fix(spawner): address PR#36 review feedback (M1+M2+S1+S2)

fix(spawner): §24 compact check must run when status=done (compact in progress)
feat(spawner): §24 compact detection via gateway log rotation events
2026-06-11 21:40:09 +08:00 · 2026-06-11 21:18:33 +08:00 · 2026-06-11 21:18:33 +08:00 · 2026-06-11 06:10:44 +00:00 · 2026-06-11 14:00:44 +08:00 · 2026-06-11 13:25:48 +08:00
52 changed files with 6052 additions and 868 deletions
@@ -1,9 +1,10 @@
 # CI 管道 — moziplus v2.0
 #
 # 触发条件：
-#   - push（非 main 分支）
 #   - pull_request（opened, synchronize）
 #
+# 注意：只保留 pull_request 触发，避免 push + pull_request 双倍触发
+#
 # Gitea v1.23.4 限制注意：
 #   - 不支持 failure() 表达式，用 always() + shell 条件判断替代
 #   - 不支持 concurrency / continue-on-error / timeout-minutes / permissions
@@ -13,10 +14,6 @@
 name: CI

 on:
-  push:
-    branches:
-      - '**'
-      - '!main'
  pull_request:
    types: [opened, synchronize]

@@ -53,7 +50,8 @@ jobs:
          /tmp/ci-venv-test/bin/pytest tests/ -m "not e2e" -x -q

  # ── Job 3: CI 失败通知 ───────────────────────────────
-  # v1.23 不支持 failure()，用 always() + shell 检查 commit status 替代
+  # 使用 needs.<job>.result 直接判断，不查询 commit status API
+  # 根因：notify 自身的 pending status 会污染 commit status 查询结果（竞态条件）
  notify-on-failure:
    runs-on: macos-arm64
    needs: [lint, test]
@@ -62,31 +60,34 @@ jobs:
      - name: Check results and notify
        env:
          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+          LINT_RESULT: ${{ needs.lint.result }}
+          TEST_RESULT: ${{ needs.test.result }}
        run: |
-          # 查询当前 commit 的 status
-          STATUS=$(curl -sf \
-            -H "Authorization: token $GITEA_TOKEN" \
-            "${{ gitea.api_url }}/repos/${{ gitea.repository }}/commits/${{ gitea.sha }}/status" \
-            | python3 -c "import sys,json; print(json.load(sys.stdin).get('state',''))" 2>/dev/null || echo "")
+          echo "Lint result: $LINT_RESULT"
+          echo "Test result: $TEST_RESULT"

-          echo "Commit status: $STATUS"
-
-          if [ "$STATUS" != "success" ]; then
-            echo "CI failed or status unknown, sending notification..."
+          # 只有 lint 或 test 明确失败时才发通知
+          if [ "$LINT_RESULT" = "failure" ] || [ "$TEST_RESULT" = "failure" ]; then
+            echo "CI has failures, sending notification..."

            # 如果是 PR 事件，写评论通知
            PR_NUMBER="${{ gitea.event.pull_request.number }}"
            if [ -n "$PR_NUMBER" ]; then
+              # 构建失败摘要
+              FAILED_JOBS=""
+              [ "$LINT_RESULT" = "failure" ] && FAILED_JOBS="${FAILED_JOBS}lint "
+              [ "$TEST_RESULT" = "failure" ] && FAILED_JOBS="${FAILED_JOBS}test "
+
              curl -sf -X POST \
                -H "Authorization: token $GITEA_TOKEN" \
                -H "Content-Type: application/json" \
                "${{ gitea.api_url }}/repos/${{ gitea.repository }}/issues/${PR_NUMBER}/comments" \
-                -d "{\"body\": \"[CI] 失败\\n\\n分支: ${{ gitea.ref_name }}\\n触发 commit: \`${{ gitea.sha }}\`\\n请检查 CI 日志并修复。\"}" \
+                -d "{\"body\": \"[CI] 失败\\n\\n分支: ${{ gitea.ref_name }}\\n触发 commit: \`${{ gitea.sha }}\`\\n失败 Job: ${FAILED_JOBS}\\n请检查 CI 日志并修复。\"}" \
                || echo "Failed to post PR comment"
              echo "PR comment posted."
            else
              echo "Not a PR event, skipping PR comment."
            fi
          else
-            echo "CI passed, no notification needed."
+            echo "No explicit failures (results: lint=$LINT_RESULT, test=$TEST_RESULT), no notification needed."
          fi
@@ -23,16 +23,16 @@ jobs:

      - name: Setup Python
        run: |
-          python3 -m venv .venv
-          .venv/bin/pip install --quiet -r requirements.txt
+          python3 -m venv /tmp/ci-venv-deploy
+          /tmp/ci-venv-deploy/bin/pip install --quiet flake8 fastapi pydantic pyyaml uvicorn requests pytest pytest-asyncio httpx

      - name: Lint
        run: |
-          .venv/bin/flake8 src/ --max-line-length=120 --extend-ignore=E501
+          /tmp/ci-venv-deploy/bin/flake8 src/ --max-line-length=120 --extend-ignore=E501

      - name: Unit & Integration Tests
        run: |
-          .venv/bin/pytest tests/ -m "not e2e" -x -q
+          /tmp/ci-venv-deploy/bin/pytest tests/ -m "not e2e" -x -q

  # ── Job 2: 部署 ─────────────────────────────────────
  deploy:
@@ -83,3 +83,66 @@ jobs:
          else
            echo "Deploy succeeded."
          fi
+
+  # ── Job 4: 部署成功通知 ──────────────────────────────
+  notify-deploy-success:
+    runs-on: macos-arm64
+    needs: [ci, deploy]
+    if: always()
+    steps:
+      - name: Notify deploy success
+        env:
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+          DEPLOY_RESULT: ${{ needs.deploy.result }}
+        run: |
+          if [ "$DEPLOY_RESULT" != "success" ]; then
+            echo "Deploy did not succeed (result: $DEPLOY_RESULT), skipping success notification."
+            exit 0
+          fi
+
+          echo "Deploy succeeded, sending notification..."
+
+          API_URL="${{ gitea.api_url }}"
+          REPO="${{ gitea.repository }}"
+          COMMIT_SHA="${{ gitea.sha }}"
+
+          # 查询关联的 merged PR 作者
+          PR_AUTHOR=$(curl --max-time 5 -sf \
+            -H "Authorization: token $GITEA_TOKEN" \
+            "$API_URL/repos/$REPO/pulls?state=closed&sort=updated&order=desc&limit=10" | \
+            python3 -c "
+import json, sys
+sha = '$COMMIT_SHA'
+for pr in json.load(sys.stdin):
+    merge_sha = pr.get('merge_commit_sha', '') or ''
+    if merge_sha.startswith(sha) or sha.startswith(merge_sha):
+        print(pr['user']['login'])
+        break
+" 2>/dev/null || echo "")
+
+          # 确定通知对象
+          if [ -n "$PR_AUTHOR" ]; then
+            NOTIFY_TO="$PR_AUTHOR"
+          else
+            # direct push 场景通知 jiangwei-infra
+            NOTIFY_TO="jiangwei-infra"
+          fi
+
+          # 发送 Mail 通知
+          MAIL_TITLE="[CD] 部署成功: $(echo $COMMIT_SHA | cut -c1-8)"
+          MAIL_TEXT="部署成功。Commit: ${COMMIT_SHA}"
+
+          curl --max-time 5 -s -X POST http://localhost:8083/api/mail \
+            -H "Content-Type: application/json" \
+            -d "{\"from\":\"system\",\"to\":\"$NOTIFY_TO\",\"title\":\"$MAIL_TITLE\",\"text\":\"$MAIL_TEXT\",\"type\":\"inform\"}" \
+            || echo "Mail notification failed (non-blocking)"
+
+          # 同时通知 pangtong-fujunshi（如果 PR 作者不是 pangtong）
+          if [ "$NOTIFY_TO" != "pangtong-fujunshi" ]; then
+            curl --max-time 5 -s -X POST http://localhost:8083/api/mail \
+              -H "Content-Type: application/json" \
+              -d "{\"from\":\"system\",\"to\":\"pangtong-fujunshi\",\"title\":\"$MAIL_TITLE\",\"text\":\"$MAIL_TEXT\",\"type\":\"inform\"}" \
+              || echo "Mail notification failed (non-blocking)"
+          fi
+
+          echo "Deploy success notification sent to: $NOTIFY_TO"
@@ -233,20 +233,27 @@ def _revive_session(agent_id: str) -> bool:
                pass
 ```

-### 4.5 O5: compact 扫描条件收紧
+### 4.5 O5: compact 检测（§24 rotation-only v3）

-当前 compact 扫描在 status 非 idle/done/unknown/None 时都触发，范围过宽。
+§24 设计文档：`docs/design/24-compact-detection-fix.md`

-**改后**：只在 status 为 running 或 compacting 相关时扫描：
+**检测方法**：读 gateway 日志尾部 2MB，按 sessionKey 过滤 `[compaction] rotated active transcript` 事件。
+如果最近的 rotation 事件在 120s 窗口内 → 视为 compact 循环进行中（可能还在 post-compact retry）。
+
+旧方法 `_check_recent_compaction_jsonl`（扫描 session jsonl 的 `type=compaction` 事件）保留作为 fallback。

 ```python
-# 只在这些状态下检查 compact
-if result["status"] in ("running",) and sf:
-    result["recent_compact"] = AgentSpawner._check_recent_compaction_jsonl(sf)
+# §24 v3: compact 检测优先用 gateway 日志 rotation 事件
+if result["status"] not in ("idle", "unknown", None):
+    session_key = f"agent:{agent_id}:main"
+    result["recent_compact"] = AgentSpawner._check_compact_in_progress_gateway(
+        session_key)
+    if not result["recent_compact"] and sf:
+        result["recent_compact"] = AgentSpawner._check_recent_compaction_jsonl(sf)
 ```

 注：Gateway 的 sessions.json status 实际值主要是 `idle/running/timeout/failed`。
-`running` 时检查 compact 有意义（agent turn 执行中可能触发 compact）。
+非空闲状态（`running`/`timeout`/`failed`）时检查 compact 有意义。
 其他状态不需要检查。

 ## 五、改动范围
@@ -110,8 +110,8 @@ TCP 握手只能检测进程端口是否监听，无法检测 Gateway **业务

 | 编号 | 条件 | outcome | 可恢复？ | 处理 |
 |------|------|---------|----------|------|
-| A12 | exit=0 + task_status ∈ {done, review} | completed | — | 正常完成 |
-| A13 | exit=0 + task_status ∉ {done, review} | agent_error | ❌ | 标 failed + 原因写黑板 |
+| A12 | ~~已合并到 A13 revised~~ | — | — | 见下方 A13 revised |
+| **A13 revised** | exit=0（无 JSON 输出） | completed | — | 信任进程退出码，exit=0 即正常完成。旧逻辑按 task_status 区分，非终态判 agent_error → 导致 inform Mail 永不标 done，与 dispatcher inform auto-done 形成死循环 |
 | **A14** | exit=130 (SIGINT) 或 exit=143 (SIGTERM) | interrupted | ✅ | retry |
 | **A15** | exit≠0 + stderr 含 network 关键字 | gateway_unreachable | ✅ | retry + cooldown 30s |
 | **A16** | exit≠0 + stderr 含 compact 关键字 | compact_interrupted | ✅ | retry + cooldown 60s |
@@ -1,6 +1,6 @@
 # 三国团队工具链与开发流程设计

-> **状态**: v3.1 — P3 端到端验证通过 + 调研结论写入 + Review API 枚举值修正
+> **状态**: v3.3 — #19 上下文四层改造合并 + CI 修复 + A13 修订
 > **作者**: 庞统（副军师）🐦
 > **评审**: 司马懿（仲达）🗡️
 > **日期**: 2026-06-06
@@ -33,26 +33,173 @@
 | 项 | 配置 |
 |----|------|
 | 地址 | `http://192.168.2.154:3000` |
-| 版本 | v1.23.4 |
-| 认证 | HTTP + token（待配置） |
-| 权限 | cfdaily 用户；姜维持有 admin 权限（启用 Actions、分支保护等） |
+| 版本 | v1.26.2（2026-06-11 从 v1.23.4 升级） |
+| 认证 | HTTP + token；admin 账号（姜维持有） |
+| 权限 | 姜维持有 admin 权限（启用 Actions、分支保护、org webhook 等） |
+| 数据库 | SQLite3 |
+| 部署方式 | Docker（NAS 群晖），数据卷 `/volume2/@docker/volumes/gitea-data/_data` |

 ### 2.2 CI/CD：Gitea Actions

 | 项 | 配置 |
 |----|------|
-| Runner | Mac mini 裸机，act-runner（Go 二进制） |
+| Runner | Mac mini 裸机，gitea-runner v1.0.8（通过 PM2 管理 `sanguo-act-runner`） |
 | 配置文件 | `.gitea/workflows/*.yml`，每个项目自管 |
-| 语法 | 兼容 GitHub Actions（v1.23.4 已验证支持） |
-| 触发 | push / PR / tag |
+| 语法 | 兼容 GitHub Actions（v1.26.2 已验证支持 concurrency groups） |
+| 触发 | push / PR / tag / workflow_dispatch |
+| v1.26 新增 | concurrency groups、re-run failed jobs、可配置 GITEA_TOKEN 权限 |
+| 仍不支持 | `failure()`、`continue-on-error`、`timeout-minutes` |

-### 2.3 部署目标
+### 2.4 Gitea 基础设施 Setup 记录（2026-06-11 姜维）
+
+> 以下为 Gitea 从 v1.23.4 升级到 v1.26.2 的完整操作记录，作为未来参考。
+
+#### 2.4.1 升级 v1.23.4 → v1.26.2
+
+**升级原因**：v1.23.4 不支持 concurrency groups，导致双倍触发问题无根因解法。
+
+**升级步骤**：
+1. 备份：`docker exec sanguo_gitea gitea dump -c /data/gitea/conf/app.ini -f /data/gitea/gitea-backup-pre-v126.zip`（765MB）
+2. 拉取镜像：Mac 上 skopeo 下载 → python docker SDK 远程 load（群晖 Docker Hub 太慢）
+3. 停止旧容器 + rename 保留回滚
+4. 启动新容器（数据库自动迁移 Migration[312]→[326]，含 concurrency #323）
+5. 验证：API + Web UI + 仓库数据 + 用户数据
+
+**踩坑：群晖内核 3.10 + git 2.52 不兼容**：
+- 根因：git 2.52 使用 `getrandom(2)` syscall，群晖内核 3.10.108 不支持（3.17 才加入）
+- 症状：`git push` 报 `unable to create temporary file: Function not implemented`
+- 修复：entrypoint 脚本在容器启动时自动从本地缓存降级 git 到 2.45.4
+- 持久化：`/data/entrypoint-wrapper.sh` + `/data/git-2.45.4-r0.apk` 在数据卷里，容器重建不丢失
+- 群晖内核**无法通过 DSM 升级**，内核版本跟硬件型号绑定
+
+**完整重建命令**：
+```bash
+docker -H tcp://192.168.2.154:2375 run -d \
+  --name sanguo_gitea \
+  --restart=always \
+  -p 3000:3000 \
+  -p 2221:22 \
+  -v /volume2/@docker/volumes/gitea-data/_data:/data \
+  -e GITEA__database__DB_TYPE=sqlite3 \
+  -e GITEA__database__PATH=/data/gitea/gitea.db \
+  -e GITEA__server__ROOT_URL=http://192.168.2.154:3000/ \
+  --entrypoint /bin/sh \
+  gitea/gitea:1.26.2 \
+  -c '/data/entrypoint-wrapper.sh'
+```
+
+#### 2.4.2 act_runner 升级 v0.2.11 → v1.0.8
+
+**升级原因**：v0.2.11 的 multi-step job 执行有 bug，Setup Python 和 Lint step 被跳过。
+
+**升级步骤**：
+1. 下载 `gitea-runner-1.0.8-darwin-arm64`（从 gitea.com releases）
+2. `codesign --force --sign -` 重签（macOS Gatekeeper 会 SIGKILL 未签名的二进制）
+3. 替换 `/Users/chufeng/bin/act_runner`
+4. PM2 restart `sanguo-act-runner`
+
+**注意**：act_runner 通过 **PM2** 管理（`sanguo-act-runner`），不是 launchd。launchd plist 仅为备份。
+
+**PM2 常用命令**：
+```bash
+pm2 restart sanguo-act-runner   # 重启
+pm2 logs sanguo-act-runner      # 查看日志
+pm2 show sanguo-act-runner      # 详情
+```
+
+#### 2.4.3 CI Workflow 配置
+
+**三个 workflow 文件**：
+
+| 文件 | 触发 | concurrency | 说明 |
+|------|------|-------------|------|
+| `ci.yml` | `pull_request` | `group: ci-${{ gitea.ref }}, cancel-in-progress: true` | 同一 PR 新 push 自动取消旧 run |
+| `deploy.yml` | `push to main` | `group: deploy-${{ gitea.ref }}, cancel-in-progress: false` | 部署排队不取消 |
+| `e2e.yml` | `workflow_dispatch` | `group: e2e-${{ gitea.ref }}, cancel-in-progress: true` | 手动触发 |
+
+**Branch Protection（main 分支）**：
+- 禁止直接 push
+- status check：`CI / lint (pull_request)` 必须通过
+- 至少 1 人 Review
+
+**⚠️ 踩坑**：v1.26 上报的 commit status context 格式变了：
+- 旧格式：`lint`
+- 新格式：`CI / lint (pull_request)`
+- branch protection 必须用新格式匹配，否则 merge 报 "Not all required status checks successful"
+
+#### 2.4.4 Org Webhook 配置
+
+- **对象**：Gitea 组织 `sanguo` webhook id=28
+- **URL**：`http://192.168.2.153:8083/webhook/gitea`
+- **事件**：16 个（push/issues/PR/PR review 等）
+
+**⚠️ 踩坑**：Gitea v1.26 的 PATCH hooks API，只传 `{"active": true}` 会把 events 重置为 `["push"]`。**必须每次 PATCH 都带上完整的 events 列表。**
+
+**临时措施（已恢复）**：2026-06-10 曾临时关闭 webhook（CI 错误大爆炸期间），2026-06-11 已恢复。
+
+#### 2.4.5 凭据管理
+
+| 凭据 | 用途 | 持有者 |
+|------|------|--------|
+| Gitea admin:cf7561523 | 仓库管理、branch protection、org webhook | 姜维 |
+| Gitea PAT (jiangwei-infra) | API 操作、git clone/push | 姜维 |
+| Gitea PAT (cfdaily) | CI workflow 中的 git 操作 | CI secrets |
+
+#### 2.4.6 备份与回滚
+
+| 项目 | 路径 | 说明 |
+|------|------|------|
+| Gitea 数据库备份 | `/data/gitea/gitea-backup-pre-v126.zip` | 升级前 dump |
+| 旧容器 | 已清理 | 升级验证通过后 `docker rm` |
+| 变更记录 | `~/.openclaw/workspace-jiangwei/changes/gitea-emergency-2026-06-10.md` | 完整操作日志 |

 | 环境 | 位置 | 说明 |
 |------|------|------|
 | Mac mini 本机 | `~/.sanguo_projects/<project>/` | 主力开发和运行环境 |
 | NAS Docker | `192.168.2.154` | 部分服务（Gitea、回测等） |

+#### 2.4.7 Gitea 迁移验证记录（2026-06-11 司马懿）
+
+> 验证 Gitea 从 gitee 迁移完成后的状态。所有验证在 2026-06-11 完成。
+
+**仓库迁移状态**：
+
+| 项目 | Gitea 仓库 | 开发目录 | 远程地址 | gitee 残留 |
+|------|-----------|---------|---------|----------|
+| sanguo_moziplus_v2 | `sanguo/sanguo_moziplus_v2` | `~/.openclaw/sanguo_projects/sanguo_moziplus_v2/` | `http://192.168.2.154:3000/sanguo/sanguo_moziplus_v2.git` | ✅ 无 |
+| sanguo_quant_live | `sanguo/sanguo_quant_live` | `~/.openclaw/sanguo_projects/sanguo_quant_live/` | `http://192.168.2.154:3000/sanguo/sanguo_quant_live.git` | ✅ 无 |
+| sanguo_vnpy | `sanguo/sanguo_vnpy` | `~/.openclaw/sanguo_projects/sanguo_vnpy/` | `http://192.168.2.154:3000/sanguo/sanguo_vnpy.git` | ✅ 无 |
+
+**验证方法**：在 3 个开发目录分别执行 `git remote -v`，确认 origin 指向 gitea 且无 gitee remote。
+
+**CI 管道验证**：
+
+| 验证项 | 结果 | 备注 |
+|--------|------|------|
+| PR #33 Lint 修复 CI 通过 | ✅ | flake8 全通过 |
+| CD pipeline (deploy.yml) 合并 | ✅ | 含 CI + deploy + notify-deploy-failure 三个 job |
+| Branch protection 生效 | ✅ | main 分支需 CI 通过 + 1 人 Review 才能合并 |
+| Gitea squash merge 兼容 | ✅ | `merge_commit_sha` 在 squash merge 下仍等于 gitea.sha |
+
+**工具链事件中枢验证**：
+
+| Webhook → Mail 流 | 验证结果 |
+|-------------------|--------|
+| PR opened → Review 请求 Mail | ✅ 司马懿收到 PR #30-#35 的 Review 请求 |
+| PR review → 结果 Mail | ✅ 张飞/庞统收到 Review 结果通知 |
+| Issue assigned → 指派 Mail | ✅ （E2E 验证通过） |
+| CI 失败评论 → 通知 Mail | ✅ （E2E 验证通过） |
+| PR synchronize → reviewer 重审 Mail | ✅ 新增（§23） |
+| Review COMMENTED → PR 作者通知 | ✅ 新增（§23） |
+
+**Agent Gitea 凭据**（各 Agent 自行持有 PAT）：
+
+| Agent | Gitea 用户名 | PAT 用途 |
+|-------|-------------|--------:|
+| simayi-challenger | simayi-challenger | PR Review 提交 |
+| pangtong-fujunshi | pangtong-fujunshi | PR 创建/合并、代码 push |
+| jiangwei-infra | jiangwei-infra | 基础设施配置（admin 级操作） |
+
 ---

 ## §3. 分支策略
@@ -157,7 +304,7 @@ Open → In Progress → Review → Closed

 每个项目在 `.gitea/workflows/ci.yml` 自定义具体步骤，但遵循统一骨架。

-> **注**：Gitea Actions v1.23.4 不支持 `paths` 过滤触发条件。通过路径判断放在 job 级别的 `if` 条件中，使用确定支持的语法。（M4 修订）
+> **注**：Gitea Actions v1.26.2 不支持 `paths` 过滤触发条件。通过路径判断放在 job 级别的 `if` 条件中，使用确定支持的语法。（M4 修订）

