)
+
+    col.save(all_facts, memory)
+    return all_facts
+```
+
+将 `run_collection` 也放入 `env_collector.py`。
+
+**Step 5: 运行测试**
+
+```bash
+pytest tests/test_env_collector.py -v
+# Expected: 6/6 PASSED
+```
+
+**Step 6: Commit**
+
+```bash
+git add src/rockchip_agents/tools/env_collector.py \
+        src/rockchip_agents/tools/project_memory.py \
+        tests/test_env_collector.py
+git commit -m "feat: add EnvCollector core class with local+SSH collection"
+```
+
+---
+
+### Task 2: scripts/run.py — collect-env 子命令
+
+**Files:**
+- Modify: `scripts/run.py`
+- Test: `tests/test_env_collector.py`（扩展 CLI 测试）
+
+**Step 1: 写失败测试**
+
+```python
+# 追加到 tests/test_env_collector.py
+
+def test_run_collection_prints_summary(tmp_path, capsys):
+    """run_collection 完整运行后应返回包含 local/device 信息的 facts dict。"""
+    from rockchip_agents.tools.env_collector import run_collection
+    from rockchip_agents.config import AgentsConfig, ClaudeConfig, SchedulerConfig, FeishuConfig
+
+    cfg = AgentsConfig(
+        projects={},
+        scheduler=SchedulerConfig(),
+        claude=ClaudeConfig(api_key="fake"),
+        feishu=FeishuConfig(),
+        devices={},
+    )
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    with patch.dict(os.environ, {"RKNN_MODEL_PATH": "/data/m"}, clear=False):
+        facts = run_collection(cfg, mem)
+    assert "local.envvar.RKNN_MODEL_PATH" in facts
+    assert "collected_at" in mem.get_facts("__env__")
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+pytest tests/test_env_collector.py::test_run_collection_prints_summary -v
+# Expected: FAIL — run_collection 函数未集成完整
+```
+
+**Step 3: 在 scripts/run.py 添加 collect-env 子命令**
+
+在 `sub.add_parser("projects", ...)` 之后添加：
+
+```python
+sub.add_parser("collect-env", help="收集本机+SSH设备环境信息（模型路径/磁盘/包）")
+```
+
+在 `elif args.cmd == "projects":` 之后添加：
+
+```python
+elif args.cmd == "collect-env":
+    from rockchip_agents.config import load_config
+    from rockchip_agents.tools.env_collector import run_collection
+    from rockchip_agents.tools.project_memory import ProjectMemory
+    cfg = load_config()
+    mem = ProjectMemory()
+    facts = run_collection(cfg, mem)
+
+    # 打印摘要
+    env_vars = {k: v for k, v in facts.items() if k.startswith("local.envvar.")}
+    print(f"\n[本机]")
+    if env_vars:
+        print(f"  ENV: {len(env_vars)} 条含模型关键词变量")
+        for k, v in list(env_vars.items())[:3]:
+            print(f"    {k.split('.')[-1]}={v}")
+    for ext in ("rknn", "onnx", "pt", "pb"):
+        val = facts.get(f"local.models.{ext}", "")
+        if val:
+            cnt = len(val.split(","))
+            print(f"  *.{ext}: {cnt} 个文件")
+
+    for key in sorted(facts):
+        if ".error" in key:
+            name = key.split(".")[0]
+            print(f"\n[{name}]")
+            print(f"  ⚠ {facts[key]}")
+        elif ".disk" in key and not key.startswith("local"):
+            name = key.split(".")[0]
+            host = ""
+            from rockchip_agents.config import load_config as _lc
+            _cfg = _lc()
+            dev = _cfg.devices.get(name)
+            if dev:
+                host = f" @ {dev.host}"
+            print(f"\n[{name}{host}]")
+            rknn_val = facts.get(f"{name}.models.rknn", "")
+            if rknn_val:
+                print(f"  *.rknn: {len(rknn_val.split(','))} 个文件")
+            print(f"  磁盘: {facts[key][:60]}")
+            pkgs = facts.get(f"{name}.packages", "")
+            if pkgs:
+                print(f"  关键包: {pkgs[:80]}")
+
+    ts = facts.get("collected_at", "")
+    print(f"\n已保存到 ProjectMemory (__env__)  [{ts}]")
+```
+
+**Step 4: 运行测试**
+
+```bash
+pytest tests/test_env_collector.py -v
+# Expected: 7/7 PASSED
+```
+
+**Step 5: 手动验证 CLI**
+
+```bash
+python scripts/run.py collect-env
+# Expected: 输出 [本机] ENV: N 条含模型关键词变量 + 文件统计
+```
+
+**Step 6: Commit**
+
+```bash
+git add scripts/run.py tests/test_env_collector.py
+git commit -m "feat: add collect-env CLI subcommand"
+```
+
+---
+
+### Task 3: scheduler.py — 24h 定时 job + 启动后台线程
+
+**Files:**
+- Modify: `src/rockchip_agents/core/scheduler.py`
+- Test: `tests/test_scheduler.py`（已有，追加测试）
+
+**Step 1: 写失败测试**
+
+```python
+# 追加到 tests/test_scheduler.py（或新建 tests/test_env_scheduler.py）
+
+def test_scheduler_has_env_collect_job(monkeypatch):
+    """start() 添加的 APScheduler jobs 应包含 env_collect（id='env_collect'）。"""
+    import types
+    added_jobs = []
+
+    class FakeScheduler:
+        def add_job(self, fn, trigger, **kwargs):
+            added_jobs.append(kwargs.get("id"))
+        def start(self):
+            raise KeyboardInterrupt  # 立即退出
+
+    monkeypatch.setattr(
+        "rockchip_agents.core.scheduler.BlockingScheduler",
+        lambda: FakeScheduler()
+    )
+    # stub 其他副作用
+    monkeypatch.setattr("rockchip_agents.core.scheduler._scan_and_enqueue", lambda: None)
+    monkeypatch.setattr("rockchip_agents.core.scheduler.TaskQueue", MagicMock())
+    monkeypatch.setattr("rockchip_agents.core.scheduler.Watchdog", MagicMock())
+    monkeypatch.setattr("rockchip_agents.core.scheduler.load_config", MagicMock(
+        return_value=MagicMock(scheduler=MagicMock(interval_hours=1), devices={}, queues={})
+    ))
+
+    from rockchip_agents.core.scheduler import start
+    try:
+        start()
+    except (KeyboardInterrupt, Exception):
+        pass
+
+    assert "env_collect" in added_jobs, f"未找到 env_collect job，jobs={added_jobs}"
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+pytest tests/test_scheduler.py::test_scheduler_has_env_collect_job -v
+# Expected: AssertionError — env_collect not in added_jobs
+```
+
+**Step 3: 修改 scheduler.py**
+
+在 `start()` 函数中，导入并添加 `_collect_env` 包装函数（在函数顶部附近添加）：
+
+```python
+def _collect_env() -> None:
+    """后台静默收集环境信息，失败不影响主流程。"""
+    try:
+        from rockchip_agents.config import load_config
+        from rockchip_agents.tools.env_collector import run_collection
+        from rockchip_agents.tools.project_memory import ProjectMemory
+        cfg = load_config()
+        run_collection(cfg, ProjectMemory())
+        logger.info("EnvCollector 完成收集")
+    except Exception:
+        logger.exception("EnvCollector 收集失败（静默跳过）")
+```
+
+在 `start()` 函数内 `scheduler.add_job(_enqueue_market_intel_task, ...)` 之后添加：
+
+```python
+scheduler.add_job(
+    _collect_env, "interval", hours=24, id="env_collect",
+    next_run_time=datetime.now() + timedelta(minutes=30),
+)
+```
+
+在 `_scan_and_enqueue()` 调用（启动时立即扫描一次）之后，启动一个 daemon 后台线程触发首次收集（不阻塞调度器）：
+
+```python
+# 启动时后台触发一次环境收集（daemon，不阻塞）
+import threading
+t = threading.Thread(target=_collect_env, daemon=True, name="env-collector-init")
+t.start()
+```
+
+**Step 4: 运行测试**
+
+```bash
+pytest tests/test_scheduler.py -v
+# Expected: 原有测试 + 新增测试均 PASSED
+```
+
+**Step 5: Commit**
+
+```bash
+git add src/rockchip_agents/core/scheduler.py tests/test_scheduler.py
+git commit -m "feat: add EnvCollector 24h scheduler job + startup daemon thread"
+```
+
+---
+
+### Task 4: developer.py — _build_device_context 注入已知模型摘要
+
+**Files:**
+- Modify: `src/rockchip_agents/agents/developer.py`
+- Test: `tests/test_developer.py`（已有，追加测试）
+
+**Step 1: 写失败测试**
+
+```python
+# 追加到 tests/test_developer.py
+
+def test_build_device_context_injects_env_models(tmp_path):
+    """_build_device_context 在 __env__ 有数据时应注入已知模型摘要。"""
+    from rockchip_agents.config import AgentsConfig, ClaudeConfig, SchedulerConfig, FeishuConfig
+    from rockchip_agents.agents.developer import DeveloperAgent
+    from rockchip_agents.tools.project_memory import ProjectMemory
+
+    cfg = AgentsConfig(
+        projects={},
+        scheduler=SchedulerConfig(),
+        claude=ClaudeConfig(api_key="fake"),
+        feishu=FeishuConfig(),
+        devices={},
+    )
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    mem.set_fact("__env__", "local.models.rknn", "/data/a.rknn,/data/b.rknn", role="env-collector")
+    mem.set_fact("__env__", "rk3588.models.rknn", "/home/pi/model.rknn", role="env-collector")
+    mem.set_fact("__env__", "rk3588.disk", "/dev/root 29G 8G 19G", role="env-collector")
+    mem.set_fact("__env__", "collected_at", "2026-03-09T10:00:00", role="env-collector")
+
+    agent = DeveloperAgent(cfg, memory=mem)
+    ctx = agent._build_device_context()
+
+    assert "已知模型" in ctx
+    assert "/data/a.rknn" in ctx or "2 个" in ctx
+    assert "rk3588" in ctx
+    assert "19G" in ctx or "29G" in ctx
+
+
+def test_build_device_context_no_env_data(tmp_path):
+    """__env__ 无数据时，_build_device_context 不注入模型摘要（不崩溃）。"""
+    from rockchip_agents.config import AgentsConfig, ClaudeConfig, SchedulerConfig, FeishuConfig
+    from rockchip_agents.agents.developer import DeveloperAgent
+    from rockchip_agents.tools.project_memory import ProjectMemory
+
+    cfg = AgentsConfig(
+        projects={}, scheduler=SchedulerConfig(),
+        claude=ClaudeConfig(api_key="fake"), feishu=FeishuConfig(), devices={},
+    )
+    mem = ProjectMemory(db_path=tmp_path / "m.db")
+    agent = DeveloperAgent(cfg, memory=mem)
+    ctx = agent._build_device_context()  # 不应抛异常
+    assert "已知模型" not in ctx
+```
+
+**Step 2: 运行确认失败**
+
+```bash
+pytest tests/test_developer.py::test_build_device_context_injects_env_models \
+       tests/test_developer.py::test_build_device_context_no_env_data -v
+# Expected: FAIL — AssertionError: "已知模型" not in ctx
+```
+
+**Step 3: 修改 developer.py _build_device_context()**
+
+在 `_build_device_context()` 方法末尾 `return "\n".join(lines)` 之前，追加以下代码：
+
+```python
+# 注入已知模型摘要（从 ProjectMemory __env__ 读取）
+if self._memory is not None:
+    env_facts = self._memory.get_facts("__env__")
+    if env_facts:
+        lines.append("【已知模型文件（避免重复下载）】")
+
+        # 本机
+        local_rknn = env_facts.get("local.models.rknn", "")
+        local_onnx = env_facts.get("local.models.onnx", "")
+        local_env_vars = {
+            k.split(".")[-1]: v
+            for k, v in env_facts.items()
+            if k.startswith("local.envvar.")
+        }
+        lines.append("本机:")
+        if local_rknn:
+            lines.append(f"  *.rknn: {local_rknn}")
+        if local_onnx:
+            lines.append(f"  *.onnx: {local_onnx}")
+        for var_name, var_val in local_env_vars.items():
+            lines.append(f"  {var_name}={var_val}（来自环境变量）")
+
+        # 各 SSH 设备
+        device_names = set()
+        for key in env_facts:
+            parts = key.split(".")
+            if len(parts) >= 2 and parts[0] not in ("local", "collected"):
+                device_names.add(parts[0])
+
+        for dev_name in sorted(device_names):
+            dev_cfg = self._config.devices.get(dev_name)
+            host_info = f" ({dev_cfg.host})" if dev_cfg else ""
+            lines.append(f"\n{dev_name}{host_info}:")
+            error = env_facts.get(f"{dev_name}.error", "")
+            if error:
+                lines.append(f"  ⚠ {error}")
+                continue
+            rknn = env_facts.get(f"{dev_name}.models.rknn", "")
+            if rknn:
+                lines.append(f"  *.rknn: {rknn}")
+            disk = env_facts.get(f"{dev_name}.disk", "")
+            if disk:
+                lines.append(f"  磁盘: {disk[:60]}")
+            pkgs = env_facts.get(f"{dev_name}.packages", "")
+            if pkgs:
+                lines.append(f"  已装: {pkgs[:80]}")
+
+        collected_at = env_facts.get("collected_at", "")
+        if collected_at:
+            lines.append(f"\n（更新于 {collected_at}，执行 `python scripts/run.py collect-env` 可刷新）")
+        lines.append("")
+```
+
+**Step 4: 运行测试**
+
+```bash
+pytest tests/test_developer.py -v
+# Expected: 全部 PASSED（含两个新测试）
+```
+
+**Step 5: Commit**
+
+```bash
+git add src/rockchip_agents/agents/developer.py tests/test_developer.py
+git commit -m "feat: inject known model summary into DeveloperAgent device context"
+```
+
+---
+
+### Task 5: 全量验证
+
+**Step 1: 运行全量测试**
+
+```bash
+pytest tests/ -q
+# Expected: 全部通过（含 test_env_collector.py 的 8+ 个测试）
+```
+
+**Step 2: 检查测试覆盖关键路径**
+
+确认以下场景有测试：
+- [ ] `collect_local` 过滤环境变量关键词
+- [ ] `collect_local` 递归扫描 + 跳过 venv
+- [ ] `collect_device` SSH 成功（mock）
+- [ ] `collect_device` SSH 离线 → error key
+- [ ] `save()` 写入 ProjectMemory
+- [ ] `env-collector` WRITE_ROLES 权限
+- [ ] scheduler 有 `env_collect` job
+- [ ] `_build_device_context` 注入模型摘要
+- [ ] `_build_device_context` 无 __env__ 数据时静默
+
+**Step 3: 手动冒烟测试 CLI**
+
+```bash
+# 应能正常运行并打印摘要
+python scripts/run.py collect-env
+
+# 验证数据写入
+python scripts/run.py facts
+# Expected: 显示 __env__ 下的 local.models.* / local.envvar.* 等条目
+```
+
+**Step 4: Commit**
+
+```bash
+git add .
+git commit -m "test: verify EnvCollector full integration"
+```
+
+---
+
+## 实施注意事项
+
+1. **`_run_on_device` 占位符**：`EnvCollector.collect_device()` 内部调用 `self._run_on_device()`，这个方法在 `EnvCollector` 类中 raise `NotImplementedError`。`run_collection()` 函数在调用前通过 monkeypatching（`col._run_on_device = _run`）注入真实的 `DeviceAgent.run_command`。测试同样用此机制 mock。
+
+2. **WRITE_ROLES 扩展**：`project_memory.py` 中 `WRITE_ROLES` 需添加 `"env-collector"`，否则 `set_fact()` 会抛 `PermissionError`。
+
+3. **`_memory` 字段检查**：`DeveloperAgent` 的 `__init__` 接收 `memory: ProjectMemory | None = None`，注入代码需先检查 `if self._memory is not None`。
+
+4. **scheduler import**：`_collect_env()` 函数内部延迟 import（`from rockchip_agents.tools.env_collector import run_collection`），避免在模块加载时引入不必要依赖。
+
+5. **设备 workspace 容错**：`find` 命令的 `remote_cwd` 用 `dev.workspace`，若设备 workspace 不存在，`find` 会返回错误但不影响其他命令执行。
@@ -0,0 +1,170 @@
+# Visual Tester 设计文档
+
+## 概述
+
+基于 Playwright 截图对比的 UI 视觉回归测试系统，支持多级断点恢复，集成至 Dashboard 看板，自动发现 `/data/company` 下所有 Web UI。
+
+---
+
+## 1. 目录结构 & /data/company 迁移
+
+### /data/company 目录布局
+
+```
+/data/company/
+├── visual-tester/                  # 新建独立服务
+│   ├── src/visual_tester/
+│   │   ├── runner.py              # Playwright 执行引擎 + checkpoint
+│   │   ├── differ.py              # PIL/pixelmatch 差异计算
+│   │   ├── discovery.py           # 自动发现 Web UI
+│   │   └── db.py                  # SQLite 数据层
+│   ├── baselines/                 # {project}/{name}.png 基线截图
+│   ├── diffs/                     # {run_id}/{case_id}_diff.png 差异图
+│   ├── checkpoints/               # {run_id}.json 会话恢复文件
+│   ├── data/visual.db
+│   ├── requirements.txt
+│   └── README.md
+├── agents/   → symlink /data/rockchip/agents
+├── yolo/     → symlink /data/rockchip/yolo
+└── ...（其他项目符号链接）
+```
+
+### configs/projects.yaml 迁移
+
+- `scan_dirs` 从 `/data/rockchip` 更改为 `/data/company`
+- 删除旧的显式 `projects:` 条目（由 scan_dirs 自动发现替代）
+- `data/tasks.db` 历史数据保留（不清除），仅新扫描基于新路径
+
+---
+
+## 2. 数据模型 & 断点恢复机制
+
+### SQLite 表结构（visual.db）
+
+```sql
+CREATE TABLE test_runs (
+    id TEXT PRIMARY KEY,
+    project TEXT NOT NULL,
+    started_at TEXT NOT NULL,
+    completed_at TEXT,
+    status TEXT NOT NULL,   -- running | done | interrupted
+    total INTEGER DEFAULT 0,
+    passed INTEGER DEFAULT 0,
+    failed INTEGER DEFAULT 0,
+    pending_review INTEGER DEFAULT 0
+);
+
+CREATE TABLE test_cases (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    run_id TEXT NOT NULL,
+    project TEXT NOT NULL,
+    url TEXT NOT NULL,
+    name TEXT NOT NULL,
+    status TEXT NOT NULL,   -- pending | running | passed | failed | needs_review | skipped
+    screenshot_path TEXT,
+    baseline_path TEXT,
+    diff_path TEXT,
+    diff_pct REAL,
+    FOREIGN KEY (run_id) REFERENCES test_runs(id)
+);
+
+CREATE TABLE baselines (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    project TEXT NOT NULL,
+    url TEXT NOT NULL,
+    name TEXT NOT NULL,
+    path TEXT NOT NULL,
+    commit_hash TEXT DEFAULT '',
+    created_at TEXT NOT NULL,
+    is_current INTEGER DEFAULT 1
+);
+
+CREATE TABLE approvals (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    case_id INTEGER NOT NULL,
+    action TEXT NOT NULL,       -- approve | reject
+    reason TEXT DEFAULT '',
+    approved_by TEXT DEFAULT 'human',
+    approved_at TEXT NOT NULL,
+    FOREIGN KEY (case_id) REFERENCES test_cases(id)
+);
+```
+
+### 四级断点恢复
+
+| 级别 | 触发场景 | 恢复策略 |
+|------|---------|---------|
+| **测试套件恢复** | 进程崩溃 / 手动中断 | 跳过 status≠pending 的 case，重置 running→pending |
+| **基线版本管理** | 误操作覆盖基线 | `baselines.is_current=0` 软删除，可随时切回 |
+| **差异审批流** | 截图差异需人工确认 | status=needs_review 暂停队列，Dashboard 审批后继续 |
+| **Claude 会话恢复** | 断网 / 关机 / token 耗尽 | `checkpoints/{run_id}.json` 记录 last_case_id + progress |
+
+Checkpoint JSON 格式：
+```json
+{
+  "run_id": "abc123",
+  "last_case_id": 42,
+  "progress": "15/30",
+  "status": "interrupted",
+  "claude_session_id": "sess_xxx",
+  "paused_at": "2026-03-09T10:30:00"
+}
+```
+
+---
+
+## 3. 自动发现 & Runner 工作流 & Dashboard 集成
+
+### 自动发现优先级（discovery.py）
+
+1. `CLAUDE.md` 中 `## 服务端口` / `## Web UI` 段落提取 URL
+2. `README.md` 中 `http://localhost:XXXX` 正则扫描
+3. `docker-compose.yml` 中 `ports: ["XXXX:XXXX"]` 提取宿主机端口
+
+发现结果缓存至 `visual.db`，支持手动覆盖。
+
+### Runner 工作流（runner.py）
+
+```
+runner.run(project, resume=run_id)
+    → 加载/创建 test_run
+    → 筛选 pending cases（断点恢复跳过已完成）
+    → for each case:
+          screenshot via Playwright → differ.compare()
+          diff_pct < threshold → passed
+          diff_pct > threshold AND has baseline → needs_review（暂停等待审批）
+          no baseline → auto-set as baseline (passed)
+          timeout/error → failed
+          写 checkpoint
+    → 更新 run status → done / interrupted
+```
+
+差异阈值：默认 `diff_threshold=0.5%`，可按项目配置。
+
+### Dashboard 新增 Visual 标签页
+
+**新增 API 端点：**
+
+```
+GET  /api/visual/runs                    # 最近 20 条运行记录
+GET  /api/visual/runs/{run_id}/cases     # 该次运行的所有 case
+POST /api/visual/runs                    # 触发新一轮测试（异步）
+POST /api/visual/cases/{id}/approve      # 批准差异 → 更新基线，status=passed
+POST /api/visual/cases/{id}/reject       # 拒绝差异 → status=failed，保留旧基线
+GET  /api/visual/baselines/{project}     # 列出该项目所有基线（含历史）
+POST /api/visual/runs/{run_id}/resume    # 从断点继续
+```
+
+**前端视图：**
+- 运行列表：项目 / 时间 / 进度条 / 状态徽章
+- Case 列表：三栏对比（基线 | 当前 | 差异）+ 审批按钮
+- 需审批的 case 在顶部 "⚠ N 个截图待审批" 聚合
+
+---
+
+## 实施原则
+
+- **TDD**：先写失败测试，再实现
+- **YAGNI**：不提前实现 AI 差异分析、多分辨率适配等
+- **断点即一等公民**：每个 case 完成后立即 flush checkpoint
+- **轻依赖**：PIL（pillow）+ Playwright，无额外服务依赖
@@ -0,0 +1,7 @@
+fastapi>=0.110.0
+uvicorn[standard]>=0.27.0
+websockets>=12.0
+paramiko>=3.0.0
+pyyaml>=6.0
+pillow>=10.0
+numpy>=1.24
@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Rockchip Agents Dashboard 独立服务"""
+"""NMFS Agents Dashboard 独立服务"""
 from __future__ import annotations