 ```yaml
 name: CI
@@ -2765,3 +2912,378 @@ Gitea v1.23.4 自带完整的 CI 管理界面：
 | §16.8 #10 | Gitea v1.23.4 review payload 调研结论（姜维 2026-06-08）：Gitea v1.23.4 review payload 只有 `type` + `content`，没有 `state`/`body`/`user`，这不是 org vs repo 差异而是 Gitea 设计。v1.24.0 格式不变。双格式兼容是防御性编码，保持现状 |
 | §16.8 #11 | Spawner compact 检测窗口修复：窗口 300s→900s，尾部读取 50KB→1MB。实测长对话中 compact 记录被推出窗口导致漏检 |
 | §16.8 #12 | inform 类型 Mail crash 误标 done bug 修复：`_mail_auto_complete` 增加 outcome 感知，inform 用白名单（completed/claimed/no_reply）控制 done 标记。spawner crash cooldown 300s→60s |
+
+---
+
+### 一、问题诊断
+
+#### 1.1 E2E 真实场景测试暴露的三个断层
+
+主公在 moziplus-v2 仓库创建了 Issue #32（添加 /api/stats 端点），指派张飞。链条在第一步就断了。
+
+| 断层 | 现象 | 根因 |
+|------|------|------|
+| **Agent 不知道该做什么** | 张飞收到 Issue 指派 Mail，回复"已阅"就结束了 | Mail 模板（issue_assigned.md）5 行信息，无流程引导；spawn prompt 说"已阅即可" |
+| **Agent 去错了仓库** | 张飞去读了 sanguo_moziplus_v2 平台代码，而不是空的实验仓库 moziplus-v2 | Mail 模板没有仓库 clone URL，张飞凭习惯去了开发目录 |
+| **Agent 在 Control UI 提问** | 张飞遇到问题直接在 Control UI 问主公，没有去 Gitea Issue 评论 | 没有任何地方引导"有疑问去 Gitea Issue 评论" |
+| **Agent 不知道怎么协作** | 张飞判断任务需要澄清，但不知道该怎么请求澄清 | 没有"做不了→在 Issue 评论 / Mail 庞统"的回退路径 |
+| **跨 Agent @mention 无法通知** | 张飞在 Issue 评论 @赵云，赵云收不到通知 | issue_comment handler 只处理 [CI] 评论，@mention 被忽略 |
+
+#### 1.2 根因：工具链在四层架构中的断层
+
+| 层 | 应该有 | 实际有 | Gap |
+|---|---|---|---|
+| **L0 铁律** | — | — | 无需改动 |
+| **L1 角色** | 工具链协作行为规范（所有 Agent 共享） | 无 | AGENTS.md 没有工具链相关内容 |
+| **L2 引擎注入** | 事件上下文（仓库 clone URL、Gitea API、Issue/PR 详情） | Mail 模板只有 5 行摘要 | 缺仓库信息和流程引导 |
+| **L3 被动参考** | 技术细节（分支命名、commit 规范、PR 创建方式） | git-workflow 等 Skill 已存在但没人触发 | Agent 不知道该加载哪个 Skill |
+
+---
+
+### 二、改造方案：四层归属
+
+#### 2.1 分层原则
+
+| 层 | 放什么 | 不放什么 | 理由 |
+|---|---|---|---|
+| **L0** | 不放 | — | 工具链不是安全底线 |
+| **L1** | 协作行为规范：收到什么通知该做什么、遇到问题怎么办 | 技术细节（分支命名、commit 格式） | 行为规范是团队常识，每个 Agent 都要知道 |
+| **L2** | 事件上下文：仓库 clone URL、Gitea API URL、Issue/PR 链接、动态信息 | 固定的协作流程 | 动态信息每次不同，由 Mail 模板 + spawn 时注入 |
+| **L3** | 技术细节：git-workflow、code-review 等 Skill 全文 | — | 按需加载，Agent 知道"我要提 PR"后自己读 |
+
+#### 2.2 各层具体内容
+
+##### L1：AGENTS.md 加工具链协作行为段（所有 Agent 统一）
+
+```markdown
+## 工具链协作（Gitea）
+
+收到 Gitea 事件通知（Issue 指派、Review 请求、CI 失败等）时，按以下流程操作：
+
+### 基本流程
+- **Issue 指派** → clone 仓库 → 开分支 → 编码 → 提 PR（参考 git-workflow Skill）
+- **Review 请求** → 读 PR diff（Gitea API）→ 提交 Review（参考 code-review Skill）
+- **Review 通过** → 等 merge
+- **Review 驳回** → 看 review body → 修代码 → 重新 push
+- **CI 失败** → 看错误摘要 → 修代码 → push（自动重触发 CI）
+- **部署失败** → 查 deploy 日志 → 修复
+
+### 协作规则
+- **有疑问？** 在 Gitea Issue 下评论，不要在 Control UI 或 Mail 里问
+- **需要别人帮忙？** 在 Issue 评论中 @mention 对应 Agent（如 @zhaoyun-data）
+- **做不了？** 回复 Mail 说明原因和建议的接手人
+- **获取完整上下文** → 用 Gitea API 拉取 Issue 详情和评论，不要只看 Mail 里的快照
+
+### Gitea API 速查
+> 其中 `{owner}/{repo}` 替换为实际仓库，如 `sanguo/sanguo_moziplus_v2`
+- Issue 详情: GET /api/v1/repos/{owner}/{repo}/issues/{number}
+- Issue 评论: GET /api/v1/repos/{owner}/{repo}/issues/{number}/comments
+- PR diff: GET /api/v1/repos/{owner}/{repo}/pulls/{number}.diff
+- 提交 Review: POST /api/v1/repos/{owner}/{repo}/pulls/{number}/reviews
+```
+
+**改动范围**：6 个 Agent 的 AGENTS.md 各加一段（内容统一）。
+
+##### L2：Mail 模板精简 + 事件上下文注入
+
+**原则**：模板只放摘要 + 链接 + 仓库信息，不写固定步骤（步骤在 L1）。
+
+**issue_assigned.md** 改为：
+
+```markdown
+Issue 指派
+
+Issue: {issue_url}
+标题: {issue_title}
+标签: {labels}
+
+📋 获取完整上下文（先读再动手）：
+- Issue 详情: GET {gitea_api}/repos/{repo}/issues/{issue_number}
+- Issue 评论: GET {gitea_api}/repos/{repo}/issues/{issue_number}/comments
+
+仓库: {repo_clone_url}
+建议分支: feat/issue-{issue_number}-{brief}
+```
+
+**review_request.md** 改为：
+
+```markdown
+PR Review 请求
+
+PR: {pr_url}
+标题: {pr_title}
+作者: {pr_author}
+分支: {branch}
+风险级别: {risk_level}
+
+📋 获取完整上下文：
+- PR diff: GET {gitea_api}/repos/{repo}/pulls/{pr_number}.diff
+- PR 文件列表: GET {gitea_api}/repos/{repo}/pulls/{pr_number}/files
+```
+
+**review_result.md** 改为：
+
+```markdown
+Review {result}
+
+PR: {pr_url}
+标题: {pr_title}
+审查者: {reviewer}
+
+{review_body}
+```
+
+**ci_failure.md** 改为：
+
+```markdown
+CI 失败
+
+Issue: {issue_url}
+分支: {branch}
+
+错误摘要:
+{error_summary}
+
+📋 CI 日志: {gitea_url}/{repo}/actions
+修复后 push 会自动重触发 CI。
+```
+
+**deploy_failure.md** 改为：
+
+```markdown
+部署失败
+
+仓库: {repo}
+Commit: {commit_sha}
+
+📋 排查步骤:
+- CI 日志: {gitea_url}/{repo}/actions
+- 服务器: pm2 logs {service_name}
+```
+
+**L2 代码改动**（`toolchain_routes.py`）：
+
+1. 从 Webhook payload 的 `repository` 对象提取 `clone_url` 和 `html_url`
+2. `render_template()` 传入新变量：`gitea_api`、`gitea_url`、`repo_clone_url`
+3. 所有模板变量统一补齐
+
+##### L3：Skill 按需加载（不改 Skill 本身）
+
+git-workflow、code-review 等 Skill 保持不变。
+
+L1 的协作行为段里会引用 Skill 名称（"参考 git-workflow Skill"），Agent 收到 Mail 后根据 L1 的引导自主加载对应 Skill。
+
+**不改 Skill 路由机制**——靠 L1 的文案触发 Agent 的 Skill 路由器匹配。
+
+---
+
+### 三、新增功能：issue_comment @mention 通知
+
+#### 3.1 设计
+
+当前 `_handle_issue_comment` 只处理 `[CI]` 前缀评论。扩展为：
+
+```
+issue_comment 事件
+  ├── 含 [CI] / CI 失败 → 原有 CI 失败通知逻辑
+  └── 含 @username → 解析 @mention → Mail 通知被 @的 Agent
+```
+
+#### 3.2 实现
+
+**`toolchain_routes.py` 新增 `_handle_issue_comment_mention()`**：
+
+```python
+AGENT_IDS = {
+    "zhangfei-dev", "guanyu-dev", "zhaoyun-data",
+    "jiangwei-infra", "simayi-challenger", "pangtong-fujunshi",
+}
+
+# 前缀映射：@张飞 → zhangfei-dev
+# 中文名映射：Agent 在 Gitea Issue 评论中可能用中文名 @mention
+# 英文短名映射：Agent 可能用不带 -dev/-infra 后缀的短名
+AGENT_ALIAS = {
+    "张飞": "zhangfei-dev",
+    "关羽": "guanyu-dev",
+    "赵云": "zhaoyun-data",
+    "姜维": "jiangwei-infra",
+    "司马懿": "simayi-challenger",
+    "庞统": "pangtong-fujunshi",
+    "pangtong": "pangtong-fujunshi",
+    "simayi": "simayi-challenger",
+    "zhangfei": "zhangfei-dev",
+    "guanyu": "guanyu-dev",
+    "zhaoyun": "zhaoyun-data",
+    "jiangwei": "jiangwei-infra",
+}
+
+def extract_mentions(body: str, sender: str) -> list[str]:
+    """从评论 body 中提取 @mention 的 Agent ID"""
+    candidates = re.findall(r"@([a-zA-Z\u4e00-\u9fa5][a-zA-Z0-9\u4e00-\u9fff-]*)", body)
+    result = set()
+    for c in candidates:
+        # 精确匹配
+        if c in AGENT_IDS:
+            result.add(c)
+        # 前缀/别名匹配
+        elif c in AGENT_ALIAS:
+            result.add(AGENT_ALIAS[c])
+        else:
+            # 前缀模糊匹配：@zhangfei → zhangfei-dev
+            for aid in AGENT_IDS:
+                if aid.startswith(c):
+                    result.add(aid)
+                    break
+    # 过滤掉评论者自己
+    result.discard(sender)
+    return list(result)
+```
+
+**新增 mention 通知模板** `templates/toolchain/mention.md`：
+
+```markdown
+你在 Issue 中被 @mention
+
+Issue: {issue_url}
+评论者: {commenter}
+评论内容:
+{comment_body}
+
+📋 获取完整上下文：
+- Issue 详情: GET {gitea_api}/repos/{repo}/issues/{issue_number}
+- Issue 评论: GET {gitea_api}/repos/{repo}/issues/{issue_number}/comments
+```
+
+**改动 `_handle_issue_comment`**：
+
+```python
+async def _handle_issue_comment(payload):
+    comment = payload.get("comment", {})
+    body = comment.get("body", "")
+    sender = comment.get("user", {}).get("login", "")
+    repo = _repo_fullname(payload)
+    issue = payload.get("issue", {})
+
+    # 原有 CI 失败逻辑（不变）
+    if "[CI]" in body or "CI 失败" in body:
+        # ... 原有逻辑 ...
+
+    # 新增：@mention 通知
+    mentions = extract_mentions(body, sender)
+    if mentions:
+        issue_number = issue.get("number", 0)
+        issue_title = issue.get("title", "")
+        text = render_template("mention", {
+            "repo": repo,
+            "issue_number": str(issue_number),
+            "issue_url": issue.get("html_url", ""),
+            "commenter": sender,
+            "comment_body": body[:500],
+            "gitea_api": "http://192.168.2.154:3000/api/v1",
+        })
+        title = f"@mention: {issue_title} ({repo}#{issue_number})"
+        for agent_id in mentions:
+            _send_mail(agent_id, title, text)
+```
+
+#### 3.3 去重
+
+- 同一条评论 @多人：每人一封 Mail（不同 to，内容相同）
+- 同一事件 org webhook + repo webhook 双触发：现有 delivery UUID 去重机制覆盖
+- 同一人被 @多次：`extract_mentions` 返回 set，自动去重
+
+---
+
+### 四、Mail Spawn Prompt 改造
+
+#### 4.1 问题
+
+当前工具链 Mail 走 Mail 通道，spawn prompt 是：
+
+```
+你收到一封飞鸽传书(纯通知)。
+发件者: system
+主题: Issue 指派: xxx
+内容: [工具链模板]
+已阅即可。
+```
+
+"已阅即可"直接让 Agent 不做事。
+
+#### 4.2 方案
+
+**不改 MAIL_INFORM_TEMPLATE / MAIL_REQUEST_TEMPLATE 本身**（那是 Mail 系统通用的）。
+
+改为：**工具链 Mail 使用 `type=request`（而不是默认的 inform）**。
+
+在 `_send_mail()` 中，工具链事件创建的 Mail 默认 `performative=request`，这样 Agent 收到时走 `MAIL_REQUEST_TEMPLATE`，知道需要处理。
+
+具体改动在 `_send_mail()` 函数或其调用处：工具链路由调用 `_send_mail` 时传入 `performative="request"`。
+
+**⚠️ 验证要点**：改为 request 后，Agent spawn prompt 变为 "请处理以下请求"，需确认：
+1. Agent 不再把工具链 Mail 当纯通知忽略
+2. Agent 能正确处理「已阅型」工具链事件（如 CI 失败通知——不需要回复，但需要知道）
+3. 对已关闭 PR/Issue 的延迟通知，Agent 不会尝试去处理
+
+验证方法：部署后发一条 Issue 指派 Mail，观察 Agent 行为是否符合预期。
+
+---
+
+### 五、完整改动清单
+
+| # | 改什么 | 改动内容 | 层 | 风险 |
+|---|--------|---------|---|------|
+| 1 | 6 个 Agent 的 `AGENTS.md` | 加"工具链协作"段（内容统一） | L1 | 低（纯追加） |
+| 2 | `templates/toolchain/issue_assigned.md` | 精简 + 加仓库上下文 + Gitea API 引导 | L2 | 低 |
+| 3 | `templates/toolchain/review_request.md` | 精简 + 加 Gitea API 引导 | L2 | 低 |
+| 4 | `templates/toolchain/review_result.md` | 精简 | L2 | 低 |
+| 5 | `templates/toolchain/ci_failure.md` | 精简 + 加 CI 日志链接 | L2 | 低 |
+| 6 | `templates/toolchain/deploy_failure.md` | 精简 + 加排查步骤 | L2 | 低 |
+| 7 | **新建** `templates/toolchain/mention.md` | @mention 通知模板 | L2 | 低 |
+| 8 | `src/api/toolchain_routes.py` | 提取 clone_url/html_url 传入模板；issue_comment 增加 @mention 解析；工具链 Mail 改为 request 类型 | L2 | 中 |
+| 9 | 不改 | git-workflow 等 Skill 保持不变 | L3 | — |
+| 10 | 不改 | daemon 核心逻辑、BootstrapBuilder、Skill 路由 | — | — |
+
+---
+
+### 六、验证方案
+
+#### 6.1 单元验证
+
+| 验证点 | 方法 |
+|--------|------|
+| `extract_mentions()` 提取 `@zhangfei-dev` | unit test |
+| `extract_mentions()` 别名匹配 `@张飞` → zhangfei-dev | unit test |
+| `extract_mentions()` 前缀匹配 `@zhangfei` → zhangfei-dev | unit test |
+| `extract_mentions()` 过滤自己 | unit test |
+| 模板渲染新变量不报错 | unit test |
+
+#### 6.2 真实场景 E2E 验证
+
+重复 Issue #32 的场景：
+1. 创建 Issue 指派张飞
+2. **验证**：张飞收到的 Mail 含 clone URL + Gitea API 引导
+3. **验证**：张飞 spawn 后知道该做什么（L1 AGENTS.md 有流程引导）
+4. **验证**：张飞有疑问时去 Gitea Issue 评论（而不是 Control UI）
+5. 在 Issue 评论 @赵云
+6. **验证**：赵云收到 @mention Mail
+
+---
+
+### 七、不做的事（标记为后续）
+
+| 标记 | 描述 | 原因 |
+|------|------|------|
+| 后续-1 | Agent 离开工具链讨论后，是否有意识回到工具链 | 需要更多真实场景观察 |
+| 后续-2 | 工具链使用标准在所有 Agent 间的一致性验证 | L1 统一段落是第一步，需要 E2E 验证 |
+| 后续-3 | Mail 通道接入 BootstrapBuilder L2 注入 | 改动大，当前方案（L1 统一段落 + 模板引导）够用 |
+| 后续-4 | Skill 路由器自动触发（引擎注入） | 改动 daemon 核心，当前靠 L1 文案触发 |
+
+---
+
+### 八、变更记录
+
+| 日期 | 版本 | 变更 |
+|------|------|------|
+| 2026-06-09 | v1.0 | 初版：E2E 真实场景暴露问题 → 四层改造方案 + @mention 通知 + Mail type 改造 |
+
@@ -1,372 +0,0 @@
-# #19 工具链事件中枢 — 上下文四层改造方案
-
-> 版本: v1.0
-> 日期: 2026-06-09
-> 作者: 庞统（副军师）
-> 状态: 待主公确认
-> 前置: #13 工具链与开发流程 §16, #05 上下文四层架构
-> 来源: E2E 真实场景测试暴露的三个断层
-
---
-
-## 一、问题诊断
-
-### 1.1 E2E 真实场景测试暴露的三个断层
-
-主公在 moziplus-v2 仓库创建了 Issue #32（添加 /api/stats 端点），指派张飞。链条在第一步就断了。
-
-| 断层 | 现象 | 根因 |
-|------|------|------|
-| **Agent 不知道该做什么** | 张飞收到 Issue 指派 Mail，回复"已阅"就结束了 | Mail 模板（issue_assigned.md）5 行信息，无流程引导；spawn prompt 说"已阅即可" |
-| **Agent 去错了仓库** | 张飞去读了 sanguo_moziplus_v2 平台代码，而不是空的实验仓库 moziplus-v2 | Mail 模板没有仓库 clone URL，张飞凭习惯去了开发目录 |
-| **Agent 在 Control UI 提问** | 张飞遇到问题直接在 Control UI 问主公，没有去 Gitea Issue 评论 | 没有任何地方引导"有疑问去 Gitea Issue 评论" |
-| **Agent 不知道怎么协作** | 张飞判断任务需要澄清，但不知道该怎么请求澄清 | 没有"做不了→在 Issue 评论 / Mail 庞统"的回退路径 |
-| **跨 Agent @mention 无法通知** | 张飞在 Issue 评论 @赵云，赵云收不到通知 | issue_comment handler 只处理 [CI] 评论，@mention 被忽略 |
-
-### 1.2 根因：工具链在四层架构中的断层
-
-| 层 | 应该有 | 实际有 | Gap |
-|---|---|---|---|
-| **L0 铁律** | — | — | 无需改动 |
-| **L1 角色** | 工具链协作行为规范（所有 Agent 共享） | 无 | AGENTS.md 没有工具链相关内容 |
-| **L2 引擎注入** | 事件上下文（仓库 clone URL、Gitea API、Issue/PR 详情） | Mail 模板只有 5 行摘要 | 缺仓库信息和流程引导 |
-| **L3 被动参考** | 技术细节（分支命名、commit 规范、PR 创建方式） | git-workflow 等 Skill 已存在但没人触发 | Agent 不知道该加载哪个 Skill |
-
---
-
-## 二、改造方案：四层归属
-
-### 2.1 分层原则
-
-| 层 | 放什么 | 不放什么 | 理由 |
-|---|---|---|---|
-| **L0** | 不放 | — | 工具链不是安全底线 |
-| **L1** | 协作行为规范：收到什么通知该做什么、遇到问题怎么办 | 技术细节（分支命名、commit 格式） | 行为规范是团队常识，每个 Agent 都要知道 |
-| **L2** | 事件上下文：仓库 clone URL、Gitea API URL、Issue/PR 链接、动态信息 | 固定的协作流程 | 动态信息每次不同，由 Mail 模板 + spawn 时注入 |
-| **L3** | 技术细节：git-workflow、code-review 等 Skill 全文 | — | 按需加载，Agent 知道"我要提 PR"后自己读 |
-
-### 2.2 各层具体内容
-
-#### L1：AGENTS.md 加工具链协作行为段（所有 Agent 统一）
-
-```markdown
-## 工具链协作（Gitea）
-
-收到 Gitea 事件通知（Issue 指派、Review 请求、CI 失败等）时，按以下流程操作：
-
-### 基本流程
- **Issue 指派** → clone 仓库 → 开分支 → 编码 → 提 PR（参考 git-workflow Skill）
- **Review 请求** → 读 PR diff（Gitea API）→ 提交 Review（参考 code-review Skill）
- **Review 通过** → 等 merge
- **Review 驳回** → 看 review body → 修代码 → 重新 push
- **CI 失败** → 看错误摘要 → 修代码 → push（自动重触发 CI）
- **部署失败** → 查 deploy 日志 → 修复
-
-### 协作规则
- **有疑问？** 在 Gitea Issue 下评论，不要在 Control UI 或 Mail 里问
- **需要别人帮忙？** 在 Issue 评论中 @mention 对应 Agent（如 @zhaoyun-data）
- **做不了？** 回复 Mail 说明原因和建议的接手人
- **获取完整上下文** → 用 Gitea API 拉取 Issue 详情和评论，不要只看 Mail 里的快照
-
-### Gitea API 速查
- Issue 详情: GET /api/v1/repos/{owner}/{repo}/issues/{number}
- Issue 评论: GET /api/v1/repos/{owner}/{repo}/issues/{number}/comments
- PR diff: GET /api/v1/repos/{owner}/{repo}/pulls/{number}.diff
- 提交 Review: POST /api/v1/repos/{owner}/{repo}/pulls/{number}/reviews
-```
-
-**改动范围**：6 个 Agent 的 AGENTS.md 各加一段（内容统一）。
-
-#### L2：Mail 模板精简 + 事件上下文注入
-
-**原则**：模板只放摘要 + 链接 + 仓库信息，不写固定步骤（步骤在 L1）。
-
-**issue_assigned.md** 改为：
-
-```markdown
-Issue 指派
-
-Issue: {issue_url}
-标题: {issue_title}
-标签: {labels}
-
-📋 获取完整上下文（先读再动手）：
- Issue 详情: GET {gitea_api}/repos/{repo}/issues/{issue_number}
- Issue 评论: GET {gitea_api}/repos/{repo}/issues/{issue_number}/comments
-
-仓库: {repo_clone_url}
-建议分支: feat/issue-{issue_number}-{brief}
-```
-
-**review_request.md** 改为：
-
-```markdown
-PR Review 请求
-
-PR: {pr_url}
-标题: {pr_title}
-作者: {pr_author}
-分支: {branch}
-风险级别: {risk_level}
-
-📋 获取完整上下文：
- PR diff: GET {gitea_api}/repos/{repo}/pulls/{pr_number}.diff
- PR 文件列表: GET {gitea_api}/repos/{repo}/pulls/{pr_number}/files
-```
-
-**review_result.md** 改为：
-
-```markdown
-Review {result}
-
-PR: {pr_url}
-标题: {pr_title}
-审查者: {reviewer}
-
-{review_body}
-```
-
-**ci_failure.md** 改为：
-
-```markdown
-CI 失败
-
-Issue: {issue_url}
-分支: {branch}
-
-错误摘要:
-{error_summary}
-
-📋 CI 日志: {gitea_url}/{repo}/actions
-修复后 push 会自动重触发 CI。
-```
-
-**deploy_failure.md** 改为：
-
-```markdown
-部署失败
-
-仓库: {repo}
-Commit: {commit_sha}
-
-📋 排查步骤:
- CI 日志: {gitea_url}/{repo}/actions
- 服务器: pm2 logs {service_name}
-```
-
-**L2 代码改动**（`toolchain_routes.py`）：
-
-1. 从 Webhook payload 的 `repository` 对象提取 `clone_url` 和 `html_url`
-2. `render_template()` 传入新变量：`gitea_api`、`gitea_url`、`repo_clone_url`
-3. 所有模板变量统一补齐
-
-#### L3：Skill 按需加载（不改 Skill 本身）
-
-git-workflow、code-review 等 Skill 保持不变。
-
-L1 的协作行为段里会引用 Skill 名称（"参考 git-workflow Skill"），Agent 收到 Mail 后根据 L1 的引导自主加载对应 Skill。
-
-**不改 Skill 路由机制**——靠 L1 的文案触发 Agent 的 Skill 路由器匹配。
-
---
-
-## 三、新增功能：issue_comment @mention 通知
-
-### 3.1 设计
-
-当前 `_handle_issue_comment` 只处理 `[CI]` 前缀评论。扩展为：
-
-```
-issue_comment 事件
-  ├── 含 [CI] / CI 失败 → 原有 CI 失败通知逻辑
-  └── 含 @username → 解析 @mention → Mail 通知被 @的 Agent
-```
-
-### 3.2 实现
-
-**`toolchain_routes.py` 新增 `_handle_issue_comment_mention()`**：
-
-```python
-AGENT_IDS = {
-    "zhangfei-dev", "guanyu-dev", "zhaoyun-data",
-    "jiangwei-infra", "simayi-challenger", "pangtong-fujunshi",
-}
-
-# 前缀映射：@张飞 → zhangfei-dev
-AGENT_ALIAS = {
-    "张飞": "zhangfei-dev",
-    "关羽": "guanyu-dev",
-    "赵云": "zhaoyun-data",
-    "姜维": "jiangwei-infra",
-    "司马懿": "simayi-challenger",
-    "庞统": "pangtong-fujunshi",
-    "pangtong": "pangtong-fujunshi",
-    "simayi": "simayi-challenger",
-    "zhangfei": "zhangfei-dev",
-    "guanyu": "guanyu-dev",
-    "zhaoyun": "zhaoyun-data",
-    "jiangwei": "jiangwei-infra",
-}
-
-def extract_mentions(body: str, sender: str) -> list[str]:
-    """从评论 body 中提取 @mention 的 Agent ID"""
-    candidates = re.findall(r"@([a-zA-Z\u4e00-\u9fa5][a-zA-Z0-9\u4e00-\u9fff-]*)", body)
-    result = set()
-    for c in candidates:
-        # 精确匹配
-        if c in AGENT_IDS:
-            result.add(c)
-        # 前缀/别名匹配
-        elif c in AGENT_ALIAS:
-            result.add(AGENT_ALIAS[c])
-        else:
-            # 前缀模糊匹配：@zhangfei → zhangfei-dev
-            for aid in AGENT_IDS:
-                if aid.startswith(c):
-                    result.add(aid)
-                    break
-    # 过滤掉评论者自己
-    result.discard(sender)
-    return list(result)
-```
-
-**新增 mention 通知模板** `templates/toolchain/mention.md`：
-
-```markdown
-你在 Issue 中被 @mention
-
-Issue: {issue_url}
-评论者: {commenter}
-评论内容:
-{comment_body}
-
-📋 获取完整上下文：
- Issue 详情: GET {gitea_api}/repos/{repo}/issues/{issue_number}
- Issue 评论: GET {gitea_api}/repos/{repo}/issues/{issue_number}/comments
-```
-
-**改动 `_handle_issue_comment`**：
-
-```python
-async def _handle_issue_comment(payload):
-    comment = payload.get("comment", {})
-    body = comment.get("body", "")
-    sender = comment.get("user", {}).get("login", "")
-    repo = _repo_fullname(payload)
-    issue = payload.get("issue", {})
-
-    # 原有 CI 失败逻辑（不变）
-    if "[CI]" in body or "CI 失败" in body:
-        # ... 原有逻辑 ...
-
-    # 新增：@mention 通知
-    mentions = extract_mentions(body, sender)
-    if mentions:
-        issue_number = issue.get("number", 0)
-        issue_title = issue.get("title", "")
-        text = render_template("mention", {
-            "repo": repo,
-            "issue_number": str(issue_number),
-            "issue_url": issue.get("html_url", ""),
-            "commenter": sender,
-            "comment_body": body[:500],
-            "gitea_api": "http://192.168.2.154:3000/api/v1",
-        })
-        title = f"@mention: {issue_title} ({repo}#{issue_number})"
-        for agent_id in mentions:
-            _send_mail(agent_id, title, text)
-```
-
-### 3.3 去重
-
- 同一条评论 @多人：每人一封 Mail（不同 to，内容相同）
- 同一事件 org webhook + repo webhook 双触发：现有 delivery UUID 去重机制覆盖
- 同一人被 @多次：`extract_mentions` 返回 set，自动去重
-
---
-
-## 四、Mail Spawn Prompt 改造
-
-### 4.1 问题
-
-当前工具链 Mail 走 Mail 通道，spawn prompt 是：
-
-```
-你收到一封飞鸽传书(纯通知)。
-发件者: system
-主题: Issue 指派: xxx
-内容: [工具链模板]
-已阅即可。
-```
-
-"已阅即可"直接让 Agent 不做事。
-
-### 4.2 方案
-
-**不改 MAIL_INFORM_TEMPLATE / MAIL_REQUEST_TEMPLATE 本身**（那是 Mail 系统通用的）。
-
-改为：**工具链 Mail 使用 `type=request`（而不是默认的 inform）**。
-
-在 `_send_mail()` 中，工具链事件创建的 Mail 默认 `performative=request`，这样 Agent 收到时走 `MAIL_REQUEST_TEMPLATE`，知道需要处理。
-
-具体改动在 `_send_mail()` 函数或其调用处：工具链路由调用 `_send_mail` 时传入 `performative="request"`。
-
---
-
-## 五、完整改动清单
-
-| # | 改什么 | 改动内容 | 层 | 风险 |
-|---|--------|---------|---|------|
-| 1 | 6 个 Agent 的 `AGENTS.md` | 加"工具链协作"段（内容统一） | L1 | 低（纯追加） |
-| 2 | `templates/toolchain/issue_assigned.md` | 精简 + 加仓库上下文 + Gitea API 引导 | L2 | 低 |
-| 3 | `templates/toolchain/review_request.md` | 精简 + 加 Gitea API 引导 | L2 | 低 |
-| 4 | `templates/toolchain/review_result.md` | 精简 | L2 | 低 |
-| 5 | `templates/toolchain/ci_failure.md` | 精简 + 加 CI 日志链接 | L2 | 低 |
-| 6 | `templates/toolchain/deploy_failure.md` | 精简 + 加排查步骤 | L2 | 低 |
-| 7 | **新建** `templates/toolchain/mention.md` | @mention 通知模板 | L2 | 低 |
-| 8 | `src/api/toolchain_routes.py` | 提取 clone_url/html_url 传入模板；issue_comment 增加 @mention 解析；工具链 Mail 改为 request 类型 | L2 | 中 |
-| 9 | 不改 | git-workflow 等 Skill 保持不变 | L3 | — |
-| 10 | 不改 | daemon 核心逻辑、BootstrapBuilder、Skill 路由 | — | — |
-
---
-
-## 六、验证方案
-
-### 6.1 单元验证
-
-| 验证点 | 方法 |
-|--------|------|
-| `extract_mentions()` 提取 `@zhangfei-dev` | unit test |
-| `extract_mentions()` 别名匹配 `@张飞` → zhangfei-dev | unit test |
-| `extract_mentions()` 前缀匹配 `@zhangfei` → zhangfei-dev | unit test |
-| `extract_mentions()` 过滤自己 | unit test |
-| 模板渲染新变量不报错 | unit test |
-
-### 6.2 真实场景 E2E 验证
-
-重复 Issue #32 的场景：
-1. 创建 Issue 指派张飞
-2. **验证**：张飞收到的 Mail 含 clone URL + Gitea API 引导
-3. **验证**：张飞 spawn 后知道该做什么（L1 AGENTS.md 有流程引导）
-4. **验证**：张飞有疑问时去 Gitea Issue 评论（而不是 Control UI）
-5. 在 Issue 评论 @赵云
-6. **验证**：赵云收到 @mention Mail
-
---
-
-## 七、不做的事（标记为后续）
-
-| 标记 | 描述 | 原因 |
-|------|------|------|
-| 后续-1 | Agent 离开工具链讨论后，是否有意识回到工具链 | 需要更多真实场景观察 |
-| 后续-2 | 工具链使用标准在所有 Agent 间的一致性验证 | L1 统一段落是第一步，需要 E2E 验证 |
-| 后续-3 | Mail 通道接入 BootstrapBuilder L2 注入 | 改动大，当前方案（L1 统一段落 + 模板引导）够用 |
-| 后续-4 | Skill 路由器自动触发（引擎注入） | 改动 daemon 核心，当前靠 L1 文案触发 |
-
---
-
-## 八、变更记录
-
-| 日期 | 版本 | 变更 |
-|------|------|------|
-| 2026-06-09 | v1.0 | 初版：E2E 真实场景暴露问题 → 四层改造方案 + @mention 通知 + Mail type 改造 |
@@ -0,0 +1,102 @@
+# §21. Handler 注册后 E2E 验证
+
+> 日期：2026-06-11
+> 状态：已完成 ✅
+> 目标：验证 Task 五层架构重构（Step 2-5）+ review 修复后，Mail/Toolchain 路径端到端工作
+
+## 前置条件
+
+- Daemon 版本：commit 83694ad（含 handler 注册 + import 修复 + SKILL_BASE_PATH 修复）
+- Handler 注册日志：
+  ```
+  Registered task type handler: task (virtual_project=None)
+  Registered task type handler: mail (virtual_project=_mail)
+  Registered task type handler: toolchain (virtual_project=_toolchain)
+  ```
+- Gitea org webhook (ID=28)：姜维启用，事件订阅含 issues/pull_request/pull_request_review 等 16 个事件
+- 测试仓库：sanguo/sanguo_moziplus_v2
+
+## 验证结果
+
+### 一、Mail Handler（✅ 全部通过）
+
+| # | 步骤 | 验证点 | 结果 | Mail ID |
+|---|------|--------|------|---------|
+| 1 | 发 inform 邮件给 zhangfei-dev | ticker 发现 `_mail` 虚拟项目 | ✅ `handler auto-working` | mail-1781106713261 |
+| 2 | zhangfei-dev 回复 | handler verify (inform_auto) → done | ✅ `verify passed (inform_auto), marked done` | — |
+| 3 | 回复邮件给 pangtong | handler auto-working + done | ✅ `verify passed (inform_auto), marked done` | mail-1781106736388 |
+
+**关键验证**：
+- ✅ `virtual_projects()` 返回 `["_mail", "_toolchain"]`（注册前为空）
+- ✅ handler `pre_spawn` (auto-working) 生效，不是旧的 `_mail_auto_working`
+- ✅ guardrail 跳过 `_mail`（`is_handler_task=True`）
+- ✅ inform 类型自动标 done，request 类型检查回复
+
+### 二、Toolchain — Issue 指派（✅ 通过）
+
+| # | 步骤 | 验证点 | 结果 | Mail ID |
+|---|------|--------|------|---------|
+| 1 | 创建 Issue #28，assignee=zhangfei-dev | webhook 触发 + Mail 通知 | ✅ | mail-1781107087549 |
+
+**Webhook 路径**：Gitea → org webhook → `POST /webhook/gitea` → 签名验证 → `_handle_issues` → `_send_mail(zhangfei-dev, ...)`
+
+**注意**：Issue #27 创建时 webhook 未启用，未触发。Issue #28 创建时 webhook 已启用，正常触发。
+
+### 三、Toolchain — PR Review（✅ 通过）
+
+| # | 步骤 | 验证点 | 结果 | Mail ID |
+|---|------|--------|------|---------|
+| 1 | 创建 PR #30 | webhook 触发 + Review 请求 Mail | ✅ | mail-1781107538823 |
+| 2 | simayi-challenger 提交 COMMENT review | Review 结果通知 PR 作者 | ✅ `Review 通过 ✓` | mail-1781107650433 |
+
+**Webhook 路径**：
+- PR opened: Gitea → `_handle_pull_request` → `_send_mail(simayi-challenger, "Review 请求")`
+- PR review: Gitea → `_handle_pull_request_review` → `_send_mail(pangtong-fujunshi, "Review 通过 ✓")`
+
+### 四、CI 失败评论（⚠️ 触发但重复）
+
+| # | 步骤 | 验证点 | 结果 | Mail ID |
+|---|------|--------|------|---------|
+| 1 | push 空 commit → CI lint 失败 | CI 失败通知 | ✅ 但收到 2 封重复 Mail | mail-1781107563991, mail-1781107560933 |
+
+**已知问题**：和上次 E2E（§18）相同——org webhook + repo webhook 双触发。上次已加去重机制（delivery UUID + content sha256），但 CI 失败场景似乎仍触发 2 封。**非新问题，待姜维统一 org/repo webhook 后解决。**
+
+### 五、负面测试（❌ 未执行）
+
+| 步骤 | 说明 | 状态 |
+|------|------|------|
+| REQUEST_CHANGES review | review 驳回通知 PR 作者 | 未测（仲达提交的是 COMMENT 而非 REQUEST_CHANGES） |
+| 已关闭 Issue CI 评论 | closed issue 不触发 Mail | 未测 |
+| 部署失败 Issue | 双收件人通知 | 未测 |
+| 幂等测试 | 同 delivery ID 重发 | §18 已验证，未重测 |
+
+## 阻塞/问题记录
+
+### 已解决
+
+| 问题 | 说明 |
+|------|------|
+| Org webhook 事件列表被 Gitea API 重置 | 姜维修复：PATCH webhook 只传 active:true 会重置 events，必须带完整事件列表 |
+| PR Review 无法用 PR 作者 token 提交 | Gitea 不允许 self-review，请仲达用 simayi token 提交 |
+
+### 遗留
+
+| 问题 | 严重度 | 说明 |
+|------|--------|------|
+| CI 失败 Mail 重复 | 🟡 | org webhook + repo webhook 双触发，§18 已记录 |
+| REQUEST_CHANGES 未验证 | 🟢 | 下次 E2E 补测 |
+
+## 测试清理
+
+- ✅ Issue #27、#28 已关闭
+- ✅ PR #29、#30 已关闭
+- ✅ 分支 `test/e2e-1781107119`、`test/e2e-pr-1781107530` 已删除
+- ✅ 本地切回 main 分支
+
+## 结论
+
+**Handler 注册后 Mail 和 Toolchain 核心流程端到端验证通过。** 关键修复（handler 注册、review verdict、SKILL_BASE_PATH）均已生效。
+
+下一步：
+- Task review 路径 E2E（明天，需要普通任务 → executor → review → verdict → done）
+- CI 失败重复 Mail 根治（需姜维统一 org/repo webhook）
@@ -0,0 +1,81 @@
+# 22 — CD 生产环境落地方案
+
+> 状态：草案，待评审
+> 作者：庞统
+> 日期：2026-06-11
+
+## 背景
+
+CD 管道已在测试仓库 `sanguo/mojiplus-v2` 验证通过（run#282 全链路 success）。
+现需将 CD 落地到生产仓库 `sanguo/sanguo_moziplus_v2`。
+
+### 当前状态
+
+| 组件 | 状态 |
+|---|---|
+| `deploy.sh` | ✅ 完整（rsync + build + pm2 restart + health check + deploy history） |
+| `deploy.yml` | ⚠️ deploy job 已调用 deploy.sh，但缺少成功通知 |
+| CI 失败通知 | ✅ ci.yml → PR comment → webhook → Mail |
+| Deploy 失败通知 | ✅ deploy.yml → Issue → webhook → Mail |
+| Deploy 成功通知 | ❌ 缺失 |
+
+## 改动方案
+
+### 改动 1：deploy.yml 增加 deploy 成功通知
+
+在 deploy job 最后增加一个 step，deploy 成功后：
+1. 从 Gitea API 查询触发 commit 关联的 merged PR
+2. 获取 PR 作者
+3. 通过 Mail API 发送成功通知给 PR 作者 + pangtong-fujunshi
+4. 如果是 direct push（非 PR merge），只通知 jiangwei-infra + pangtong-fujunshi
+
+**文件**：`.gitea/workflows/deploy.yml`
+
+**改动范围**：deploy job 内新增 1 个 step（约 30 行 shell）
+
+**关键逻辑**：
+```bash
+# 查询关联 PR
+PR_INFO=$(curl -sf \
+  -H "Authorization: token $GITEA_TOKEN" \
+  "$API_URL/repos/$REPO/pulls?state=closed&limit=5" | \
+  python3 -c "
+import json,sys
+for pr in json.load(sys.stdin):
+    if pr.get('merge_commit_sha','') == '$COMMIT_SHA':
+        print(pr['user']['login'])
+        break
+" 2>/dev/null || echo "")
+
+# 发 Mail
+curl -s -X POST http://localhost:8083/api/mail \
+  -H "Content-Type: application/json" \
+  -d "{\"from\":\"daemon\",\"to\":\"$PR_AUTHOR\",\"title\":\"...\",\"text\":\"...\",\"type\":\"inform\"}"
+```
+
+**约束**：
+- 使用 `if: always()` + shell 判断 `needs.deploy.result == "success"`，确保只在成功时执行
+- GITEA_TOKEN 通过 secrets 注入
+- Mail API 调用超时 5 秒，失败不影响部署结果
+
+### 不改的文件
+
+| 文件 | 原因 |
+|---|---|
+| `src/api/toolchain_routes.py` | 不新增 webhook 事件，deploy 成功通知在 yml 内闭环 |
+| `src/daemon/toolchain_handler.py` | 不涉及 |
+| `templates/toolchain/*.md` | 不新增模板，通知内容直接在 shell 中构建 |
+| `scripts/deploy.sh` | 已完整，不需改动 |
+
+## 影响范围
+
+- **风险**：低。只在 deploy job 末尾追加通知 step，不修改已有的 deploy/notify 逻辑
+- **回退**：删除新增 step 即可
+- **测试**：push main 后观察 deploy workflow 执行结果
+
+## E2E 验证计划
+
+1. 在生产仓库创建测试分支，push → PR → merge → 触发 deploy
+2. 验证 deploy 成功后 Mail 通知到达 PR 作者
+3. 验证部署文件同步到 `~/.sanguo_projects/sanguo_moziplus_v2/`
+4. 验证 health check 通过
@@ -0,0 +1,144 @@
+# §23 — 工具链事件中枢补全：PR 全生命周期通知
+
+> 状态：草案，待评审
+> 作者：庞统
+> 日期：2026-06-11
+> 框架：基于 §20 Task Type Architecture + §13 工具链设计
+
+## 背景
+
+### 问题
+
+工具链事件中枢（`toolchain_routes.py`）当前只覆盖了 PR 生命周期中约一半的交互节点。review 驳回后 PR 作者修改代码，没有机制通知 reviewer 重新 review——流程在这里断链。
+
+### 当前覆盖
+
+| 事件节点 | handler | 模板 | 状态 |
+|---|---|---|---|
+| PR 创建 → 通知 reviewer | `_handle_pull_request` (opened) | `review_request.md` | ✅ |
+| Review 通过 → 通知 PR 作者 | `_handle_pull_request_review` (APPROVED) | `review_result.md` | ✅ |
+| Review 驳回 → 通知 PR 作者 | `_handle_pull_request_review` (REQUEST_CHANGES) | `review_result.md` | ✅ |
+| Issue 指派 → 通知被指派人 | `_handle_issues` (assigned) | `issue_assigned.md` | ✅ |
+| CI 失败评论 → 通知 | `_handle_issue_comment` ([CI]) | `ci_failure.md` | ✅ |
+| 部署失败 Issue → 通知 | `_handle_issues` (opened+"部署失败") | `deploy_failure.md` | ✅ |
+
+### 缺失节点
+
+| # | 事件节点 | Gitea 事件 | 优先级 | 理由 |
+|---|---|---|---|---|
+| E1 | PR 更新（push 新 commit）→ 通知 reviewer | `pull_request.synchronize` | **高** | review 驳回→修改→重 review 的关键闭环 |
+| ~~E2~~ | ~~PR 合并通知~~ | ~~已删除~~ | ~~—~~ | ~~和 §22 CD 成功通知重叠，已删~~ |
+| E3 | Review 评论（COMMENTED）→ 通知 PR 作者 | `pull_request_review` (COMMENTED) | 中 | reviewer 讨论提问，作者应知道 |
+| E4 | PR 上普通评论 → 通知相关人 | `issue_comment` (on PR) | 低 | 非关键路径 |
+
+## 方案
+
+### 框架对齐
+
+按 §20 Task Type Architecture，新增事件处理遵循：
+1. `_EVENT_HANDLERS` 映射 → 路由到对应 handler 函数
+2. handler 提取变量 → `render_template()` 渲染模板
+3. `_TEMPLATE_MAP` 注册模板名 → `templates/toolchain/` 下新建模板文件
+4. 通知目标通过 Gitea username → `to_agent_id()` 映射
+
+### 新增 Handler 1：`_handle_pull_request_synchronize`
+
+**触发**：`pull_request` 事件 + `action=synchronize`（PR 分支有新 push）
+
+**通知对象**：PR 的 reviewer（从 PR 的 `requested_reviewers` 或最近一次 non-COMMENTED review 的提交者）
+
+**实现**：
+
+修改 `_handle_pull_request` 的 action 过滤，从只处理 `opened` 扩展为同时处理 `synchronize`：
+
+```python
+async def _handle_pull_request(payload: Dict[str, Any]) -> None:
+    action = payload.get("action", "")
+    if action == "opened":
+        await _handle_pr_opened(payload)
+    elif action == "synchronize":
+        await _handle_pr_synchronize(payload)