 import subprocess, sys as _sys
@@ -70,7 +70,7 @@ def _watcher(app_state, loop: asyncio.AbstractEventLoop) -> None:


 def main() -> None:
-    from rockchip_agents.dashboard.api import make_app
+    from nmfs_agents.dashboard.api import make_app

    app = make_app(db_path=DB_PATH, config_dir=CONFIG_DIR,
                   global_claude_dir=GLOBAL_CLAUDE_DIR)
@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Rockchip Agents 启动入口"""
+"""NMFS Agents 启动入口"""
 from __future__ import annotations

 import argparse
@@ -9,7 +9,7 @@ logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name


 def main() -> None:
-    parser = argparse.ArgumentParser(description="Rockchip Agents")
+    parser = argparse.ArgumentParser(description="NMFS Agents")
    sub = parser.add_subparsers(dest="cmd")
    sub.add_parser("scan",     help="扫描所有项目，打印发现的任务（不执行）")
    sub.add_parser("run",      help="扫描 → 入队 → 执行所有任务（一键完成）")
@@ -33,8 +33,8 @@ def main() -> None:
    args = parser.parse_args()

    if args.cmd == "scan":
-        from rockchip_agents.core.scanner import Scanner
-        from rockchip_agents.config import load_config
+        from nmfs_agents.core.scanner import Scanner
+        from nmfs_agents.config import load_config
        cfg = load_config()
        tasks = Scanner(cfg).scan_all()
        if not tasks:
@@ -44,23 +44,23 @@ def main() -> None:

    elif args.cmd == "run":
        import asyncio
-        from rockchip_agents.agents.manager import ManagerAgent
-        from rockchip_agents.config import load_config
+        from nmfs_agents.agents.manager import ManagerAgent
+        from nmfs_agents.config import load_config
        cfg = load_config()
        stats = asyncio.run(ManagerAgent(cfg).run_cycle())
        print(f"扫描入队: {stats['enqueued']} 个，执行完成: {stats['executed']} 个")

    elif args.cmd == "run-next":
        import asyncio
-        from rockchip_agents.core.executor import Executor
-        from rockchip_agents.config import load_config
+        from nmfs_agents.core.executor import Executor
+        from nmfs_agents.config import load_config
        cfg = load_config()
        ran = asyncio.run(Executor(cfg).run_next())
        print("已执行" if ran else "队列为空")

    elif args.cmd == "status":
-        from rockchip_agents.agents.manager import ManagerAgent
-        from rockchip_agents.config import load_config
+        from nmfs_agents.agents.manager import ManagerAgent
+        from nmfs_agents.config import load_config
        cfg = load_config()
        status = ManagerAgent(cfg).get_status()
        if not status:
@@ -70,12 +70,12 @@ def main() -> None:
                print(f"  {state}: {count}")

    elif args.cmd == "start":
-        from rockchip_agents.core.scheduler import start
+        from nmfs_agents.core.scheduler import start
        start()

    elif args.cmd == "arch-versions":
-        from rockchip_agents.config import load_config
-        from rockchip_agents.tools.arch_version import ArchVersion
+        from nmfs_agents.config import load_config
+        from nmfs_agents.tools.arch_version import ArchVersion
        cfg = load_config()
        proj_cfg = cfg.projects.get(args.project)
        if not proj_cfg:
@@ -93,8 +93,8 @@ def main() -> None:
                    print(f"  {v.version_id}{mark}  [{v.status:12s}] [{passed}]  {v.created_at[:16]}  {v.description}")

    elif args.cmd == "arch-restore":
-        from rockchip_agents.config import load_config
-        from rockchip_agents.tools.arch_version import ArchVersion
+        from nmfs_agents.config import load_config
+        from nmfs_agents.tools.arch_version import ArchVersion
        cfg = load_config()
        proj_cfg = cfg.projects.get(args.project)
        if not proj_cfg:
@@ -105,7 +105,7 @@ def main() -> None:
            print(f"已将 {args.project} 恢复到 {args.version_id}")

    elif args.cmd == "goals":
-        from rockchip_agents.tools.project_memory import ProjectMemory
+        from nmfs_agents.tools.project_memory import ProjectMemory
        goals = ProjectMemory().list_goals()
        if not goals:
            print("暂无项目目标记录（运行 productizer 或 architect 任务后自动写入）")
@@ -115,8 +115,8 @@ def main() -> None:
                print(f"  {g.goal_text}")

    elif args.cmd == "facts":
-        from rockchip_agents.tools.project_memory import ProjectMemory
-        from rockchip_agents.config import load_config
+        from nmfs_agents.tools.project_memory import ProjectMemory
+        from nmfs_agents.config import load_config
        mem = ProjectMemory()
        cfg = load_config()
        any_found = False
@@ -131,13 +131,13 @@ def main() -> None:
            print("暂无项目事实记录（运行 architect 任务后自动写入）")

    elif args.cmd == "set-fact":
-        from rockchip_agents.tools.project_memory import ProjectMemory
+        from nmfs_agents.tools.project_memory import ProjectMemory
        mem = ProjectMemory()
        mem.set_fact(args.project, args.key, args.value, role="developer")
        print(f"已写入: [{args.project}] {args.key} = {args.value}")

    elif args.cmd == "projects":
-        from rockchip_agents.config import load_config
+        from nmfs_agents.config import load_config
        cfg = load_config()
        print(f"{'项目名':20s}  {'类型':8s}  {'模式':8s}  路径")
        print("-" * 70)
@@ -146,9 +146,9 @@ def main() -> None:
            print(f"{name:20s}  {p.kind:8s}  {p.mode:8s}  {p.path}{dw}")

    elif args.cmd == "collect-env":
-        from rockchip_agents.config import load_config
-        from rockchip_agents.tools.env_collector import run_collection
-        from rockchip_agents.tools.project_memory import ProjectMemory
+        from nmfs_agents.config import load_config
+        from nmfs_agents.tools.env_collector import run_collection
+        from nmfs_agents.tools.project_memory import ProjectMemory
        cfg = load_config()
        mem = ProjectMemory()
        facts = run_collection(cfg, mem)
@@ -196,8 +196,8 @@ def main() -> None:
        print(f"\n已保存到 ProjectMemory (__env__)  [{ts}]")