+    # 其他 action 忽略
+```
+
+新增 `_handle_pr_synchronize`：
+1. 从 payload 取 PR 信息（number、title、author、head sha）
+2. 查询最近一次 review（Gitea API `GET /repos/{owner}/{repo}/pulls/{number}/reviews`）取 reviewer
+3. 如果没有 review 记录（首次 push 后 reviewer 还没 review），跳过（opened 事件已经通知过了）
+4. 渲染 `review_updated.md` 模板，发送 Mail 给 reviewer
+
+**关键设计决策**：
+- 不用 `requested_reviewers`（可能为空），用最近 review 的提交者
+- 只在有 review 历史时才通知（避免 opened + synchronize 重复通知）
+- Mail from 用 `system`
+
+### ~~Handler 2：PR 合并通知~~ — 已删除
+
+> 司马懿 review 指出与 §22 CD 成功通知重叠。CD 成功通知已隐含合并信息，无需单独发 merged 通知。
+
+### 新增 Handler 3：review COMMENTED 处理
+
+**触发**：`pull_request_review` 事件 + `state=COMMENTED`
+
+**通知对象**：PR 作者（不是 reviewer）
+
+**实现**：
+
+修改现有 `_handle_pull_request_review`，当前逻辑是"非 COMMENTED 才通知"，改为 COMMENTED 也通知，但用不同模板：
+
+```python
+    # 现有逻辑：非 COMMENTED 通知 PR 作者
+    if state in ("APPROVED", "REQUEST_CHANGES"):
+        template_name = "review_result"
+    elif state == "COMMENTED":
+        template_name = "review_comment"
+    else:
+        return  # PENDING 等忽略
+```
+
+### 新增模板
+
+| 模板文件 | 变量 | 说明 |
+|---|---|---|
+| `review_updated.md` | repo, pr_number, pr_title, pr_author, branch, new_sha, reviewer | PR 有新 commit，请重新 review |
+| ~~`pr_merged.md`~~ | ~~已删除~~ | ~~—~~ |
+| `review_comment.md` | repo, pr_number, pr_title, reviewer, comment_body | reviewer 提交了评论 |
+
+### `_EVENT_HANDLERS` 无需改动
+
+`synchronize` 和 `closed` 都是 `pull_request` 事件的 action 子类型，已映射到 `_handle_pull_request`。COMMENTED 是 `pull_request_review` 的 state 子类型，已映射到 `_handle_pull_request_review`。
+
+所以 **`_EVENT_HANDLERS` 不需要修改**，只需修改 handler 内部的 action/state 分发逻辑。
+
+### 不做的事
+
+| 项 | 理由 |
+|---|---|
+| E4 PR 上普通评论通知 | 低优，非关键路径，后续按需加 |
+| Issue 关闭通知 | 低优，关怀性质 |
+| reviewer 从 `requested_reviewers` 取 | 不可靠（可能为空），用最近 review 记录更稳定 |
+
+## 改动范围
+
+| 文件 | 改动 |
+|---|---|
+| `src/api/toolchain_routes.py` | 修改 `_handle_pull_request`（扩展 action 分发）+ 新增 `_handle_pr_synchronize` + 修改 `_handle_pull_request_review`（支持 COMMENTED） |
+| `templates/toolchain/review_updated.md` | 新增 |
+| ~~`templates/toolchain/pr_merged.md`~~ | ~~已删除~~ |
+| `templates/toolchain/review_comment.md` | 新增 |
+| `src/daemon/toolchain_templates.py` | `_TEMPLATE_MAP` 新增 3 个映射 |
+| `docs/design/23-toolchain-pr-lifecycle.md` | 本文档 |
+
+## 验证计划
+
+在 `sanguo/moziplus-v2` 测试仓库上 E2E 验证：
+
+1. **synchronize**：创建 PR → review 驳回 → push 新 commit → 验证 reviewer 收到"请重新 review" Mail
+~~2. merged~~：已删除
+3. **COMMENTED**：review 提交纯评论 → 验证 PR 作者收到通知
+
+## 风险评估
+
+- **风险等级**：低。新增事件处理，不修改现有 handler 逻辑
+- **幂等性**：复用现有 `_is_duplicate` 机制
+- **性能**：synchronize handler 有一次 Gitea API 调用（查 review 历史），频率低（只在 push 后触发）
@@ -0,0 +1,205 @@
+# §24 — Compact 检测方案修正
+
+> 状态：v3（rotation-only），待实施
+> 作者：庞统
+> 日期：2026-06-11
+> 框架：基于 §07 Spawner Acquire-First
+> 评审：仲达 3 轮评审（v1 trajectory → v2 gateway precheck → v3 rotation-only）
+
+## 1. 问题
+
+### 1.1 现象
+
+2026-06-11 14:02，pangtong main session 正在做 compaction（13:59:26 开始，14:06:00 结束，耗时 ~6.5 分钟），但 spawner Phase 2 检查时 `compact=False`，仍然 spawn 了新进程处理 Mail，导致两个 agent turn 撞车。
+
+### 1.2 根因
+
+当前 compact 检测方法 `_check_recent_compaction_jsonl` 扫描 session jsonl，查找 `type == "compaction"` 事件。这是 compact **完成后**才写入的摘要记录，compact **进行中**时不存在 → 漏检。
+
+同时 Gateway 触发 compact 时先把 session 标为 `done`，所以 `status=running + lock_pid_alive` 检查也无效。14:02:11 实际状态：`status=done lock_pid_alive=False compact=False`——三个检查全部漏过。
+
+## 2. 方案：Rotation-Only 检测（v3）
+
+### 2.1 核心洞察（仲达 v2 评审）
+
+v2 方案依赖 `[context-overflow-precheck]` route=compact 作为开始标志。但实测数据：
+
+| Agent | Rotation 事件 | 有 Precheck | 无 Precheck |
+|-------|:---:|:---:|:---:|
+| pangtong | 7 | 3 | 4 |
+| simayi | 3 | 0 | 3 |
+
+**10 次 compact 只有 3 次有 precheck，覆盖率 30%。** 原因：post-compact retry 触发的后续 compact 不经过 precheck 日志路径。
+
+**结论**：开始标志不可靠。反转检测逻辑——只用可靠的 rotation 事件作为信号。
+
+### 2.2 Rotation 事件
+
+Gateway 日志中 `[compaction] rotated active transcript after compaction (sessionKey=...)` 事件：
+- **100% 覆盖率**：全天 10 次 compact 全部有 rotation 事件
+- **含 sessionKey**：可以精确匹配目标 session
+- **JSON 格式**：易解析
+
+### 2.3 检测逻辑
+
+```
+1. 读 gateway 日志（当天 + 昨天尾部）
+2. 按目标 sessionKey 过滤 compact 相关事件
+3. 从后往前找最后一条 rotation 事件：
+   a. 如果 rotation 事件在窗口内（< 120s）→ compact=True
+      （刚完成一轮 compact，可能还在 post-compact retry 循环中）
+   b. 无 rotation 事件或超出时间窗口 → compact=False
+
+**注意：此方案仅检查 rotation 事件，不检查 model.completed 等其他事件。**
+这是有意为之的保守策略：不检查正常 turn 事件意味着 compact 完成后的
+120s 内都可能被误判为 compact 进行中，但误判代价低（仅 skip 一轮 ticker），
+宁可多拦也不漏放。
+```
+
+**为什么 rotation + 时间窗口就够了？**
+- compact 后 Gateway 会 retry prompt
+- 如果 retry 又触发 overflow → 又一轮 compact → 又一个 rotation 事件
+- 如果 retry 成功 → 正常 turn → 新的 session.started / model.completed 事件
+- 所以「最近一个事件是 rotation 且时间很近」= compact 循环还在进行
+
+### 2.4 时间窗口选择
+
+compact 通常耗时 1-10 分钟。post-compact retry 如果又触发 compact，间隔通常 <60 秒。
+
+- **窗口太短（如 30s）**：可能漏掉 compact 结束后正在 retry 但还没触发下一轮的场景
+- **窗口太长（如 900s）**：compact 完成后正常工作很久了还误判
+- **推荐 120s**：compact 循环中两次 rotation 间隔通常 <60s，120s 有足够余量
+
+误判代价低（skip 一轮 ticker），所以宁可多拦也不漏放。
+
+## 3. 改动范围
+
+| 文件 | 改动 | 行数估计 |
+|------|------|---------|
+| `spawner.py` | 新增 `_check_compact_in_progress_gateway()` | ~40 行 |
+| `spawner.py` | `_check_session_state()` 调用新方法，替换旧方法 | ~5 行 |
+| `spawner.py` | 日志路径配置化 | ~5 行 |
+| `docs/design/07-spawner-acquire-first.md` | §4.5 O5 更新 | ~10 行 |
+| `docs/design/24-compact-detection-fix.md` | 本文档 | 已有 |
+
+**总计 ~60 行代码改动。**
+
+## 4. 实现细节
+
+### 4.1 核心方法
+
+```python
+def _check_compact_in_progress_gateway(self, session_key: str, window_seconds: int = 120) -> bool:
+    """检查 gateway 日志，判断指定 session 是否刚完成 compact（可能在 retry 循环中）。
+    
+    检测逻辑：如果目标 session 最近一个事件是 rotation 且在窗口内，视为 compact 进行中。
+    """
+    log_paths = self._get_recent_gateway_logs()
+    if not log_paths:
+        return False
+    
+    now = datetime.now(timezone.utc)
+    window_start = now - timedelta(seconds=window_seconds)
+    
+    last_rotation_time = None
+    
+    for log_path in log_paths:
+        if not os.path.exists(log_path):
+            continue
+        
+        with open(log_path, 'rb') as f:
+            # 读尾部 2MB
+            f.seek(0, 2)
+            size = f.tell()
+            f.seek(max(0, size - 2 * 1024 * 1024))
+            
+            for raw_line in f:
+                try:
+                    obj = json.loads(raw_line)
+                except (json.JSONDecodeError, ValueError):
+                    continue
+                
+                msg = obj.get("message", "")
+                ts_str = obj.get("time", "")
+                
+                # 只看包含目标 sessionKey 的事件
+                if session_key not in msg:
+                    continue
+                
+                # rotation 事件
+                if "[compaction] rotated active transcript" in msg:
+                    try:
+                        event_time = datetime.fromisoformat(ts_str)
+                        if last_rotation_time is None or event_time > last_rotation_time:
+                            last_rotation_time = event_time
+                    except (ValueError, TypeError):
+                        continue
+    
+    if last_rotation_time is not None:
+        return last_rotation_time >= window_start
+    
+    return False
+```
+
+### 4.2 日志路径
+
+```python
+def _get_recent_gateway_logs(self) -> list:
+    """获取当天和昨天的 gateway 日志路径"""
+    log_dir = os.environ.get("OPENCLAW_LOG_DIR", "/tmp/openclaw")
+    today = datetime.now().strftime("%Y-%m-%d")
+    yesterday = (datetime.now() - timedelta(days=1)).strftime("%Y-%m-%d")
+    paths = []
+    for d in [today, yesterday]:
+        p = os.path.join(log_dir, f"openclaw-{d}.log")
+        if os.path.exists(p):
+            paths.append(p)
+    return paths
+```
+
+### 4.3 Phase 2 集成
+
+```python
+# 在 _check_session_state 中，不依赖 status，直接检查
+compact = self._check_compact_in_progress_gateway(session_key)
+if not compact:
+    compact = self._check_recent_compaction_jsonl(...)  # fallback
+    
+if compact:
+    blockers.append(("session_compacting", None))
+```
+
+## 5. 边界情况
+
+| 边界情况 | 处理 |
+|---------|------|
+| 日志文件不存在 | 返回 False（fallback 到旧方法） |
+| 跨天 compact | 同时检查昨天日志尾部 |
+| compact 失败（无 rotation） | rotation 事件不会出现 → 检测不到 → 回退到旧方法 |
+| 误判（compact 完成后正常工作中） | 时间窗口 120s 内可能被误判，但代价低（skip 一轮 ticker）。不检查正常 turn 事件，是保守策略 |
+
+## 6. 测试验证
+
+### 6.1 单元测试
+
+- `_check_compact_in_progress_gateway`：
+  - rotation 事件在窗口内 → True
+  - rotation 事件超出窗口 → False
+  - 无 rotation 事件 → False
+  - 日志不存在 → False
+  - sessionKey 不匹配 → False
+
+### 6.2 集成验证
+
+- `pytest -m "not e2e"` 全量测试
+
+## 7. 关联设计
+
+- §07 Spawner Acquire-First（§4.5 O5 compact 扫描条件收紧）
+- §08 Classify Outcome Optimization（compact_hanging 处理）
+
+## 8. 评审记录
+
+- **v1**：trajectory jsonl 间接推断 → 仲达指出 trajectoryPath 不可用、需多文件等 3 个问题
+- **v2**：gateway 日志 precheck 开始标志 → 仲达指出开始标志覆盖率仅 30%，建议 rotation-only
+- **v3**：rotation-only（当前版本）→ 仲达已确认方向，待代码实现后再审
@@ -0,0 +1,224 @@
+# v3.0 vs HEAD 背靠背 Review — 庞统
+
+**日期**: 2026-06-11
+**范围**: v3.0 tag → HEAD（6 commits, Step 2-5 Task 五层架构重构）
+**对比**: `git diff v3.0..HEAD` + 安装目录代码验证
+
+---
+
+## Part A: v3.0 逻辑丢失检查
+
+### 方法论
+v3.0 → HEAD 的重构将 `_mail_*` 硬编码逻辑统一为 handler 架构（TaskTypeRegistry + BaseTaskHandler）。核心变更：
+- dispatcher.py: `_mail_on_checks_passed` / `_mail_on_complete` → `_handler_on_checks_passed` / `_handler_on_complete`
+- spawner.py: `_build_mail_prompt` → handler.build_prompt
+- ticker.py: `_mail_check_reply` → handler.check_completion, `_mail` 硬编码 → `TaskTypeRegistry.virtual_projects()`
+
+### 检查结果
+
+| # | 文件 | v3.0 逻辑 | 当前状态 | 严重度 | 说明 |
+|---|------|----------|---------|--------|------|
+| 1 | dispatcher.py | `_legacy_on_complete` 中 review verdict 处理（approved→done, 非 approved→@mention assignee） | **缺失** | 🔴 | 新版 `_legacy_on_complete` 在 `_is_review=True` 时只有 crash rollback，**没有 verdict 判断逻辑**。review agent 完成后任务永远不会从 review→done。**仅影响非 handler 项目（_general）**。handler 项目（_mail/_toolchain）的 review 由 TaskHandler.post_complete 正确处理 |
+| 2 | dispatcher.py | `_mail_auto_working` / `_mail_auto_complete` / `_mail_revert_to_pending` 方法 | 保留但主流程不再调用 | 🟢 | 方法体仍存在（标记为 deprecated），主流程改走 handler.pre_spawn / handler.post_complete。正常的重构 |
+| 3 | dispatcher.py | spawn 失败回退 `working→pending` | **逻辑改进** | 🟢 | v3.0 用 `_mail_revert_to_pending`（只处理 _mail），新版用通用 DB 操作处理所有 handler 项目 |
+| 4 | spawner.py | `_build_mail_prompt` 精简模板 | **替换为 handler.build_prompt** | 🟢 | MailHandler 使用 PromptSection 组装，功能更完整 |
+| 5 | spawner.py | `_build_api_section` 中 mail 直接 done | **替换为 handler.target_success_status** | 🟢 | 等价实现 |
+| 6 | ticker.py | `_mail` 硬编码虚拟项目 | **替换为 TaskTypeRegistry.virtual_projects()** | 🟢 | 正常重构，可扩展 |
+| 7 | ticker.py | `_mail_check_reply` 兜底（超时检查） | **替换为 handler.check_completion** | 🟢 | 等价实现，缩进正确 |
+| 8 | ticker.py | `_dispatch_reviews` 跳过 `_mail` | **替换为 handler 检查** | 🟢 | 等价 |
+
+### 🔴 严重问题 #1 详解
+
+**位置**: `dispatcher.py` L250-260 `_legacy_on_complete`
+
+**v3.0 逻辑**（已删除）:
+```python
+if _is_review:
+    if _task_db and outcome in ("completed", "session_revived"):
+        # 读 verdict
+        if verdict == "approved":
+            _dispatcher._mark_task_status(_task_db, _task_id, "done")
+        else:
+            # @mention assignee + 保持 review
+            bb.add_comment(_task_id, "daemon", f"@{assignee} 审查结论: {verdict_str}")
+```
+
+**当前逻辑**:
+```python
+def _legacy_on_complete(aid, outcome):
+    if outcome in ROLLBACK_CURRENT_AGENT_OUTCOMES and _task_db:
+        _dispatcher._rollback_current_agent(_task_db, _task_id, aid)
+    if not _is_review:  # ← review 时什么都不做
+        _dispatcher._task_auto_complete(_task_id, _task_db)
+```
+
+**影响**: `_dispatch_reviews` (ticker.py:1307) 对非 handler 项目会 dispatch review agent。review agent 完成后走 `_legacy_on_complete`，但 `_is_review=True` 时逻辑为空。任务永远停在 `review` 状态。
+
+**修复方案**: 在 `_legacy_on_complete` 中补充 review verdict 处理逻辑，或让非 handler 项目也走 TaskHandler（注册 `_general` 到 TaskTypeRegistry）。
+
+---
+
+## Part B: 专题 01-13 设计编码一致性
+
+### 专题 01: 四相循环（不参考实现，只检查设计遗漏）
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | §3.3 Spawn Prompt 框架（任务+约束+API+准则+完成标准） | ✅ BootstrapBuilder + PromptSection 实现 | ✅ | |
+| 2 | §3.4 @mention 通知机制 | ✅ `_process_mentions` + `mention_queue` | ✅ | |
+| 3 | §4 庞统 Review 机制（三问） | ✅ review agent + verdict 处理 | ✅ | |
+
+**设计遗漏**: 无明显遗漏。
+
+### 专题 02: Main Session + Delegation
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | 3.1 投递到 Main Session | ✅ `use_main_session=True` 参数 | ✅ | |
+| 2 | 3.2 Delegation（subagent-delegation skill） | ✅ 外部 skill，不在此代码库 | ✅ | |
+| 3 | 3.3 续杯机制 | ✅ `use_main_session=True` + session 复用 | ✅ | |
+| 4 | 4.1 投递消息格式 | ✅ dispatcher 构建 | ✅ | |
+| 5 | 4.3 消息优先级与中断策略 | ❌ 无优先级队列 | ⚠️ | 设计描述了优先级但未实现，非关键 |
+| 6 | 4.4 Subagent 背压控制 | ❌ 无显式背压 | ⚠️ | 靠 counter 间接控制 |
+
+### 专题 03: Prompt 进化
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | 3.1 广播认领模板改写 | ✅ PromptSection 组装 | ✅ | |
+| 2 | P4 群体智能（Boids） | ✅ agent 自主决策 | ✅ | 设计原则，非具体代码 |
+| 3 | P6 反静默降级 | ❌ 无 scope reduction detection 自动机制 | ⚠️ | 设计原则，未自动实现 |
+| 4 | P7 经验闭环 | ❌ 无 IMPROVE 阶段自动触发 | ⚠️ | P4 级待实现 |
+
+### 专题 04: 黑板协作模型
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | 3.1 assignee 降级为显示字段，路由走 @mention | 🟡 assignee 仍做直接路由 | ⚠️ | router.py L160-166 仍有 assignee 快速路径。设计说 Phase 1 双轨并行，当前停在 Phase 1。未迁移到 Phase 2 |
+| 2 | 3.2 @mention 语义增强（mention_queue + comment_type） | ✅ 已实现 | ✅ | |
+| 3 | 3.3 多人协作模式（co_assignees） | ❌ 无 co_assignees 字段 | ❌ | 数据库无此列 |
+| 4 | 3.4 信息关联模型（output↔comment link） | ❌ 无关联字段 | ❌ | outputs 表无 comment_id 列 |
+| 5 | 3.5 层级查询 API | ✅ parent_task 支持 | ✅ | |
+
+**总结**: 3.3 和 3.4 设计了但未实现。3.1 停在 Phase 1。
+
+### 专题 05: 上下文四层架构
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | L0 铁律层 | ✅ 通过 workspace 文件注入 | ✅ | |
+| 2 | L1 角色层 | ✅ SOUL.md / IDENTITY.md | ✅ | |
+| 3 | L2 引擎注入层 | ✅ BootstrapBuilder | ✅ | |
+| 4 | L3 被动参考层 | ❌ 无 _inject_wiki_knowledge | ❌ | wiki 知识注入未实现 |
+
+### 专题 06: PM2 Crash 恢复
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | 4.1 总体流程（_startup_recover） | ✅ ticker.py:1614 | ✅ | |
+| 2 | 4.2 claimed 状态恢复 | ✅ | ✅ | |
+| 3 | 4.2 working 状态恢复 | ✅ `_recover_working_task` | ✅ | |
+| 4 | 4.2 review 状态恢复 | ✅ `_recover_review_task` | ✅ | |
+| 5 | 设计提到 7 个恢复方法 | 🟡 只看到 2 个公开方法 | ⚠️ | 可能在内部逻辑中覆盖，需详细检查 |
+
+### 专题 07: Spawner Acquire-First
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | Phase 0: Pre-acquire 修复 | ✅ L499-512 | ✅ | |
+| 2 | Phase 1: Counter acquire | ✅ L516-521 | ✅ | |
+| 3 | Phase 2: Session check | ✅ L523-568 | ✅ | |
+| 4 | Phase 2.5: 假死修复 | ✅ L557-568 | ✅ | |
+| 5 | O1: lock PID 死 + running 假死 | ✅ | ✅ | |
+| 6 | O4: revive 清理 lock 文件 | ✅ | ✅ | |
+
+### 专题 08: Classify Outcome 优化
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | A0-A17 判定树 | ✅ `_classify_outcome` 方法 | ✅ | |
+| 2 | A9 api_error 特殊路径 | ✅ api_retry_count | ✅ | |
+| 3 | A14-A17 可恢复 retry + cooldown 60s | ✅ cooldown_seconds + set_cooldown | ✅ | |
+| 4 | Gateway Watchdog | ✅ 外部脚本 | ✅ | |
+| 5 | Registry 逻辑删除 | ✅ | ✅ | |
+
+### 专题 09: Rebuttal + Goal Gate
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | 2.1 Rebuttal 自动化（review 非 approved → @mention assignee） | ✅ task_handler.py handle_review_complete + ticker.py _rebuttal_on_complete | ✅ | |
+| 2 | 2.1 防止无限循环（max 2 轮） | ✅ RebuttalManager.MAX_ROUNDS | ✅ | |
+| 3 | 2.2 目标一致性 Gate | ❌ 无 goal gate 自动检查 | ⚠️ | 设计为 Agent 端行为，非 Daemon 侧 |
+| 4 | _task_on_complete 改动（design §2.1 代码改动） | 🟡 已移到 handler | ✅ | 重构后的等价位置 |
+
+### 专题 10: T3 需求探索 + 黑板展示
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | A2: 需求探索过程写黑板 comments | ✅ 后端支持 comment_type | ✅ | |
+| 2 | A3: TaskModal 实时刷新 | ✅ SSE comment_added/checkpoint_resolved | ✅ | |
+| 3 | D1: 砍掉 AI 摘要 | ✅ 黑板直投前端 | ✅ | |
+| 4 | D2: SSE 只做通知 | ✅ 前端按需拉数据 | ✅ | |
+
+### 专题 11: 上下文四层重设计
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | L2 操作规范型 6 个 skill 全文注入 | ❌ BootstrapBuilder 只注入通用 prompt，无 skill 全文注入 | ❌ | 设计 §2.3 要求将 6 个操作规范型 skill（blackboard-executor, code-review 等）全文注入 L2，bootstrap.py 无此逻辑 |
+| 2 | L3 _inject_wiki_knowledge | ❌ 完全未实现 | ❌ | |
+| 3 | review_protocols/ 目录 | ❌ 目录不存在 | ❌ | |
+| 4 | 2.3 提到的 handoff.schema.json | ❌ 不存在 | ❌ | |
+
+**总结**: 专题 11 大部分 L2/L3 改造未实现。BootstrapBuilder 做了基础框架但缺少 skill 注入和知识注入。
+
+### 专题 12: Pipeline 设计
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | §3 Pipeline 注册表（pipeline 字段） | ❌ 无 pipeline 数据结构 | ❌ | |
+| 2 | §4 路由逻辑更新（task_type 路由） | ❌ router.py 无 task_type 路由 | ❌ | |
+| 3 | §8 Pipeline 引擎 + PipelineRegistry | ❌ 不存在 | ❌ | |
+| 4 | §8.2 状态流转校验 | ❌ 无 flow_rules | ❌ | |
+| 5 | §9 实施路线标记为 "待实现" | — | — | 设计文档本身就标记为 TODO |
+
+**总结**: Pipeline 整个设计未实施。设计文档 §9 自身标记为待实现。
+
+### 专题 13: 工具链开发工作流（不参考实现，只检查设计遗漏）
+
+| # | 设计描述 | 代码状态 | 一致性 | 说明 |
+|---|---------|---------|--------|------|
+| 1 | §16 工具链事件中枢 | ✅ toolchain_routes.py + toolchain_handler.py | ✅ | |
+| 2 | Gitea webhook 处理 | ✅ 5 模板 + 去重 | ✅ | |
+| 3 | CI 前缀 [CI] | ✅ | ✅ | |
+| 4 | §5 CI/CD 管道设计 | 🟡 Gitea Actions 为主，非 Daemon 侧 | ✅ | |
+
+**设计遗漏**: 无明显遗漏。
+
+---
+
+## 汇总
+
+### 🔴 严重（需修复）
+
+| # | 问题 | 影响 |
+|---|------|------|
+| A1 | `_legacy_on_complete` review verdict 处理丢失 | 非 handler 项目（_general）的 review agent 完成后任务永远停在 review 状态 |
+
+### 🟡 中等（设计-代码不一致，可后续处理）
+
+| # | 专题 | 设计描述 | 实际状态 |
+|---|------|---------|---------|
+| B4-1 | 04 黑板协作 | 3.1 assignee 降级 Phase 2 | 停在 Phase 1 |
+| B4-3 | 04 黑板协作 | 3.3 co_assignees 多人协作 | 未实现 |
+| B4-4 | 04 黑板协作 | 3.4 output↔comment 关联 | 未实现 |
+| B5-4 | 05 上下文层 | L3 wiki 知识注入 | 未实现 |
+| B11-1 | 11 上下文重设计 | L2 操作规范型 skill 全文注入 | 未实现 |
+| B11-2 | 11 上下文重设计 | handoff.schema.json | 未实现 |
+| B11-3 | 11 上下文重设计 | review_protocols/ 目录 | 未实现 |
+| B12 | 12 Pipeline | 整个 Pipeline 引擎 | 未实现（设计自标 TODO） |
+
+### 🟢 正常（重构等价或设计已标记待实现）
+
+- _mail_* 方法 deprecated 但保留（平滑迁移）
+- handler 架构统一替代硬编码（等价实现）
+- 专题 01/02/03/06/07/08/09/10/13 无严重不一致
@@ -0,0 +1,707 @@
+# v3.0 vs HEAD 背靠背 Review — 司马懿
+
+> **日期**: 2026-06-10 (v2)  
+> **范围**: v3.0 tag → HEAD（6 commits, +1584/-134 行, 9 个文件）  
+> **方法**: `git diff v3.0..HEAD` 逐文件逐行比对 + v3.0 源码 `git show v3.0:` 回溯验证  
+> **独立判断**: 不参考庞统 review，独立产出后比对
+
+---
+
+## 总览
+
+v3.0 → HEAD 的核心改动是 **Step 2-5 五层架构重构**：
+
+| 层 | 新增/改动 | 说明 |
+|---|---------|------|
+| Protocol + Registry | `task_type_registry.py`（已有，未改） | `TaskTypeHandler` Protocol + `TaskTypeRegistry` |
+| 基类 | `base_task_handler.py`（新增 +183） | `BaseTaskHandler` — crash/verify/mark/notify 统一流程 |
+| Handler × 3 | `task_handler.py`（+378）、`mail_handler.py`（+210）、`toolchain_handler.py`（+277） | 各自实现 `build_prompt` / `verify_completion` / `post_complete` |
+| 引擎接入 | `dispatcher.py`（-95/+58）、`spawner.py`（+38）、`ticker.py`（+31/-27） | `_mail_*` 硬编码 → `TaskTypeRegistry` 查表 |
+| 设计文档 | `step5-impact-analysis.md`（+324）、`step5-audit-report.md`（+74） | 影响分析 + 双重审计 |
+
+**核心结论**：架构方向正确，但 **handler 注册初始化缺失导致所有 handler 路径为死代码**，实际运行仍走 `_legacy_on_complete` 旧路径。旧路径中 review verdict 处理被删除，造成 **非 handler 项目的 review 流程失效**。
+
+---
+
+## Part A: v3.0 逻辑丢失检查
+
+### 方法论
+
+逐文件追踪 v3.0 中每个 `_mail` / `_task` / `project_id == "_mail"` 分支，验证 HEAD 中是否存在等价实现。分三层检查：
+1. **功能等价**：新代码是否完整覆盖旧逻辑
+2. **路径可达**：新代码是否会被实际执行（handler 注册？legacy fallback？）
+3. **行为一致**：边界条件、异常处理是否等价
+
+### 检查结果
+
+#### A1 🔴 致命：dispatcher.py — review verdict 处理丢失
+
+**v3.0 逻辑**（`dispatcher.py` L253-308 `_task_on_complete`）：
+```python
+if _is_review:
+    if _task_db and outcome in ("completed", "session_revived"):
+        # 读 verdict
+        review = conn.execute(
+            "SELECT verdict FROM reviews WHERE task_id=? ORDER BY created_at DESC LIMIT 1",
+            (_task_id,)).fetchone()
+        if review and review["verdict"] == "approved":
+            _dispatcher._mark_task_status(_task_db, _task_id, "done")
+        else:
+            # 非 approved → @mention assignee + 保持 review
+            bb.add_comment(_task_id, "daemon", f"@{assignee} 审查结论: {verdict_str}")
+```
+
+**HEAD 逻辑**（`dispatcher.py` L246-258 `_legacy_on_complete`）：
+```python
+def _legacy_on_complete(aid, outcome):
+    if outcome in ROLLBACK_CURRENT_AGENT_OUTCOMES and _task_db:
+        _dispatcher._rollback_current_agent(_task_db, _task_id, aid)
+    if not _is_review:  # ← review 时整个 if 被跳过
+        _dispatcher._task_auto_complete(_task_id, _task_db)
+```
+
+**分析**：
+- `_legacy_on_complete` 在 `_is_review=True` 时**什么也不做**——无 verdict 读取、无 done 标记、无 @mention 通知
+- `TaskHandler.handle_review_complete()` 方法有完整 verdict 处理，但 handler 未注册（见 A2），此代码不可达
+- **影响**：所有非 handler 项目（`_general` 等）的 review agent 完成后，任务永远停在 `review` 状态
+
+**补充**：rebuttal 路径不受影响——`_rebuttal_on_complete` 在 `ticker.py` L756-790 独立定义，直接读 verdict 并处理，不经过 `_legacy_on_complete`。
+
+---
+
+#### A2 🔴 致命：Handler 注册初始化缺失
+
+**证据**：
+```bash
+$ grep -rn "TaskTypeRegistry.register" src/
+# 零结果
+```
+
+`TaskTypeRegistry.register()` 在整个代码库中**从未被调用**。`TaskHandler` / `MailHandler` / `ToolchainHandler` 类已定义但从未实例化和注册。
+
+**后果链**：
+1. `TaskTypeRegistry.get_by_project()` 永远返回 `None`
+2. 所有 `if handler:` 分支不进入 → 走 `else` / fallback 路径
+3. `TaskTypeRegistry.virtual_projects()` 返回空列表 → `_mail` / `_toolchain` 不被 ticker 自动发现
+
+**各路径受影响分析**：
+
+| 路径 | dispatcher | spawner | ticker | 实际走什么 |
+|------|-----------|---------|--------|----------|
+| Mail `_mail` | `handler=None` → `_legacy_on_complete` | `handler=None` → 旧 `_build_mail_prompt` | `virtual_projects()` 空 → **_mail 不被 tick** | 旧路径（无 handler），但 **ticker 不扫描 _mail** |
+| Task `_general` | `handler=None` → `_legacy_on_complete` | `handler=None` → BootstrapBuilder | 不涉及 handler | 旧路径，但 review 处理被删（A1） |
+| Toolchain `_toolchain` | N/A | N/A | `virtual_projects()` 空 → **_toolchain 不被 tick** | **完全不可达** |
+
+**⚠️ A2 导致 ticker 不再扫描 `_mail` 虚拟项目**，这是 v3.0 有、HEAD 丢失的行为——v3.0 中 `_mail` 硬编码在 ticker L218-229，HEAD 改为 `TaskTypeRegistry.virtual_projects()` 但注册为空。
+
+**需要添加的初始化代码**（缺失）：
+```python
+# 应在 server.py 或 bootstrap.py 的启动流程中
+from src.daemon.task_handler import TaskHandler
+from src.daemon.mail_handler import MailHandler
+from src.daemon.toolchain_handler import ToolchainHandler
+
+TaskTypeRegistry.register(TaskHandler())
+TaskTypeRegistry.register(MailHandler())
+TaskTypeRegistry.register(ToolchainHandler())
+```
+
+---
+
+#### A3 🟡 中等：dispatcher.py — 旧 `_mail_*` 方法成为死代码
+
+**v3.0**：`_mail_auto_working` / `_mail_auto_complete` / `_mail_check_reply` / `_mail_revert_to_pending` 被 `dispatch()` 主流程调用。
+
+**HEAD**：这些方法仍保留在 dispatcher.py 中（L628-860），但主流程已改走 handler 路径。由于 handler 未注册，主流程走 `_legacy_on_complete`（无 handler 分支），也不调用这些方法。
+
+**结论**：方法体保留但无外部调用者，属于死代码。不影响当前运行（因为 `_legacy_on_complete` 有独立的 executor 逻辑），但增加维护混淆。
+
+---
+
+#### A4 🟢 低：dispatcher.py — spawn 失败回退等价
+
+**v3.0**：`self._mail_revert_to_pending(task.id, db_path)` — 调独立方法。
+**HEAD**：内联代码（L309-327），`BEGIN IMMEDIATE` + 状态检查 + `UPDATE ... SET status='pending'`。
+
+**等价**：新版逻辑更通用（不限于 `_mail`，任何 handler 项目都可回退）。
+
+---
+
+#### A5 🟢 低：dispatcher.py — `_legacy_dispatch` 路径 handler 化
+
+**v3.0**：`is_mail_legacy = project_id.get("project_id") == "_mail"`
+**HEAD**：`handler_legacy = TaskTypeRegistry.get_by_project(project_id_legacy)`
+
+**等价**：`handler_legacy` 为 None 时跳过 pre_spawn，与 v3.0 中 `is_mail_legacy=False` 行为一致。`_legacy_dispatch` 本身仅在 `router=None` 时触发，当前配置不会进入。
+
+---
+
+#### A6 🟢 低：spawner.py — prompt 构建双路径
+
+**v3.0**：`if project_id == "_mail": return self._build_mail_prompt(...)` → 走 BootstrapBuilder。
+**HEAD**：`handler = TaskTypeRegistry.get_by_project(project_id)` → `if handler: return handler.build_prompt(ctx)` → else 走 BootstrapBuilder。
+
+**分析**：
+- handler 未注册时，等价于 v3.0（走 BootstrapBuilder）
+- handler 注册后，Task/Mail/Toolchain 走新 PromptSection 路径
+- **注意**：新旧路径的 Skill 注入策略不同——旧路径（BootstrapBuilder）**全文注入** Skill，新路径（RoleSkillSection）只给**索引+引导语**。这可能导致 handler 注册后 Agent 行为变化
+
+---
+
+#### A7 🟢 低：spawner.py — `_build_api_section` success_status
+
+**v3.0**：`success_status = '"done"' if project_id == "_mail" else '"review"'`
+**HEAD**：`success_status = '"done"' if handler.target_success_status == "done" else '"review"'`
+
+**等价**：handler 未注册时走 else 分支 → `'"review"'`，与 v3.0 非 mail 项目一致。
+
+---
+
+#### A8 🟡 中等：spawner.py — retry prompt 仍用硬编码
+
+**v3.0**：`is_mail = project_id == "_mail"` → 用 `MAIL_RETRY_PROMPT` 模板。
+**HEAD**：同样 `is_mail = project_id == "_mail"` 硬编码（L1128），未改走 handler。
+
+**影响**：不影响功能（retry prompt 正确），但与设计文档 §6 不一致。属于 Step 5 未覆盖的遗留点。
+
+---
+
+#### A9 🟢 低：ticker.py — 虚拟项目扫描
+
+**v3.0**：硬编码 `_mail` 扫描。
+**HEAD**：`TaskTypeRegistry.virtual_projects()` 循环。
+
+**分析**：逻辑正确，但注册为空时 `_mail` 不被扫描（见 A2）。注册后自动发现 `_mail` + `_toolchain`，比 v3.0 更可扩展。
+
+---
+
+#### A10 🟢 低：ticker.py — assignee 清空条件
+
+**v3.0**：`if self._current_project_id == "_mail":` → 不清空 assignee。
+**HEAD**：`handler = TaskTypeRegistry.get_by_project(...); if handler:` → 不清空。
+
+**等价**：handler 未注册时，非 handler 项目正常清空 assignee。
+
+---
+
+#### A11 🟢 低：ticker.py — 跳过 claimed 状态
+
+**v3.0**：`if project_id == "_mail":` → 跳过 claimed，直接 working。
+**HEAD**：`handler = TaskTypeRegistry.get_by_project(project_id); if handler:` → 跳过。
+
+**等价**。
+
+---
+
+#### A12 🟢 低：ticker.py — review dispatch 跳过
+
+**v3.0**：`if project_id == "_mail": return []`
+**HEAD**：`handler = TaskTypeRegistry.get_by_project(project_id); if handler: return []`
+
+**等价**。
+
+---
+
+#### A13 🟢 低：ticker.py — 超时检查幻觉门控
+
+**v3.0**：
+```python
+if self._current_project_id == "_mail":
+    has_reply = self._mail_check_reply(task.id, db_path)
+    if has_reply:
+        ... # mark done
+```
+
+**HEAD**：
+```python
+handler = TaskTypeRegistry.get_by_project(self._current_project_id)
+if handler and handler.check_completion(task.id, db_path):
+    ... # mark done
+```
+
+**等价**：`MailHandler.check_completion` 内部调 `_check_reply`，查询语义与 v3.0 的 `_mail_check_reply` 完全一致（`SELECT id FROM tasks WHERE id != ? AND must_haves LIKE ?`）。
+
+**缩进问题**：HEAD L1483 `if handler and handler.check_completion(...):` 后续 body 缩进 5 级（28 空格），与同级代码不一致。不影响运行，但增加维护混淆。
+
+---
+
+#### A14 🟢 低：ticker.py — startup recovery 虚拟项目列表
+
+**v3.0**：`for virtual_id in ("_general", "_mail"):`
+**HEAD**：`virtual_ids = ["_general"] + TaskTypeRegistry.virtual_projects()`
+
+**等价**：注册为空时只有 `_general`，注册后自动包含 `_mail` + `_toolchain`。
+
+---
+
+### Part A 汇总
+
+| 严重度 | 数量 | 项目 |
+|--------|------|------|
+| 🔴 致命 | 2 | A1 review verdict 丢失, A2 handler 未注册 |
+| 🟡 中等 | 2 | A3 死代码未清理, A8 retry prompt 硬编码 |
+| 🟢 低 | 10 | A4~A7, A9~A14 |
+
+**A1+A2 联合根因分析**：
+
+设计意图是 handler 注册后 review 走 `TaskHandler.post_complete` → `handle_review_complete`。但注册代码缺失导致：
+1. 所有项目走 `_legacy_on_complete`（旧路径）
+2. 旧路径中 review 处理被删除（信任 handler 会处理）
+3. review agent 完成后无任何后续动作
+
+**同时**，ticker 不再扫描 `_mail` 虚拟项目（原来硬编码扫描），`_mail` 项目的 pending 任务无人处理。
+
+---
+
+## Part B: 13 个重点专题设计-编码一致性
+
+逐专题检查设计文档描述与 HEAD 代码的一致性。标记：
+- ✅ 一致
+- ⚠️ 设计已标注未实施/Phase N（不算差异）
+- ❌ 设计承诺但代码不一致
+- 🟡 部分一致
+
+---
+
+### B1: 专题 01 四相循环
+
+**设计文档**：`01-four-phase-loop.md` — PRD Phase 1~4 完整实现方案
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B1-1 | §3.3 Spawn Prompt 框架（任务+约束+API+准则+完成标准） | BootstrapBuilder + PromptSection 实现 | ✅ |
+| B1-2 | §3.4 @mention 通知机制 | `_process_mentions` + `mention_queue` | ✅ |
+| B1-3 | §4 Review 机制（verdict → done/notify） | `TaskHandler.handle_review_complete`（handler 未注册）+ `_rebuttal_on_complete`（ticker 独立） | ⚠️ handler 路径不可达，但 rebuttal 路径完整 |
+
+---
+
+### B2: 专题 02 Main Session + Delegation
+
+**设计文档**：`02-main-session-delegation.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B2-1 | §3.1 投递到 Main Session | `use_main_session=True` 参数 | ✅ |
+| B2-2 | §3.3 续杯机制 | `use_main_session=True` + session 复用 | ✅ |
+| B2-3 | §4.3 消息优先级与中断策略 | 无优先级队列 | ⚠️ 设计描述但未标注 Phase |
+| B2-4 | §4.4 Subagent 背压控制 | 无显式背压，靠 counter 间接控制 | ⚠️ |
+
+---
+
+### B3: 专题 03 Prompt 进化
+
+**设计文档**：`03-prompt-evolution.md` — 从 SOP 到任务式指挥
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B3-1 | §3.1 广播认领模板改写 | PromptSection 组装（新路径）+ BootstrapBuilder（旧路径） | ✅ |
+| B3-2 | P6 反静默降级 | 无 `scope-reduction-detection` 自动机制 | ⚠️ 设计原则，未强制实施 |
+| B3-3 | P7 经验闭环 | 无 IMPROVE 阶段自动触发 | ⚠️ |
+
+---
+
+### B4: 专题 04 黑板协作模型
+
+**设计文档**：`04-blackboard-collaboration-model.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B4-1 | §3.1 assignee 降级为显示字段，路由走 @mention | `router.py` L160-166 仍有 assignee 快速路径 | ⚠️ 设计说 Phase 1 双轨并行，Phase 2 废弃。当前停在 Phase 1 |
+| B4-2 | §3.2 @mention 语义增强 | `mention_queue` + `comment_type` 已实现 | ✅ |
+| B4-3 | §3.3 多人协作 `co_assignees` | 数据库无此字段 | ⚠️ Phase 3 |
+| B4-4 | §3.4 output↔comment 关联 | 无关联字段 | ⚠️ Phase 2 |