    elif args.cmd == "research":
-        from rockchip_agents.config import load_config
-        from rockchip_agents.core.queue import TaskQueue, Task
+        from nmfs_agents.config import load_config
+        from nmfs_agents.core.queue import TaskQueue, Task
        cfg = load_config()
        q = TaskQueue()
        research_tasks = [
@@ -0,0 +1,359 @@
+"""
+edge-platform 全链路任务种子脚本。
+
+向三个队列注入初始任务，覆盖所有 Agent 角色，建立完整的前后依赖上下文。
+
+使用方法：
+    source venv/bin/activate
+    python scripts/seed_edge_platform.py [--dry-run]
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+PROJECT = "edge-platform"
+PLATFORM_DESC = "边端 AI 推理服务平台（RK3588 NPU + RTP 流媒体），目标：FPS 60+，延迟 <80ms"
+
+# ── 初始项目目标（productizer 写入，其他 agent 读取）────────────────────────
+INITIAL_GOALS = {
+    "productizer": "边端 AI 推理服务平台达到生产可用：内存稳定、NPU三核并行、RTP延迟<80ms",
+    "architect": "整合 YOLO/MPP/RTP 技术栈，实现从摄像头到客户端的完整推理链路",
+}
+
+# ── 初始知识库（agent 可直接读取）───────────────────────────────────────────
+INITIAL_FACTS = {
+    "fps_baseline": "28",
+    "rtp_latency_baseline_ms": "180",
+    "npu_core_mode": "NPU_CORE_0（单核，需改为 NPU_CORE_ALL）",
+    "known_bugs": "内存泄漏,错误处理缺失,模型热加载缺失,RTP延迟高",
+    "test_cmd": "pytest tests/ -v",
+    "model_path": "/data/models/yolo11n.rknn",
+}
+
+
+def seed_memory(mem: ProjectMemory) -> None:
+    """写入初始项目目标和知识库，为 agent 提供起点上下文。"""
+    for role, goal in INITIAL_GOALS.items():
+        try:
+            mem.set_goal(PROJECT, goal, role=role)
+        except Exception:
+            pass  # 已存在则跳过
+    for key, val in INITIAL_FACTS.items():
+        try:
+            mem.set_fact(PROJECT, key, val, role="productizer")
+        except Exception:
+            pass
+
+
+def build_tasks() -> list[tuple[str, list[Task]]]:
+    """构建所有队列的初始任务列表。
+
+    返回 [(queue_name, [Task, ...]), ...]
+    """
+    # ── project_delivery 队列任务 ──────────────────────────────────────────
+    delivery_tasks = [
+        # 1. Productizer 先做成熟度评估，设定长期目标
+        Task(
+            project=PROJECT,
+            type="productize",
+            title="初始成熟度评估：建立产品基线",
+            context=(
+                f"{PLATFORM_DESC}\n\n"
+                "请评估当前项目成熟度（0-100分），重点关注：\n"
+                "- 功能完整性：推理 API/RTP 推流/错误处理\n"
+                "- 性能基线：FPS 28（目标 60），延迟 180ms（目标 80ms）\n"
+                "- 测试覆盖：test_api.py / test_inference.py / test_streamer.py\n"
+                "输出 [记忆] edge-platform: maturity_score=N，设定产品目标。"
+            ),
+            priority=1,
+            mode="report",
+            agent_role="productizer",
+            initiator="seed-script",
+        ),
+
+        # 2. Architect 做跨项目分析，派发开发任务
+        Task(
+            project="ALL",
+            type="architect",
+            title="全局架构分析：edge-platform 与现有项目整合",
+            context=(
+                "重点分析新加入的 edge-platform 项目与现有项目的整合机会：\n"
+                "- edge-platform（推理服务）+ yolo（检测模型）的集成路径\n"
+                "- edge-platform + embedding（向量检索）的语义搜索扩展\n"
+                "- edge-platform + claude-wx（飞书通知）的告警接入\n"
+                "- RTP 延迟问题与 os-base 调研的关联\n\n"
+                "必须输出以下任务：\n"
+                "[开发] edge-platform: 修复推理引擎内存泄漏——将模型加载移至 __init__，添加 release()\n"
+                "[开发] edge-platform: 修复 NPU 单核问题——改 NPU_CORE_ALL 并测试 FPS\n"
+                "[开发] edge-platform: 添加 /infer 错误处理——try/except base64/reshape 异常返回 400\n"
+                "[测试] edge-platform: 运行全量测试套件并报告失败的 test_api.py 断言\n"
+                "[记忆] edge-platform: arch_integration=yolo+embedding+claude-wx 三路整合方案已分析"
+            ),
+            priority=2,
+            mode="report",
+            agent_role="architect",
+            initiator="seed-script",
+        ),
+
+        # 3. Developer 修复 P0 内存泄漏（最高优先级）
+        Task(
+            project=PROJECT,
+            type="fix_bug",
+            title="[修复] P0 内存泄漏：推理引擎模型重复加载",
+            context=(
+                "文件：src/inference.py\n\n"
+                "问题：RknnInferenceEngine.infer() 每次调用都执行 _load_model()，"
+                "导致 RKNN 对象未释放，长时运行内存持续增长（500次请求 +200MB）。\n\n"
+                "修复方案：\n"
+                "1. 在 __init__ 中调用 _load_model() 并保存到 self._rknn\n"
+                "2. infer() 直接使用 self._rknn\n"
+                "3. 添加 __del__ 或 close() 方法释放 rknn 对象\n"
+                "4. 运行 tests/test_inference.py::test_memory_no_growth_over_requests 验证\n\n"
+                "完成后输出 [SPAWN:tester] 验证: 内存泄漏修复验证，运行 test_memory_no_growth"
+            ),
+            priority=1,
+            mode="auto",
+            agent_role="developer",
+            initiator="seed-script",
+            discussion=f"来自 seed-script 初始化：{PLATFORM_DESC}，P0 内存泄漏需优先修复",
+        ),
+
+        # 4. Developer 修复 NPU 单核问题
+        Task(
+            project=PROJECT,
+            type="fix_bug",
+            title="[修复] P1 NPU 单核 → 三核并行，目标 FPS 60+",
+            context=(
+                "文件：src/inference.py\n\n"
+                "问题：init_runtime(core_mask=RKNNLite.NPU_CORE_0) 只使用单核，"
+                "三核空闲，FPS 仅 28（目标 60+）。\n\n"
+                "修复方案：\n"
+                "1. 改为 core_mask=RKNNLite.NPU_CORE_ALL\n"
+                "2. 在 infer() 末尾记录 fps 并更新 [记忆] edge-platform: fps_actual=N\n"
+                "3. 更新 FPS_TARGET=60 断言\n\n"
+                "完成后输出 [SPAWN:system-tester] 验证: 板端 FPS 回归测试，验证 NPU_CORE_ALL 效果"
+            ),
+            priority=2,
+            mode="auto",
+            agent_role="developer",
+            initiator="seed-script",
+        ),
+
+        # 5. Developer 修复错误处理
+        Task(
+            project=PROJECT,
+            type="fix_bug",
+            title="[修复] P1 /infer 端点错误处理——base64/reshape 异常返回 400",
+            context=(
+                "文件：src/server.py\n\n"
+                "问题：/infer 端点无 try/except，无效 base64 或尺寸不符直接 500。\n\n"
+                "修复方案：\n"
+                "1. 用 try/except 包装 base64.b64decode，抛出 HTTPException(400)\n"
+                "2. 用 try/except 包装 np.reshape，抛出 HTTPException(400)\n"
+                "3. 在 infer() 调用处捕获推理异常，返回 HTTPException(422)\n"
+                "4. 运行 tests/test_api.py 所有用例（含 test_infer_invalid_base64_returns_error）\n\n"
+                "完成后输出 [SPAWN:tester] 验证: 错误处理修复验证，运行 test_api.py"
+            ),
+            priority=2,
+            mode="auto",
+            agent_role="developer",
+            initiator="seed-script",
+        ),
+
+        # 6. System Tester 全系统测试（含视觉推理验证）
+        Task(
+            project=PROJECT,
+            type="test",
+            title="全系统测试：功能+性能+板端验证",
+            context=(
+                f"{PLATFORM_DESC}\n\n"
+                "测试清单：\n"
+                "1. 功能测试：pytest tests/ -v（本机），重点关注失败用例\n"
+                "2. 性能测试：连续调用 /infer 100 次，统计 avg_fps 和 p95 延迟\n"
+                "3. 板端测试（可选）：ssh pi@192.168.123.181 运行 pytest tests/\n"
+                "4. 内存稳定性：验证 500 次请求后内存无增长\n\n"
+                "发现问题请输出 [需求] edge-platform: <具体问题描述>\n"
+                "性能基线请输出 [记忆] edge-platform: fps_system_test=N，latency_p95_ms=N"
+            ),
+            priority=3,
+            mode="auto",
+            agent_role="system-tester",
+            initiator="seed-script",
+        ),
+    ]
+
+    # ── insight 队列任务 ───────────────────────────────────────────────────
+    insight_tasks = [
+        # Planner 制定市场洞察计划
+        Task(
+            project=PROJECT,
+            type="plan",
+            title="边端 AI 推理服务市场洞察规划",
+            context=(
+                f"{PLATFORM_DESC}\n\n"
+                "请制定市场洞察调研计划：\n"
+                "- 边端 AI 推理服务（Edge Inference as a Service）市场规模\n"
+                "- RK3588 平台在工业/安防/运动分析场景的竞争格局\n"
+                "- 实时视频推理服务的付费意愿调研\n"
+                "输出结构化调研任务给 market-pm 执行。"
+            ),
+            priority=3,
+            mode="auto",
+            agent_role="planner",
+            initiator="seed-script",
+        ),
+
+        # Market PM 分析市场机会
+        Task(
+            project=PROJECT,
+            type="market_research",
+            title="边端 AI 推理服务市场分析：运动/工业/安防三方向",
+            context=(
+                f"{PLATFORM_DESC}\n\n"
+                "请分析以下三个目标市场的机会：\n\n"
+                "1. 【运动】实时动作识别（跑步/球类）+ 教练反馈，B 端体育机构\n"
+                "   - 竞品：Camera IQ、Hudl，RK3588 本地化的差异点\n"
+                "   - 付费场景：赛事直播增强、训练质量评估\n\n"
+                "2. 【工业/安防】边端目标检测，私有部署无需云端\n"
+                "   - 客户痛点：数据隐私、网络依赖、云计算费用\n"
+                "   - 定价逻辑：一次性硬件+软件授权 vs 订阅\n\n"
+                "3. 【AI 安全/隐私】端侧推理保证数据不出设备\n"
+                "   - 与 info-privacy 项目的结合点\n\n"
+                "输出格式：\n"
+                "[调研报告] __market__.运动.opportunity=<市场机会摘要>\n"
+                "[调研报告] __market__.工业.opportunity=<市场机会摘要>\n"
+                "[调研报告] __market__.AI安全.opportunity=<市场机会摘要>"
+            ),
+            priority=3,
+            mode="auto",
+            agent_role="market-pm",
+            initiator="seed-script",
+        ),
+    ]
+
+    # ── base_opt 队列任务 ──────────────────────────────────────────────────
+    base_opt_tasks = [
+        # RTP 调研（直接关联 edge-platform 的 RTP 延迟问题）
+        Task(
+            project="rtp-research",
+            type="research",
+            title="RTP 延迟优化调研：针对 edge-platform MPP+RTP 瓶颈",
+            context=(
+                "关联项目：edge-platform（src/streamer.py）\n"
+                "当前瓶颈：MPP 编码器每帧重新初始化，端到端延迟 ~180ms，目标 <80ms。\n\n"
+                "请重点调研：\n"
+                "1. MPP 编码器会话保持（mpp_enc_cfg_init/mpp_enc_ref_cfg_init 复用）\n"
+                "2. RTP 包分片策略（MTU=1500 下 H264 NAL 单元分割）\n"
+                "3. SO_PRIORITY + SO_SNDBUF 对 UDP 延迟的影响\n"
+                "4. 零拷贝 DMA buffer → RTP payload 路径\n\n"
+                "输出 [优化任务] os-engineer: <具体 sysctl/MPP API 优化步骤>"
+            ),
+            priority=2,
+            mode="auto",
+            agent_role="rtp-researcher",
+            initiator="seed-script",
+            discussion=f"edge-platform RTP 延迟 180ms，超过目标 80ms",
+        ),
+
+        # Net 调研（API 服务网络调优）
+        Task(
+            project="net-research",
+            type="research",
+            title="网络栈调优调研：edge-platform 推理 API 响应时间优化",
+            context=(
+                "关联项目：edge-platform（src/server.py，FastAPI on uvicorn）\n"
+                "当前问题：高并发下 API 响应时间不稳定（p99 > 200ms）。\n\n"
+                "请调研：\n"
+                "1. uvicorn worker 数量与 RK3588 CPU 核心绑定策略\n"
+                "2. TCP_NODELAY + SO_REUSEPORT 对 FastAPI 延迟的影响\n"
+                "3. 推理服务与 RTP 推流共用网卡时的 QoS 优先级划分\n"
+                "4. 内核 net.core.somaxconn / net.ipv4.tcp_tw_reuse 参数建议\n\n"
+                "输出具体 sysctl 命令和 uvicorn 启动参数建议\n"
+                "输出 [优化任务] os-engineer: <网络参数配置步骤>"
+            ),
+            priority=2,
+            mode="auto",
+            agent_role="net-researcher",
+            initiator="seed-script",
+        ),
+
+        # Kernel 调研（NPU + API 并发调度）
+        Task(
+            project="kernel-analysis",
+            type="research",
+            title="内核调度调研：NPU 推理与 RTP 推流的资源竞争",
+            context=(
+                "关联项目：edge-platform（推理 API + RTP 推流并发运行）\n"
+                "当前问题：NPU 推理（大核 A76）与 RTP 编码（MPP/VPU）争抢内存带宽，"
+                "高负载下 FPS 从 28 降至 15。\n\n"
+                "请调研：\n"
+                "1. RK3588 NPU/VPU/ISP 内存带宽分配（IOMMU/CMA 配置）\n"
+                "2. isolcpus=4-7（A76 大核）专用于推理进程\n"
+                "3. SCHED_FIFO 实时优先级给 RTP 推流线程\n"
+                "4. CPUFreq governor=performance 对 NPU 推理 FPS 的影响\n\n"
+                "输出 [优化任务] dev-kernel: <内核参数/Kconfig 建议>"
+            ),
+            priority=2,
+            mode="auto",
+            agent_role="kernel-analyzer",
+            initiator="seed-script",
+        ),
+    ]
+
+    return [
+        ("project_delivery", delivery_tasks),
+        ("insight", insight_tasks),
+        ("base_opt", base_opt_tasks),
+    ]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="seed edge-platform tasks")
+    parser.add_argument("--dry-run", action="store_true", help="只显示，不写入")
+    args = parser.parse_args()
+
+    # 初始化 memory
+    mem = ProjectMemory()
+
+    # 构建并注入初始队列任务
+    queue_configs = [
+        ("project_delivery", "data/tasks.db"),
+        ("insight", "data/insight_tasks.db"),
+        ("base_opt", "data/base_opt_tasks.db"),
+    ]
+    task_map = dict(build_tasks())
+
+    total = 0
+    for queue_name, db_path in queue_configs:
+        tasks = task_map.get(queue_name, [])
+        if not tasks:
+            continue
+        print(f"\n[{queue_name}] ({db_path}) — {len(tasks)} 个任务：")
+        for t in tasks:
+            print(f"  [{t.agent_role:16s}] {t.title[:60]}")
+        if not args.dry_run:
+            q = TaskQueue(db_path=Path(db_path))
+            for t in tasks:
+                q.enqueue(t)
+            total += len(tasks)
+
+    if not args.dry_run:
+        # 写入初始 memory
+        seed_memory(mem)
+        print(f"\n✓ 已写入 {total} 个初始任务")
+        print("✓ 已写入 edge-platform 项目目标和知识库")
+        print("\n启动调度器执行：bash scripts/start_daemon.sh")
+        print("监控进度：    bash scripts/watch_daemon.sh")
+    else:
+        print(f"\n[dry-run] 共 {sum(len(v) for v in task_map.values())} 个任务，未写入")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+# NMFS Agents 后台守护启动脚本
+# 用法：bash scripts/start_daemon.sh [--hours 24]
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(dirname "$SCRIPT_DIR")"
+LOG_DIR="$PROJECT_ROOT/logs"
+PID_FILE="$PROJECT_ROOT/logs/daemon.pid"
+LOG_FILE="$LOG_DIR/daemon_$(date '+%Y%m%d_%H%M%S').log"
+
+mkdir -p "$LOG_DIR"
+
+# ── 检查是否已在运行 ────────────────────────────────────────────
+if [[ -f "$PID_FILE" ]]; then
+    OLD_PID=$(cat "$PID_FILE")
+    if kill -0 "$OLD_PID" 2>/dev/null; then
+        echo "⚠️  守护进程已在运行（PID=$OLD_PID），如需重启请先执行："
+        echo "    kill $OLD_PID && rm $PID_FILE"
+        exit 1
+    else
+        echo "旧 PID 文件残留，已清理"
+        rm -f "$PID_FILE"
+    fi
+fi
+
+# ── 激活虚拟环境 ────────────────────────────────────────────────
+VENV="$PROJECT_ROOT/venv"
+if [[ ! -f "$VENV/bin/activate" ]]; then
+    echo "❌ 未找到虚拟环境: $VENV/bin/activate"
+    echo "请先执行: python3.10 -m venv venv && pip install -r requirements.txt"
+    exit 1
+fi
+source "$VENV/bin/activate"
+
+# ── 启动调度器（后台 nohup）────────────────────────────────────
+echo "🚀 启动 NMFS Agents 守护进程"
+echo "   日志文件: $LOG_FILE"
+echo "   PID 文件: $PID_FILE"
+
+nohup python "$PROJECT_ROOT/scripts/run.py" start \
+    >> "$LOG_FILE" 2>&1 &
+
+DAEMON_PID=$!
+echo "$DAEMON_PID" > "$PID_FILE"
+echo "✅ 守护进程已启动（PID=$DAEMON_PID）"
+echo ""
+echo "查看日志:    tail -f $LOG_FILE"
+echo "查看状态:    $VENV/bin/python $PROJECT_ROOT/scripts/run.py status"
+echo "查看目标:    $VENV/bin/python $PROJECT_ROOT/scripts/run.py goals"
+echo "停止进程:    kill $DAEMON_PID"
@@ -0,0 +1,400 @@
+#!/usr/bin/env python3
+"""
+24小时项目自动化运行测试 - 基于 Playwright 浏览器自动化
+
+用法：
+    python scripts/test_24h.py --hours 24 --project /data/rockchip/logodetect-x86
+    python scripts/test_24h.py --hours 1 --project /data/rockchip/logodetect-x86 --dry-run
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import time
+import urllib.request
+from datetime import datetime, timedelta
+from pathlib import Path
+
+
+DASHBOARD_URL = "http://localhost:9080"
+REPORT_DIR = Path("tmp/test_24h_report")
+CHECK_INTERVAL = 900  # 每 15 分钟一次
+
+
+# ── 辅助函数 ─────────────────────────────────────────────────────────
+
+def _api(method: str, path: str, body: dict | None = None) -> dict:
+    url = DASHBOARD_URL + path
+    data = json.dumps(body).encode() if body else None
+    headers = {"Content-Type": "application/json"} if data else {}
+    req = urllib.request.Request(url, data=data, headers=headers, method=method)
+    with urllib.request.urlopen(req, timeout=10) as r:
+        return json.loads(r.read())
+
+
+def _get_project_tasks(project_name: str) -> list[dict]:
+    return _api("GET", f"/api/tasks?project={project_name}")
+
+
+def _task_stats(tasks: list[dict]) -> dict:
+    statuses = [t["status"] for t in tasks]
+    return {
+        "total":    len(tasks),
+        "pending":  statuses.count("pending"),
+        "running":  statuses.count("running"),
+        "done":     statuses.count("done"),
+        "failed":   statuses.count("failed"),
+        "vetoed":   statuses.count("vetoed"),
+        "waiting":  statuses.count("waiting_approval"),
+    }
+
+
+# ── Phase 1: 浏览器 UI 操作（Playwright）─────────────────────────────
+
+def setup_via_browser(
+    page,  # playwright Page
+    project_path: str,
+    linked_project: str,
+    project_name: str,
+) -> dict:
+    """通过浏览器 UI 完整设置项目（导入 + 生成任务 + 启动执行器）"""
+    from playwright.sync_api import expect
+
+    print(f"[setup] 打开 Dashboard: {DASHBOARD_URL}")
+    page.goto(DASHBOARD_URL)
+    page.wait_for_timeout(2000)
+
+    # 截图：初始状态
+    _screenshot(page, "00_initial")
+
+    # 点击「导入项目」
+    print("[setup] 点击「导入项目」")
+    page.click("button:has-text('导入项目')")
+    page.wait_for_timeout(500)
+
+    # 填写路径（触发项目名自动推断）
+    print(f"[setup] 填写路径: {project_path}")
+    # 第 1 个 input = 目录路径
+    page.locator("input").nth(0).fill(project_path)
+    page.wait_for_timeout(500)
+
+    # 填写关联项目（第 3 个 input，placeholder 含 logo-collect）
+    if linked_project:
+        print(f"[setup] 填写关联项目: {linked_project}")
+        page.locator("input").nth(2).fill(linked_project)
+
+    # 确认 mode = auto
+    mode_select = page.locator("select")
+    mode_select.select_option("auto")
+
+    # 点击「预览任务」展开模板
+    try:
+        page.click("button:has-text('预览任务')")
+        page.wait_for_timeout(300)
+    except Exception:
+        pass
+
+    _screenshot(page, "01_import_modal_filled")
+
+    # 点击「导入并生成任务」
+    print("[setup] 提交导入表单")
+    page.click("button:has-text('导入并生成任务')")
+    page.wait_for_timeout(3000)
+
+    _screenshot(page, "02_import_success")
+
+    # 点击「完成」关闭成功弹窗
+    try:
+        page.click("button:has-text('完成')")
+        page.wait_for_timeout(1000)
+    except Exception:
+        pass
+
+    # 切换到目标项目 Tab
+    print(f"[setup] 切换到项目: {project_name}")
+    try:
+        page.click(f"button:has-text('{project_name}')")
+        page.wait_for_timeout(500)
+    except Exception:
+        pass
+
+    _screenshot(page, "03_project_kanban")
+
+    # 启动执行器（点击「开始执行」）
+    print("[setup] 启动执行器")
+    try:
+        page.click("button:has-text('开始执行')")
+        page.wait_for_timeout(2000)
+    except Exception as e:
+        print(f"[WARN] 启动执行器按钮未找到: {e}")
+
+    _screenshot(page, "04_executor_started")
+
+    # 收集初始任务统计
+    tasks = _get_project_tasks(project_name)
+    stats = _task_stats(tasks)
+    print(f"[setup] 初始任务统计: {stats}")
+    return stats
+
+
+# ── Phase 2: 24 小时监控循环 ─────────────────────────────────────────
+
+def monitoring_loop(
+    page,
+    project_name: str,
+    total_seconds: int,
+    dry_run: bool = False,
+) -> list[dict]:
+    """每 CHECK_INTERVAL 秒检查一次，记录任务进展"""
+    metrics: list[dict] = []
+    start_time = time.time()
+    check_num = 0
+    interval = 60 if dry_run else CHECK_INTERVAL  # dry-run 用 1 分钟
+
+    print(f"\n[monitor] 开始监控，总时长 {total_seconds/3600:.1f}h，间隔 {interval}s")
+
+    while time.time() - start_time < total_seconds:
+        check_num += 1
+        elapsed_s = time.time() - start_time
+        elapsed_h = elapsed_s / 3600
+
+        print(f"\n[monitor] 第 {check_num} 次检查 ({elapsed_h:.2f}h / {total_seconds/3600:.1f}h)")
+
+        # 刷新页面，截图
+        try:
+            page.goto(DASHBOARD_URL)
+            try:
+                page.click(f"button:has-text('{project_name}')")
+            except Exception:
+                pass
+            page.wait_for_timeout(2000)
+            _screenshot(page, f"monitor_{check_num:04d}_{elapsed_h:.1f}h")
+        except Exception as e:
+            print(f"[WARN] 截图失败: {e}")
+
+        # 收集任务统计
+        try:
+            tasks = _get_project_tasks(project_name)
+            stats = _task_stats(tasks)
+            exec_status = _api("GET", "/api/executor/status")
+
+            record = {
+                "check": check_num,
+                "elapsed_h": round(elapsed_h, 2),
+                "timestamp": datetime.utcnow().isoformat(),
+                "tasks": stats,
+                "executor": exec_status,
+                "done_titles": [t["title"] for t in tasks if t["status"] == "done"],
+                "failed_titles": [t["title"] for t in tasks if t["status"] == "failed"],
+            }
+            metrics.append(record)
+            print(f"  任务: {stats}")
+            print(f"  执行器: running={exec_status.get('running')}, "
+                  f"active={exec_status.get('active')}")
+
+        except Exception as e:
+            print(f"[WARN] 指标收集失败: {e}")
+
+        # 检查执行器是否停止，若停止则重启
+        try:
+            exec_st = _api("GET", "/api/executor/status")
+            if not exec_st.get("running") and exec_st.get("pending", 0) > 0:
+                print("[monitor] 执行器已停止，重新启动...")
+                _api("POST", "/api/executor/start")
+        except Exception:
+            pass
+
+        # 等待下次检查
+        next_check = start_time + check_num * interval
+        wait_s = max(0, next_check - time.time())
+        if wait_s > 0 and time.time() - start_time < total_seconds:
+            print(f"  下次检查：{wait_s:.0f}s 后")
+            time.sleep(wait_s)
+
+    return metrics
+
+
+# ── Phase 3: 评估报告生成 ─────────────────────────────────────────────
+
+def generate_report(
+    project_name: str,
+    project_path: str,
+    linked_project: str,
+    metrics: list[dict],
+    total_hours: float,
+    page,
+) -> Path:
+    """生成 Markdown 评估报告"""
+    if not metrics:
+        print("[report] 无指标数据")
+        return REPORT_DIR / "report.md"
+
+    first = metrics[0]["tasks"]
+    last  = metrics[-1]["tasks"]
+
+    tasks_done   = last["done"]
+    tasks_total  = last["total"]
+    tasks_failed = last["failed"]
+    completion_pct = round(tasks_done / max(tasks_total, 1) * 100, 1)
+
+    # 收集所有 done 任务标题
+    all_done = []
+    for m in metrics:
+        for title in m.get("done_titles", []):
+            if title not in all_done:
+                all_done.append(title)
+
+    all_failed = []
+    for m in metrics:
+        for title in m.get("failed_titles", []):
+            if title not in all_failed:
+                all_failed.append(title)
+
+    # 执行器在线率
+    online_checks = sum(1 for m in metrics if m.get("executor", {}).get("running"))
+    online_pct = round(online_checks / max(len(metrics), 1) * 100, 1)
+
+    report_lines = [
+        f"# {total_hours:.0f}h 自动化运行评估报告",
+        f"",
+        f"**项目**: `{project_name}` ({project_path})",
+        f"**关联项目**: `{linked_project or '无'}`",
+        f"**报告时间**: {datetime.now().strftime('%Y-%m-%d %H:%M')}",
+        f"**运行时长**: {total_hours:.1f} 小时",
+        f"",
+        f"---",
+        f"",
+        f"## 执行摘要",
+        f"",
+        f"| 指标 | 数值 |",
+        f"|------|------|",
+        f"| 总任务数 | {tasks_total} |",
+        f"| 完成任务 | {tasks_done} ({completion_pct}%) |",
+        f"| 失败任务 | {tasks_failed} |",
+        f"| 待执行   | {last['pending']} |",
+        f"| 执行器在线率 | {online_pct}% |",
+        f"",
+        f"## 已完成任务",
+        f"",
+    ]
+    for t in all_done:
+        report_lines.append(f"- ✅ {t}")
+
+    if all_failed:
+        report_lines += ["", "## 失败任务", ""]
+        for t in all_failed:
+            report_lines.append(f"- ❌ {t}")
+
+    report_lines += [
+        "",
+        "## 进度时间线",
+        "",
+        "| 时间点 | 完成 | 失败 | 待执行 |",
+        "|--------|------|------|--------|",
+    ]
+    for m in metrics[::4] + ([metrics[-1]] if metrics else []):  # 每4条取1条
+        s = m["tasks"]
+        report_lines.append(
+            f"| {m['elapsed_h']:.1f}h | {s['done']} | {s['failed']} | {s['pending']} |"
+        )
+
+    report_lines += [
+        "",
+        "## 截图目录",
+        f"",
+        f"见 `{REPORT_DIR}/` 目录，共 {len(list(REPORT_DIR.glob('*.png')))} 张截图。",
+        "",
+        "---",
+        f"*由 `scripts/test_24h.py` 自动生成*",
+    ]
+
+    report_path = REPORT_DIR / "report.md"
+    report_path.write_text("\n".join(report_lines), encoding="utf-8")
+    print(f"\n[report] 报告已生成: {report_path}")
+    return report_path
+
+
+# ── 主入口 ────────────────────────────────────────────────────────────
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="24h 项目自动化运行测试")
+    parser.add_argument("--project", default="/data/rockchip/logodetect-x86",
+                        help="项目目录路径")
+    parser.add_argument("--linked", default="/data/test-platform/logo-collect",
+                        help="关联项目路径（用于集成任务）")
+    parser.add_argument("--hours", type=float, default=24.0,
+                        help="运行时长（小时），默认 24")
+    parser.add_argument("--dry-run", action="store_true",
+                        help="快速测试模式（1 分钟间隔，5 分钟总时长）")
+    parser.add_argument("--headless", action="store_true", default=True,
+                        help="无头模式运行浏览器（默认开启）")
+    parser.add_argument("--no-headless", dest="headless", action="store_false",
+                        help="显示浏览器窗口（需要显示器）")
+    args = parser.parse_args()
+
+    project_path = args.project
+    project_name = Path(project_path).name
+    linked_project = args.linked
+    total_hours = 5 / 60 if args.dry_run else args.hours  # dry-run = 5 分钟
+    total_seconds = int(total_hours * 3600)
+
+    REPORT_DIR.mkdir(parents=True, exist_ok=True)
+    metrics_path = REPORT_DIR / "metrics.jsonl"
+
+    print("=" * 60)
+    print(f"NMFS Agents 自动化测试")
+    print(f"  项目: {project_name} ({project_path})")
+    print(f"  关联: {linked_project}")
+    print(f"  时长: {total_hours:.2f}h ({'dry-run' if args.dry_run else '正式'})")
+    print(f"  报告: {REPORT_DIR}/")
+    print("=" * 60)
+
+    try:
+        from playwright.sync_api import sync_playwright
+    except ImportError:
+        print("[ERROR] playwright 未安装，运行: pip install playwright && playwright install chromium")
+        raise SystemExit(1)
+
+    with sync_playwright() as pw:
+        browser = pw.chromium.launch(headless=args.headless)
+        ctx = browser.new_context(viewport={"width": 1440, "height": 900})
+        page = ctx.new_page()
+
+        try:
+            # Phase 1: 设置
+            print("\n=== Phase 1: 浏览器 UI 设置 ===")
+            setup_via_browser(page, project_path, linked_project, project_name)
+
+            # Phase 2: 监控
+            print("\n=== Phase 2: 监控循环 ===")
+            metrics = monitoring_loop(page, project_name, total_seconds, args.dry_run)
+
+            # 保存原始指标
+            with open(metrics_path, "w") as f:
+                for m in metrics:
+                    f.write(json.dumps(m, ensure_ascii=False) + "\n")
+
+            # Phase 3: 生成报告
+            print("\n=== Phase 3: 生成评估报告 ===")
+            _screenshot(page, "final_state")
+            report_path = generate_report(
+                project_name, project_path, linked_project,
+                metrics, total_hours, page,
+            )
+            print(f"\n完成！评估报告: {report_path}")
+
+        finally:
+            browser.close()
+
+
+def _screenshot(page, name: str) -> None:
+    try:
+        path = REPORT_DIR / f"{name}.png"
+        page.screenshot(path=str(path), full_page=False)
+        print(f"  [screenshot] {path.name}")
+    except Exception as e:
+        print(f"  [WARN] 截图失败 {name}: {e}")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+# 实时监控 NMFS Agents 守护进程
+# 用法：bash scripts/watch_daemon.sh
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(dirname "$SCRIPT_DIR")"
+PID_FILE="$PROJECT_ROOT/logs/daemon.pid"
+VENV="$PROJECT_ROOT/venv"
+
+if [[ ! -f "$PID_FILE" ]]; then
+    echo "守护进程未运行（PID 文件不存在）"
+    echo "启动方法: bash scripts/start_daemon.sh"
+    exit 0
+fi
+
+PID=$(cat "$PID_FILE")
+if ! kill -0 "$PID" 2>/dev/null; then
+    echo "守护进程已停止（PID=$PID 不存在）"
+    rm -f "$PID_FILE"
+    exit 0
+fi
+
+echo "═══════════════════════════════════════════"
+echo " NMFS Agents 监控  PID=$PID"
+echo " $(date '+%Y-%m-%d %H:%M:%S')"
+echo "═══════════════════════════════════════════"
+
+# 任务队列状态
+echo ""
+echo "── 任务队列状态 ─────────────────────────────"
+source "$VENV/bin/activate"
+python "$PROJECT_ROOT/scripts/run.py" status 2>/dev/null || echo "（获取失败）"
+
+# 项目目标
+echo ""
+echo "── 项目目标记忆 ─────────────────────────────"
+python "$PROJECT_ROOT/scripts/run.py" goals 2>/dev/null || echo "（获取失败）"
+
+# memory-analyze 版本历史
+echo ""
+echo "── memory-analyze 架构版本 ──────────────────"
+python "$PROJECT_ROOT/scripts/run.py" arch-versions memory-analyze 2>/dev/null || echo "（获取失败）"
+
+# 最新日志
+echo ""
+echo "── 最新日志（最近 30 行）────────────────────"
+LATEST_LOG=$(ls -t "$PROJECT_ROOT/logs"/daemon_*.log 2>/dev/null | head -1)
+if [[ -n "$LATEST_LOG" ]]; then
+    echo "  文件: $LATEST_LOG"
+    tail -30 "$LATEST_LOG"
+else
+    echo "（暂无日志文件）"
+fi
@@ -0,0 +1,14 @@
+[Unit]
+Description=EdgeValidatorService for RK3588
+After=network.target
+
+[Service]
+Type=simple
+User=pi
+WorkingDirectory=/home/pi/Desktop/edge_validator
+ExecStart=/home/pi/Desktop/edge_validator/venv/bin/uvicorn main:app --host 0.0.0.0 --port 8899
+Restart=always
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target
@@ -0,0 +1,207 @@
+"""EdgeValidatorService: 部署在 RK3588 上，为开发机 agents 提供 HTTP 接口。
+
+启动：uvicorn main:app --host 0.0.0.0 --port 8899
+"""
+from __future__ import annotations
+
+import base64
+import glob
+import re
+import subprocess
+import time
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+from fastapi import FastAPI, HTTPException
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+
+app = FastAPI(title="EdgeValidatorService", version="1.0.0")
+app.add_middleware(CORSMiddleware, allow_origins=["*"],
+                   allow_methods=["*"], allow_headers=["*"])
+
+# 设备配置（本机 + 接入的 MCU）
+SERIAL_DEVICES: dict[str, dict] = {}
+
+# 推理日志路径
+LOG_PATHS = {
+    "default": "/tmp/infer.log",
+    "yolo": "/tmp/yolo_infer.log",
+}
+
+
+class RunTestRequest(BaseModel):
+    cmd: str
+    cwd: str = "/home/pi/Desktop"
+    timeout: int = 120
+
+
+class FlashRequest(BaseModel):
+    firmware_b64: str
+    offset: str = "0x0"
+    port: str = "/dev/ttyUSB0"