+| B4-5 | §3.5 层级查询 API | `parent_task` 支持 | ✅ |
+
+---
+
+### B5: 专题 05 上下文四层架构
+
+**设计文档**：`05-context-layers.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B5-1 | L0 铁律层 | workspace 文件注入（SOUL.md/IDENTITY.md 等） | ✅ |
+| B5-2 | L1 角色层 | SOUL.md / IDENTITY.md | ✅ |
+| B5-3 | L2 引擎注入层 | BootstrapBuilder 实现 | ✅ |
+| B5-4 | L3 被动参考层（wiki knowledge） | 无 `_inject_wiki_knowledge` | ⚠️ 设计标注为 Phase 2 |
+
+---
+
+### B6: 专题 06 PM2 Crash 恢复
+
+**设计文档**：`06-pm2-crash-recovery.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B6-1 | §4.1 总体流程 `_startup_recover` | ticker.py L1614 有启动恢复 | ✅ |
+| B6-2 | §4.2 claimed 状态恢复 | ✅ | ✅ |
+| B6-3 | §4.2 working 状态恢复 `_recover_working_task` | ✅ | ✅ |
+| B6-4 | §4.2 review 状态恢复 `_recover_review_task` | ✅ | ✅ |
+
+---
+
+### B7: 专题 07 Spawner Acquire-First
+
+**设计文档**：`07-spawner-acquire-first.md` — #07.1 已实施, #07.2 已实施
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B7-1 | Phase 0 Pre-acquire 修复 | spawner.py L499-512 | ✅ |
+| B7-2 | Phase 1 Counter acquire | spawner.py L516-521 | ✅ |
+| B7-3 | Phase 2 Session check | spawner.py L523-568 | ✅ |
+| B7-4 | Phase 2.5 假死修复 | spawner.py L557-568 | ✅ |
+
+---
+
+### B8: 专题 08 Classify Outcome 优化
+
+**设计文档**：`08-classify-outcome-optimization.md` — 已实施 ✅
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B8-1 | A0-A17 判定树 | `_classify_outcome` 方法 | ✅ |
+| B8-2 | A9 api_error 特殊路径 | `api_retry_count` | ✅ |
+| B8-3 | A14-A17 可恢复 retry + cooldown 60s | `cooldown_seconds` + `set_cooldown` | ✅ |
+
+---
+
+### B9: 专题 09 Rebuttal + Goal Gate
+
+**设计文档**：`09-rebuttal-and-goal-gate.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B9-1 | §2.1 Rebuttal 自动化（review 非 approved → @mention assignee） | `review.py RebuttalManager` + `ticker.py _rebuttal_on_complete` | ✅ |
+| B9-2 | §2.1 防止无限循环（max 2 轮） | `RebuttalManager.MAX_ROUNDS = 2` | ✅ |
+| B9-3 | §2.2 目标一致性 Gate | 无自动 goal gate 检查 | ⚠️ 设计为 Agent 端行为，非 Daemon 侧 |
+
+---
+
+### B10: 专题 10 T3 需求探索 + 黑板展示
+
+**设计文档**：`10-t3-requirement-exploration-and-blackboard-display.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B10-1 | A2 需求探索过程写黑板 comments | 后端支持 `comment_type` | ✅ |
+| B10-2 | A3 TaskModal 实时刷新 | SSE `comment_added` / `checkpoint_resolved` | ✅ |
+| B10-3 | D1 砍掉 AI 摘要 | 黑板直投前端 | ✅ |
+| B10-4 | D2 SSE 只做通知 | 前端按需拉数据 | ✅ |
+
+---
+
+### B11: 专题 11 上下文四层重设计
+
+**设计文档**：`11-context-layers-redesign.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B11-1 | §2.3 L2 操作规范型 6 个 Skill 全文注入 | `BootstrapBuilder` 有 `ROLE_SKILL_MAP` + `_read_skill` 全文注入 ✅；`task_handler.py RoleSkillSection` 只给索引+引导语 ⚠️ | 🟡 **双路径并存**，策略矛盾 |
+| B11-2 | §2.3 `handoff.schema.json` | 不存在 | ⚠️ Phase 3 |
+| B11-3 | §2.3 `review_protocols/` 目录 | 不存在，但 `review-quality` Skill 文件存在 | ⚠️ 设计文档 §三归属表已改归类为 L3 Skill |
+| B11-4 | §6 Phase 3 Step 6-8 BootstrapBuilder 改造 | 已完成（ROLE_SKILL_MAP + _read_skill） | ✅ |
+| B11-5 | §2.3 token 预算 ~600 tokens | bootstrap.py 有 warn 但不截断 | 🟡 有告警无硬限制 |
+
+**B11 关键发现**：新旧路径的 Skill 注入策略矛盾——
+- 旧路径（BootstrapBuilder）：**全文注入** Skill（`_read_skill` 读文件全文）
+- 新路径（RoleSkillSection）：**只给索引**（"请用 read 工具读取 SKILL.md"）
+- 设计文档 §2.3 要求 "A 类 Skill 全文注入"
+- handler 注册后会从旧路径切换到新路径，导致 **Skill 从全文注入降级为索引提示**
+
+这是一个 **隐性回归**：注册 handler 后 Agent 获取的操作规范信息量大幅减少。
+
+---
+
+### B12: 专题 12 Pipeline 设计
+
+**设计文档**：`12-pipeline-design.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B12-1 | §3 Pipeline 注册表 | 不存在 | ⚠️ 设计 §9 标注 Phase 2 |
+| B12-2 | §4 路由逻辑 task_type | router.py 无 task_type 路由 | ⚪ |
+| B12-3 | §8 PipelineRegistry | 不存在 | ⚪ |
+| B12-4 | §10.1 task_type 默认值改 None | `blackboard_routes.py` 已为 None（v3.0 已修） | ⚪ **已实施** |
+| B12-5 | §10.2 广播计数器修正 | `_broadcast_tracker` + `BroadcastRound` 已实现 | ⚪ **已实施** |
+
+**B12 结论**：Pipeline 主体未实施符合设计路线图（Phase 2），但 §10 Phase 1 的两个 bug fix 明确标注为"立做"却未执行。
+
+---
+
+### B13: 专题 13 工具链与开发工作流
+
+**设计文档**：`13-toolchain-and-dev-workflow.md`
+
+| # | 设计要求 | 代码现状 | 判定 |
+|---|---------|---------|------|
+| B13-1 | §16 工具链事件中枢 | `toolchain_handler.py` + `toolchain_templates.py` | ✅ |
+| B13-2 | Gitea webhook 处理 | 5 模板 + 去重 | ✅ |
+| B13-3 | CI 前缀 `[CI]` | ✅ | ✅ |
+
+---
+
+### Part B 汇总
+
+| 判定 | 数量 | 主要项目 |
+|------|------|---------|
+| ✅ 一致 | 21 | B1-1, B1-2, B2-1/2, B3-1, B4-2/5, B5-1/2/3, B6-1~4, B7-1~4, B8-1~3, B9-1/2, B10-1~4, B11-4, B13-1~3 |
+| 🟡 部分一致 | 3 | B11-1 双路径策略矛盾, B11-5 token 预算无硬限制 |
+| ⚠️ 设计标注未实施 | 10 | B1-3（handler 不可达）, B2-3/4, B3-2/3, B4-1/3/4, B5-4, B9-3, B11-2/3 |
+| ❌ 设计承诺未交付 | 0 | — |
+
+---
+
+## Step 5 审计报告偏差项验证
+
+`step5-audit-report.md` 列出 6 项偏差（D1-D6）。逐项验证 HEAD 代码：
+
+| # | 审计描述 | HEAD 实际状态 | 判定 |
+|---|---------|-------------|------|
+| D1 | pre_spawn 返回值未检查 | **已修复**：`if not _handler.pre_spawn(...): raise RuntimeError("handler_pre_spawn_failed")` | ✅ 已修 |
+| D2 | PromptContext 缺少 from_agent/mail_type | **已修复**：spawner L289-296 从 must_haves JSON 提取 | ✅ 已修 |
+| D3 | inform outcome 白名单缺失 | 未修复。但影响极小——CRASH_OUTCOMES 由基类处理，剩余异常 outcome 罕见 | 🟢 可接受 |
+| D4 | retry prompt 仍用 `is_mail` 硬编码 | **未修复**：spawner L1128 仍硬编码 `is_mail = project_id == "_mail"` | 🟡 遗留 |
+| D5 | _check_reply 语义差异 | **已修复**：MailHandler._check_reply 用 `SELECT id FROM tasks WHERE id != ? AND must_haves LIKE ?`，与 v3.0 一致 | ✅ 已修 |
+| D6 | 标 done 重试机制 | **已修复**：`BaseTaskHandler._mark_task_status` 有 3 次重试 | ✅ 已修 |
+
+**结论**：D1/D2/D5/D6 已在后续 commit 修复，D3 可接受，D4 是遗留项。
+
+---
+
+## 与庞统 Review 的背靠背比对
+
+| 维度 | 司马懿 | 庞统 | 差异分析 |
+|------|--------|------|---------|
+| **致命问题** | A1 review verdict 丢失 + A2 handler 未注册 | 仅 #1 review verdict 丢失 | **关键差异**：庞统未将 handler 未注册列为致命问题。庞统认为 `_legacy_on_complete` 仍可运行所以只关注 review 路径。但我认为 **ticker 不再扫描 `_mail`** 是 v3.0 有、HEAD 丢失的行为，这比 review 路径更严重——Mail 系统完全停止工作 |
+| **_mail tick 丢失** | 明确指出 A2 导致 ticker 不扫描 `_mail` | 未提及 | 庞统漏检了 `virtual_projects()` 返回空时 `_mail` 不被 tick 的后果 |
+| **Skill 注入降级** | B11-1 发现新旧路径策略矛盾 | 未提及 | 庞统未分析 handler 注册后 Skill 注入策略的变化 |
+| **D1/D2/D5 修复状态** | 验证了审计报告的修复项 | 未验证 | 庞统的 review 早于修复 commit |
+| **专题覆盖范围** | 13 个全覆盖 | 13 个全覆盖 | 一致 |
+| **D4 retry 硬编码** | A8 独立发现 | B4 独立发现 | 独立确认 |
+
+---
+
+## 建议优先级
+
+| 优先级 | 项目 | 说明 |
+|--------|------|------|
+| **P0** | A2 handler 注册初始化 | 添加 `TaskTypeRegistry.register()` 启动代码。**这是所有 handler 路径的前提** |
+| **P0** | A1 review verdict 处理 | P0 修复后自然恢复（`TaskHandler.handle_review_complete` 生效）。**但如果不想立即注册 handler**，需先在 `_legacy_on_complete` 中恢复 review verdict 逻辑作为 interim fix |
+| **P1** | B11-1 Skill 注入策略对齐 | 注册 handler 后 RoleSkillSection 只给索引——需确认这是有意降级还是遗漏。如果是遗漏，RoleSkillSection 应全文注入 |
+| **P1** | A8 retry prompt handler 化 | spawner L1128 `is_mail` 硬编码改走 handler |
+| **P2** | A3 死代码清理 | `_mail_auto_working` / `_mail_auto_complete` 等方法在 handler 注册后确认不再需要再删除 |
+| ~~P2~~ | ~~B12-4/5~~ | ~~已验证 v3.0 tag 时已修复，不需要再改~~ |
+| **P3** | ticker.py L1483 缩进对齐 | 风格问题 |
+
+---
+
+## Cross-Check：庞统 Review 逐项验证
+
+对庞统 `review-v3-vs-head-pangtong.md` 的每个发现，独立验证：**是不是真问题？根因对不对？修复方案对不对？**
+
+---
+
+### Part A 逐项验证
+
+#### 庞统 #1 🔴 review verdict 丢失
+
+**庞统判定**：致命。"仅影响非 handler 项目（_general）"。
+
+**我的验证**：
+
+1. **是致命问题** ✅ —— `_legacy_on_complete` 在 `_is_review=True` 时确实什么都不做。
+
+2. **"仅影响非 handler 项目" — 表述不准确**。实际情况更复杂：
+   - `_general` 项目确实受影响（走 `_legacy_on_complete`）
+   - 但 `_mail` / `_toolchain` 不受影响，**不是因为"handler 正确处理"**，而是因为它们**根本不走 review 流程**（ticker 中 `_dispatch_reviews` 对 handler 项目 `return []`）
+   - 庞统说"handler 项目（_mail/_toolchain）的 review 由 TaskHandler.post_complete 正确处理"——**这个说法有误导性**。TaskHandler 不是 `_mail`/`_toolchain` 的 handler，它们各自的 handler（MailHandler/ToolchainHandler）没有 `handle_review_complete` 方法。它们不走 review 是因为设计上就不走。
+
+3. **庞统的修复方案有隐藏缺陷**。庞统说"让非 handler 项目也走 TaskHandler（注册 `_general` 到 TaskTypeRegistry）"。但 `TaskTypeRegistry.get_by_project()` 匹配的是 `handler.virtual_project`，而 TaskHandler 的 `virtual_project = None`。所以：
+   - `get_by_project("_general")` → 遍历所有 handler，检查 `h.virtual_project == "_general"` → TaskHandler 的 `virtual_project` 是 `None` → **不匹配** → 返回 `None`
+   - 即使注册了 TaskHandler，`_general` 项目仍然走 `_legacy_on_complete`
+   - 庞统的修复方案需要**额外改 TaskHandler.virtual_project 或 registry 匹配逻辑**，但他没指出这一点
+
+**结论**：问题是真的，严重度判定正确。但影响范围描述和修复方案都不完整。
+
+---
+
+#### 庞统 #2 🟢 旧 `_mail_*` 方法保留
+
+**庞统判定**：正常重构，方法体保留标记为 deprecated。
+
+**我的验证**：
+
+1. **方法体确实保留** ✅（dispatcher.py L628-860）
+2. **但"标记为 deprecated"不对**——代码中没有 `@deprecated` 装饰器或注释。这些方法就是安静地躺在那里，没有任何标记告诉维护者"别用了"
+3. **我标 🟡 中等而非 🟢**的原因：无 deprecated 标记 + 主流程不再调用 = 未来维护者容易误用
+
+**结论**：问题不大，但庞统多给了信息（"标记为 deprecated"）——代码中实际没有标记。
+
+---
+
+#### 庞统 #3 🟢 spawn 失败回退
+
+**庞统判定**：逻辑改进。
+
+**我的验证**：✅ 确认等价，新版更通用。
+
+---
+
+#### 庞统 #4-5 🟢 spawner prompt/api_section
+
+**庞统判定**：等价实现。
+
+**我的验证**：✅ 确认等价。
+
+---
+
+#### 庞统 #6 🟢 ticker `_mail` → `virtual_projects()`
+
+**庞统判定**：正常重构，可扩展。
+
+**我的验证**：**这是庞统最大的漏检**。
+
+庞统只看了代码方向（硬编码 → 注册表），**没有检查注册表是否为空**。
+
+实际运行时 `TaskTypeRegistry.virtual_projects()` 返回空列表 → `_mail` 不被 ticker 扫描。这是一个 **v3.0 有、HEAD 丢失的行为**——v3.0 中 `_mail` 硬编码在 ticker L218-229，HEAD 中完全消失。
+
+后果：所有 Mail 任务的 pending → claimed → working 流程中断，整个飞鸽传书系统停止工作。
+
+这不是"正常重构"，是**致命回归**。
+
+---
+
+#### 庞统 #7-8 🟢 ticker check_reply / dispatch_reviews
+
+**庞统判定**：等价实现。
+
+**我的验证**：✅ 确认等价。但 #7 说"缩进正确"——实际 ticker.py L1483 有缩进不一致（28 空格 vs 同级 24 空格），不影响运行但增加维护混淆。
+
+---
+
+### Part B 逐专题验证
+
+#### 专题 01-03：无分歧
+
+庞统的检查和我的结论一致。设计原则未强制实施属于正常。
+
+---
+
+#### 专题 04：庞统更严格
+
+庞统把 B4-3（co_assignees）和 B4-4（output↔comment）标 ❌，我标 ⚪（Phase 2/3）。
+
+庞统的判定更严格——"设计了但没实现就是不一致" vs 我的"设计自身标注了 Phase，未实施是预期的"。两种视角都有道理，**不算错误**。
+
+---
+
+#### 专题 05：判定标准差异
+
+庞统把 B5-4（L3 wiki 知识注入）标 ❌。我标 ⚪（Phase 2）。
+
+同专题 04，判定标准差异。
+
+---
+
+#### 专题 06：庞统更细致
+
+庞统多了 B6-5"设计提到 7 个恢复方法只看到 2 个公开方法"——这是一个合理的疑问，我没有提出。
+
+---
+
+#### 专题 07-10：无分歧
+
+---
+
+#### 专题 11：庞统全标 ❌ 是错的
+
+庞统 B11-1 说"BootstrapBuilder 只注入通用 prompt，无 skill 全文注入"。
+
+**我验证了代码**：
+```python
+# bootstrap.py L29
+ROLE_SKILL_MAP = {
+    "executor": "blackboard-executor",
+    "reviewer": "blackboard-reviewer",
+    ...
+}
+
+# bootstrap.py L68-72
+skill_name = self.ROLE_SKILL_MAP.get(role)
+if skill_name:
+    skill_content = self._read_skill(skill_name)  # 读全文
+    if skill_content:
+        sections.append(skill_content)
+```
+
+**BootstrapBuilder 有 Skill 全文注入**。庞统说"无 skill 全文注入"与代码不符。他可能只看了 `task_handler.py` 的 RoleSkillSection（确实只给索引），没有看 `bootstrap.py` 的旧路径。
+
+**实际情况**：双路径并存。旧路径（BootstrapBuilder）全文注入，新路径（RoleSkillSection）只给索引。handler 注册后从旧路径切换到新路径，Skill 信息量降级。这才是真正的问题。
+
+---
+
+#### 专题 12：我之前的 B12-4/5 判定有误
+
+我在 Part B 中说"B12-4 task_type 默认值仍为 `\"coding\"`"和"B12-5 广播计数器 retry_count 不递增"是 Phase 1 承诺未交付。
+
+**cross-check 时我重新验证了代码**：
+
+- **B12-4**：`blackboard_routes.py` L138 已是 `body.get("task_type", None)`，**默认值已经是 None**。v3.0 tag 中也是 None。设计文档 §10.1 的 bug fix 可能在 v3.0 之前就修了，或者设计文档基于旧版本写的。**不是问题**，我之前的判定有误。
+
+- **B12-5**：`ticker.py` 中 `_broadcast_tracker` + `BroadcastRound` + `round_number >= 3` 升级庞统的机制已实现。`mark_mention_retry` 有 `retry_count = retry_count + 1`。设计 §10.2 描述的问题已在 v3.0 或更早修复。**不是问题**，我之前的判定有误。
+
+庞统对专题 12 的判定（"设计文档 §9 自身标记为待实现"）比我准确。
+
+**修正我的报告**：Part B 中 B12-4 和 B12-5 应从 ❌ 改为 ⚪（设计自标 Phase 2，主体未实施是预期的）。
+
+---
+
+#### 专题 13：无分歧
+
+---
+
+### 庞统未引用 Step 5 审计报告
+
+庞统的 review 完全没引用 `step5-audit-report.md`（v3.0..HEAD diff 中新增的文件）。这意味着 D1/D2/D5 的修复状态未经庞统验证。我逐项验证了 D1/D2/D5 **已修复**，D4 **未修复**（retry 硬编码），D3 **可接受**，D6 **已修复**。
+
+---
+
+### 庞统漏检的额外行为回归
+
+handler 未注册还导致一个庞统完全没提到的问题：
+
+**guardrail 回归**。v3.0 中 dispatcher L127-128：
+```python
+is_mail = project_config.get("project_id") == "_mail" if project_config else False
+if self.guardrails and not is_mail:
+```
+
+HEAD dispatcher L128-131：
+```python
+handler = TaskTypeRegistry.get_by_project(project_config.get("project_id", "") ...)
+is_handler_task = handler is not None
+if self.guardrails and not is_handler_task:
+```
+
+handler 未注册 → `is_handler_task = False` → **`_mail` 项目也要过 guardrail 检查了**。v3.0 中 `_mail` 是跳过 guardrail 的。这可能导致某些 Mail 任务被 guardrail 拦截。
+
+---
+
+### Cross-Check 总结
+
+| 维度 | 庞统 review 质量 |
+|------|-----------------|
+| **致命问题发现** | 发现 A1 ✅，漏检 A2（handler 注册 + ticker 不可达 + guardrail 回归）❌ |
+| **根因分析** | A1 根因正确。修复方案不完整（没指出 TaskHandler.virtual_project=None 导致注册也匹配不到 `_general`） |
+| **Part B 专题覆盖** | 13/13 全覆盖 ✅ |
+| **Part B 事实准确性** | B11 "无 skill 全文注入"与代码不符 ❌。B12 比我准确 ✅ |
+| **Part B 多给信息** | #2 说"标记为 deprecated"但代码无标记 ⚠️ |
+| **Part B 更严格处** | B04-3/4 标 ❌（合理），B06-5 恢复方法数量疑问（合理） |
+| **审计报告验证** | 未引用，未验证 D1-D6 修复状态 |
+| **遗漏的行为回归** | guardrail 对 `_mail` 的回归 |
+
+**我的自我修正**：B12-4/5 判定有误，应改为 ⚪。v3.0 tag 时这两个问题已修复，设计文档描述的是更早期的问题。
+
+---
+
+*— 司马懿 仲达，质量总监 🗡️*
@@ -0,0 +1,74 @@
+# Step 5 双重审计报告
+
+## 摘要
+- 设计一致性检查项: 8
+- 特殊逻辑覆盖检查项: 22
+- 一致/覆盖: 24
+- **偏差/遗漏: 6（严重 3 / 轻微 3）**
+
+---
+
+## 偏差/遗漏清单
+
+| # | 维度 | 设计要求 / 旧逻辑 | 代码实际 | 严重程度 | 建议 |
+|---|------|-------------------|---------|---------|------|
+| **D1** | B1.2 pre_spawn | 旧 `_mail_on_checks_passed`: `if not _mail_auto_working(): raise RuntimeError` — pre_spawn 失败时中止 spawn | 新 `_handler_on_checks_passed`: `_handler.pre_spawn(...)` 返回值未检查，`handler_marked_working = True` 无条件执行 | **严重** | 改为 `if not _handler.pre_spawn(...): raise RuntimeError("handler_pre_spawn_failed")` |
+| **D2** | B3.1 PromptContext | 旧 `_build_mail_prompt` 从 must_haves JSON 解析 `from_agent` 和 `performative` 传入模板 | 新 `spawner._build_spawn_message` 构建 PromptContext 时缺少 `from_agent` 和 `mail_type`，均为空字符串 | **严重** | 从 `must_haves` JSON 提取 `from` 和 `performative` 填入 PromptContext |
+| **D3** | B1.3 inform outcome 白名单 | 旧 `_mail_auto_complete`: inform 类型有 outcome 白名单 `{"completed", "claimed", "no_reply"}`，不在白名单的 outcome 跳过 auto-done | 新 `MailHandler.verify_completion`: inform 始终返回 True，不检查 outcome | **轻微** | CRASH_OUTCOMES 已被基类处理。剩余异常 outcome（session_revived/api_error/fallback_timeout）极少出现，且旧逻辑不标 done 只是等 ticker 重投，最终效果差异不大。但严格对齐需要加白名单检查 |
+| **D4** | A. 设计 §6 retry 逻辑 | 设计文档要求 retry 逻辑中 `handler = TaskTypeRegistry.get_by_project(project_id); if handler: return handler.build_retry_prompt(...)` | spawner L1118-1130 重试 prompt 仍用 `is_mail = project_id == "_mail"` 硬编码 | **轻微** | 当前不影响运行（旧的 `_build_mail_prompt` 仍保留且可用），但与设计文档不一致 |
+| **D5** | B1.5 _check_reply 语义差异 | 旧 `_mail_check_reply`: `SELECT id FROM tasks WHERE id != ? AND must_haves LIKE ?` — 检查是否有其他任务的 must_haves 包含当前 task_id（即 in_reply_to 匹配） | 新 `MailHandler._check_reply`: `SELECT COUNT(*) FROM comments WHERE task_id=? AND author != 'daemon' AND comment_type != 'system'` — 检查当前任务是否有非系统 comment | **严重** | 两个查询语义完全不同。旧逻辑检查的是 **mail 表的回复任务**（通过 must_haves 中 in_reply_to 关联），新逻辑检查的是 **当前任务的 comments**。这可能导致 request 类型邮件的幻觉门控行为不同 |
+| **D6** | B1.3 标 done 重试机制 | 旧 `_mail_auto_complete`: 标 done 时外层有 `for attempt in range(3)` 循环 | 新 `BaseTaskHandler._mark_task_status`: H1 修复后已有 3 次重试 | **轻微** | ✅ 已修复，但注意旧代码标 done 和标 failed 是分开的重试循环，新代码统一走 `_mark_task_status`。行为等价 |
+
+---
+
+## 一致确认项
+
+### A. 设计一致性
+
+| # | 维度 | 检查点 | 结果 |
+|---|------|--------|------|
+| A1 | §6 dispatcher | classify_outcome 后调 handler.post_complete | ✅ on_complete 闭包替换为 handler.post_complete |
+| A2 | §6 dispatcher | on_checks_passed → handler.pre_spawn | ✅ _handler_on_checks_passed 调用 handler.pre_spawn（但返回值未检查，见 D1） |
+| A3 | §6 dispatcher | guardrail 跳过 → handler 判断 | ✅ `is_handler_task = handler is not None` |
+| A4 | §6 spawner | _build_prompt → handler.build_prompt | ✅ handler 路径调用 handler.build_prompt(ctx) |
+| A5 | §6 spawner | _build_api_section → handler 查询 | ✅ handler 存在时 success_status 从 handler.target_success_status 获取 |
+| A6 | §6 ticker | 虚拟项目扫描 → registry.virtual_projects() | ✅ 循环 `TaskTypeRegistry.virtual_projects()` |
+| A7 | §6 ticker | check_completion → handler.check_completion | ✅ 超时检查中调 `handler.check_completion(task.id, db_path)` |
+| A8 | §6 兼容期 | 设计说"兼容期保留旧逻辑" | ✅ 无 handler 的项目走旧路径（legacy_on_complete） |
+
+### B. 特殊逻辑覆盖
+
+| # | 维度 | 检查点 | 结果 |
+|---|------|--------|------|
+| B1 | 1.1 guardrail | handler 项目跳过，_general 等走 guardrail | ✅ |
+| B2 | 1.2 _mail_auto_working | `BEGIN IMMEDIATE` + status 检查 + 标 working | ✅ `_auto_mark_working` 完全一致 |
+| B3 | 1.3 request 无回复 → 标 failed + notify | ✅ MailHandler.on_failure 调 `_mark_task_status(failed)` + `notify_mail_failed` |
+| B4 | 1.4 _mail_revert_to_pending | spawn 失败回退 working → pending | ✅ Exception handler 中有 `BEGIN IMMEDIATE` + 状态检查回退 |
+| B5 | 1.6 Task review verdict 读取 | approved → done | ✅ handle_review_complete |
+| B6 | 1.6 Task review 非 approved → @mention assignee + 保持 review | ✅ H3 修复后保持 review + INSERT comment with comment_type='review' |
+| B7 | 1.6 Task executor 三信号验证 | output/comment/terminal status → review | ✅ verify_completion 完全一致 |
+| B8 | 1.7 Legacy dispatch 路径 | handler 替代 is_mail_legacy | ✅ handler_legacy 查注册表 |
+| B9 | 2.1 _transition_status assignee 清空 | handler 项目不清空 | ✅ |
+| B10 | 2.2 跳过 claimed 状态 | handler 项目跳过 claimed 直接 working | ✅ |
+| B11 | 2.3 _dispatch_reviews 跳过 | handler 项目不走 review | ✅ |
+| B12 | 2.5 startup recovery | `_general` + virtual_projects() | ✅ 不会重复扫描 |
+| B13 | 3.1 _build_api_section | handler 存在时正确获取 success_status | ✅ |
+| B14 | B4.1 TaskHandler.post_complete | 区分 executor/review 流程 | ✅ 通过读 DB status 判断 |
+| B15 | B4.2 MailHandler.post_complete | 基类统一流程 | ✅ |
+| B16 | B4.3 ToolchainHandler.post_complete | 基类统一流程 | ✅ |
+| B17 | B1.5 _check_reply 异常保守处理 | 旧: return True（保守）/ 新: return False | 见 D5 |
+| B18 | CRASH_OUTCOMES 集合 | 与旧 ROLLBACK_CURRENT_AGENT_OUTCOMES 一致 | ✅ 完全一致 |
+| B19 | B2.1 _toolchain ticker 扫描 | _toolchain 会被 ticker 扫描 | ✅ _toolchain 有 blackboard.db 时会被 tick_project 处理 |
+| B20 | B2.3 handler 项目都跳过 claimed | _toolchain 也跳过 | ✅ 所有 handler 项目统一处理 |
+
+---
+
+## 修复优先级
+
+| 优先级 | # | 修复内容 |
+|--------|---|---------|
+| **P0** | D1 | dispatcher _handler_on_checks_passed 检查 pre_spawn 返回值 |
+| **P0** | D2 | spawner PromptContext 从 must_haves 提取 from_agent 和 mail_type |
+| **P0** | D5 | MailHandler._check_reply 恢复旧查询语义（检查 must_haves 中的 in_reply_to） |
+| P1 | D3 | inform outcome 白名单（可选，影响极小） |
+| P2 | D4 | retry prompt 用 handler 路径替代硬编码 |
@@ -0,0 +1,324 @@
+# Step 5 引擎接入 — 影响分析与逐点对照
+
+## 方法论
+
+逐行审查 dispatcher.py / spawner.py / ticker.py 中所有 `is_mail` / `_mail` / `project_id == "_mail"` 分支，
+对照 handler 实现，确认每个特殊处理的去向。
+
+---
+
+## 一、dispatcher.py（985 行）
+
+### 1.1 Guardrail 跳过（L127-129）
+
+```python
+is_mail = project_config.get("project_id") == "_mail" if project_config else False
+if self.guardrails and not is_mail:
+    violations = self.guardrails.check_task(task)
+```
+
+**特殊处理**：Mail 不做 guardrail 检查。
+
+**Handler 覆盖**：设计文档 D6 "skip_guardrail 从接口删除，guardrail 自己判断"。Step 5 改为：`if self.guardrails and handler is None`（无 handler 时走 guardrail），或者用 handler.virtual_project 判断。handler 存在时跳过 guardrail。
+
+**改动**：`is_mail` → `TaskTypeRegistry.get_by_project(project_id) is not None`
+
+---
+
+### 1.2 Mail on_checks_passed（L194-213）
+
+```python
+on_checks_passed = None
+_mail_marked_working = False
+if is_mail and db_path:
+    def _mail_on_checks_passed():
+        nonlocal _mail_marked_working
+        if not _disp._mail_auto_working(_task_id, _mail_db):
+            raise RuntimeError("mail_auto_working_failed")
+        _mail_marked_working = True
+    on_checks_passed = _mail_on_checks_passed
+```
+
+**特殊处理**：Mail spawn 前通过 on_checks_passed 回调标 working，标记成功后才 spawn，spawn 失败回退。
+
+**Handler 覆盖**：MailHandler.pre_spawn 调用 `_auto_mark_working`，和 `_mail_auto_working` 逻辑完全一致。
+
+**改动**：
+- `on_checks_passed` 改为调用 `handler.pre_spawn(task_id, db_path)`
+- `_mail_marked_working` 标记保留，用于 Exception 回退
+
+---
+
+### 1.3 Mail on_complete（L224-238）
+
+```python
+if is_mail:
+    def _mail_on_complete(aid, outcome):
+        _dispatcher._mail_auto_complete(_task_id, aid, _mail_db, _must_haves, outcome=outcome)
+    on_complete = _mail_on_complete
+```
+
+**特殊处理**：Mail on_complete 调用 `_mail_auto_complete`（含 inform/request 分支、幻觉门控、重试 3 次、失败通知）。
+
+**Handler 覆盖**：MailHandler 使用基类 post_complete 统一流程（crash→verify→mark→notify）。但现有 `_mail_auto_complete` 有几个细节差异需要注意：
+
+| 现有逻辑 | Handler 覆盖 | 差异 |
+|---------|-------------|------|
+| request 无回复 → 重试 3 次标 failed | on_failure 标 failed + notify | ⚠️ 缺少 3 次重试 |
+| inform 只在特定 outcome 标 done | verify 始终返回 True → 基类标 done | ✅ 简化了，合理 |
+| 标 done 重试 3 次 | _mark_task_status 单次 | ⚠️ 缺少重试 |
+| notify_mail_failed | on_failure 中调用 notify_mail_failed | ✅ 一致 |
+
+**⚠️ 关键发现**：现有代码标状态时有 **重试 3 次** 机制（防止 DB 锁），handler 的 `_mark_task_status` 只做一次。需要把重试逻辑补到 `_mark_task_status` 或在 handler 层加。
+
+**改动**：on_complete 改为调用 `handler.post_complete(task_id, agent_id, outcome, db_path)`
+
+---
+
+### 1.4 Task on_complete（L241-310）
+
+```python
+else:
+    def _task_on_complete(aid, outcome):
+        # #07.2: crash 回退
+        if outcome in ROLLBACK_CURRENT_AGENT_OUTCOMES and _task_db:
+            _dispatcher._rollback_current_agent(_task_db, _task_id, aid)
+        
+        if _is_review:
+            if outcome in ("completed", "session_revived"):
+                # 读 verdict → approved 标 done / 非 approved @mention assignee
+            else:
+                logger.warning("review agent outcome=%s, NOT marking done", outcome)
+        else:
+            # executor: 三信号验证 → 标 review
+            _dispatcher._task_auto_complete(_task_id, _task_db)
+```
+
+**特殊处理清单**：
+
+1. **#07.2 crash 回退**：executor 和 review 都回退 current_agent → assignee
+2. **review 分支**：outcome 必须是 "completed" 或 "session_revived" 才走 verdict 读取
+3. **review verdict 读取**：approved → done，非 approved → @mention assignee + 保持 review
+4. **review @mention**：通过 Blackboard.add_comment，comment_type="review"
+5. **executor 分支**：走 _task_auto_complete → 三信号验证 → review
+
+**Handler 覆盖**：
+- crash 回退：✅ BaseTaskHandler.post_complete 第一步
+- review verdict：⚠️ **TaskHandler.handle_review_complete 存在但未被 dispatcher 调用**。现有 dispatcher 直接在闭包里做了，不走 handler。
+- @mention：⚠️ handler 用 `conn.execute("INSERT INTO comments")` 直接插入，dispatcher 用 `Blackboard.add_comment`（会做更多处理，如 comment_type="review"）
+- executor 三信号：✅ TaskHandler.verify_completion
+
+**⚠️ 关键发现**：
+1. dispatcher 的 review @mention 用 `bb.add_comment(..., comment_type="review")`，handler 直接 INSERT 不带 comment_type。需要修复 handler。
+2. dispatcher 对 review outcome 有白名单检查（只处理 "completed"/"session_revived"），handler 的 post_complete 没有 outcome 白名单——crash 已在基类处理，其他 outcome 都会走 verify。
+3. dispatcher review 非 approved 时**保持 review 状态**，handler 的 handle_review_complete 标回 working。这是**行为差异**。
+
+**改动**：需要先修复 handler 的 review 分支，再替换 on_complete。
+
+---
+
+### 1.5 Mail spawn 失败回退（L355-358）
+
+```python
+except Exception as e:
+    if _mail_marked_working:
+        self._mail_revert_to_pending(task.id, db_path)
+```
+
+**特殊处理**：spawn 失败（subprocess 启动失败）回退 working → pending。
+
+**Handler 覆盖**：❌ handler 没有这个。这是 dispatcher 级别的异常处理，和 handler 无关。但 toolchain 也需要类似逻辑。
+
+**改动**：保留在 dispatcher 中，改为 `_mail_marked_working` → `handler_marked_working`。
+
+---
+
+### 1.6 Legacy dispatch（L584-660）
+
+```python
+is_mail_legacy = project_config.get("project_id") == "_mail"
+if is_mail_legacy:
+    if not self._mail_auto_working(task.id, db_path_legacy):
+        return error
+```
+
+**特殊处理**：legacy 路径（router=None 时触发）也有 mail 特殊处理。
+
+**Handler 覆盖**：同 1.2/1.3，用 handler 替代。
+
+**改动**：同样用 handler.pre_spawn 和 handler.post_complete 替代。
+
+---
+
+### 1.7 现有 Mail 辅助方法（L658-870）
+
+`_mail_auto_working` / `_mail_revert_to_pending` / `_mail_auto_complete` / `_mail_check_reply`
+
+**改动**：Step 5 不删这些方法（安全起见保留，标记 deprecated），只改调用方。确认稳定后再删。
+
+---
+
+## 二、spawner.py（1704 行）
+
+### 2.1 _build_prompt 中的 mail 分支（L282-284）
+
+```python
+if project_id == "_mail":
+    return self._build_mail_prompt(task_id, title, description, must_haves, agent_id)
+```
+
+**特殊处理**：Mail 用专用精简模板。
+
+**Handler 覆盖**：MailHandler.build_prompt 通过 PromptComposer 拼 3 个 section。
+
+**改动**：查注册表 → handler.build_prompt(context)。需要构建 PromptContext 传入。
+
+---
+
+### 2.2 _build_api_section（L321-325）
+
+```python
+success_status = '"done"' if project_id == "_mail" else '"review"'
+```
+
+**特殊处理**：Mail 的 success_status 是 done。
+
+**Handler 覆盖**：已由 handler 的 PromptSection 处理（TaskApiSection hardcode review，MailApiSection 不含 status 回写指令）。
+
+**改动**：如果 handler 存在，跳过 _build_api_section（handler.build_prompt 已包含）。
+
+---
+
+### 2.3 classify_outcome 中的 handler 调用
+
+spawner 在 classify_outcome 后调 on_complete(outcome)。on_complete 是 dispatcher 传入的闭包。
+
+**改动**：on_complete 闭包改为调用 handler.post_complete。spawner 本身不直接查注册表。
+
+---
+
+## 三、ticker.py（1897 行）
+
+### 3.1 虚拟项目扫描（L218-229）
+
+```python
+mail_db = Path(self.registry.root) / "_mail" / "blackboard.db"
+if mail_db.exists() and "_mail" not in active_projects:
+    pr = await self._tick_project("_mail", {...})
+```
+
+**特殊处理**：_mail 硬编码扫描。
+
+**Handler 覆盖**：TaskTypeRegistry.virtual_projects() 返回 ["_toolchain", "_mail"]。
+
+**改动**：循环 `TaskTypeRegistry.virtual_projects()` 替代硬编码。_toolchain 如果也需要 ticker 扫描就自动发现。但需确认 _toolchain 是否需要 ticker——当前 toolchain 任务创建和完成都在 toolchain_routes.py 中处理，可能不需要 ticker 扫描。
+
+---
+
+### 3.2 _transition_status 中 mail assignee 不清空（L953-960）
+
+```python
+if new_status == "pending":
+    if self._current_project_id == "_mail":
+        # Mail 的 assignee 是收件人，永不清空
+        conn.execute("UPDATE tasks SET status=?, updated_at=? WHERE id=?", ...)
+    else:
+        conn.execute("UPDATE tasks SET status=?, assignee=NULL, ...", ...)
+```
+
+**特殊处理**：Mail 重置到 pending 时不清空 assignee（assignee 是收件人）。
+
+**Handler 覆盖**：❌ handler 不管 ticker 的状态转换逻辑。这是 ticker 内部逻辑。
+
+**改动**：用 `TaskTypeRegistry.get_by_project(project_id)` 判断替代硬编码。
+
+---
+
+### 3.3 Mail 跳过 claimed 状态（L1029-1043）
+
+```python
+if project_id == "_mail":
+    conn.execute("UPDATE tasks SET current_agent=? WHERE id=?", ...)
+    # 跳过 claimed，直接 working
+```
+
+**特殊处理**：Mail 不走 claimed 中间态（已在 dispatcher 中标 working）。
+
+**Handler 覆盖**：handler.pre_spawn 的 _auto_mark_working 跳过了 claimed。
+
+**改动**：用 handler 判断替代硬编码。
+
+---
+
+### 3.4 _dispatch_reviews 跳过 mail（L1304）
+
+```python
+if project_id == "_mail":
+    return []
+```
+
+**特殊处理**：Mail 不走 review 流程。
+
+**Handler 覆盖**：MailHandler.target_success_status = "done"，不走 review。但 ticker 的 _dispatch_reviews 是看项目级。
+
+**改动**：用 handler 判断。
+
+---
+
+### 3.5 Mail 幻觉门控兜底（L1474-1492）
+
+```python
+if self._current_project_id == "_mail":
+    has_reply = self._mail_check_reply(task.id, db_path)
+    if has_reply:
+        # working → done
+```
+
+**特殊处理**：Ticker 超时检查时，如果 mail 有回复，标 done 而非 failed。
+
+**Handler 覆盖**：❌ handler 的 check_completion 只返回 bool，不做状态标记。
+
+**改动**：调用 handler.check_completion 替代 _mail_check_reply。状态标记逻辑保留在 ticker 中。
+
+---
+
+### 3.6 _mail_check_reply（L1555-1575）
+
+和 dispatcher 版本一致。
+
+**改动**：用 handler.check_completion 替代。
+
+---
+
+### 3.7 虚拟项目 init + recovery 扫描（L1625-1643）
+
+```python
+for virtual_id in ("_general", "_mail"):
+    ...
+    # _mail 项目不清空 assignee
+```
+
+**改动**：virtual_projects() + _general 硬编码。
+
+---
+
+## 四、Handler 缺陷（需在 Step 5 前修复）
+
+| # | 缺陷 | 影响 | 修复方案 |
+|---|------|------|---------|
+| H1 | BaseTaskHandler._mark_task_status 无重试 | DB 锁时标状态失败，任务卡住 | 加 3 次重试（和 dispatcher 现有行为一致） |
+| H2 | TaskHandler.handle_review_complete 中 @mention 不带 comment_type="review" | review comment 无类型标记 | INSERT 加 comment_type |
+| H3 | dispatcher review 非 approved 保持 review 状态，handler 标 working | **行为差异** | handler 改为保持 review 状态（和 dispatcher 一致） |
+| H4 | dispatcher review outcome 有白名单（"completed"/"session_revived"），handler 无 | crash 之外的异常 outcome 也会走 verify | handler 的 post_complete 已在基类处理 crash，其余 outcome 走 verify 是合理的 |
+
+**H3 最关键**——dispatcher review 非 approved 保持 review 状态（等 assignee 自己处理），handler 标 working 会触发 ticker 重新 dispatch executor，这不是预期行为。
+
+## 五、改动策略
+
+**不删旧代码，只改调用方**：
+1. dispatcher 中 is_mail → handler 判断，on_checks_passed/on_complete → handler.pre_spawn/post_complete
+2. spawner 中 _build_prompt → handler.build_prompt
+3. ticker 中虚拟项目扫描 → registry.virtual_projects()，mail 特殊判断 → handler 判断
+4. 旧方法（_mail_auto_working 等）标记 @deprecated 保留，不删
+
+**先修 handler 缺陷（H1-H3），再改引擎**。
@@ -59,7 +59,10 @@ async def list_tasks(project_id: str,
                     assignee: Optional[str] = None,
                     parent_task: Optional[str] = None):
    bb = _bb(project_id)
-    tasks = bb.list_tasks(status=status, assignee=assignee, parent_task=parent_task)
+    tasks = bb.list_tasks(
+        status=status,
+        assignee=assignee,
+        parent_task=parent_task)
    return {"tasks": [_task_to_dict(t) for t in tasks]}