+
+
+# ── 端点 ──────────────────────────────────────────────────────────
+
+@app.get("/health")
+def health():
+    return {"status": "ok", "timestamp": datetime.utcnow().isoformat()}
+
+
+@app.get("/devices")
+def list_devices():
+    """列出识别到的串口设备"""
+    serial_ports = glob.glob("/dev/ttyUSB*") + glob.glob("/dev/ttyACM*")
+    return {"rk3588": "linux/ssh", "serial_ports": serial_ports,
+            "configured_mcus": list(SERIAL_DEVICES.keys())}
+
+
+@app.get("/metrics/{device_name}")
+def get_metrics(device_name: str):
+    if device_name == "rk3588":
+        return _linux_metrics()
+    if device_name in SERIAL_DEVICES:
+        return _mcu_metrics(SERIAL_DEVICES[device_name])
+    raise HTTPException(404, f"未知设备: {device_name}")
+
+
+@app.post("/run-test/{device_name}")
+def run_test(device_name: str, req: RunTestRequest):
+    if device_name == "rk3588":
+        return _run_local(req)
+    if device_name in SERIAL_DEVICES:
+        return _run_serial(SERIAL_DEVICES[device_name], req)
+    raise HTTPException(404, f"未知设备: {device_name}")
+
+
+@app.post("/screenshot")
+def screenshot():
+    """截取 RK3588 屏幕（需要 scrot 或 ffmpeg）"""
+    import tempfile
+    import os
+    tmp = tempfile.mktemp(suffix=".png")
+    r = subprocess.run(["scrot", tmp], capture_output=True, timeout=10)
+    if r.returncode != 0:
+        r = subprocess.run(
+            f"cat /dev/fb0 | ffmpeg -vcodec rawvideo -f rawvideo "
+            f"-pix_fmt rgb32 -s 1920x1080 -i - -f image2 -vcodec png {tmp} -y",
+            shell=True, capture_output=True, timeout=15,
+        )
+    if r.returncode != 0:
+        raise HTTPException(500, "截图失败，请确认 scrot 或 ffmpeg 已安装")
+    with open(tmp, "rb") as f:
+        img_b64 = base64.b64encode(f.read()).decode()
+    os.unlink(tmp)
+    return {"image_b64": img_b64, "format": "png"}
+
+
+@app.get("/logs/{device_name}/{n}")
+def get_logs(device_name: str, n: int = 50):
+    log_path = LOG_PATHS.get(device_name, LOG_PATHS["default"])
+    try:
+        r = subprocess.run(["tail", f"-{n}", log_path],
+                           capture_output=True, text=True, timeout=5)
+        return {"lines": r.stdout.splitlines(), "file": log_path}
+    except Exception as e:
+        return {"lines": [], "error": str(e), "file": log_path}
+
+
+# ── 内部函数 ──────────────────────────────────────────────────────
+
+def _run_local(req: RunTestRequest) -> dict:
+    t0 = time.monotonic()
+    r = subprocess.run(
+        req.cmd, shell=True, capture_output=True, text=True,
+        cwd=req.cwd, timeout=req.timeout,
+    )
+    duration = time.monotonic() - t0
+    stdout = r.stdout
+    pass_ = r.returncode == 0 and not re.search(r"\b(FAILED|ERROR)\b", stdout + r.stderr)
+    return {"rc": r.returncode, "stdout": stdout, "stderr": r.stderr,
+            "duration_s": round(duration, 2), "pass_": pass_}
+
+
+def _run_serial(dev: dict, req: RunTestRequest) -> dict:
+    try:
+        import serial  # type: ignore
+    except ImportError:
+        return {"rc": 1, "stdout": "", "stderr": "pyserial 未安装",
+                "duration_s": 0, "pass_": False}
+    t0 = time.monotonic()
+    lines: list[str] = []
+    try:
+        ser = serial.Serial(dev["port"], dev.get("baud", 115200), timeout=req.timeout)
+        ser.write((req.cmd + "\r\n").encode())
+        deadline = time.monotonic() + req.timeout
+        while time.monotonic() < deadline:
+            line = ser.readline().decode(errors="replace").strip()
+            if line:
+                lines.append(line)
+            if line.startswith(("PASS", "FAIL", "ERROR")):
+                break
+        ser.close()
+    except Exception as e:
+        return {"rc": 1, "stdout": "\n".join(lines), "stderr": str(e),
+                "duration_s": time.monotonic() - t0, "pass_": False}
+    stdout = "\n".join(lines)
+    rc = 0 if any(l.startswith("PASS") for l in lines) else 1
+    return {"rc": rc, "stdout": stdout, "stderr": "",
+            "duration_s": round(time.monotonic() - t0, 2), "pass_": rc == 0}
+
+
+def _linux_metrics() -> dict:
+    m: dict = {"device": "rk3588", "timestamp": datetime.utcnow().isoformat()}
+    try:
+        with open("/proc/stat") as f:
+            parts = f.readline().split()
+        user, nice, sys_, idle = int(parts[1]), int(parts[2]), int(parts[3]), int(parts[4])
+        total = user + nice + sys_ + idle
+        m["cpu_pct"] = round(100 * (user + sys_) / total, 1) if total else 0
+    except Exception:
+        m["cpu_pct"] = None
+    try:
+        r = subprocess.run(["free", "-m"], capture_output=True, text=True, timeout=3)
+        for line in r.stdout.splitlines():
+            if line.startswith("Mem:"):
+                parts = line.split()
+                m["mem_used_mb"] = float(parts[2])
+                break
+    except Exception:
+        m["mem_used_mb"] = None
+    try:
+        r = subprocess.run(
+            "tail -10 /tmp/infer.log 2>/dev/null | grep -oP 'fps=\\K[0-9.]+' | tail -1",
+            shell=True, capture_output=True, text=True, timeout=3,
+        )
+        raw = r.stdout.strip()
+        m["fps"] = float(raw) if raw else None
+    except Exception:
+        m["fps"] = None
+    return m
+
+
+def _mcu_metrics(dev: dict) -> dict:
+    m: dict = {"device": dev.get("name", "mcu"),
+               "timestamp": datetime.utcnow().isoformat()}
+    try:
+        import serial  # type: ignore
+        ser = serial.Serial(dev["port"], dev.get("baud", 115200), timeout=3)
+        ser.write(b"GET_METRICS\r\n")
+        for _ in range(20):
+            line = ser.readline().decode(errors="replace").strip()
+            if line.startswith("METRICS:"):
+                for kv in line[8:].split():
+                    k, _, v = kv.partition("=")
+                    if k == "fps":
+                        m["fps"] = float(v)
+                    elif k == "latency_ms":
+                        m["inference_latency_ms"] = float(v)
+                    elif k == "ram_kb":
+                        m["ram_used_kb"] = int(v)
+                break
+        ser.close()
+    except Exception as e:
+        m["error"] = str(e)
+    return m
@@ -0,0 +1,3 @@
+fastapi>=0.110.0
+uvicorn[standard]>=0.29.0
+pyserial>=3.5
@@ -0,0 +1,418 @@
+from __future__ import annotations
+
+import logging
+import os
+import re
+import subprocess
+from pathlib import Path
+from typing import Optional
+
+from nmfs_agents.agents.developer import (
+    AgentResult, _parse_cli_result, _make_drop_root_preexec,
+    _run_with_log, _task_log_path,
+)
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.arch_version import ArchVersion
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+TIMEOUT = 3600
+TEST_TIMEOUT = 120
+
+ARCHITECT_SYSTEM = """你是一位技术架构师，负责分析多个软件项目，识别整合机会，提出架构改进建议。
+
+分析维度：
+- 功能互补性：哪些项目的能力可以组合，产生新价值
+- 技术共性：是否有可复用的基础设施或数据管道
+- 产品可能性：整合后能解决什么实际问题
+
+请用中文输出结构化分析报告，末尾按格式列出任务（每行一条）：
+- [预研] 任务名称  ← 跨项目探索、新功能研究
+- [优化] <项目名>: 优化描述  ← 对特定项目的架构改进（项目名须在上面列出的项目中）
+- [开发] <项目名>: 任务描述  ← 派发具体开发任务给 developer（可多条，多开发者并行执行）
+- [测试] <项目名>: 验证要点  ← 派发集成测试任务给 tester
+- [重构] <项目名>: 重构内容  ← 派发代码重构任务给 dev-refactor
+- [内核开发] <项目名>: 任务  ← 派发内核驱动开发任务给 dev-kernel
+- [目标] <项目名>: 目标描述  ← 更新项目长期目标（仅当需要修订时）
+- [记忆] <项目名>: key=value, key2=value2  ← 记录项目关键技术事实
+- [后续] 自动续链: 描述  ← 基于本次分析发现，列出1-3条应立即跟进的技术调研或验证任务
+
+## 技能配置能力（Skills）
+如分析中发现 Agent 工具/能力需要调整，可直接修改配置文件：
+- MCP 工具配置：`/data/company/agents/configs/mcp_tools.yaml`（角色工具授权）
+- Agent 配置：`/data/company/agents/configs/agents.yaml`（队列、并发、角色列表）
+- 每次修改配置文件后，在 [记忆] 中记录变更原因：key=config_update, value=<变更摘要>
+"""
+
+
+class ArchitectAgent:
+    def __init__(self, config: AgentsConfig,
+                 queue: Optional[TaskQueue] = None,
+                 memory: Optional[ProjectMemory] = None) -> None:
+        self._config = config
+        self._queue = queue or TaskQueue()
+        self._memory = memory or ProjectMemory()
+
+    # ── entry point ───────────────────────────────────────────────────────────
+
+    def run(self, task: Task) -> AgentResult:
+        """Route to optimize flow or cross-project analysis based on task type."""
+        if not self._verify_claude_binary():
+            return AgentResult(status="failed", summary="claude CLI 不可达，bootstrap 验证失败")
+        if task.type == "arch_optimize":
+            return self._optimize(task)
+        return self._analyze(task)
+
+    @staticmethod
+    def _verify_claude_binary() -> bool:
+        """验证 claude CLI 可达性（bootstrap 检查）。"""
+        try:
+            result = subprocess.run(
+                ["claude", "--version"],
+                capture_output=True, text=True, timeout=10,
+            )
+            if result.returncode == 0:
+                logger.info("claude CLI 可达: %s", result.stdout.strip())
+                return True
+            logger.warning("claude --version 返回非零码 %d: %s", result.returncode, result.stderr.strip())
+            return False
+        except (FileNotFoundError, subprocess.TimeoutExpired) as e:
+            logger.error("claude CLI 不可达: %s", e)
+            return False
+
+    # ── cross-project analysis ────────────────────────────────────────────────
+
+    def _analyze(self, task: Task) -> AgentResult:
+        """Generate architecture report and enqueue [预研]/[优化] tasks."""
+        summaries = self._collect_summaries()
+        cmd = [
+            "claude", "--print",
+            "--output-format", "json",
+            "--no-session-persistence",
+            "--model", self._config.claude.model,
+            "--allowedTools", "Read,Glob,Grep",
+            "--append-system-prompt", ARCHITECT_SYSTEM,
+            "--", self._build_prompt(summaries),
+        ]
+        try:
+            _env = self._build_env()
+            result = _run_with_log(
+                cmd, None, _env, TIMEOUT, _task_log_path(task.id),
+                preexec_fn=_make_drop_root_preexec(_env),
+            )
+            if result.status == "done":
+                self._extract_and_enqueue(result.summary, task.project)
+                logger.info("架构分析完成，已生成 %d 字报告", len(result.summary))
+            return result
+        except Exception as e:
+            logger.error("ArchitectAgent 分析失败: %s", e)
+            return AgentResult(status="failed", summary=str(e))
+
+    # ── iterative optimization with version control ───────────────────────────
+
+    def _optimize(self, task: Task) -> AgentResult:
+        """Apply an architectural optimization with snapshot/test/rollback cycle."""
+        proj_cfg = self._config.projects.get(task.project)
+        if not proj_cfg:
+            return AgentResult(status="failed",
+                               summary=f"项目 {task.project!r} 未在配置中找到")
+
+        arch_v = ArchVersion(proj_cfg.path)
+        version_id = arch_v.snapshot(task.title)
+
+        # Apply optimization
+        cmd = [
+            "claude", "--print",
+            "--output-format", "json",
+            "--no-session-persistence",
+            "--model", self._config.claude.model,
+            "--dangerously-skip-permissions",
+            "--append-system-prompt", ARCHITECT_SYSTEM,
+            "--", self._build_optimize_prompt(task),
+        ]
+        try:
+            _env = self._build_env()
+            apply_result = _run_with_log(
+                cmd, str(proj_cfg.path), _env, TIMEOUT, _task_log_path(task.id),
+                preexec_fn=_make_drop_root_preexec(_env),
+            )
+        except Exception as e:
+            logger.error("arch_optimize 执行失败: %s", e)
+            try:
+                arch_v.rollback(version_id, str(e))
+            except Exception:
+                pass
+            return AgentResult(status="failed", summary=str(e))
+
+        if apply_result.status == "failed":
+            arch_v.rollback(version_id, apply_result.summary)
+            return AgentResult(
+                status="failed",
+                summary=f"[{version_id}] 优化执行失败，已回滚:\n{apply_result.summary}",
+            )
+
+        # Validate with tests
+        test_res = self._run_tests(proj_cfg.path)
+        if test_res["passed"]:
+            arch_v.promote(version_id, test_res["output"])
+            return AgentResult(
+                status="done",
+                summary=(
+                    f"[{version_id}] 优化成功并晋升\n"
+                    f"{apply_result.summary}\n\n"
+                    f"测试: {test_res['output'][:300]}"
+                ),
+            )
+        else:
+            arch_v.rollback(version_id, test_res["output"])
+            return AgentResult(
+                status="failed",
+                summary=(
+                    f"[{version_id}] 优化后测试失败，已回滚\n"
+                    f"{test_res['output'][:400]}"
+                ),
+            )
+
+    # ── helpers ───────────────────────────────────────────────────────────────
+
+    def _collect_summaries(self) -> dict[str, str]:
+        return {
+            name: self._read_project_summary(proj_cfg.path)
+            for name, proj_cfg in self._config.projects.items()
+        }
+
+    def _read_project_summary(self, path: Path) -> str:
+        for fname in ("README.md", "CLAUDE.md"):
+            f = path / fname
+            if not f.exists():
+                continue
+            lines = [l.strip() for l in f.read_text(errors="ignore").splitlines()
+                     if l.strip() and not l.startswith("#")]
+            if lines:
+                return " ".join(lines[:3])[:300]
+        return f"（{path.name}，无说明文档）"
+
+    def _build_prompt(self, summaries: dict[str, str]) -> str:
+        proj_names = "、".join(summaries.keys())
+        lines = [f"以下是当前所有项目的功能描述（可用项目名：{proj_names}）：\n"]
+        for name, desc in summaries.items():
+            goal_ctx = self._memory.get_context_for_prompt(name, role="architect")
+            lines.append(f"**{name}**: {desc}")
+            if goal_ctx.strip():
+                lines.append(goal_ctx.strip())
+        lines.append(
+            "\n请分析这些项目的能力，识别整合机会，给出产品架构建议。\n"
+            "在报告末尾按以下格式列出任务（每行一条）：\n"
+            "- [预研] 任务名称\n"
+            f"- [优化] <项目名>: 优化描述  （项目名须为：{proj_names} 之一）\n"
+            f"- [开发] <项目名>: 具体开发任务描述  （可多条，并行派发给 developer）\n"
+            f"- [测试] <项目名>: 集成测试验证要点  （派发给 tester）\n"
+            f"- [重构] <项目名>: 重构目标和范围  （派发给 dev-refactor）\n"
+            f"- [内核开发] <项目名>: 内核/驱动任务  （派发给 dev-kernel）\n"
+            f"- [目标] <项目名>: 更新后的项目长期目标描述  （可选，仅当目标需要修订时）"
+        )
+        return "\n".join(lines)
+
+    def _build_optimize_prompt(self, task: Task) -> str:
+        proj_cfg = self._config.projects.get(task.project)
+        proj_desc = proj_cfg.description if proj_cfg else ""
+        return (
+            f"你是 {task.project} 项目的架构优化 Agent。\n"
+            f"项目: {task.project}，描述: {proj_desc}\n"
+            f"优化任务: {task.title}\n"
+            f"背景: {task.context}\n\n"
+            f"请对该项目进行架构层面的优化：\n"
+            f"1. 分析现有代码结构和性能瓶颈\n"
+            f"2. 实施具体的架构改进（直接修改代码文件）\n"
+            f"3. 确保改动向后兼容，不破坏现有接口\n"
+            f"4. 输出改动摘要\n"
+            f"请直接修改代码文件，不要只输出报告。"
+        )
+
+    def _build_env(self) -> dict:
+        env = os.environ.copy()
+        if self._config.claude.api_key:
+            env["ANTHROPIC_API_KEY"] = self._config.claude.api_key
+        env.pop("CLAUDECODE", None)
+        env.pop("CLAUDE_CODE_ENTRYPOINT", None)
+        return env
+
+    def _run_tests(self, project_path: Path) -> dict:
+        """Run pytest in the project directory. Returns {passed, output}."""
+        tests_dir = project_path / "tests"
+        if not tests_dir.exists():
+            return {"passed": True, "output": "（无测试目录，跳过验证）"}
+
+        for pytest_bin in [str(project_path / "venv" / "bin" / "pytest"), "pytest"]:
+            try:
+                r = subprocess.run(
+                    [pytest_bin, "tests/", "-q", "--tb=short"],
+                    capture_output=True, text=True,
+                    timeout=TEST_TIMEOUT, cwd=str(project_path),
+                )
+                return {
+                    "passed": r.returncode == 0,
+                    "output": (r.stdout + r.stderr)[:1000],
+                }
+            except FileNotFoundError:
+                continue
+            except subprocess.TimeoutExpired:
+                return {"passed": False, "output": f"测试超时（{TEST_TIMEOUT}s）"}
+        return {"passed": False, "output": "未找到 pytest 可执行文件"}
+
+    def _extract_and_enqueue(self, report_text: str, requester: str) -> None:
+        """Parse [预研], [优化], [目标], and [记忆] tags from report."""
+        # [预研] → cross-project research task (report mode)
+        for m in re.finditer(r"\[预研\]\s+(.+)", report_text):
+            title = m.group(1).strip()[:100]
+            self._queue.enqueue(Task(
+                project="research", type="architect", title=title,
+                context=f"由 Architect Agent 从架构分析中提取（来源: {requester}）",
+                priority=3, mode="report", agent_role="architect",
+                initiator="architect",
+                discussion=f"架构分析报告中的预研项（来源项目: {requester}）",
+            ))
+            logger.info("预研任务入队: %s", title)
+
+        # [优化] <project>: <desc> → arch_optimize task
+        for m in re.finditer(r"\[优化\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            description = m.group(2).strip()[:100]
+            if project not in self._config.projects:
+                logger.warning("优化任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            self._queue.enqueue(Task(
+                project=project,
+                type="arch_optimize",
+                title=f"[架构优化] {description}",
+                context=f"由 Architect Agent 提出（来源: {requester}）\n{description}",
+                priority=2,
+                mode=proj_cfg.mode,
+                agent_role="architect",
+                initiator="architect",
+                discussion=f"Architect Agent 在分析 {requester} 时提出的优化建议",
+            ))
+            logger.info("架构优化任务入队: [%s] %s", project, description)
+
+        # [开发] <project>: <desc> → developer task (multiple allowed for parallel dispatch)
+        for m in re.finditer(r"\[开发\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            description = m.group(2).strip()[:100]
+            if project not in self._config.projects:
+                logger.warning("开发任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            self._queue.enqueue(Task(
+                project=project, type="fix_bug",
+                title=f"[开发] {description}",
+                context=f"由 Architect 指派（来源分析: {requester}）\n{description}",
+                priority=2, mode=proj_cfg.mode, agent_role="developer",
+                initiator="architect",
+                discussion=f"Architect 分析 {requester} 时指派的开发任务:\n{description}",
+            ))
+            logger.info("开发任务入队: [%s] %s", project, description)
+
+        # [测试] <project>: <desc> → tester task
+        for m in re.finditer(r"\[测试\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            description = m.group(2).strip()[:100]
+            if project not in self._config.projects:
+                logger.warning("测试任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            self._queue.enqueue(Task(
+                project=project, type="test_improvement",
+                title=f"[测试] {description}",
+                context=f"由 Architect 指派（来源分析: {requester}）\n{description}",
+                priority=2, mode=proj_cfg.mode, agent_role="tester",
+                initiator="architect",
+                discussion=f"Architect 分析 {requester} 时指派的测试任务:\n{description}",
+            ))
+            logger.info("测试任务入队: [%s] %s", project, description)
+
+        # [重构] <project>: <desc> → dev-refactor task
+        for m in re.finditer(r"\[重构\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            description = m.group(2).strip()[:100]
+            if project not in self._config.projects:
+                logger.warning("重构任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            self._queue.enqueue(Task(
+                project=project, type="refactor",
+                title=f"[重构] {description}",
+                context=f"由 Architect 指派（来源分析: {requester}）\n{description}",
+                priority=3, mode=proj_cfg.mode, agent_role="dev-refactor",
+                initiator="architect",
+                discussion=f"Architect 分析 {requester} 时指派的重构任务:\n{description}",
+            ))
+            logger.info("重构任务入队: [%s] %s", project, description)
+
+        # [内核开发] <project>: <desc> → dev-kernel task
+        for m in re.finditer(r"\[内核开发\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            description = m.group(2).strip()[:100]
+            if project not in self._config.projects:
+                logger.warning("内核开发任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            self._queue.enqueue(Task(
+                project=project, type="fix_bug",
+                title=f"[内核开发] {description}",
+                context=f"由 Architect 指派（来源分析: {requester}）\n{description}",
+                priority=2, mode=proj_cfg.mode, agent_role="dev-kernel",
+                initiator="architect",
+                discussion=f"Architect 分析 {requester} 时指派的内核开发任务:\n{description}",
+            ))
+            logger.info("内核开发任务入队: [%s] %s", project, description)
+
+        # [目标] <project>: <goal> → update project memory (architect can write)
+        for m in re.finditer(r"\[目标\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            goal_text = m.group(2).strip()[:300]
+            if project not in self._config.projects:
+                logger.warning("目标更新项目 %r 不在配置中，跳过", project)
+                continue
+            try:
+                self._memory.set_goal(project, goal_text, role="architect")
+                logger.info("项目目标已更新 by architect: [%s]", project)
+            except Exception as e:
+                logger.warning("[%s] 写入项目目标失败: %s", project, e)
+
+        # [记忆] <project>: key=value, key2=value2 → update project facts
+        for m in re.finditer(r"\[记忆\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            kv_text = m.group(2).strip()
+            if project not in self._config.projects:
+                logger.warning("记忆更新项目 %r 不在配置中，跳过", project)
+                continue
+            for kv in kv_text.split(","):
+                kv = kv.strip()
+                if "=" in kv:
+                    key, _, value = kv.partition("=")
+                    key = key.strip()[:50]
+                    value = value.strip()[:200]
+                    if key and value:
+                        try:
+                            self._memory.set_fact(project, key, value, "architect")
+                            logger.info("项目事实已记录 [%s] %s=%s", project, key, value)
+                        except Exception as e:
+                            logger.warning("[%s] 写入项目事实失败: %s", project, e)
+
+        # [后续] 自动续链: <desc> → 立即跟进的调研/验证任务（入 project_delivery，低优先级）
+        for m in re.finditer(r"\[后续\][^\[：:]*[：:]\s*(.+)", report_text):
+            desc = m.group(1).strip()[:120]
+            self._queue.enqueue(Task(
+                project=requester if requester in self._config.projects else "research",
+                type="tech_support",
+                title=f"[后续] {desc}",
+                context=f"由 Architect 在分析中自动续链（来源: {requester}）\n{desc}",
+                priority=4,
+                mode="auto",
+                agent_role="architect",
+                initiator="architect",
+            ))
+            logger.info("后续任务入队: %s", desc)
@@ -0,0 +1,25 @@
+from __future__ import annotations
+
+from nmfs_agents.agents.architect import ArchitectAgent
+
+BASE_ARCHITECT_SYSTEM = """你是基础技术研发组的架构师，专注平台级技术栈整体设计。
+
+职责（区别于项目组 architect）：
+- 分析 OS/算法/硬件 三个方向的技术债和瓶颈
+- 识别跨方向整合机会（如防抖算法与 GPS 惯导融合）
+- 评估新硬件对现有基础能力的影响
+- 为 os-engineer / algo-researcher 提出具体 [优化] 任务
+
+输出格式：
+- [优化] os-engineer: 具体优化建议
+- [优化] algo-researcher: 具体优化建议
+- [记忆] __platform__: key=value（记录架构决策）
+"""
+
+
+class BaseArchitectAgent(ArchitectAgent):
+    """基础技术研发组：OS/算法/硬件跨方向架构分析"""
+
+    def _build_prompt(self, summaries: dict[str, str]) -> str:
+        base = super()._build_prompt(summaries)
+        return BASE_ARCHITECT_SYSTEM + "\n\n" + base
@@ -0,0 +1,55 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+BASE_VALIDATOR_SYSTEM = """你是基础技术研发组的验证工程师，负责算法和OS的基线验证。
+
+验证维度：
+1. 精度测试：检出率、误检率、漏检率，与 __hardware__ 基线对比
+2. 性能测试：FPS、推理延迟（ms）、NPU利用率、CPU/内存占用
+3. 稳定性测试：长时间运行（≥1h）、内存泄漏检测、异常恢复
+4. 硬件约束验证：读取 __hardware__ facts，确认算法满足硬件限制
+
+边端验证方式：
+- 通过 EdgeValidatorService HTTP API 执行测试（见上方设备信息）
+- 或通过 SSH 直接执行：sshpass -p '...' ssh pi@<host> "cd <workspace> && pytest ..."
+
+输出格式：
+- [验证结果] project=<name> fps=<N> precision=<0.xx> pass=true/false
+- [验证退化] project=<name> metric=fps current=<N> baseline=<M> drop_pct=<P>
+- [记忆] __hardware__: <project>.fps_baseline=<N>  （首次验证或基线更新时）
+- [需求] <project>: <具体优化需求>  （当 fps drop >10% 或 precision drop >5% 时）
+
+退化阈值（自动生成 [需求]）：
+- FPS 下降超过 10%
+- 精度（precision/recall）下降超过 5 个百分点
+- 内存使用增长超过 20%
+
+请用中文输出，数值精确，结论明确。
+"""
+
+
+class BaseValidatorAgent(DeveloperAgent):
+    """基础技术组：算法精度 + 性能 + 稳定性 + 硬件约束验证"""
+
+    def __init__(
+        self,
+        config: AgentsConfig,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        super().__init__(config, memory=memory)
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        hw_facts = self._memory.get_facts("__hardware__")
+        if hw_facts:
+            hw_lines = "\n".join(f"  {k}: {v}" for k, v in hw_facts.items())
+            hw_section = f"\n【硬件约束（来自 hw-engineer）】\n{hw_lines}\n"
+        else:
+            hw_section = "\n【硬件约束】暂无，请查阅 __hardware__ facts\n"
+        return base + hw_section + self._build_device_context() + "\n" + BASE_VALIDATOR_SYSTEM
@@ -0,0 +1,132 @@
+from __future__ import annotations
+
+"""boss.py — L0 Boss Agent（CEO/战略决策者）。
+
+职责：
+- 每周接收各部门长汇报，制定战略方向
+- 读取四域知识库（tech/market/management/delivery）做决策参考
+- 不介入日常开发细节，聚焦跨部门协作和长期规划
+- 接收 [越级汇报] 标签的重要信息（重大技术风险/市场机会）
+- 输出 [决策] [方向] [协调] 标签驱动后续行动
+"""
+
+import logging
+import re
+from typing import Optional
+
+from nmfs_agents.agents.developer import AgentResult, DeveloperAgent
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.knowledge_base import KnowledgeBase
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+BOSS_SYSTEM = """你是 NMFS AI Company 的 CEO/Boss（L0 战略决策层）。
+
+## 职责定位
+- 接收基础技术部、项目交付部、市场洞察部三个部门长的周报
+- 制定公司技术战略和市场方向
+- 协调跨部门资源分配和优先级
+- 处理 [越级汇报] 重要信息（技术风险/市场机会/跨组织问题）
+- **不介入**具体代码、测试、日常开发细节
+
+## 决策维度
+1. **技术战略**：基础技术积累是否与市场需求对齐？
+2. **市场机会**：当前研发方向是否抓住了最重要的市场窗口？
+3. **资源效率**：三个部门的投入产出比是否合理？
+4. **风险管控**：是否存在技术债务、人员短板、外部威胁？
+5. **长期护城河**：我们的核心壁垒是什么，如何持续加固？
+
+## 输出格式（必须包含）
+- [决策] <具体决策或批准事项>
+- [方向] <调整优先级或战略方向>
+- [协调] <需要跨部门协作的事项，说明哪两个部门>
+- [记忆] kb.tech: 关键技术决策  或  kb.market: 市场判断  或  kb.management: 管理决策
+- [后续] <下周需要重点跟进的1-3件事>
+
+## 工作模式
+1. 读取四域知识库最新动态（上方已注入）
+2. 综合分析三部门现状
+3. 识别最重要的2-3个战略问题
+4. 给出明确决策，不模糊，不推诿
+5. 记录关键决策到知识库
+
+## 沟通原则
+- 与部门长通过任务队列异步沟通
+- 接受 [越级汇报] 但不鼓励绕过中间层（除非紧急）
+- 决策有依据，不独断专行
+"""
+
+
+class BossAgent(DeveloperAgent):
+    """L0 战略决策 Boss Agent。"""
+
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+        knowledge_base: Optional[KnowledgeBase] = None,
+    ) -> None:
+        super().__init__(config, memory=memory)
+        self._queue = queue or TaskQueue()
+        self._kb = knowledge_base or KnowledgeBase()
+
+    def run(self, task: Task) -> AgentResult:
+        result = super().run(task)
+        if result.status == "done":
+            self._extract_kb_facts(result.summary)
+            self._dispatch_coordination(result.summary, task)
+        return result
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        # 注入四域知识库摘要
+        kb_sections = []
+        for domain in ("tech", "market", "management", "delivery"):
+            snippet = self._kb.context_for_agent(domain, limit=5)
+            kb_sections.append(snippet)
+        kb_ctx = "\n\n".join(kb_sections)
+        stats = self._kb.summary_by_domain()
+        stats_str = " | ".join(f"{d}: {stats.get(d, 0)}条" for d in ("tech", "market", "management", "delivery"))
+        return (
+            base
+            + f"\n\n【知识库统计】{stats_str}\n\n"
+            + kb_ctx
+            + "\n\n"
+            + BOSS_SYSTEM
+        )
+
+    def _extract_kb_facts(self, report_text: str) -> None:
+        """解析 [记忆] kb.{domain}: 内容，写入知识库。"""
+        for m in re.finditer(r"\[记忆\]\s+kb\.(\w+):\s*(.+)", report_text):
+            domain, content = m.group(1), m.group(2).strip()
+            if domain in ("tech", "market", "management", "delivery"):
+                self._kb.store(
+                    domain=domain,
+                    category="boss-decision",
+                    title=f"Boss决策 {__import__('datetime').datetime.now().strftime('%Y-%m-%d')}",
+                    content=content,
+                    source="boss",
+                )
+                logger.info("Boss 写入知识库 [%s]: %s", domain, content[:80])
+
+    def _dispatch_coordination(self, report_text: str, task: Task) -> None:
+        """解析 [协调] 标签，创建跨部门协调任务。"""
+        for m in re.finditer(r"\[协调\]\s*(.+)", report_text):
+            desc = m.group(1).strip()[:200]
+            logger.info("Boss 协调任务: %s", desc)
+            # 协调任务写入 project_delivery 队列，交部门长处理
+            self._queue.enqueue(Task(
+                project="company",
+                type="coordination",
+                title=f"[Boss协调] {desc[:60]}",
+                context=f"Boss 战略协调指令：{desc}",
+                priority=2,
+                mode=task.mode,
+                agent_role="tech-dept-leader",
+                parent_task_id=task.id,
+                initiator="boss",
+                discussion=f"来自 Boss 周报决策",
+            ))
@@ -11,9 +11,9 @@ from typing import Optional

 from pathlib import Path

-from rockchip_agents.config import AgentsConfig
-from rockchip_agents.core.queue import Task
-from rockchip_agents.tools.project_memory import ProjectMemory
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task
+from nmfs_agents.tools.project_memory import ProjectMemory

 logger = logging.getLogger(__name__)

@@ -185,7 +185,7 @@ class DeveloperAgent:

    def _build_cmd(self, task: Task) -> tuple[list[str], list[str]]:
        """返回 (命令列表, 需要清理的临时文件列表)"""
-        from rockchip_agents.config import load_mcp_tools
+        from nmfs_agents.config import load_mcp_tools
        mcp_cfg = load_mcp_tools()
        registry: dict[str, str] = mcp_cfg.get("registry", {})
        role_tools: list[str] = mcp_cfg.get("roles", {}).get(task.agent_role, [])
@@ -200,7 +200,7 @@ class DeveloperAgent:
        tmp_files: list[str] = []

        if task.mode == "report":
-            cmd += ["--allowedTools", "Read,Glob,Grep"]
+            cmd += ["--allowedTools", "Read,Glob,Grep,WebFetch,WebSearch"]
        else:
            # auto/confirm 模式：跳过 claude 内部权限对话框
            cmd += ["--dangerously-skip-permissions"]
@@ -258,13 +258,26 @@ class DeveloperAgent:
            "只有 sequentialthinking 调用完成后才能调用 Write/Edit/Bash。\n\n"
        )

+        spawn_hint = "" if task.mode == "report" else (
+            "【工作流协作】完成代码修改并确认有效变更后，在输出末尾追加：\n"
+            "[SPAWN:tester] 验证: <简述改动要点和测试重点>\n"
+            "（若无实质代码改动，则不输出此标签）\n\n"
+        )
+
+        discussion_section = (
+            f"【任务背景/沟通纪要】\n{task.discussion.strip()}\n\n"
+            if task.discussion and task.discussion.strip() else ""
+        )
+
        return (
            f"{plan_first}"
+            f"{spawn_hint}"
            f"你是 {task.project} 项目的 {task.agent_role} Agent。\n"
            f"项目路径: {proj_path}。\n"
            f"项目描述: {proj_desc}。\n"
            f"当前模式: {task.mode} — {mode_desc}\n"
            f"{mem_section}"
+            f"{discussion_section}"
            f"请用中文回复，保持简洁，聚焦问题。"
        )