@@ -79,10 +82,12 @@ async def get_task(project_id: str, task_id: str,
            result["outputs_count"] = detail.get("outputs_count", 0)
            result["review_status"] = detail.get("review_status")
            result["latest_event_detail"] = detail.get("latest_event_detail")
-        result["comments"] = [dict(c.__dict__) for c in bb.get_comments(task_id)]
+        result["comments"] = [dict(c.__dict__)
+                              for c in bb.get_comments(task_id)]
        result["outputs"] = [dict(o.__dict__) for o in bb.get_outputs(task_id)]
        result["reviews"] = [dict(r.__dict__) for r in bb.get_reviews(task_id)]
-        result["decisions"] = [dict(d.__dict__) for d in bb.get_decisions(task_id)]
+        result["decisions"] = [dict(d.__dict__)
+                               for d in bb.get_decisions(task_id)]
        result["events"] = q.task_events(task_id)
        result["experiences"] = q.task_experiences(task_id)
    return result
@@ -134,7 +139,8 @@ async def create_task(project_id: str, body: Dict[str, Any]):
        priority=body.get("priority", 5),
        assignee=assignee,
        assigned_by=body.get("assigned_by", "user"),
-        depends_on=json.dumps(body["depends_on"]) if "depends_on" in body else None,
+        depends_on=json.dumps(
+            body["depends_on"]) if "depends_on" in body else None,
        parent_task=body.get("parent_task"),
        risk_level=body.get("risk_level", "standard"),
        stage=body.get("stage"),
@@ -175,7 +181,8 @@ async def _generate_title(description: str) -> str | None:
        resp = client.chat.completions.create(
            model=model,
            messages=[
-                {"role": "system", "content": "你是一个任务标题生成器。根据用户的需求描述，生成一个简洁的中文标题（5-15字），只输出标题，不要任何其他内容。"},
+                {"role": "system",
+                 "content": "你是一个任务标题生成器。根据用户的需求描述，生成一个简洁的中文标题（5-15字），只输出标题，不要任何其他内容。"},
                {"role": "user", "content": description[:500]},
            ],
            max_tokens=50,
@@ -187,7 +194,8 @@ async def _generate_title(description: str) -> str | None:
            return title
    except Exception as e:
        import logging
-        logging.getLogger("moziplus-v2").warning(f"Title generation failed: {e}")
+        logging.getLogger(
+            "moziplus-v2").warning(f"Title generation failed: {e}")
    return None


@@ -205,7 +213,8 @@ async def task_progress(project_id: str, task_id: str):
 async def claim_task(project_id: str, task_id: str, body: Dict[str, Any]):
    bb = _bb(project_id)
    if not bb.claim_task(task_id, body["agent"]):
-        raise HTTPException(409, "Claim failed (already claimed or wrong assignee)")
+        raise HTTPException(
+            409, "Claim failed (already claimed or wrong assignee)")
    return {"ok": True}


@@ -273,10 +282,20 @@ def _init_agent_ids():
        return
    try:
        import yaml
-        cfg_path = os.path.join(os.path.dirname(__file__), "..", "..", "config", "default.yaml")
+        cfg_path = os.path.join(
+            os.path.dirname(__file__),
+            "..",
+            "..",
+            "config",
+            "default.yaml")
        with open(cfg_path) as f:
            cfg = yaml.safe_load(f)
-        _KNOWN_AGENT_IDS = list(cfg.get("daemon", {}).get("agent_profiles", {}).keys())
+        _KNOWN_AGENT_IDS = list(
+            cfg.get(
+                "daemon",
+                {}).get(
+                "agent_profiles",
+                {}).keys())
    except Exception:
        _KNOWN_AGENT_IDS = []

@@ -285,7 +304,10 @@ def _extract_mentions(text: str) -> list:
    """从文本中自动提取 @agent-id 格式的 mention"""
    import re
    _init_agent_ids()
-    candidates = set(re.findall(r'@([a-z][a-z0-9]*(?:-[a-z][a-z0-9]*)+)', text))
+    candidates = set(
+        re.findall(
+            r'@([a-z][a-z0-9]*(?:-[a-z][a-z0-9]*)+)',
+            text))
    return [a for a in candidates if a in _KNOWN_AGENT_IDS]