@@ -0,0 +1,152 @@
+from __future__ import annotations
+
+"""group_leader.py — L2/L3 组长基类。
+
+职责：
+- 协调组内成员工作，不直接写业务代码
+- 解析成员输出的 [越级汇报] 标签并转发
+- 周期性生成组内状态摘要，向 L1 汇报
+- 根据进展动态调整成员任务优先级
+- 组长自身通过 SPAWN 派发任务给具体成员
+"""
+
+import logging
+import re
+from typing import Optional
+
+from nmfs_agents.agents.developer import AgentResult, DeveloperAgent
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+GROUP_LEADER_SYSTEM = """你是一个技术组的组长（L2 级别）。
+
+## 职责范围
+- 协调组内成员（L3）工作，解决跨成员依赖和阻塞
+- 向上向 L1 部门长汇报组内进展
+- 识别并上报重大技术风险或机会（用 [越级汇报] 标签）
+- **不直接写业务代码**，但可以做技术方案评审
+
+## 组内管理原则
+- 管理而非控制：给方向和支持，不微管理每行代码
+- 识别瓶颈：哪个成员被卡住了？原因是什么？
+- 资源调配：某个任务比其他任务更紧急时，调整优先级
+- 知识流通：成员 A 的发现对成员 B 有价值，主动传递
+
+## 输出格式
+- [状态] 组内整体进展（1-3 句话）
+- [阻塞] <成员/任务>: <阻塞原因>（如有）
+- [调度] <成员角色>: <新的优先级任务>（通过 SPAWN 派发）
+- [向上汇报] <向 L1 部门长的关键信息>
+- [越级汇报] <重大风险或机会>（直达 Boss，谨慎使用）
+- [后续] <下轮组内重点>
+
+## 工作模式
+1. 读取组内各成员最近记忆和任务结果
+2. 识别进展、阻塞、风险
+3. 调整优先级，派发协调任务
+4. 生成向上汇报摘要
+"""
+
+
+class GroupLeaderAgent(DeveloperAgent):
+    """L2/L3 组长 Agent 基类。子类设置 GROUP_NAME / MEMBER_ROLES / DEPT_LEADER_ROLE。"""
+
+    GROUP_NAME: str = "通用组"
+    MEMBER_ROLES: list[str] = []
+    DEPT_LEADER_ROLE: str = "tech-dept-leader"
+
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        super().__init__(config, memory=memory)
+        self._queue = queue or TaskQueue()
+
+    def run(self, task: Task) -> AgentResult:
+        result = super().run(task)
+        if result.status == "done":
+            self._handle_escalation(result.summary, task)
+            self._dispatch_spawned_tasks(result.summary, task)
+        return result
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        # 注入组内成员最近记忆
+        member_section = self._build_member_context()
+        return base + member_section + "\n\n" + GROUP_LEADER_SYSTEM
+
+    def _build_member_context(self) -> str:
+        """汇总组内各成员的 ProjectMemory facts，供组长参考。"""
+        lines = [f"\n【{self.GROUP_NAME} 成员状态汇总】"]
+        for role in self.MEMBER_ROLES:
+            facts = self._memory.get_facts(role) or {}
+            if facts:
+                fact_strs = [f"{k}={v}" for k, v in list(facts.items())[:5]]
+                lines.append(f"  {role}: {', '.join(fact_strs)}")
+            else:
+                lines.append(f"  {role}: (暂无记忆记录)")
+        return "\n".join(lines) + "\n"
+
+    def _handle_escalation(self, report_text: str, task: Task) -> None:
+        """将 [越级汇报] 内容转发至 Boss 队列。"""
+        for m in re.finditer(r"\[越级汇报\]\s*(.+)", report_text):
+            content = m.group(1).strip()[:300]
+            logger.warning("[%s] 越级汇报触发: %s", self.GROUP_NAME, content[:80])
+            self._queue.enqueue(Task(
+                project="company",
+                type="escalation",
+                title=f"[越级汇报][{self.GROUP_NAME}] {content[:60]}",
+                context=content,
+                priority=1,
+                mode=task.mode,
+                agent_role="boss",
+                parent_task_id=task.id,
+                initiator=task.agent_role,
+                discussion=f"来自 {self.GROUP_NAME} 组长越级汇报",
+            ))
+
+    def _dispatch_spawned_tasks(self, report_text: str, task: Task) -> None:
+        """解析 [调度] 标签，为指定成员入队任务。"""
+        for m in re.finditer(r"\[调度\]\s*([a-zA-Z0-9_-]+):\s*(.+)", report_text):
+            role, desc = m.group(1).strip(), m.group(2).strip()[:200]
+            if role not in self.MEMBER_ROLES:
+                logger.debug("[%s] 调度目标 %s 不在组内，跳过", self.GROUP_NAME, role)
+                continue
+            self._queue.enqueue(Task(
+                project=task.project,
+                type="code_improve",
+                title=f"[{self.GROUP_NAME}组长调度] {desc[:60]}",
+                context=desc,
+                priority=task.priority,
+                mode=task.mode,
+                agent_role=role,
+                parent_task_id=task.id,
+                initiator=task.agent_role,
+                discussion=f"{self.GROUP_NAME} 组长协调",
+            ))
+            logger.info("[%s] 组长调度 %s: %s", self.GROUP_NAME, role, desc[:60])
+
+
+# ── 具体组长实现 ────────────────────────────────────────────────────
+
+
+class OsBaseLeaderAgent(GroupLeaderAgent):
+    """OS-Base 组组长（短距通信/内核底层/视频子系统/网络优化）。"""
+
+    GROUP_NAME = "OS-Base"
+    MEMBER_ROLES = ["rtp-researcher", "net-researcher", "kernel-analyzer",
+                    "dev-kernel", "os-engineer"]
+    DEPT_LEADER_ROLE = "tech-dept-leader"
+
+
+class VisionAlgoLeaderAgent(GroupLeaderAgent):
+    """视觉算法组组长（EIS/光流/RKNN 推理优化）。"""
+
+    GROUP_NAME = "视觉算法"
+    MEMBER_ROLES = ["algo-antishake", "algo-vision"]
+    DEPT_LEADER_ROLE = "tech-dept-leader"
@@ -0,0 +1,21 @@
+from __future__ import annotations
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.core.queue import Task
+
+
+class HwEngineerAgent(DeveloperAgent):
+    """芯片选型、外设接口设计、BOM 评审、功耗分析、Datasheet 查阅专职 Agent"""
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        hw_guidance = (
+            "\n【硬件工程规范】\n"
+            "- 芯片选型：从算力(TOPS)、功耗(TDP)、接口(PCIe/USB/MIPI)、成本、供货稳定性五维评估\n"
+            "- Rockchip 芯片系列：RK3588(8TOPS)、RV1106(0.5TOPS)、RK3562(1TOPS)\n"
+            "- BOM 评审：关注关键器件替代方案，避免单一货源风险\n"
+            "- 功耗分析：区分静态功耗与动态功耗，评估散热需求\n"
+            "- Datasheet 查阅：优先使用 fetch 工具获取最新官方文档\n"
+            "- 接口设计：注意信号完整性、阻抗匹配、ESD 保护\n"
+        )
+        return base + hw_guidance
@@ -0,0 +1,20 @@
+from __future__ import annotations
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.core.queue import Task
+
+
+class KernelDevAgent(DeveloperAgent):
+    """Linux 内核模块、驱动开发、Kconfig/Device Tree 专职 Agent"""
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        kernel_guidance = (
+            "\n【内核开发规范】\n"
+            "- 修改 DTS/DTSI 前必须了解上下游节点关系\n"
+            "- 内核模块编译：make -C /lib/modules/$(uname -r)/build M=$(pwd) modules\n"
+            "- 驱动调试：dmesg -T | tail -50；insmod/rmmod 前确认依赖\n"
+            "- Kconfig 变更需同步更新 defconfig\n"
+            "- 内存安全：检查 NULL 指针、use-after-free、buffer overflow\n"
+        )
+        return base + self._build_device_context() + kernel_guidance
@@ -0,0 +1,20 @@
+from __future__ import annotations
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.core.queue import Task
+
+
+class LowlevelDevAgent(DeveloperAgent):
+    """BSP、Bootloader（U-Boot）、裸机/RTOS、硬件 bring-up 专职 Agent"""
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        lowlevel_guidance = (
+            "\n【底层开发规范】\n"
+            "- U-Boot 编译：make CROSS_COMPILE=aarch64-linux-gnu- rockchip_rk3588_defconfig && make\n"
+            "- 裸机调试优先使用 UART 串口输出，不依赖 OS\n"
+            "- 地址空间：严格区分物理地址与虚拟地址，注意 MMU 开关时机\n"
+            "- 硬件 bring-up：先验证电源时序，再验证时钟，最后验证外设\n"
+            "- RTOS 任务优先级设计：中断 > 实时任务 > 普通任务\n"
+        )
+        return base + self._build_device_context() + lowlevel_guidance
@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Optional
+
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.executor import Executor
+from nmfs_agents.core.queue import TaskQueue
+from nmfs_agents.core.scanner import Scanner
+from nmfs_agents.tools.callback_server import CallbackServer
+
+logger = logging.getLogger(__name__)
+
+
+class ManagerAgent:
+    """顶层协调器：扫描 → 入队 → 执行 → 汇报。"""
+
+    def __init__(self, config: AgentsConfig,
+                 queue: Optional[TaskQueue] = None) -> None:
+        self._config = config
+        self._queue = queue or TaskQueue()
+        self._executor = Executor(config, queue=self._queue)
+        self._callback_server: Optional[CallbackServer] = None
+
+    def start_callback_server(self) -> None:
+        if not self._callback_server:
+            self._callback_server = CallbackServer(
+                port=self._config.feishu.callback_port,
+                queue=self._queue,
+            )
+            self._callback_server.start()
+
+    def stop_callback_server(self) -> None:
+        if self._callback_server:
+            self._callback_server.stop()
+            self._callback_server = None
+
+    def scan_and_enqueue(self) -> int:
+        """扫描所有项目，将发现的任务入队，返回新增任务数。"""
+        scanner = Scanner(self._config)
+        results = scanner.scan_all()
+        count = 0
+        for r in results:
+            from nmfs_agents.core.queue import Task
+            tid = self._queue.enqueue(Task(
+                project=r.project, type=r.type, title=r.title,
+                context=r.context, priority=r.priority,
+                mode=r.mode, agent_role=r.agent_role,
+            ))
+            if tid > 0:
+                count += 1
+        logger.info("本轮扫描新增任务: %d", count)
+        return count
+
+    async def run_cycle(self) -> dict:
+        """执行一次完整循环，返回统计信息。"""
+        enqueued = self.scan_and_enqueue()
+        executed = await self._executor.run_all_pending()
+        return {"enqueued": enqueued, "executed": executed}
+
+    def get_status(self) -> dict:
+        """返回任务队列各状态计数。"""
+        conn = self._queue._conn()
+        try:
+            rows = conn.execute(
+                "SELECT status, COUNT(*) as cnt FROM tasks GROUP BY status"
+            ).fetchall()
+        finally:
+            conn.close()
+        return {row["status"]: row["cnt"] for row in rows}
@@ -0,0 +1,32 @@
+from __future__ import annotations
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.core.queue import Task
+
+MEDIA_PRODUCER_SYSTEM = """你是产品洞察与市场落地组的媒体制作师，随时可被触发，兼顾技术演示和市场展示。
+
+技术演示侧：
+- 使用 OpenCV（python3 + cv2）在推理结果图上绘制检测框、置信度、平台水印
+- 命令示例：python3 -c "import cv2; img=cv2.imread('input.jpg'); ..."
+- 输出至 assets/demo/
+
+市场展示侧：
+- 生成产品效果图：Pillow 合成硬件外观图 + 算法效果标注
+- Demo 视频剪辑：ffmpeg -i input.mp4 -vf "drawtext=text='RK3588':..." output.mp4
+- 输出至 assets/market/
+
+工作流：
+1. 确认输入素材路径（推理图/视频/产品图）
+2. 选择合适处理方式（cv2/ffmpeg/Pillow）
+3. 执行并保存输出文件
+4. 报告生成的文件路径和简要说明
+
+注意：随时可触发，不依赖测试结果，同时服务技术和市场两种用途。
+"""
+
+
+class MediaProducerAgent(DeveloperAgent):
+    """产品洞察与市场落地组：技术演示图 + 产品效果图 + Demo 视频"""
+
+    def _build_context(self, task: Task) -> str:
+        return super()._build_context(task) + "\n" + MEDIA_PRODUCER_SYSTEM
@@ -0,0 +1,45 @@
+from __future__ import annotations
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.core.queue import Task
+
+OS_ENGINEER_SYSTEM = """你是项目研发与交付组的系统性能剖析工程师，专注识别和消除系统瓶颈。
+
+职责：不做基础搭建，只做性能问题的深度剖析与优化建议。
+
+四大剖析领域：
+
+【内存分析】
+- 工具：smem -r（进程内存排行）、valgrind --leak-check=full、pmap -x <pid>
+- 目标：定位内存泄漏、OOM killer 触发链、大 allocation 热点
+- 输出：内存占用趋势图 + 泄漏函数调用链
+
+【网络分析】
+- 工具：iperf3 -s/-c（吞吐测试）、ss -tulpn（连接状态）、tc qdisc show（排队策略）
+- 目标：带宽瓶颈、延迟尖峰、包重传率
+- 输出：吞吐/延迟基准数据 + 建议调整的内核参数
+
+【视觉库分析】
+- 工具：OpenCV 帧处理计时（cv2.getTickCount）、NEON/SIMD 利用率（perf stat -e simd）
+- 目标：Mat 内存分配热点、colorspace 转换开销、resize/warp 算子效率
+- 输出：帧处理时序分解 + 高耗时算子替换建议
+
+【AI 算子分析】
+- 工具：RKNN profiler（rknn_server --profile）、rknpu-debug、perf record + flamegraph
+- 目标：NPU 算子耗时排行、内存搬运开销（DDR ↔ NPU SRAM）、算子融合机会
+- 输出：算子热点火焰图 + 量化/融合优化建议
+
+输出规范：
+- 报告：docs/perf_reports/YYYY-MM-DD-{project}.md（含测量命令 + 原始数据 + 建议）
+- 基线：[记忆] __hardware__: perf.{category}.bottleneck={desc}
+- 优化任务：[需求] {project}: 基于剖析结果的具体优化项
+
+请用中文，数据精确，建议可操作。
+"""
+
+
+class OsEngineerAgent(DeveloperAgent):
+    """基础技术研发组：RTOS + Linux OS 基础层"""
+
+    def _build_context(self, task: Task) -> str:
+        return super()._build_context(task) + self._build_device_context() + "\n" + OS_ENGINEER_SYSTEM
@@ -0,0 +1,232 @@
+from __future__ import annotations
+
+import logging
+import os
+import re
+from typing import Optional
+
+from nmfs_agents.agents.developer import (
+    AgentResult, _make_drop_root_preexec,
+    _run_with_log, _task_log_path,
+)
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+TIMEOUT = 600  # planner 是 report 模式，只读分析
+
+_DEFAULT_PLATFORM_GOAL = (
+    "构建 RK3588 NPU 平台上的模块化 AI 推理生态，"
+    "持续提升各推理模块成熟度，探索跨模块整合产品价值。"
+)
+
+PLANNER_SYSTEM = """你是 RK3588 AI 生态团队的产品经理。
+
+你的职责：
+1. 分析平台目标与当前项目状态的差距
+2. 识别跨模块整合机会，提出高价值的下一步需求
+3. 不重复已经在执行或等待中的任务
+
+输出格式（报告末尾，每行一条）：
+- [需求] <项目名>: 具体需求描述（现有项目优化，直接入队执行）
+- [集成] 跨模块整合需求描述（直接入队执行）
+- [目标] <项目名>: 更新后的项目目标（仅当需要修订时）
+- [平台目标] 更新后的平台整体目标（仅当需要修订时）
+- [选题] 新方向描述：价值分析（≥30字）
+
+[选题] 仅用于：
+✓ 当前项目列表中完全不存在的全新技术领域
+✓ 需要新建项目的方向
+✓ 超出现有平台目标范围的探索性方向
+✗ 现有项目的优化/修复 → 用 [需求]
+✗ 现有能力的组合整合 → 用 [集成]
+
+去重原则：若相似需求已在队列中，请勿重复生成。
+请用中文输出，保持简洁。
+"""
+
+
+class PlannerAgent:
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        self._config = config
+        self._queue = queue or TaskQueue()
+        self._memory = memory or ProjectMemory()
+
+    def run(self, task: Task) -> AgentResult:
+        self._ensure_platform_goal()
+        prompt = self._build_prompt()
+        cmd = [
+            "claude", "--print",
+            "--output-format", "json",
+            "--no-session-persistence",
+            "--model", self._config.claude.model,
+            "--allowedTools", "Read,Glob,Grep",
+            "--append-system-prompt", PLANNER_SYSTEM,
+            "--", prompt,
+        ]
+        try:
+            env = self._build_env()
+            result = _run_with_log(
+                cmd, None, env, TIMEOUT, _task_log_path(task.id),
+                preexec_fn=_make_drop_root_preexec(env),
+            )
+            if result.status == "done":
+                self._extract_and_enqueue(result.summary)
+                logger.info("PlannerAgent 分析完成，报告 %d 字", len(result.summary))
+            return result
+        except Exception as e:
+            logger.error("PlannerAgent 运行失败: %s", e)
+            return AgentResult(status="failed", summary=str(e))
+
+    # ── helpers ───────────────────────────────────────────────────────────────
+
+    def _ensure_platform_goal(self) -> None:
+        """首次运行时初始化平台目标。"""
+        if not self._memory.get_goal("__platform__"):
+            try:
+                self._memory.set_goal("__platform__", _DEFAULT_PLATFORM_GOAL, role="planner")
+                logger.info("平台目标已初始化")
+            except Exception as e:
+                logger.warning("平台目标初始化失败: %s", e)
+
+    def _build_prompt(self) -> str:
+        platform_ctx = self._memory.get_context_for_prompt("__platform__", role="planner")
+        proj_names = "、".join(self._config.projects.keys())
+        lines = [
+            f"以下是当前 RK3588 AI 生态团队的状态（可用项目：{proj_names}）：\n",
+            platform_ctx.strip(),
+            "",
+        ]
+        for name, proj_cfg in self._config.projects.items():
+            proj_ctx = self._memory.get_context_for_prompt(name, role="planner")
+            desc = proj_cfg.description if hasattr(proj_cfg, "description") else ""
+            lines.append(f"**{name}**: {desc}")
+            if proj_ctx.strip():
+                lines.append(proj_ctx.strip())
+        # 附上当前队列摘要（防重复生成）
+        pending_titles = self._get_pending_titles()
+        if pending_titles:
+            lines.append(f"\n当前待处理任务（{len(pending_titles)} 条，避免重复）：")
+            for t in pending_titles[:15]:
+                lines.append(f"  - {t}")
+        lines.append(
+            "\n请分析平台目标与当前项目状态的差距，提出 3~5 条最高价值的需求或整合机会。"
+        )
+        return "\n".join(lines)
+
+    def _get_pending_titles(self) -> list[str]:
+        try:
+            conn = self._queue._conn()
+            rows = conn.execute(
+                "SELECT title FROM tasks WHERE status IN ('pending','running') ORDER BY id DESC LIMIT 30"
+            ).fetchall()
+            conn.close()
+            return [r["title"] for r in rows]
+        except Exception:
+            return []
+
+    def _is_duplicate(self, title: str) -> bool:
+        """模糊去重：与任何 pending/running 任务 title 共同前缀超 40 字或完整标题相互包含则视为重复。"""
+        key = title[:40]
+        for existing in self._get_pending_titles():
+            # 两个标题取各自前 40 字，若较短一方与另一方的前缀匹配则视为重复
+            existing_key = existing[:40]
+            min_len = min(len(key), len(existing_key))
+            if min_len > 0 and key[:min_len] == existing_key[:min_len]:
+                return True
+        return False
+
+    def _extract_and_enqueue(self, report_text: str) -> None:
+        # [需求] <project>: <desc>
+        for m in re.finditer(r"\[需求\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            desc = m.group(2).strip()[:120]
+            if project not in self._config.projects:
+                logger.warning("需求任务项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            title = f"[需求] {desc}"
+            if self._is_duplicate(title):
+                logger.info("需求重复跳过: %s", title)
+                continue
+            self._queue.enqueue(Task(
+                project=project, type="feature", title=title,
+                context=f"由 PlannerAgent 从差距分析中提取\n{desc}",
+                priority=3, mode=proj_cfg.mode,
+                agent_role="developer", initiator="planner",
+                discussion="PlannerAgent 基于平台目标-项目现状差距分析生成",
+            ))
+            logger.info("需求任务入队: [%s] %s", project, title)
+
+        # [集成] <desc>  （跨项目，无 project 前缀）
+        for m in re.finditer(r"\[集成\]\s+(.+)", report_text):
+            desc = m.group(1).strip()[:120]
+            title = f"[集成] {desc}"
+            if self._is_duplicate(title):
+                logger.info("集成需求重复跳过: %s", title)
+                continue
+            self._queue.enqueue(Task(
+                project="research", type="architect", title=title,
+                context=f"由 PlannerAgent 提出的跨模块整合机会\n{desc}",
+                priority=3, mode="report",
+                agent_role="architect", initiator="planner",
+                discussion="PlannerAgent 跨项目整合分析",
+            ))
+            logger.info("集成任务入队: %s", title)
+
+        # [目标] <project>: <goal>
+        for m in re.finditer(r"\[目标\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            goal_text = m.group(2).strip()[:300]
+            if project not in self._config.projects:
+                logger.warning("目标更新项目 %r 不在配置中，跳过", project)
+                continue
+            try:
+                self._memory.set_goal(project, goal_text, role="planner")
+                logger.info("项目目标已更新 by planner: [%s]", project)
+            except Exception as e:
+                logger.warning("[%s] 写入项目目标失败: %s", project, e)
+
+        # [平台目标] <goal>
+        for m in re.finditer(r"\[平台目标\]\s+(.+)", report_text):
+            goal_text = m.group(1).strip()[:300]
+            try:
+                self._memory.set_goal("__platform__", goal_text, role="planner")
+                logger.info("平台目标已更新 by planner")
+            except Exception as e:
+                logger.warning("写入平台目标失败: %s", e)
+
+        # [选题] <desc>：新方向/新课题，写入 topic_pending 待人工审批
+        for m in re.finditer(r"\[选题\]\s+(.+?)(?:[：:]\s*(.+))?$", report_text, re.MULTILINE):
+            desc = m.group(1).strip()[:120]
+            context = (m.group(2) or "").strip()[:300]
+            title = f"[选题] {desc}"
+            # 检查是否已有同标题的 topic_pending（去重）
+            existing = self._queue.get_topic_pending_tasks()
+            if any(t.title == title for t in existing):
+                logger.info("选题重复跳过: %s", title)
+                continue
+            tid = self._queue.enqueue_topic(Task(
+                project="research", type="topic", title=title,
+                context=context or f"由 PlannerAgent 提出的新课题方向\n{desc}",
+                priority=3, mode="report",
+                agent_role="planner", initiator="planner",
+                discussion="PlannerAgent 识别为新方向/新课题，需人工确认后执行",
+            ))
+            if tid > 0:
+                logger.info("选题入队待审批 #%d: %s", tid, title)
+
+    def _build_env(self) -> dict:
+        env = os.environ.copy()
+        if self._config.claude.api_key:
+            env["ANTHROPIC_API_KEY"] = self._config.claude.api_key
+        env.pop("CLAUDECODE", None)
+        env.pop("CLAUDE_CODE_ENTRYPOINT", None)
+        return env
@@ -0,0 +1,143 @@
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Optional
+
+from nmfs_agents.agents.developer import AgentResult
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MaturityItem:
+    name: str
+    score: int       # 实际得分
+    max_score: int   # 满分
+    present: bool
+    note: str = ""
+
+
+@dataclass
+class MaturityReport:
+    project: str
+    score: int           # 总分 0-100
+    items: list[MaturityItem] = field(default_factory=list)
+    suggestions: list[str] = field(default_factory=list)
+
+    def summary_text(self) -> str:
+        lines = [f"# {self.project} 成熟度报告", f"总分: {self.score}/100", ""]
+        lines.append("## 各项评分")
+        for item in self.items:
+            icon = "✅" if item.present else "❌"
+            lines.append(f"{icon} {item.name}: {item.score}/{item.max_score}"
+                         + (f" — {item.note}" if item.note else ""))
+        if self.suggestions:
+            lines.append("\n## 改进建议")
+            for s in self.suggestions:
+                lines.append(f"- {s}")
+        return "\n".join(lines)
+
+
+class ProductizerAgent:
+    """基于规则评估项目产品化成熟度，生成改进建议。"""
+
+    CHECKS: list[tuple[str, int]] = [
+        ("README.md",       20),   # 项目说明文档
+        ("RELEASE.md",      15),   # 发布记录
+        ("pyproject.toml",  10),   # 打包配置
+        ("tests/",          20),   # 测试目录
+        ("DEVELOP.md",      10),   # 开发指南
+        ("src/",            10),   # 标准源码结构
+        ("configs/",         5),   # 配置文件
+        ("scripts/",         5),   # 工具脚本
+        ("Makefile",         5),   # 构建入口
+    ]
+
+    def __init__(self, config: AgentsConfig,
+                 memory: Optional[ProjectMemory] = None) -> None:
+        self._config = config
+        self._memory = memory or ProjectMemory()
+        self.last_report: Optional[MaturityReport] = None
+
+    def run(self, task: Task) -> AgentResult:
+        proj_cfg = self._config.projects.get(task.project)
+        if not proj_cfg:
+            return AgentResult(status="failed", summary=f"项目不存在: {task.project}")
+
+        report = self._evaluate(task.project, proj_cfg.path)
+        self.last_report = report
+        summary = report.summary_text()
+        logger.info("[%s] 成熟度评分: %d/100", task.project, report.score)
+
+        # Write initial project goal if not already set (productizer can always update)
+        self._refresh_goal(task.project, proj_cfg.path, report)
+
+        return AgentResult(status="done", summary=summary)
+
+    def _evaluate(self, project: str, path: Path) -> MaturityReport:
+        items: list[MaturityItem] = []
+        total = 0
+
+        for name, max_score in self.CHECKS:
+            target = path / name
+            present = target.exists()
+            score = max_score if present else 0
+
+            # 额外细粒度检查
+            note = ""
+            if present and name == "README.md":
+                text = target.read_text(errors="ignore")
+                has_quickstart = any(kw in text for kw in ["快速开始", "Quick Start", "quickstart", "## 安装", "## Install"])
+                if not has_quickstart:
+                    score = max_score // 2
+                    note = "缺少快速开始章节"
+            if present and name == "tests/":
+                test_files = list(target.rglob("test_*.py")) + list(target.rglob("*_test.py"))