@@ -397,7 +419,8 @@ async def write_output(project_id: str, task_id: str, body: Dict[str, Any]):
        )
        os.makedirs(artifacts_dir, exist_ok=True)
        # 安全文件名
-        safe_name = "".join(c if c.isalnum() or c in "._-" else "_" for c in title)
+        safe_name = "".join(
+            c if c.isalnum() or c in "._-" else "_" for c in title)
        if not safe_name:
            safe_name = "output"
        file_path = os.path.join(artifacts_dir, safe_name)
@@ -12,7 +12,9 @@ from typing import Optional
 from src.blackboard.operations import Blackboard
 from src.utils import get_data_root

-router = APIRouter(prefix="/api/projects/{project_id}/tasks/{task_id}/checkpoints", tags=["checkpoints"])
+router = APIRouter(
+    prefix="/api/projects/{project_id}/tasks/{task_id}/checkpoints",
+    tags=["checkpoints"])


 # ── 请求模型 ──
@@ -50,10 +52,12 @@ def list_checkpoints(project_id: str, task_id: str):


@router.post("")
-def create_checkpoint(project_id: str, task_id: str, req: CreateCheckpointRequest):
+def create_checkpoint(project_id: str, task_id: str,
+                      req: CreateCheckpointRequest):
    """Agent 创建 checkpoint"""
    if req.type not in ("verify", "decision", "action"):
-        raise HTTPException(status_code=400, detail=f"Invalid checkpoint type: {req.type}")
+        raise HTTPException(status_code=400,
+                            detail=f"Invalid checkpoint type: {req.type}")

    bb = _bb(project_id)
    # 验证 task 存在
@@ -73,10 +77,15 @@ def create_checkpoint(project_id: str, task_id: str, req: CreateCheckpointReques


@router.post("/{checkpoint_id}/approve")
-def approve_checkpoint(project_id: str, task_id: str, checkpoint_id: str, req: ResolveCheckpointRequest):
+def approve_checkpoint(project_id: str, task_id: str,
+                       checkpoint_id: str, req: ResolveCheckpointRequest):
    """用户通过 checkpoint → 自动推进 task 状态"""
    bb = _bb(project_id)
-    result = bb.resolve_checkpoint(checkpoint_id, "approve", req.resolved_by, req.note)
+    result = bb.resolve_checkpoint(
+        checkpoint_id,
+        "approve",
+        req.resolved_by,
+        req.note)
    if result is None:
        raise HTTPException(status_code=404, detail="Checkpoint not found")
    if "error" in result:
@@ -97,10 +106,15 @@ def approve_checkpoint(project_id: str, task_id: str, checkpoint_id: str, req: R


@router.post("/{checkpoint_id}/reject")
-def reject_checkpoint(project_id: str, task_id: str, checkpoint_id: str, req: ResolveCheckpointRequest):
+def reject_checkpoint(project_id: str, task_id: str,
+                      checkpoint_id: str, req: ResolveCheckpointRequest):
    """用户驳回 checkpoint → task 回到 working"""
    bb = _bb(project_id)
-    result = bb.resolve_checkpoint(checkpoint_id, "reject", req.resolved_by, req.note)
+    result = bb.resolve_checkpoint(
+        checkpoint_id,
+        "reject",
+        req.resolved_by,
+        req.note)
    if result is None:
        raise HTTPException(status_code=404, detail="Checkpoint not found")
    if "error" in result:
@@ -34,7 +34,8 @@ def _get_valid_agents() -> set:
        except Exception:
            pass
    # fallback：硬编码
-    return {"zhangfei-dev", "guanyu-dev", "zhaoyun-data", "jiangwei-infra", "pangtong-fujunshi", "simayi-challenger"}
+    return {"zhangfei-dev", "guanyu-dev", "zhaoyun-data",
+            "jiangwei-infra", "pangtong-fujunshi", "simayi-challenger"}


 router = APIRouter(prefix="/api/mail", tags=["mail"])
@@ -98,7 +99,10 @@ async def list_mail(
 ):
    """Mail 列表（按时间倒序）"""
    bb = _bb()
-    tasks = bb.list_tasks(status=status, assignee=to_agent, assigned_by=from_agent)
+    tasks = bb.list_tasks(
+        status=status,
+        assignee=to_agent,
+        assigned_by=from_agent)

    mails = []
    for t in tasks:
@@ -229,7 +233,10 @@ async def send_mail(body: Dict[str, Any]):
        to_agent = body.get("to", "").strip()
        corrected_to = orig_from  # 回复方向固定: reply → original sender
        if to_agent and to_agent != corrected_to:
-            auto_corrected = {"field": "to", "original": to_agent, "corrected": corrected_to}
+            auto_corrected = {
+                "field": "to",
+                "original": to_agent,
+                "corrected": corrected_to}
        to_agent = corrected_to
    else:
        # --- A2: to 必填（非回复场景） ---
@@ -256,7 +263,8 @@ async def send_mail(body: Dict[str, Any]):
    conversation_id = body.get("conversation_id")
    if not conversation_id and original:
        try:
-            orig_meta = json.loads(original.must_haves) if original.must_haves else {}
+            orig_meta = json.loads(
+                original.must_haves) if original.must_haves else {}
            conversation_id = orig_meta.get("conversation_id")
        except Exception:
            pass
@@ -311,10 +319,12 @@ async def delete_mail(prefix: Optional[str] = Query(None)):
    for t in tasks:
        if t.title and t.title.startswith(prefix):
            if t.status not in ("cancelled",):
-                bb.update_task_status(t.id, "cancelled", agent="mail-cleanup-api")
+                bb.update_task_status(
+                    t.id, "cancelled", agent="mail-cleanup-api")
                deleted_ids.append(t.id)

-    return {"ok": True, "deleted_count": len(deleted_ids), "deleted_ids": deleted_ids}
+    return {"ok": True, "deleted_count": len(
+        deleted_ids), "deleted_ids": deleted_ids}


@router.patch("/{mail_id}")
@@ -31,8 +31,10 @@ async def list_projects():
        if db_path.exists():
            try:
                conn = sqlite3.connect(str(db_path), timeout=5)
-                total = conn.execute("SELECT COUNT(*) FROM tasks WHERE status != 'cancelled'").fetchone()[0]
-                active = conn.execute("SELECT COUNT(*) FROM tasks WHERE COALESCE(archived,0)=0").fetchone()[0]
+                total = conn.execute(
+                    "SELECT COUNT(*) FROM tasks WHERE status != 'cancelled'").fetchone()[0]
+                active = conn.execute(
+                    "SELECT COUNT(*) FROM tasks WHERE COALESCE(archived,0)=0").fetchone()[0]
                archived = total - active
                conn.close()
                info['task_count'] = active
@@ -45,8 +47,10 @@ async def list_projects():
    if general_db.exists() and "_general" not in projects:
        try:
            conn = sqlite3.connect(str(general_db), timeout=5)
-            total = conn.execute("SELECT COUNT(*) FROM tasks WHERE status != 'cancelled'").fetchone()[0]
-            active = conn.execute("SELECT COUNT(*) FROM tasks WHERE COALESCE(archived,0)=0").fetchone()[0]
+            total = conn.execute(
+                "SELECT COUNT(*) FROM tasks WHERE status != 'cancelled'").fetchone()[0]
+            active = conn.execute(
+                "SELECT COUNT(*) FROM tasks WHERE COALESCE(archived,0)=0").fetchone()[0]
            conn.close()
            projects["_general"] = {
                "id": "_general", "name": "一般任务", "description": "无项目归属的通用任务",
@@ -60,8 +64,10 @@ async def list_projects():
        if general_db_check.exists():
            try:
                conn = sqlite3.connect(str(general_db_check), timeout=5)
-                total = conn.execute("SELECT COUNT(*) FROM tasks WHERE status != 'cancelled'").fetchone()[0]
-                active = conn.execute("SELECT COUNT(*) FROM tasks WHERE COALESCE(archived,0)=0").fetchone()[0]
+                total = conn.execute(
+                    "SELECT COUNT(*) FROM tasks WHERE status != 'cancelled'").fetchone()[0]
+                active = conn.execute(
+                    "SELECT COUNT(*) FROM tasks WHERE COALESCE(archived,0)=0").fetchone()[0]
                conn.close()
                projects["_general"]["task_count"] = active
                projects["_general"]["task_count_total"] = total
@@ -173,7 +179,10 @@ async def move_task(project_id: str, task_id: str, body: Dict[str, Any]):
            depends_on=child.depends_on, must_haves=child.must_haves,
        )
        tgt_bb.create_task(moved_child)
-        src_bb.update_task_status(child.id, "cancelled", detail=f"Moved to {target_project}")
+        src_bb.update_task_status(
+            child.id,
+            "cancelled",
+            detail=f"Moved to {target_project}")
        moved_ids.append(child.id)

    # 移动主任务
@@ -186,7 +195,10 @@ async def move_task(project_id: str, task_id: str, body: Dict[str, Any]):
        depends_on=task.depends_on, must_haves=task.must_haves,
    )
    tgt_bb.create_task(moved_task)
-    src_bb.update_task_status(task_id, "cancelled", detail=f"Moved to {target_project}")
+    src_bb.update_task_status(
+        task_id,
+        "cancelled",
+        detail=f"Moved to {target_project}")
    moved_ids.insert(0, task_id)

    return {"ok": True, "moved_to": target_project, "moved_ids": moved_ids}
@@ -46,7 +46,8 @@ _TTL_SECONDS = 7 * 24 * 3600
 _idempotency_lock = asyncio.Lock()


-def _is_duplicate(event: str, delivery: str, payload: Optional[Dict[str, Any]] = None) -> bool:
+def _is_duplicate(event: str, delivery: str,
+                  payload: Optional[Dict[str, Any]] = None) -> bool:
    """检查 Webhook 是否重复投递，自动清理过期条目。

    双重去重策略：
@@ -56,7 +57,8 @@ def _is_duplicate(event: str, delivery: str, payload: Optional[Dict[str, Any]] =
    """
    now = time.time()
    # 清理过期条目
-    while _delivery_timestamps and (now - _delivery_timestamps[0][0]) > _TTL_SECONDS:
+    while _delivery_timestamps and (
+            now - _delivery_timestamps[0][0]) > _TTL_SECONDS:
        _, key = _delivery_timestamps.pop(0)
        _delivery_cache.discard(key)

@@ -75,9 +77,14 @@ def _is_duplicate(event: str, delivery: str, payload: Optional[Dict[str, Any]] =
        # 取 body 或 content，优先 body（webhookNotifier 格式）
        content = review.get("body", "") or review.get("content", "")
        content_hash = hashlib.sha256(content.encode()).hexdigest()[:16]
-        content_key = f"content:{event}:{pr_num}:{sender}:{content_hash}"
+        review_id = review.get("id", "")
+        content_key = f"content:{event}:{pr_num}:{sender}:{review_id}:{content_hash}"
        if content_key in _delivery_cache:
-            logger.info("Content-based duplicate detected: %s PR#%s by %s", event, pr_num, sender)
+            logger.info(
+                "Content-based duplicate detected: %s PR#%s by %s",
+                event,
+                pr_num,
+                sender)
            return True
        _delivery_cache.add(content_key)
        _delivery_timestamps.append((now, content_key))
@@ -137,8 +144,16 @@ async def _fetch_pr_files(repo: str, pr_number: int) -> Tuple[List[str], str]:
            last_error = str(e)
            if attempt < 2:
                await asyncio.sleep(0.5 * (attempt + 1))
-                logger.warning("Retry %d/3 fetching PR files: %s/pulls/%d", attempt + 1, repo, pr_number)
-    logger.warning("Failed to fetch PR files after 3 retries: %s/pulls/%d - %s", repo, pr_number, last_error)
+                logger.warning(
+                    "Retry %d/3 fetching PR files: %s/pulls/%d",
+                    attempt + 1,
+                    repo,
+                    pr_number)
+    logger.warning(
+        "Failed to fetch PR files after 3 retries: %s/pulls/%d - %s",
+        repo,
+        pr_number,
+        last_error)
    return [], f"获取文件列表失败（重试3次）: {last_error}"


@@ -244,14 +259,21 @@ def _repo_fullname(payload: Dict[str, Any]) -> str:


 async def _handle_pull_request(payload: Dict[str, Any]) -> None:
-    """处理 pull_request 事件：opened → 通知 simayi-challenger。"""
+    """处理 pull_request 事件：opened → 通知 reviewer；synchronize → 通知 reviewer 重新 review。"""
    action = payload.get("action", "")
-    if action != "opened":
-        return
+    if action == "opened":
+        await _handle_pr_opened(payload)
+    elif action == "synchronize":
+        await _handle_pr_synchronize(payload)
+    # 其他 action 忽略

+
+async def _handle_pr_opened(payload: Dict[str, Any]) -> None:
+    """PR opened → 通知 simayi-challenger。"""
    pr = payload.get("pull_request")
    if not pr or not isinstance(pr, dict):
-        logger.warning("pull_request event missing pull_request field, skipping")
+        logger.warning(
+            "pull_request event missing pull_request field, skipping")
        return
    repo = _repo_fullname(payload)
    pr_number = pr.get("number", 0)
@@ -265,7 +287,8 @@ async def _handle_pull_request(payload: Dict[str, Any]) -> None:
    if fetch_error:
        file_list = f"⚠️ {fetch_error}"
    else:
-        file_list = "\n".join(f"- {f}" for f in changed_files) if changed_files else "(无文件变更)"
+        file_list = "\n".join(
+            f"- {f}" for f in changed_files) if changed_files else "(无文件变更)"

    text = render_template("review_request", {
        "repo": repo,
@@ -290,11 +313,13 @@ async def _handle_pull_request_review(payload: Dict[str, Any]) -> None:
    """
    review = payload.get("review")
    if not review or not isinstance(review, dict):
-        logger.warning("pull_request_review event missing review field, skipping")
+        logger.warning(
+            "pull_request_review event missing review field, skipping")
        return
    pr = payload.get("pull_request")
    if not pr or not isinstance(pr, dict):
-        logger.warning("pull_request_review event missing pull_request field, skipping")
+        logger.warning(
+            "pull_request_review event missing pull_request field, skipping")
        return

    # 兼容两种 payload 格式提取 state
@@ -309,18 +334,39 @@ async def _handle_pull_request_review(payload: Dict[str, Any]) -> None:
        }
        state = type_map.get(review_type, "")

-    # 只通知 APPROVED 和 REQUEST_CHANGES，跳过 COMMENTED 和其他状态
-    if state == "COMMENTED":
-        return
-
    repo = _repo_fullname(payload)
    pr_number = pr.get("number", 0)
    pr_title = pr.get("title", "")
    pr_author = pr.get("user", {}).get("login", "unknown")
    # 兼容：org webhook 的 review 没有 user，从 sender 取
-    reviewer = review.get("user", {}).get("login", "") or payload.get("sender", {}).get("login", "unknown")
+    reviewer = review.get(
+        "user",
+        {}).get(
+        "login",
+        "") or payload.get(
+            "sender",
+            {}).get(
+                "login",
+        "unknown")
    review_body = review.get("body", "") or review.get("content", "(无评论)")

+    if state == "COMMENTED":
+        # Review 评论 → 通知 PR 作者
+        review_body = review.get("body", "") or review.get("content", "(无评论)")
+        reviewer = review.get("user", {}).get("login", "") or payload.get("sender", {}).get("login", "unknown")
+
+        text = render_template("review_comment", {
+            "repo": repo,
+            "pr_number": str(pr_number),
+            "pr_title": pr_title,
+            "reviewer": reviewer,
+            "comment_body": review_body,
+        })
+
+        title = f"Review 评论: {pr_title} ({repo}#{pr_number})"
+        _send_mail(pr_author, title, text)
+        return
+
    result_map = {"APPROVED": "通过 ✓", "REQUEST_CHANGES": "驳回 ✗"}
    if state not in result_map:
        return
@@ -339,6 +385,72 @@ async def _handle_pull_request_review(payload: Dict[str, Any]) -> None:
    _send_mail(pr_author, title, text)


+async def _fetch_latest_reviewer(repo: str, pr_number: int) -> str:
+    """查询 PR 最近一次非 PENDING review 的提交者。
+
+    Returns:
+        reviewer login 或空字符串
+    """
+    if not _GITEA_TOKEN:
+        return ""
+
+    url = f"{_GITEA_BASE}/repos/{repo}/pulls/{pr_number}/reviews"
+    headers = {"Authorization": f"token {_GITEA_TOKEN}"}
+
+    try:
+        async with httpx.AsyncClient(timeout=5.0) as client:
+            resp = await client.get(url, headers=headers)
+            resp.raise_for_status()
+            reviews = resp.json()
+
+            # 取最后一个非 PENDING 的 review 的 user
+            for review in reversed(reviews):
+                state = review.get("state", "")
+                if state in ("APPROVED", "REQUEST_CHANGES", "COMMENTED"):
+                    user = review.get("user", {})
+                    return user.get("login", "")
+    except Exception as e:
+        logger.warning("Failed to fetch reviews for %s#%d: %s", repo, pr_number, e)
+
+    return ""
+
+
+async def _handle_pr_synchronize(payload: Dict[str, Any]) -> None:
+    """PR 更新（新 push）→ 通知 reviewer 重新 review。
+
+    查询最近一次 review 的提交者作为通知目标。
+    只在有 review 历史时才通知（避免和 opened 重复）。
+    """
+    pr = payload.get("pull_request")
+    if not pr or not isinstance(pr, dict):
+        return
+
+    repo = _repo_fullname(payload)
+    pr_number = pr.get("number", 0)
+    pr_title = pr.get("title", "")
+    pr_author = pr.get("user", {}).get("login", "unknown")
+    new_sha = pr.get("head", {}).get("sha", "unknown")[:12]
+
+    # 查询最近 review 的提交者
+    reviewer = await _fetch_latest_reviewer(repo, pr_number)
+    if not reviewer:
+        # 没有 review 历史，跳过（opened 事件已经通知过）
+        logger.debug("No review history for PR #%s, skipping synchronize notification", pr_number)
+        return
+
+    text = render_template("review_updated", {
+        "repo": repo,
+        "pr_number": str(pr_number),
+        "pr_title": pr_title,
+        "pr_author": pr_author,
+        "new_sha": new_sha,
+        "reviewer": reviewer,
+    })
+
+    title = f"PR 更新: {pr_title} ({repo}#{pr_number})"
+    _send_mail(reviewer, title, text)
+
+
 async def _handle_issues(payload: Dict[str, Any]) -> None:
    """处理 issues 事件：assigned → 通知被指派人；opened+部署失败 → 通知运维。"""
    action = payload.get("action", "")
@@ -365,7 +477,8 @@ async def _handle_issues(payload: Dict[str, Any]) -> None:
            logger.debug("Issue assigned but no assignee found, skipping")
            return

-        labels_list = [lbl.get("name", "") for lbl in (issue.get("labels") or [])]
+        labels_list = [lbl.get("name", "")
+                       for lbl in (issue.get("labels") or [])]
        labels = ", ".join(labels_list) if labels_list else "(无标签)"
        issue_body = issue.get("body", "(无描述)")
        brief = issue_title[:20].replace(" ", "-").lower()
@@ -416,7 +529,9 @@ async def _handle_issue_comment(payload: Dict[str, Any]) -> None:

    # 已关闭的 Issue/PR 不再发送 CI 失败通知
    if issue.get("state") == "closed":
-        logger.debug("Skipping CI failure notification for closed issue #%s", issue.get("number"))
+        logger.debug(
+            "Skipping CI failure notification for closed issue #%s",
+            issue.get("number"))
        return

    repo = _repo_fullname(payload)
@@ -484,7 +599,8 @@ async def gitea_webhook(
    # 1. 签名验证
    if not _verify_signature(body, x_gitea_signature):
        logger.warning("Webhook signature verification failed")
-        return Response(status_code=403, content="signature verification failed")
+        return Response(status_code=403,
+                        content="signature verification failed")

    # 3. 解析 payload（提前解析，用于幂等检查）
    try:
@@ -497,14 +613,18 @@ async def gitea_webhook(
    if x_gitea_event and x_gitea_delivery:
        async with _idempotency_lock:
            if _is_duplicate(x_gitea_event, x_gitea_delivery, payload):
-                logger.debug("Duplicate webhook: %s/%s", x_gitea_event, x_gitea_delivery)
+                logger.debug(
+                    "Duplicate webhook: %s/%s",
+                    x_gitea_event,
+                    x_gitea_delivery)
                return Response(status_code=200, content="duplicate")

    # 4. 查找 handler
    handler = _EVENT_HANDLERS.get(x_gitea_event or "")
    if not handler:
        logger.debug("Unhandled event type: %s", x_gitea_event)
-        return Response(status_code=200, content=f"unhandled event: {x_gitea_event}")
+        return Response(status_code=200,
+                        content=f"unhandled event: {x_gitea_event}")

    # 5. 执行 handler
    try:
@@ -132,8 +132,10 @@ def _migrate_v28(conn: sqlite3.Connection) -> None:
        resolved_by TEXT,
        resolve_note TEXT
    )""")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_checkpoints_task ON checkpoints(task_id)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_checkpoints_status ON checkpoints(status)")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_checkpoints_task ON checkpoints(task_id)")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_checkpoints_status ON checkpoints(status)")

    # 4. outputs 扩展字段（M3 成果物）
    _safe_add_column(conn, "outputs", "file_name", "TEXT")
@@ -188,18 +190,20 @@ TERMINAL_STATUSES = frozenset()  # v3.1: 无终态，全靠 VALID_TRANSITIONS
 MANUAL_STATUSES = frozenset({"cancelled", "paused", "reviewing"})