+                if not test_files:
+                    score = max_score // 2
+                    note = "tests/ 目录为空"
+                else:
+                    note = f"{len(test_files)} 个测试文件"
+
+            items.append(MaturityItem(name=name, score=score,
+                                      max_score=max_score, present=present, note=note))
+            total += score
+
+        suggestions = self._generate_suggestions(items, path)
+        return MaturityReport(project=project, score=total,
+                              items=items, suggestions=suggestions)
+
+    def _generate_suggestions(self, items: list[MaturityItem], path: Path) -> list[str]:
+        suggestions = []
+        for item in items:
+            if not item.present:
+                suggestions.append(f"添加 {item.name}（可提升 {item.max_score} 分）")
+            elif item.score < item.max_score:
+                suggestions.append(f"完善 {item.name}：{item.note}")
+        return suggestions
+
+    def _refresh_goal(self, project: str, path: Path,
+                      report: MaturityReport) -> None:
+        """Write a structured project goal to memory.
+
+        The goal captures the project's current purpose and maturity status so
+        that future agents always work with an up-to-date objective.
+        """
+        desc = self._config.projects[project].description
+        score_line = f"当前成熟度评分 {report.score}/100"
+        top_suggestions = "；".join(report.suggestions[:3]) if report.suggestions else "暂无改进建议"
+
+        if desc:
+            goal_text = f"{desc}。{score_line}。近期改进方向：{top_suggestions}。"
+        else:
+            goal_text = f"{project} 项目。{score_line}。近期改进方向：{top_suggestions}。"
+
+        try:
+            self._memory.set_goal(project, goal_text, role="productizer")
+        except Exception as e:
+            logger.warning("[%s] 写入项目目标失败: %s", project, e)
@@ -0,0 +1,68 @@
+from __future__ import annotations
+
+"""senior_dev.py — 高级研发 Agent（L3 Senior Developer）。
+
+职责：
+- 平衡「高质量代码实现」与「架构视角」
+- 可主动提出架构改进建议（[架构建议] 标签）
+- 技术决策有依据，关注先进性（通过 web search 跟踪技术前沿）
+- 普通任务以代码质量为主；发现架构问题时触发讨论
+- 可接受 senior-dev 任务，也可接受 developer 任务
+"""
+
+import logging
+from typing import Optional
+
+from nmfs_agents.agents.developer import DeveloperAgent
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+SENIOR_DEV_SYSTEM = """你是一名高级研发工程师（Senior Developer，L3）。
+
+## 与普通研发的区别
+- **架构视角**：不只写功能代码，还思考设计是否合理、可扩展
+- **技术先进性**：关注业界最新方案，通过 web search 验证是否有更好的实现
+- **代码质量**：类型注解、测试覆盖、可读性，高于普通标准
+- **主动发现**：遇到技术债务或架构隐患，主动提出而不忽视
+
+## 工作模式
+1. **理解需求**：明确本次任务边界和验收标准
+2. **技术调研**（如需要）：用 web search 看业界最优实践
+3. **实现**：写高质量代码，完整类型注解
+4. **自测**：写/完善 pytest 测试，覆盖边界情况
+5. **架构反思**：此次修改是否引入了新的技术债务？是否有更好的设计？
+
+## 输出标签
+- [架构建议] <发现的架构问题或改进机会>（会触发讨论任务）
+- [技术讨论] <值得团队讨论的技术方案>
+- [越级汇报] <重大技术风险>（谨慎使用）
+- [记忆] <项目名>: key=value（记录关键技术事实）
+- [后续] <下一步建议>
+
+## 编码规范（严格执行）
+- `from __future__ import annotations`
+- 完整类型注解；路径用 `pathlib.Path`
+- 日志用 `logging`，禁止 `print`
+- 新功能必须有对应 pytest 测试
+- 不引入未经评估的外部依赖
+"""
+
+
+class SeniorDevAgent(DeveloperAgent):
+    """高级研发 Agent：平衡代码实现与架构视角。"""
+
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        super().__init__(config, memory=memory)
+        self._queue = queue or TaskQueue()
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        return base + "\n\n" + SENIOR_DEV_SYSTEM
@@ -0,0 +1,75 @@
+from __future__ import annotations
+
+import logging
+import re
+from typing import Optional
+
+from nmfs_agents.agents.developer import AgentResult, DeveloperAgent
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+SYSTEM_TESTER_SYSTEM = """你是项目研发与交付组的系统测试工程师。
+
+测试五层能力：
+1. 功能测试：运行 pytest，SSH 板端自动执行，检查通过/失败
+2. 视觉测试：使用 Playwright 截图推理界面，分析检测框/FPS/误检，截图保存至 docs/test_reports/
+3. 性能测试：测量 FPS、推理延迟、内存占用，与目标值对比，写入 [记忆]
+4. 跨平台测试：RK3588/全志/ESP32 分平台调度执行，记录差异
+5. 测试报告：生成结构化 Markdown 报告至 docs/test_reports/YYYY-MM-DD-{project}.md
+
+工具使用：
+- Playwright：browser_navigate → 推理演示页 → browser_take_screenshot → 分析截图
+- SSH 板端：通过 Bash 执行 ssh user@host "pytest tests/ -q"
+- 性能测量：time/perf_counter 或板端 top/free 命令
+
+失败处理：
+- 测试失败 → 输出 [需求] <项目>: 具体问题描述，触发 developer 修复
+- 性能退化 → [记忆] <项目>: fps_baseline=N，记录当前基线
+
+请用中文，数值精确，测试结论明确。
+"""
+
+
+class SystemTesterAgent(DeveloperAgent):
+    """项目交付组：多层系统测试（功能+视觉+性能+跨平台+报告）"""
+
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        super().__init__(config, memory=memory)
+        self._queue = queue or TaskQueue()
+
+    def run(self, task: Task) -> AgentResult:
+        result = super().run(task)
+        if result.status == "done":
+            self._dispatch_require_tasks(result.summary, task)
+        return result
+
+    def _build_context(self, task: Task) -> str:
+        base = super()._build_context(task)
+        return base + self._build_device_context() + "\n" + SYSTEM_TESTER_SYSTEM
+
+    def _dispatch_require_tasks(self, report_text: str, task: Task) -> None:
+        """解析 [需求] 标签，将测试发现的问题派发给 developer 修复。"""
+        for m in re.finditer(r"\[需求\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            desc = m.group(2).strip()[:120]
+            child = Task(
+                project=project,
+                type="fix_bug",
+                title=f"[修复] {desc}",
+                agent_role="developer",
+                priority=task.priority,
+                mode=task.mode,
+                parent_task_id=task.id,
+                initiator="system-tester",
+                discussion=f"系统测试发现问题（任务#{task.id} {task.title}）:\n{desc}",
+            )
+            self._queue.enqueue(child)
+            logger.info("系统测试派发修复任务: project=%s, title=%s", project, child.title)
@@ -0,0 +1,51 @@
+from __future__ import annotations
+
+import logging
+
+from nmfs_agents.agents.developer import AgentResult
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task
+from nmfs_agents.tools.device_agent import DeviceAgent
+from nmfs_agents.tools.file_tools import run_command
+
+logger = logging.getLogger(__name__)
+
+
+class TesterAgent:
+    def __init__(self, config: AgentsConfig) -> None:
+        self._config = config
+
+    def run(self, task: Task) -> AgentResult:
+        proj_cfg = self._config.projects.get(task.project)
+        if not proj_cfg:
+            return AgentResult(status="failed", summary=f"项目不存在: {task.project}")
+
+        lines = []
+        # 1. 本地测试
+        local_result = run_command(
+            "source venv/bin/activate && pytest tests/ -v --tb=short -q",
+            cwd=str(proj_cfg.path),
+        )
+        lines.append(f"=== 本地测试 ===\n{local_result}")
+
+        # 2. 设备端测试（如有配置）
+        if proj_cfg.device_workspace and self._config.devices:
+            dev_name = next(iter(self._config.devices))
+            dev_cfg = self._config.devices[dev_name]
+            device = DeviceAgent(dev_cfg)
+            # 先同步
+            sync_result = device.sync(str(proj_cfg.path), proj_cfg.device_workspace)
+            lines.append(f"=== 设备同步 ===\n{sync_result}")
+            # 然后运行板端测试
+            board_result = device.run_command(
+                "pytest tests/ -v --tb=short -q",
+                proj_cfg.device_workspace,
+            )
+            lines.append(f"=== 板端测试 ({dev_name}) ===\n{board_result}")
+
+        summary = "\n\n".join(lines)
+        failed = "FAILED" in summary or "ERROR" in summary
+        return AgentResult(
+            status="failed" if failed else "done",
+            summary=summary,
+        )
@@ -0,0 +1,141 @@
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import shutil
+import tempfile
+from typing import Optional
+
+from nmfs_agents.agents.developer import (
+    AgentResult, _make_drop_root_preexec, _run_with_log, _task_log_path,
+)
+from nmfs_agents.config import AgentsConfig
+from nmfs_agents.core.queue import Task, TaskQueue
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+TIMEOUT = 600
+
+VISION_ANALYST_SYSTEM = """你是产品洞察与市场落地组的视觉分析师，基于 Claude Vision 进行多帧视觉质量分析。
+
+工作流程（5帧分析）：
+1. 调用 EdgeValidatorService 截图（优先）或 Playwright 截图
+   - EdgeValidatorService: fetch POST http://<host>:8899/screenshot（见上方设备信息）
+   - Playwright 备用: browser_navigate → browser_take_screenshot
+2. 连续截取 5 帧，间隔 2 秒
+3. 逐帧作为多模态 LLM 分析（你直接看截图）：
+   - 检测框位置合理性（有无明显偏移/错位）
+   - FPS 数值读取（从 OSD/界面文字）
+   - 漏检：应检测目标未出现检测框
+   - 误检：背景被误识别
+4. 跨帧分析：检测稳定性（同一目标的框是否抖动/消失）
+5. 存档：docs/visual_tests/YYYY-MM-DD-{project}-{seq:02d}.png
+
+输出格式：
+- [视觉结果] fps=<N> stability=HIGH/MED/LOW detected=<N> false_pos=<N> pass=true/false
+- [视觉退化] fps_trend=升/降 bbox_jitter=高/中/低
+- [需求] <项目名>: 具体视觉问题描述和改进方向
+
+结合推理日志定量验证：
+- fetch GET http://<host>:8899/logs/yolo/20
+
+请用中文，视觉描述精确，问题定位清晰。
+"""
+
+
+class VisionAnalystAgent:
+    def __init__(
+        self,
+        config: AgentsConfig,
+        queue: Optional[TaskQueue] = None,
+        memory: Optional[ProjectMemory] = None,
+    ) -> None:
+        self._config = config
+        self._queue = queue or TaskQueue()
+        self._memory = memory or ProjectMemory()
+
+    def run(self, task: Task) -> AgentResult:
+        from nmfs_agents.config import load_mcp_tools
+        mcp_cfg = load_mcp_tools()
+        registry = mcp_cfg.get("registry", {})
+
+        mcp_servers: dict = {}
+        for tool in ["playwright", "sequential-thinking"]:
+            if tool in registry:
+                parts = registry[tool].split()
+                if shutil.which(parts[0]):
+                    mcp_servers[tool] = {"command": parts[0], "args": parts[1:]}
+
+        tmp_files: list[str] = []
+        cmd = [
+            "claude", "--print",
+            "--output-format", "json",
+            "--no-session-persistence",
+            "--model", self._config.claude.model,
+            "--dangerously-skip-permissions",
+            "--append-system-prompt", VISION_ANALYST_SYSTEM,
+        ]
+        if mcp_servers:
+            with tempfile.NamedTemporaryFile(
+                mode="w", suffix=".json", delete=False, prefix="mcp_vision_"
+            ) as f:
+                json.dump({"mcpServers": mcp_servers}, f)
+                tmp_files.append(f.name)
+            cmd += ["--mcp-config", tmp_files[-1]]
+        cmd += ["--", self._build_prompt(task)]
+
+        try:
+            env = self._build_env()
+            result = _run_with_log(
+                cmd, None, env, TIMEOUT, _task_log_path(task.id),
+                preexec_fn=_make_drop_root_preexec(env),
+            )
+            if result.status == "done":
+                self._extract_needs(result.summary)
+            return result
+        except Exception as e:
+            logger.error("VisionAnalystAgent 失败: %s", e)
+            return AgentResult(status="failed", summary=str(e))
+        finally:
+            for f in tmp_files:
+                try:
+                    os.unlink(f)
+                except OSError:
+                    pass
+
+    def _build_prompt(self, task: Task) -> str:
+        proj_cfg = self._config.projects.get(task.project)
+        desc = proj_cfg.description if proj_cfg else task.project
+        return (
+            f"项目: {task.project}，描述: {desc}\n"
+            f"任务: {task.title}\n"
+            f"背景: {task.context}\n\n"
+            f"请使用 Playwright 截图并分析推理效果，输出视觉分析报告。"
+        )
+
+    def _extract_needs(self, report_text: str) -> None:
+        for m in re.finditer(r"\[需求\]\s+([\w-]+)\s*[：:]\s*(.+)", report_text):
+            project = m.group(1).strip()
+            desc = m.group(2).strip()[:120]
+            if project not in self._config.projects:
+                logger.warning("视觉分析需求项目 %r 不在配置中，跳过", project)
+                continue
+            proj_cfg = self._config.projects[project]
+            self._queue.enqueue(Task(
+                project=project, type="fix_bug",
+                title=f"[视觉分析] {desc}",
+                context=f"由 VisionAnalystAgent 发现\n{desc}",
+                priority=2, mode=proj_cfg.mode,
+                agent_role="developer", initiator="vision-analyst",
+                discussion="VisionAnalystAgent 视觉分析发现的问题",
+            ))
+
+    def _build_env(self) -> dict:
+        env = os.environ.copy()
+        if self._config.claude.api_key:
+            env["ANTHROPIC_API_KEY"] = self._config.claude.api_key
+        env.pop("CLAUDECODE", None)
+        env.pop("CLAUDE_CODE_ENTRYPOINT", None)
+        return env
@@ -0,0 +1,213 @@
+from __future__ import annotations
+
+import logging
+import os
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Optional
+
+import yaml
+
+logger = logging.getLogger(__name__)
+
+CONFIG_DIR = Path(__file__).parent.parent.parent / "configs"
+
+
+@dataclass
+class ProjectConfig:
+    path: Path
+    mode: str = "report"           # auto | confirm | report
+    kind: str = "host"             # embedded | host
+    device_workspace: Optional[str] = None
+    description: str = ""
+
+
+@dataclass
+class QueueConfig:
+    db: str                          # 相对于项目根目录的 DB 路径
+    max_concurrent: int = 3
+    scan_interval_hours: int = 1
+    roles: frozenset[str] = field(default_factory=frozenset)
+
+
+@dataclass
+class SchedulerConfig:
+    interval_hours: int = 1
+    max_concurrent: int = 3
+
+
+@dataclass
+class ClaudeConfig:
+    model: str = "claude-sonnet-4-6"
+    api_key: str = field(default_factory=lambda: os.environ.get("ANTHROPIC_API_KEY", ""))
+
+
+@dataclass
+class FeishuConfig:
+    webhook_url: str = field(default_factory=lambda: os.environ.get("FEISHU_WEBHOOK_URL", ""))
+    bot_token: str = field(default_factory=lambda: os.environ.get("FEISHU_BOT_TOKEN", ""))
+    callback_port: int = 8765
+
+
+@dataclass
+class DeviceConfig:
+    host: str = ""
+    user: str = ""
+    password: str = ""
+    workspace: str = ""
+    pre_cmd: str = ""
+    type: str = "linux"      # linux | esp32 | nordic
+    connect: str = "ssh"     # ssh | serial | gateway
+    via: str = ""            # gateway 时的中转设备名（对应 devices 中的 key）
+    port: str = ""           # 串口路径，如 /dev/ttyUSB0
+    baud: int = 115200
+
+
+@dataclass
+class AgentsConfig:
+    projects: dict[str, ProjectConfig]
+    scheduler: SchedulerConfig
+    claude: ClaudeConfig
+    feishu: FeishuConfig
+    devices: dict[str, DeviceConfig]
+    claude_wx_url: str = ""  # 可选：claude-wx 服务端 URL，用于 agent dashboard 通知
+    queues: dict[str, "QueueConfig"] = field(default_factory=dict)
+
+
+def load_config(config_dir: Path | None = None) -> AgentsConfig:
+    """Load configuration, merging auto-discovered projects with explicit overrides.
+
+    projects.yaml may contain:
+    - ``scan_dirs``: list of paths to scan for Claude-managed projects
+    - ``projects``: explicit per-project config (overrides auto-discovered values;
+      ``path`` is optional if the project is found via scan_dirs)
+    """
+    d = config_dir or CONFIG_DIR
+    with open(d / "projects.yaml") as f:
+        proj_data = yaml.safe_load(f) or {}
+    with open(d / "agents.yaml") as f:
+        agent_data = yaml.safe_load(f) or {}
+    with open(d / "devices.yaml") as f:
+        dev_data = yaml.safe_load(f) or {}
+
+    # Parse devices first (needed for device_workspace inference)
+    devices: dict[str, DeviceConfig] = {
+        name: DeviceConfig(
+            host=v.get("host", ""),
+            user=v.get("user", ""),
+            password=str(v.get("password", "")),
+            workspace=v.get("workspace", ""),
+            pre_cmd=v.get("pre_cmd", ""),
+            type=v.get("type", "linux"),
+            connect=v.get("connect", "ssh"),
+            via=v.get("via", ""),
+            port=v.get("port", ""),
+            baud=int(v.get("baud", 115200)),
+        )
+        for name, v in (dev_data.get("devices") or {}).items()
+    }
+
+    # ── step 1: auto-discover from scan_dirs ──────────────────────────────────
+    discovered: dict[str, ProjectConfig] = {}
+    scan_dirs = proj_data.get("scan_dirs") or []
+    if scan_dirs:
+        from nmfs_agents.tools.project_discovery import ProjectDiscovery
+        discovery = ProjectDiscovery(devices=devices)
+        for raw_path in scan_dirs:
+            target = Path(raw_path)
+            for disc in discovery.scan(target):
+                discovered[disc.name] = ProjectConfig(
+                    path=disc.path,
+                    mode=disc.mode,
+                    kind=disc.kind,
+                    device_workspace=disc.device_workspace,
+                    description=disc.description,
+                )
+                logger.debug("自动发现项目: %s (%s) at %s",
+                             disc.name, disc.kind, disc.path)
+
+    # ── step 1b: remove excluded projects ────────────────────────────────────
+    exclude_set = set(proj_data.get("exclude_projects") or [])
+    for name in list(discovered.keys()):
+        if name in exclude_set:
+            del discovered[name]
+            logger.debug("已排除项目: %s", name)
+
+    # ── step 2: apply explicit overrides ─────────────────────────────────────
+    explicit: dict[str, ProjectConfig] = {}
+    for name, v in (proj_data.get("projects") or {}).items():
+        base = discovered.get(name)
+        # path is optional when project was already discovered
+        raw_path = v.get("path")
+        if raw_path:
+            path = Path(raw_path)
+        elif base:
+            path = base.path
+        else:
+            logger.warning("项目 %r 无法确定路径（未在 scan_dirs 中发现，且未指定 path），跳过", name)
+            continue
+        explicit[name] = ProjectConfig(
+            path=path,
+            mode=v.get("mode", base.mode if base else "report"),
+            kind=v.get("kind", base.kind if base else "host"),
+            device_workspace=v.get("device_workspace",
+                                   base.device_workspace if base else None),
+            description=v.get("description", base.description if base else ""),
+        )
+
+    # 只启用显式配置的项目；scan_dirs 仅用于路径解析，不自动引入未列出的项目
+    projects = explicit
+
+    sched = agent_data.get("scheduler", {})
+    claude_d = agent_data.get("claude", {})
+    feishu_d = agent_data.get("feishu", {})
+
+    # 解析多队列配置
+    queues: dict[str, QueueConfig] = {}
+    for q_name, q_data in agent_data.get("queues", {}).items():
+        queues[q_name] = QueueConfig(
+            db=q_data.get("db", f"data/{q_name}_tasks.db"),
+            max_concurrent=q_data.get("max_concurrent", 3),
+            scan_interval_hours=q_data.get("scan_interval_hours", 1),
+            roles=frozenset(q_data.get("roles", [])),
+        )
+
+    return AgentsConfig(
+        projects=projects,
+        scheduler=SchedulerConfig(**sched),
+        claude=ClaudeConfig(model=claude_d.get("model", "claude-sonnet-4-6")),
+        feishu=FeishuConfig(callback_port=feishu_d.get("callback_port", 8765)),
+        devices=devices,
+        claude_wx_url=agent_data.get("claude_wx_url", ""),
+        queues=queues,
+    )
+
+
+def load_mcp_tools(config_dir: Path | None = None) -> dict:
+    """加载 MCP 工具注册表和角色工具映射（支持热更新）"""
+    d = config_dir or CONFIG_DIR
+    path = d / "mcp_tools.yaml"
+    if not path.exists():
+        return {"registry": {}, "roles": {}}
+    with open(path) as f:
+        return yaml.safe_load(f) or {"registry": {}, "roles": {}}
+
+
+def load_role_routing(config_dir: Path | None = None) -> dict[str, list[str]]:
+    """加载角色关键词路由表（返回 role → [keywords] 映射）"""
+    d = config_dir or CONFIG_DIR
+    path = d / "role_routing.yaml"
+    if not path.exists():
+        return {}
+    with open(path) as f:
+        data = yaml.safe_load(f) or {}
+    return data.get("role_keywords", {})
+
+
+def route_role(text: str, routing: dict[str, list[str]]) -> str:
+    """按关键词路由任务到合适的 agent_role，无匹配返回 dev-embedded"""
+    text_lower = text.lower()
+    for role, keywords in routing.items():
+        if any(kw.lower() in text_lower for kw in keywords):
+            return role
+    return "dev-embedded"
@@ -0,0 +1,356 @@
+from __future__ import annotations
+
+import sqlite3
+from dataclasses import dataclass, field
+from datetime import datetime, timedelta
+from pathlib import Path
+from typing import Optional
+
+DEFAULT_DB = Path(__file__).parent.parent.parent.parent / "data" / "tasks.db"
+
+
+@dataclass
+class Task:
+    project: str
+    type: str        # fix_bug | improve_perf | add_feature | code_review | productize | architect
+    title: str
+    priority: int    # 1（严重）→ 5（低）
+    mode: str        # auto | confirm | report
+    agent_role: str  # developer | tester | productizer | architect
+    context: str = ""
+    id: int = 0
+    status: str = "pending"
+    result_summary: str = ""
+    # 制约工作流扩展字段
+    retry_count: int = 0
+    parent_task_id: int = 0
+    constraint_type: str = ""   # "veto_pending"|"approval_pending"|"adversarial"
+    awaiting_role: str = ""     # 串行链中等待哪个角色
+    created_at: str = field(default_factory=lambda: datetime.utcnow().isoformat())
+    # 任务溯源字段
+    initiator: str = "system"   # 发起人：scanner / architect / manager / user / system
+    discussion: str = ""        # 沟通纪要：任务背景、决策依据、相关讨论
+    execution_host: str = "dev"  # dev | edge-api
+
+
+class TaskQueue:
+    def __init__(self, db_path: Path | None = None) -> None:
+        self._db = db_path or DEFAULT_DB
+        self._db.parent.mkdir(parents=True, exist_ok=True)
+        self._init_db()
+
+    def _conn(self) -> sqlite3.Connection:
+        conn = sqlite3.connect(self._db, isolation_level=None)  # autocommit
+        conn.row_factory = sqlite3.Row
+        return conn
+
+    def _init_db(self) -> None:
+        with self._conn() as c:
+            c.execute("PRAGMA journal_mode=WAL")
+            c.execute("""
+                CREATE TABLE IF NOT EXISTS tasks (
+                    id           INTEGER PRIMARY KEY AUTOINCREMENT,
+                    project      TEXT NOT NULL,
+                    type         TEXT NOT NULL,
+                    title        TEXT NOT NULL,
+                    context      TEXT DEFAULT '',
+                    priority     INTEGER DEFAULT 3,
+                    mode         TEXT DEFAULT 'report',
+                    agent_role   TEXT DEFAULT 'developer',
+                    status       TEXT DEFAULT 'pending',
+                    created_at   TEXT,
+                    started_at   TEXT,
+                    completed_at TEXT,
+                    result_summary TEXT DEFAULT '',
+                    retry_count     INTEGER DEFAULT 0,
+                    parent_task_id  INTEGER DEFAULT 0,
+                    constraint_type TEXT DEFAULT '',
+                    awaiting_role   TEXT DEFAULT '',
+                    initiator       TEXT DEFAULT 'system',
+                    discussion      TEXT DEFAULT ''
+                )
+            """)
+        # 迁移：为已存在的旧表添加新列（幂等）
+        with self._conn() as c:
+            for col, col_type, defval in [
+                ("retry_count",     "INTEGER", "0"),
+                ("parent_task_id",  "INTEGER", "0"),
+                ("constraint_type", "TEXT",    "''"),
+                ("awaiting_role",   "TEXT",    "''"),
+                ("initiator",       "TEXT",    "'system'"),
+                ("discussion",      "TEXT",    "''"),
+                ("execution_host",  "TEXT",    "'dev'"),
+            ]:
+                try:
+                    c.execute(f"ALTER TABLE tasks ADD COLUMN {col} {col_type} DEFAULT {defval}")
+                except sqlite3.OperationalError:
+                    pass  # 列已存在
+
+    def _row_to_task(self, row) -> Task:
+        return Task(
+            id=row["id"], project=row["project"], type=row["type"],
+            title=row["title"], context=row["context"], priority=row["priority"],
+            mode=row["mode"], agent_role=row["agent_role"], status=row["status"],
+            retry_count=int(row["retry_count"] or 0),
+            parent_task_id=row["parent_task_id"] or 0,
+            constraint_type=row["constraint_type"] or "",
+            awaiting_role=row["awaiting_role"] or "",
+            initiator=row["initiator"] or "system",
+            discussion=row["discussion"] or "",
+            execution_host=row["execution_host"] if "execution_host" in row.keys() else "dev",
+        )
+
+    def enqueue(self, task: Task) -> int:
+        """返回 task id；若已存在同 project+type+title 的 pending/running 任务则返回 -1。"""
+        with self._conn() as c:
+            exists = c.execute(
+                "SELECT id FROM tasks WHERE project=? AND type=? AND title=? AND status IN ('pending','running')",
+                (task.project, task.type, task.title),
+            ).fetchone()
+            if exists:
+                return -1
+            cur = c.execute(
+                "INSERT INTO tasks "
+                "(project,type,title,context,priority,mode,agent_role,status,created_at,"
+                "retry_count,parent_task_id,constraint_type,awaiting_role,initiator,discussion,"
+                "execution_host) "
+                "VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)",
+                (task.project, task.type, task.title, task.context,
+                 task.priority, task.mode, task.agent_role, "pending", task.created_at,
+                 task.retry_count, task.parent_task_id, task.constraint_type, task.awaiting_role,
+                 task.initiator, task.discussion, task.execution_host),
+            )
+            return cur.lastrowid
+
+    def dequeue(
+        self,
+        exclude_projects: frozenset[str] | None = None,
+        role_filter: frozenset[str] | None = None,
+    ) -> Optional[Task]:
+        """原子化取出优先级最高的 pending 任务并标记为 running。
+        exclude_projects: 跳过这些项目的任务（用于项目级暂停）。
+        role_filter: 只取这些角色的任务（用于多队列隔离）。
+        """
+        conn = self._conn()
+        try:
+            conn.execute("BEGIN EXCLUSIVE")
+            # 动态构建 WHERE 子句
+            conditions = ["status='pending'"]
+            params: list = []
+            if exclude_projects:
+                placeholders = ",".join("?" * len(exclude_projects))
+                conditions.append(f"project NOT IN ({placeholders})")
+                params.extend(tuple(exclude_projects))
+            if role_filter:
+                placeholders = ",".join("?" * len(role_filter))
+                conditions.append(f"agent_role IN ({placeholders})")
+                params.extend(tuple(role_filter))
+            where = " AND ".join(conditions)
+            row = conn.execute(
+                f"SELECT * FROM tasks WHERE {where} ORDER BY priority ASC, id ASC LIMIT 1",
+                params,
+            ).fetchone()
+            if not row:
+                conn.execute("COMMIT")
+                return None
+            conn.execute(
+                "UPDATE tasks SET status='running', started_at=? WHERE id=?",
+                (datetime.utcnow().isoformat(), row["id"]),
+            )
+            conn.execute("COMMIT")
+            t = self._row_to_task(row)
+            t.status = "running"
+            return t
+        except Exception:
+            conn.execute("ROLLBACK")
+            raise
+        finally:
+            conn.close()
+
+    def mark_done(self, task_id: int, summary: str = "") -> None:
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='done', completed_at=?, result_summary=? WHERE id=?",
+                (datetime.utcnow().isoformat(), summary, task_id),
+            )
+
+    def mark_failed(self, task_id: int, reason: str = "") -> None:
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='failed', completed_at=?, result_summary=? WHERE id=?",
+                (datetime.utcnow().isoformat(), reason, task_id),
+            )
+
+    def mark_waiting_confirm(self, task_id: int, diff: str = "") -> None:
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='waiting_confirm', result_summary=? WHERE id=?",
+                (diff, task_id),
+            )
+
+    def mark_waiting_approval(self, task_id: int, awaiting_role: str = "") -> None:
+        """进入串行审批链等待状态（不占 Semaphore）"""
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='waiting_approval', awaiting_role=? WHERE id=?",
+                (awaiting_role, task_id),
+            )
+
+    def mark_vetoed(self, task_id: int, reason: str = "") -> None:
+        """被否决：retry_count+1；第 3 次否决时标记为 failed（最多允许 2 次重试，retry_count 上限为 2）"""
+        with self._conn() as c:
+            row = c.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
+            if not row:
+                return
+            retry = int(row["retry_count"] or 0) + 1
+            if retry > 2:
+                c.execute(
+                    "UPDATE tasks SET status='failed', result_summary=? WHERE id=?",
+                    (f"否决次数超限: {reason}", task_id),
+                )
+                return
+            new_context = (
+                f"\u26a0\ufe0f 上次执行被否决（第 {retry} 次重试）\n"
+                f"原因：{reason}\n\n"
+                f"--- 原始任务上下文 ---\n"
+                f"{row['context']}"
+            )
+            c.execute(
+                "UPDATE tasks SET status='pending', retry_count=?, context=?, "
+                "awaiting_role='', constraint_type='veto_pending' WHERE id=?",
+                (retry, new_context, task_id),
+            )
+
+    def mark_adversarial_pending(self, task_id: int) -> None:
+        """对抗式提案等待 SELECT"""
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='adversarial_pending', "
+                "constraint_type='adversarial' WHERE id=?",
+                (task_id,),
+            )
+
+    def approve_waiting(self, task_id: int) -> None:
+        """审批通过：waiting_approval → done"""
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='done', awaiting_role='' WHERE id=?",
+                (task_id,),
+            )
+
+    def reset_stale_running(self, timeout_minutes: int = 120) -> int:
+        """将超时的 running 任务标记为 failed，返回处理数量。
+
+        timeout_minutes=0 表示重置所有 running 任务（用于进程重启后的恢复）。
+        """
+        with self._conn() as c:
+            if timeout_minutes == 0:
+                rows = c.execute(
+                    "SELECT id FROM tasks WHERE status='running'"
+                ).fetchall()
+            else:
+                cutoff = (datetime.utcnow() - timedelta(minutes=timeout_minutes)).isoformat()
+                rows = c.execute(
+                    "SELECT id FROM tasks WHERE status='running' AND started_at < ?",
+                    (cutoff,),
+                ).fetchall()
+            if not rows:
+                return 0
+            ids = [r["id"] for r in rows]
+            placeholders = ",".join("?" * len(ids))
+            reason = "执行超时（Watchdog 检测）" if timeout_minutes > 0 else "进程重启恢复"
+            c.execute(
+                f"UPDATE tasks SET status='failed', completed_at=?, result_summary=? "
+                f"WHERE id IN ({placeholders})",
+                [datetime.utcnow().isoformat(), reason, *ids],
+            )
+            return len(ids)
+
+    def bump_stale_priorities(self, threshold_minutes: int = 120) -> int:
+        """将长时间等待的 pending 任务优先级提升（priority-1，最低为 1），返回处理数量。"""
+        cutoff = (datetime.utcnow() - timedelta(minutes=threshold_minutes)).isoformat()
+        with self._conn() as c:
+            rows = c.execute(
+                "SELECT id, priority FROM tasks "
+                "WHERE status='pending' AND priority > 1 AND created_at < ?",
+                (cutoff,),
+            ).fetchall()
+            for r in rows:
+                c.execute(
+                    "UPDATE tasks SET priority=? WHERE id=?",
+                    (r["priority"] - 1, r["id"]),
+                )
+            return len(rows)
+
+    def get_waiting_approval_tasks(self) -> list[Task]:
+        """获取所有 waiting_approval 状态的任务"""
+        with self._conn() as c:
+            rows = c.execute(
+                "SELECT * FROM tasks WHERE status='waiting_approval'"
+            ).fetchall()
+        return [self._row_to_task(r) for r in rows]
+
+
+    def enqueue_topic(self, task: Task) -> int:
+        """将选题任务写入 topic_pending 状态（不走普通 pending 去重检查）。"""
+        with self._conn() as c:
+            exists = c.execute(
+                "SELECT id FROM tasks WHERE project=? AND title=? AND status='topic_pending'",
+                (task.project, task.title),
+            ).fetchone()
+            if exists:
+                return -1
+            cur = c.execute(
+                "INSERT INTO tasks "
+                "(project,type,title,context,priority,mode,agent_role,status,created_at,"
+                "initiator,discussion,execution_host) "
+                "VALUES (?,?,?,?,?,?,?,'topic_pending',?,?,?,?)",
+                (task.project, task.type, task.title, task.context,
+                 task.priority, task.mode, task.agent_role, task.created_at,
+                 task.initiator, task.discussion, task.execution_host),
+            )
+            return cur.lastrowid
+
+    def approve_topic(self, task_id: int) -> None:
+        """用户确认选题：topic_pending → pending"""
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='pending' WHERE id=? AND status='topic_pending'",
+                (task_id,),
+            )
+
+    def reject_topic(self, task_id: int, reason: str = "") -> None:
+        """用户否决选题：topic_pending → vetoed"""
+        with self._conn() as c:
+            c.execute(
+                "UPDATE tasks SET status='vetoed', result_summary=? WHERE id=? AND status='topic_pending'",
+                (reason, task_id),
+            )
+
+    def get_topic_pending_tasks(self) -> list[Task]:
+        """返回所有待审批选题。"""
+        with self._conn() as c:
+            rows = c.execute(
+                "SELECT * FROM tasks WHERE status='topic_pending' ORDER BY created_at DESC"
+            ).fetchall()
+        return [self._row_to_task(r) for r in rows]
+
+    def clear_project_tasks(self, project: str, statuses: list[str] | None = None) -> int:
+        """删除指定项目的任务记录，默认删除全部状态。返回删除数量。"""
+        with self._conn() as c:
+            if statuses:
+                placeholders = ",".join("?" * len(statuses))
+                rows = c.execute(
+                    f"SELECT id FROM tasks WHERE project=? AND status IN ({placeholders})",
+                    [project, *statuses],
+                ).fetchall()
+            else:
+                rows = c.execute(
+                    "SELECT id FROM tasks WHERE project=?", (project,)
+                ).fetchall()
+            if not rows:
+                return 0
+            ids = [r["id"] for r in rows]
+            placeholders = ",".join("?" * len(ids))
+            c.execute(f"DELETE FROM tasks WHERE id IN ({placeholders})", ids)
+            return len(ids)
@@ -0,0 +1,192 @@
+from __future__ import annotations
+
+import logging
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+from nmfs_agents.config import AgentsConfig, ProjectConfig
+from nmfs_agents.tools.project_memory import ProjectMemory
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ScanResult:
+    project: str
+    type: str        # fix_bug | improve_perf | code_review
+    title: str
+    context: str
+    priority: int
+    mode: str
+    agent_role: str = "developer"
+    initiator: str = "scanner"
+    discussion: str = ""
+
+
+class Scanner:
+    def __init__(self, config: AgentsConfig,
+                 memory: ProjectMemory | None = None) -> None:
+        self._config = config
+        self._memory = memory or ProjectMemory()
+        from nmfs_agents.config import load_role_routing
+        self._role_routing = load_role_routing()
+
+    def scan_all(self) -> list[ScanResult]:
+        results: list[ScanResult] = []
+        for name, proj_cfg in self._config.projects.items():
+            if not proj_cfg.path.exists():
+                logger.warning("项目路径不存在: %s", proj_cfg.path)
+                continue
+            results.extend(self.scan_project(name, proj_cfg))
+        return results
+
+    def scan_project(self, name: str, cfg: ProjectConfig) -> list[ScanResult]:
+        results: list[ScanResult] = []
+        results.extend(self._scan_readme(name, cfg))
+        results.extend(self._scan_todos(name, cfg))
+        results.extend(self._scan_claude_md(name, cfg))
+        results.extend(self._scan_description(name, cfg))
+        results.extend(self._scan_planned_features(name, cfg))
+        results.extend(self._scan_periodic(name, cfg))
+        # Append long-term goal to every task context so agents stay aligned
+        goal_ctx = self._memory.get_context_for_prompt(name)
+        if goal_ctx:
+            for r in results:
+                r.context = r.context + goal_ctx
+        return results
+
+    def _make(self, project: str, cfg: ProjectConfig, type_: str,
+              title: str, context: str, priority: int,
+              agent_role: str | None = None) -> ScanResult:
+        if agent_role is None:
+            from nmfs_agents.config import route_role
+            agent_role = route_role(f"{title} {context}", self._role_routing)
+        return ScanResult(
+            project=project, type=type_, title=title,
+            context=context, priority=priority, mode=cfg.mode,
+            agent_role=agent_role,
+        )
+
+    def _scan_readme(self, project: str, cfg: ProjectConfig) -> list[ScanResult]:
+        results = []
+        readme = cfg.path / "README.md"
+        if not readme.exists():
+            return results
+        text = readme.read_text(encoding="utf-8", errors="ignore")
+        # 查找"已知问题"章节
+        section = re.search(r"##\s*(已知问题|Known Issues?|known issues?)(.*?)(?=\n##|\Z)",
+                            text, re.DOTALL | re.IGNORECASE)
+        if not section:
+            return results
+        for line in section.group(2).splitlines():
+            line = line.strip("- *\t ")
+            if len(line) > 5:
+                results.append(self._make(project, cfg, "fix_bug",
+                                          f"[README] {line[:80]}", line, priority=2))
+        return results
+
+    _SKIP_DIRS = frozenset({
+        "venv", ".venv", ".git", "__pycache__", "node_modules",
+        ".tox", "build", "dist", "*.egg-info",
+    })
+
+    def _scan_todos(self, project: str, cfg: ProjectConfig) -> list[ScanResult]:
+        results = []
+        src_dirs = [cfg.path / "src", cfg.path / "tools", cfg.path]
+        pattern = re.compile(r"#\s*(TODO|FIXME|HACK)[:\s]+(.*)", re.IGNORECASE)
+        seen: set[str] = set()
+        for src_dir in src_dirs:
+            if not src_dir.is_dir():
+                continue
+            for py_file in self._iter_py_files(src_dir):
+                try:
+                    rel = py_file.relative_to(cfg.path)
+                    for i, line in enumerate(py_file.read_text(errors="ignore").splitlines(), 1):
+                        m = pattern.search(line)
+                        if m:
+                            key = m.group(2).strip()[:60]
+                            if key not in seen:
+                                seen.add(key)
+                                results.append(self._make(
+                                    project, cfg, "code_review",
+                                    f"TODO: {key}",
+                                    f"{rel}:{i} — {line.strip()}",
+                                    priority=4,
+                                ))
+                except OSError:
+                    pass
+        return results
+
+    def _iter_py_files(self, root: Path):
+        """递归遍历 .py 文件，跳过 venv/__pycache__ 等无关目录。"""
+        for entry in root.iterdir():
+            if entry.is_dir():
+                if entry.name in self._SKIP_DIRS or entry.name.endswith(".egg-info"):
+                    continue
+                yield from self._iter_py_files(entry)
+            elif entry.suffix == ".py":
+                yield entry
+
+    def _scan_claude_md(self, project: str, cfg: ProjectConfig) -> list[ScanResult]:
+        """从 CLAUDE.md 的关键技术点/关键约束中提取潜在优化任务。"""
+        results = []
+        cfile = cfg.path / "CLAUDE.md"
+        if not cfile.exists():
+            return results
+        text = cfile.read_text(encoding="utf-8", errors="ignore")
+        # 查找包含 ❌ 或 ⚠ 的行（表示有问题的项目）
+        for line in text.splitlines():
+            if ("❌" in line or "⚠" in line) and len(line.strip()) > 5:
+                title = re.sub(r"[❌⚠️\|\s]+", " ", line).strip()[:80]
+                if title:
+                    results.append(self._make(project, cfg, "fix_bug",
+                                              f"[CLAUDE.md] {title}", line, priority=2))
+        return results
+
+    _BUG_KEYWORDS = frozenset(["待修复", "bug", "问题", "错误", "失败", "零检出", "报错"])
+    _PERF_KEYWORDS = frozenset(["fps偏低", "fps 偏低", "性能偏低", "偏低", "延迟高", "速度慢"])
+
+    def _scan_description(self, project: str, cfg: ProjectConfig) -> list[ScanResult]:
+        """从 projects.yaml description 字段提取问题任务。"""
+        desc = cfg.description.strip()
+        if not desc:
+            return []
+        desc_lower = desc.lower()
+        if any(kw in desc_lower for kw in self._BUG_KEYWORDS):
+            return [self._make(project, cfg, "fix_bug",
+                               f"[描述] {desc[:80]}", desc, priority=2)]
+        if any(kw in desc_lower for kw in self._PERF_KEYWORDS):
+            return [self._make(project, cfg, "improve_perf",
+                               f"[描述] {desc[:80]}", desc, priority=3)]
+        return []
+
+    def _scan_periodic(self, project: str, cfg: ProjectConfig) -> list[ScanResult]:
+        """为每个项目生成兜底定期代码巡检任务（优先级最低）。"""
+        return [self._make(
+            project, cfg, "code_review",
+            "定期代码巡检",
+            f"对 {project} 项目进行整体代码质量检查，识别潜在改进点。",
+            priority=5,
+        )]
+
+    def _scan_planned_features(self, project: str, cfg: ProjectConfig) -> list[ScanResult]:
+        """从 CLAUDE.md 的 ## 规划功能 章节提取 new_feature 任务"""
+        results = []
+        cfile = cfg.path / "CLAUDE.md"
+        if not cfile.exists():
+            return results
+        text = cfile.read_text(encoding="utf-8", errors="ignore")
+        section = re.search(
+            r"##\s*规划功能(.*?)(?=\n##|\Z)", text, re.DOTALL
+        )
+        if not section:
+            return results
+        for line in section.group(1).splitlines():
+            line = line.strip("- *\t ")
+            if len(line) > 5:
+                results.append(self._make(
+                    project, cfg, "new_feature",
+                    f"[规划] {line[:80]}", line, priority=3,
+                ))
+        return results
@@ -0,0 +1,64 @@
+from __future__ import annotations
+
+import logging
+import threading
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from nmfs_agents.core.queue import TaskQueue
+
+logger = logging.getLogger(__name__)
+
+WATCHDOG_INTERVAL = 60        # 检查间隔（秒）
+TASK_TIMEOUT_MINUTES = 150    # running 任务超过此时长视为卡死（任务最大超时 120min + 30min 缓冲）
+PRIORITY_AGING_MINUTES = 120  # pending 任务等待超过此时长触发优先级提升
+
+
+class Watchdog(threading.Thread):
+    """后台看门狗线程：定期检测并修复卡死任务，防止队列长期阻塞。
+
+    功能：
+    - 超时任务检测：running > TASK_TIMEOUT_MINUTES → failed（释放 semaphore 名额）
+    - 优先级老化：pending > PRIORITY_AGING_MINUTES → priority -1（防止任务永久饥饿）
+    - 心跳日志：每次 tick 写 DEBUG 日志，便于确认 watchdog 存活
+    """
+
+    def __init__(
+        self,
+        queue: TaskQueue,
+        interval: int = WATCHDOG_INTERVAL,
+        task_timeout_minutes: int = TASK_TIMEOUT_MINUTES,
+    ) -> None:
+        super().__init__(name="Watchdog", daemon=True)
+        self._queue = queue
+        self._interval = interval
+        self._task_timeout = task_timeout_minutes
+        self._stop_event = threading.Event()
+
+    def run(self) -> None:
+        logger.info(
+            "Watchdog 已启动（检查间隔=%ds，任务超时阈值=%dm）",
+            self._interval, self._task_timeout,
+        )
+        # Event.wait(timeout) 在 set() 时立即返回 True，否则等满 timeout 秒返回 False
+        while not self._stop_event.wait(self._interval):
+            try:
+                self._tick()
+            except Exception as e:
+                logger.error("Watchdog tick 异常: %s", e)
+        logger.info("Watchdog 已停止")
+
+    def stop(self) -> None:
+        """通知 Watchdog 退出，调用后线程将在当前 sleep 结束时退出。"""
+        self._stop_event.set()
+
+    def _tick(self) -> None:
+        stale = self._queue.reset_stale_running(self._task_timeout)
+        if stale:
+            logger.warning("Watchdog: %d 个超时任务已重置为 failed", stale)
+
+        aged = self._queue.bump_stale_priorities(PRIORITY_AGING_MINUTES)
+        if aged:
+            logger.info("Watchdog: %d 个长期 pending 任务优先级已提升", aged)
+
+        logger.debug("Watchdog tick OK (stale=%d, aged=%d)", stale, aged)
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" class="iconify iconify--logos" width="31.88" height="32" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 257"><defs><linearGradient id="IconifyId1813088fe1fbc01fb466" x1="-.828%" x2="57.636%" y1="7.652%" y2="78.411%"><stop offset="0%" stop-color="#41D1FF"></stop><stop offset="100%" stop-color="#BD34FE"></stop></linearGradient><linearGradient id="IconifyId1813088fe1fbc01fb467" x1="43.376%" x2="50.316%" y1="2.242%" y2="89.03%"><stop offset="0%" stop-color="#FFEA83"></stop><stop offset="8.333%" stop-color="#FFDD35"></stop><stop offset="100%" stop-color="#FFA800"></stop></linearGradient></defs><path fill="url(#IconifyId1813088fe1fbc01fb466)" d="M255.153 37.938L134.897 252.976c-2.483 4.44-8.862 4.466-11.382.048L.875 37.958c-2.746-4.814 1.371-10.646 6.827-9.67l120.385 21.517a6.537 6.537 0 0 0 2.322-.004l117.867-21.483c5.438-.991 9.574 4.796 6.877 9.62Z"></path><path fill="url(#IconifyId1813088fe1fbc01fb467)" d="M185.432.063L96.44 17.501a3.268 3.268 0 0 0-2.634 3.014l-5.474 92.456a3.268 3.268 0 0 0 3.997 3.378l24.777-5.718c2.318-.535 4.413 1.507 3.936 3.838l-7.361 36.047c-.495 2.426 1.782 4.5 4.151 3.78l15.304-4.649c2.372-.72 4.652 1.36 4.15 3.788l-11.698 56.621c-.732 3.542 3.979 5.473 5.943 2.437l1.313-2.028l72.516-144.72c1.215-2.423-.88-5.186-3.54-4.672l-25.505 4.922c-2.396.462-4.435-1.77-3.759-4.114l16.646-57.705c.677-2.35-1.37-4.583-3.769-4.113Z"></path></svg>