 VALID_TRANSITIONS = {
-    "pending":       {"claimed", "paused", "blocked", "cancelled"},
-    "claimed":       {"working", "paused", "pending", "cancelled"},
-    "working":       {"review", "done", "blocked", "failed", "paused", "escalated", "waiting_human", "cancelled", "pending"},  # pending: Mail spawn 失败回退
-    "paused":        {"working", "claimed", "review", "escalated", "waiting_human", "cancelled"},  # 恢复到 resumed_from 记录的状态
-    "review":        {"done", "pending", "failed", "paused", "escalated", "waiting_human", "cancelled"},
-    "blocked":       {"pending", "escalated", "cancelled"},
-    "failed":        {"pending", "escalated", "cancelled"},
-    "escalated":     {"working", "pending", "paused", "cancelled"},
+    "pending": {"claimed", "paused", "blocked", "cancelled"},
+    "claimed": {"working", "paused", "pending", "cancelled"},
+    # pending: Mail spawn 失败回退
+    "working": {"review", "done", "blocked", "failed", "paused", "escalated", "waiting_human", "cancelled", "pending"},
+    # 恢复到 resumed_from 记录的状态
+    "paused": {"working", "claimed", "review", "escalated", "waiting_human", "cancelled"},
+    "review": {"done", "pending", "failed", "paused", "escalated", "waiting_human", "cancelled"},
+    "blocked": {"pending", "escalated", "cancelled"},
+    "failed": {"pending", "escalated", "cancelled"},
+    "escalated": {"working", "pending", "paused", "cancelled"},
    "waiting_human": {"working", "done", "paused", "cancelled"},
-    "done":          {"cancelled", "reviewing"},
-    "reviewing":     {"done", "working", "cancelled"},
-    "cancelled":     {"pending"},
+    "done": {"cancelled", "reviewing"},
+    "reviewing": {"done", "working", "cancelled"},
+    "cancelled": {"pending"},
 }

 COMMENT_TYPES = frozenset({
@@ -223,7 +227,8 @@ EVENT_TYPES = frozenset({

 OUTPUT_TYPES = frozenset({"code", "document", "data", "config", "other"})

-REVIEW_TYPES = frozenset({"plan_review", "output_review", "guardrail", "final_review"})
+REVIEW_TYPES = frozenset(
+    {"plan_review", "output_review", "guardrail", "final_review"})
 VERDICT_TYPES = frozenset({"approved", "rejected", "needs_revision"})

 EXPERIENCE_SOURCES = frozenset({
@@ -83,7 +83,8 @@ class Blackboard:
        """获取单个任务"""
        conn = self._conn()
        try:
-            row = conn.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
+            row = conn.execute(
+                "SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
            return Task.from_row(row) if row else None
        finally:
            conn.close()
@@ -128,7 +129,8 @@ class Blackboard:
                updates["completed_at"] = now  # paused 也记录时间用于恢复
                updates["resumed_from"] = old_status  # 记录暂停前状态
            elif new_status == "pending":
-                # 所有 →pending 转换都清空 assignee（与 ticker._transition_status L414 对齐）
+                # 所有 →pending 转换都清空 assignee（与 ticker._transition_status L414
+                # 对齐）
                updates["assignee"] = None
                updates["claimed_at"] = None
                updates["current_agent"] = None
@@ -707,7 +709,8 @@ class Blackboard:
        import uuid
        # BUG-33: 校验 payload 结构必须含 version 字段
        if not isinstance(payload, dict) or "version" not in payload:
-            raise ValueError("payload must be a dict containing 'version' field")
+            raise ValueError(
+                "payload must be a dict containing 'version' field")
        cp_id = checkpoint_id or f"cp-{uuid.uuid4().hex[:8]}"
        conn = self._conn()
        try:
@@ -964,7 +967,8 @@ class Blackboard:
        finally:
            conn.close()

-    def get_pending_mentions(self, max_retries: int = 5) -> List[Dict[str, Any]]:
+    def get_pending_mentions(
+            self, max_retries: int = 5) -> List[Dict[str, Any]]:
        """获取所有 pending 且未超过重试上限的 mentions"""
        conn = self._conn()
        try:
@@ -999,7 +1003,8 @@ class Blackboard:
        conn = self._conn()
        try:
            conn.execute("BEGIN IMMEDIATE")
-            conn.execute("UPDATE mention_queue SET retry_count=retry_count+1 WHERE id=?", (mention_id,))
+            conn.execute(
+                "UPDATE mention_queue SET retry_count=retry_count+1 WHERE id=?", (mention_id,))
            conn.commit()
            return True
        finally:
@@ -1010,7 +1015,8 @@ class Blackboard:
        conn = self._conn()
        try:
            conn.execute("BEGIN IMMEDIATE")
-            conn.execute("UPDATE mention_queue SET status='failed' WHERE id=?", (mention_id,))
+            conn.execute(
+                "UPDATE mention_queue SET status='failed' WHERE id=?", (mention_id,))
            conn.commit()
            return True
        finally:
@@ -132,7 +132,8 @@ class Queries:
        """任务详情聚合（含关联数据）"""
        conn = self._conn()
        try:
-            row = conn.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
+            row = conn.execute(
+                "SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
            if not row:
                return None
            task = dict(row)
@@ -159,7 +160,8 @@ class Queries:
        finally:
            conn.close()

-    def task_events(self, task_id: str, limit: int = 50) -> List[Dict[str, Any]]:
+    def task_events(self, task_id: str,
+                    limit: int = 50) -> List[Dict[str, Any]]:
        """任务事件列表"""
        conn = self._conn()
        try:
@@ -265,7 +267,8 @@ class Queries:
                return "review"

            # 有 working/claimed → working
-            if status_counts.get("working", 0) > 0 or status_counts.get("claimed", 0) > 0:
+            if status_counts.get("working", 0) > 0 or status_counts.get(
+                    "claimed", 0) > 0:
                return "working"

            # 有 pending → pending
@@ -337,7 +340,8 @@ class Queries:
            # 当前活跃 stage
            active_stage = None
            for sp in stage_progress:
-                if sp["active"] > 0 or (sp["total"] > 0 and sp["done"] < sp["total"]):
+                if sp["active"] > 0 or (
+                        sp["total"] > 0 and sp["done"] < sp["total"]):
                    if not active_stage and sp["done"] < sp["total"]:
                        active_stage = sp["label"]

@@ -119,7 +119,8 @@ class ProjectRegistry:
        finally:
            conn.close()

-    def list_projects(self, status: Optional[str] = None) -> Dict[str, Dict[str, Any]]:
+    def list_projects(
+            self, status: Optional[str] = None) -> Dict[str, Dict[str, Any]]:
        """列出项目"""
        conn = self._connect()
        try:
@@ -178,7 +179,8 @@ class ProjectRegistry:
            status="deleted",
        )

-    def physical_delete_project(self, project_id: str) -> Optional[Dict[str, Any]]:
+    def physical_delete_project(
+            self, project_id: str) -> Optional[Dict[str, Any]]:
        """物理删除项目（删目录 + 删 registry 条目）"""
        import shutil

@@ -260,7 +262,8 @@ class ProjectRegistry:
    # 迁移（从 _registry.yaml）
    # ===================================================================

-    def discover_sanguo_projects(self, scan_dir: Optional[Path] = None) -> List[str]:
+    def discover_sanguo_projects(
+            self, scan_dir: Optional[Path] = None) -> List[str]:
        """扫描 sanguo_projects 开发目录，自动注册正式项目"""
        scan_dir = scan_dir or Path(os.environ.get(
            "SANGUO_PROJECTS_DIR",
@@ -355,4 +358,3 @@ class ProjectRegistry:

    def reload(self) -> None:
        """兼容旧接口（SQLite 不需要 reload cache）"""
-        pass
@@ -35,7 +35,9 @@ def _get_queries(project_id: str) -> Queries:


 def build_blackboard_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(prog="blackboard", description="Agent blackboard operations")
+    parser = argparse.ArgumentParser(
+        prog="blackboard",
+        description="Agent blackboard operations")
    sub = parser.add_subparsers(dest="command")

    # read
@@ -206,7 +208,11 @@ def _cmd_comment(opts) -> int:

 def _cmd_decide(opts) -> int:
    bb = _get_bb(opts.project)
-    did = bb.add_decision(opts.task_id, opts.decider, opts.decision, opts.rationale)
+    did = bb.add_decision(
+        opts.task_id,
+        opts.decider,
+        opts.decision,
+        opts.rationale)
    print(f"Decision recorded: {did}")
    return 0

@@ -251,7 +257,8 @@ def _print_tasks(tasks, as_json: bool):


 def build_admin_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(prog="admin", description="Admin operations")
+    parser = argparse.ArgumentParser(
+        prog="admin", description="Admin operations")
    sub = parser.add_subparsers(dest="command")

    # project create
@@ -262,7 +269,7 @@ def build_admin_parser() -> argparse.ArgumentParser:
    p_pc.add_argument("--description", default="")

    # project list
-    _ = sub.add_parser("project-list", help="List projects")
+    sub.add_parser("project-list", help="List projects")

    # project archive
    p_pa = sub.add_parser("project-archive", help="Archive project")
@@ -300,7 +307,8 @@ def run_admin_cli(args: Optional[List[str]] = None) -> int:
            for pid, info in projects.items():
                status = info.get("status", "?")
                agents = ",".join(info.get("agents", []))
-                print(f"  {pid}  [{status}]  {info.get('name', '')}  agents: {agents}")
+                print(
+                    f"  {pid}  [{status}]  {info.get('name', '')}  agents: {agents}")
            return 0

        elif opts.command == "project-archive":
@@ -0,0 +1,183 @@
+"""base_task_handler.py — Task type handler 基类。
+
+收敛合理的共性能力（crash rollback + verify + mark + notify），
+子类只实现差异点。
+"""
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional
+
+from src.daemon.prompt_composer import PromptContext, PromptSection
+from src.blackboard.db import get_connection
+
+logger = logging.getLogger("moziplus-v2.handler")
+
+
+@dataclass
+class VerifyResult:
+    """验证结果"""
+    passed: bool
+    reason: str          # "has_output" / "no_reply" / "no_signal" / ...
+    evidence: str        # "output_count=1, comment_count=0"
+    can_retry: bool = True
+    retry_count: int = 0
+
+
+class BaseTaskHandler:
+    """所有 task type handler 的基类。
+
+    职责：L2 引擎注入层的业务逻辑——prompt 构建、完成验证、状态标记。
+    不管：进程生命周期、exit 分类、重试决策（这些归 spawner）。
+    """
+
+    # crash 类 outcome（进程级异常，需要 rollback）
+    CRASH_OUTCOMES = frozenset({
+        "crashed", "compact_failed", "process_crash",
+        "session_stuck", "compact_hanging",
+    })
+
+    task_type: str = ""
+    virtual_project: Optional[str] = None
+    display_name: str = ""  # 中文展示名（ticker 扫描日志用）
+
+    # === 子类必须实现 ===
+
+    def build_prompt(self, context: PromptContext) -> str:
+        """构建 L2 prompt（通过 PromptComposer 拼 section）。子类实现。"""
+        raise NotImplementedError
+
+    def verify_completion(self, task_id: str, db_path: Path) -> VerifyResult:
+        """验证任务完成质量。每个 handler 自己的验证逻辑。子类实现。"""
+        raise NotImplementedError
+
+    def target_success_status(self) -> str:
+        """验证通过后的目标状态。task='review', mail/toolchain='done'"""
+        return "review"
+
+    def get_sections(self) -> list[PromptSection]:
+        """返回此 handler 的 prompt section 列表。子类实现。"""
+        return []
+
+    # === 基类提供统一流程 ===
+
+    def pre_spawn(self, task_id: str, db_path: Path) -> bool:
+        """spawn 前业务准备。默认 True。
+        mail/toolchain override 为 auto_working。"""
+        return True
+
+    def post_complete(self, task_id: str, agent_id: str,
+                      outcome: str, db_path: Path) -> None:
+        """spawn 完成后的业务处理。统一 4 步流程：
+        1. crash 处理 → rollback current_agent
+        2. verify → 验证产出
+        3. mark → 标目标状态
+        4. notify → 失败时 on_failure
+        """
+        # 1. crash 处理（基类提供，所有 handler 继承）
+        if outcome in self.CRASH_OUTCOMES:
+            self._rollback_current_agent(db_path, task_id, agent_id)
+            return
+
+        # 2. verify
+        result = self.verify_completion(task_id, db_path)
+
+        # 3. mark
+        if result.passed:
+            self._mark_task_status(db_path, task_id, self.target_success_status())
+            logger.info("Task %s: verify passed (%s), marked %s",
+                        task_id, result.reason, self.target_success_status())
+        else:
+            # 4. notify
+            self.on_failure(task_id, agent_id, db_path, result)
+
+    def on_failure(self, task_id: str, agent_id: str,
+                   db_path: Path, verify: VerifyResult) -> None:
+        """验证失败处理。默认：标 failed。子类可 override。"""
+        self._mark_task_status(db_path, task_id, "failed")
+        logger.info("Task %s: verify failed (%s), marked failed",
+                    task_id, verify.reason)
+
+    def check_completion(self, task_id: str, db_path: Path) -> bool:
+        """ticker 级别的完成检查。默认：False。"""
+        return False
+
+    # === 内部工具方法 ===
+
+    def _rollback_current_agent(self, db_path: Path, task_id: str, agent_id: str) -> None:
+        """crash 后回退 current_agent → assignee，避免 exclude_current 卡死。
+        从 dispatcher._rollback_current_agent 迁移。"""
+        try:
+            conn = get_connection(db_path)
+            try:
+                conn.execute(
+                    "UPDATE tasks SET current_agent = "
+                    "(SELECT assignee FROM tasks WHERE id=?) "
+                    "WHERE id=? AND current_agent=?",
+                    (task_id, task_id, agent_id)
+                )
+                conn.commit()
+            finally:
+                conn.close()
+            logger.info("Task %s: rolled back current_agent from %s to assignee",
+                        task_id, agent_id)
+        except Exception as e:
+            logger.warning("Task %s: failed to rollback current_agent: %s",
+                           task_id, e)
+
+    def _mark_task_status(self, db_path: Path, task_id: str, status: str) -> None:
+        """更新任务状态 + 写审计事件（带 3 次重试，防 SQLite DB 锁）。"""
+        for attempt in range(3):
+            try:
+                conn = get_connection(db_path)
+                try:
+                    conn.execute("BEGIN IMMEDIATE")
+                    old_row = conn.execute(
+                        "SELECT status FROM tasks WHERE id=?", (task_id,)
+                    ).fetchone()
+                    old_status = old_row["status"] if old_row else "unknown"
+                    conn.execute(
+                        "UPDATE tasks SET status=?, updated_at=datetime('now') WHERE id=?",
+                        (status, task_id),
+                    )
+                    conn.execute(
+                        "INSERT INTO events (task_id, agent, event_type, payload) "
+                        "VALUES (?, 'handler', 'status_change', ?)",
+                        (task_id,
+                         f'{{"from": "{old_status}", "to": "{status}", '
+                         f'"source": "{self.task_type}_handler"}}'),
+                    )
+                    conn.commit()
+                    return
+                finally:
+                    conn.close()
+            except Exception as e:
+                logger.warning("Handler: mark %s → %s attempt %d failed: %s",
+                               task_id, status, attempt + 1, e)
+        logger.error("Handler: mark %s → %s all 3 attempts failed", task_id, status)
+
+    def _auto_mark_working(self, task_id: str, db_path: Path) -> bool:
+        """pending → working（mail/toolchain 通用）。"""
+        try:
+            conn = get_connection(db_path)
+            try:
+                conn.execute("BEGIN IMMEDIATE")
+                row = conn.execute(
+                    "SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
+                if not row or row["status"] not in ("pending", "claimed"):
+                    logger.warning("Task %s: cannot mark working (status=%s)",
+                                   task_id, row["status"] if row else "not found")
+                    return False
+                conn.execute(
+                    "UPDATE tasks SET status='working', updated_at=datetime('now') "
+                    "WHERE id=?", (task_id,))
+                conn.commit()
+                logger.info("Task %s: auto-marked working", task_id)
+                return True
+            finally:
+                conn.close()
+        except Exception as e:
+            logger.error("Task %s: failed to mark working: %s", task_id, e)
+            return False
@@ -27,12 +27,12 @@ class BootstrapBuilder:
    """L2 引擎注入层构建器（v2.1 四段式）"""

    ROLE_SKILL_MAP = {
-        "executor":          "blackboard-executor",
-        "reviewer":          "blackboard-reviewer",
-        "reviewer-simayi":   "blackboard-reviewer-simayi",
+        "executor": "blackboard-executor",
+        "reviewer": "blackboard-reviewer",
+        "reviewer-simayi": "blackboard-reviewer-simayi",
        "reviewer-pangtong": "blackboard-reviewer-pangtong",
-        "planner":           "blackboard-planner",
-        "claim":             "blackboard-claim",
+        "planner": "blackboard-planner",
+        "claim": "blackboard-claim",
    }

    # 默认从环境变量或配置读取，fallback 到默认路径
@@ -61,7 +61,9 @@ class BootstrapBuilder:

        # 段 2: 前序产出（有依赖时注入）
        if task.get("depends_on_outputs"):
-            sections.append(self._format_prior_outputs(task["depends_on_outputs"]))
+            sections.append(
+                self._format_prior_outputs(
+                    task["depends_on_outputs"]))

        # 段 3: 角色操作规范全文（通过 ROLE_SKILL_MAP 从 Skill 文件读取）
        skill_name = self.ROLE_SKILL_MAP.get(role)
@@ -133,7 +135,8 @@ class BootstrapBuilder:
        """格式化前序产出摘要（段 2）"""
        parts = ["## 前序产出"]
        for out in outputs:
-            parts.append(f"- [{out.get('task_id', '?')}] {out.get('summary', '无摘要')}")
+            parts.append(
+                f"- [{out.get('task_id', '?')}] {out.get('summary', '无摘要')}")
        return "\n".join(parts)

    def _format_constraints(self, role: str) -> str:
@@ -68,20 +68,23 @@ class ActiveAgentCounter:
        self._cooldown_until.pop(agent_id, None)
        return False

-    def set_cooldown(self, agent_id: str, seconds: Optional[float] = None) -> None:
+    def set_cooldown(self, agent_id: str,
+                     seconds: Optional[float] = None) -> None:
        """设置冷却期（默认 120 秒）"""
        cd = seconds if seconds is not None else self._default_cooldown_seconds
        self._cooldown_until[agent_id] = time.time() + cd
        logger.info("Cooldown set for %s: %.0fs (until %.0f)",
                    agent_id, cd, self._cooldown_until[agent_id])

-    async def can_acquire(self, agent_id: str, session_id: str = "main") -> bool:
+    async def can_acquire(self, agent_id: str,
+                          session_id: str = "main") -> bool:
        """三层检查：cooldown → global → per agent → per session key"""
        if self.is_cooling_down(agent_id):
            return False
        if self._global_active >= self._max_global:
            return False
-        if self._agent_active.get(agent_id, 0) >= self._max_concurrent_sessions:
+        if self._agent_active.get(
+                agent_id, 0) >= self._max_concurrent_sessions:
            return False
        key = self._make_key(agent_id, session_id)
        if self._active_keys.get(key, 0) >= self._max_per_session:
@@ -122,7 +125,8 @@ class ActiveAgentCounter:
                del self._active_keys[key]

        if agent_id in self._agent_active:
-            self._agent_active[agent_id] = max(0, self._agent_active[agent_id] - 1)
+            self._agent_active[agent_id] = max(
+                0, self._agent_active[agent_id] - 1)
            if self._agent_active[agent_id] == 0:
                del self._agent_active[agent_id]

@@ -12,7 +12,6 @@ Dispatcher 负责：
 from __future__ import annotations

 import json
-import pathlib
 import logging
 import sqlite3
 from enum import Enum
@@ -23,6 +22,7 @@ from src.blackboard.models import Task
 from src.blackboard.db import get_connection
 from src.daemon.spawner import AgentBusyError
 from src.daemon.router import AgentRouter
+from src.daemon.task_type_registry import TaskTypeRegistry

 logger = logging.getLogger("moziplus-v2.dispatcher")

@@ -64,7 +64,8 @@ class Dispatcher:
        if self._legacy_mode:
            self.registered_agents = set(registered_agents or [])
            self.capability_map = capability_map or {}
-            logger.warning("Dispatcher running in legacy mode (no AgentRouter)")
+            logger.warning(
+                "Dispatcher running in legacy mode (no AgentRouter)")

    def decide(self, task: Task, action_type: str = "") -> Dict[str, Any]:
        """调度决策（委托给 Router）
@@ -123,17 +124,23 @@ class Dispatcher:
             "status": "dispatched"|"skipped"|"error"|"blocked", "reason": str}
        """
        # 安全红线检查（调度前拦截）
-        # Mail 是 Agent 间通信，不做 guardrail 检查
-        is_mail = project_config.get("project_id") == "_mail" if project_config else False
-        if self.guardrails and not is_mail:
+        # handler 项目（_mail/_toolchain）不做 guardrail 检查
+        handler = TaskTypeRegistry.get_by_project(
+            project_config.get("project_id", "") if project_config else "")
+        is_handler_task = handler is not None
+        if self.guardrails and not is_handler_task:
            violations = self.guardrails.check_task(task)
-            critical = [v for v in violations if v.action in ("block_and_notify", "terminate_and_escalate")]
+            critical = [
+                v for v in violations if v.action in (
+                    "block_and_notify",
+                    "terminate_and_escalate")]
            if critical:
                v = critical[0]
                logger.warning("Task '%s' blocked by guardrail: %s - %s",
                               task.title, v.rule_id, v.message)
                # 写入黑板事件
-                _routing_db = Path(project_config["db_path"]) if project_config and "db_path" in project_config else self.db_path
+                _routing_db = Path(
+                    project_config["db_path"]) if project_config and "db_path" in project_config else self.db_path
                if _routing_db:
                    self._record_routing(task, {"level": DispatchLevel.BLOCKED, "agent_id": "none",
                                                "reason": v.message}, "blocked", v.message, _routing_db)
@@ -152,7 +159,8 @@ class Dispatcher:
        decision = self.decide(task, action_type)
        level = decision["level"]
        # 从 project_config 获取项目级 DB 路径（路由审计日志写入项目 DB）
-        _routing_db = Path(project_config["db_path"]) if project_config and "db_path" in project_config else None
+        _routing_db = Path(
+            project_config["db_path"]) if project_config and "db_path" in project_config else None
        agent_id = decision["agent_id"]

        # v2.7.2: counter 检查移到 spawn_full_agent 内部
@@ -160,7 +168,8 @@ class Dispatcher:

        # 本地执行
        if level == DispatchLevel.LOCAL:
-            self._record_routing(task, decision, "dispatched", None, _routing_db)
+            self._record_routing(
+                task, decision, "dispatched", None, _routing_db)
            return {
                "level": level.value,
                "agent_id": "daemon",
@@ -172,7 +181,8 @@ class Dispatcher:
        # Full Agent / Escalate spawn
        if level in (DispatchLevel.FULL_AGENT, DispatchLevel.ESCALATE):
            if not self.spawner:
-                self._record_routing(task, decision, "error", "No spawner", _routing_db)
+                self._record_routing(
+                    task, decision, "error", "No spawner", _routing_db)
                return {
                    "level": level.value,
                    "agent_id": agent_id,
@@ -182,107 +192,111 @@ class Dispatcher:
                }

            try:
-                # [v2.7.1] Mail: 标 working 移到 spawn_full_agent 内部（check 通过后、subprocess 前）
-                is_mail = project_config.get("project_id") == "_mail" if project_config else False
-                if is_mail:
-                    db_path = Path(project_config["db_path"]) if project_config and "db_path" in project_config else None
+                # [Step 5] Handler: pre_spawn + on_checks_passed 统一
+                project_id = project_config.get("project_id", "") if project_config else ""
+                handler = TaskTypeRegistry.get_by_project(project_id)
+                db_path = Path(
+                    project_config["db_path"]) if project_config and "db_path" in project_config else None

-                # on_checks_passed: 所有检查通过后才标 working，检查失败不标
+                # on_checks_passed: handler 项目在 check 通过后调用 handler.pre_spawn
                on_checks_passed = None
-                _mail_marked_working = False
-                if is_mail and db_path:
+                handler_marked_working = False
+                if handler and db_path:
                    _task_id = task.id
-                    _mail_db = db_path
-                    _disp = self
+                    _handler_db = db_path
+                    _handler = handler

-                    def _mail_on_checks_passed():
-                        nonlocal _mail_marked_working
-                        if not _disp._mail_auto_working(_task_id, _mail_db):
-                            raise RuntimeError("mail_auto_working_failed")
-                        _mail_marked_working = True
-                    on_checks_passed = _mail_on_checks_passed
+                    def _handler_on_checks_passed():
+                        nonlocal handler_marked_working
+                        if not _handler.pre_spawn(_task_id, _handler_db):
+                            raise RuntimeError("handler_pre_spawn_failed")
+                        handler_marked_working = True
+                    on_checks_passed = _handler_on_checks_passed

                # 构建 spawn message
                message = self._build_spawn_message(task, agent_id, project_config,
-                                                    mode=decision.get("mode", ""),
+                                                    mode=decision.get(
+                                                        "mode", ""),
                                                    spawn_type=action_type or "executor")

-                # v2.7.2: on_complete 只含业务逻辑，不含 counter.release
-                # counter.release 由 spawn_full_agent 内部的 wrapped_on_complete 保证
+                # [Step 5] Handler: on_complete 统一走 handler.post_complete
+                # 保留旧路径作为 fallback（无 handler 的项目）
                on_complete = None
-                if is_mail:
+                if handler:
                    _task_id = task.id
-                    _mail_db = db_path
-                    _must_haves = task.must_haves or ""
-                    _dispatcher = self
+                    _handler_db = db_path
+                    _handler = handler

-                    def _mail_on_complete(aid, outcome):
-                        # 幻觉门控：检查是否有回复，自动标 done/failed
+                    def _handler_on_complete(aid, outcome):
                        try:
-                            _dispatcher._mail_auto_complete(_task_id, aid, _mail_db, _must_haves, outcome=outcome)
+                            _handler.post_complete(
+                                _task_id, aid, outcome, _handler_db)
                        except Exception as e:
-                            logger.error("Mail %s: on_complete error: %s", _task_id, e)
-                    on_complete = _mail_on_complete
+                            logger.error(
+                                "Handler %s: on_complete error: %s", _task_id, e)
+                    on_complete = _handler_on_complete
                else:
-                    # #02: Task 路径也加 on_complete（幻觉门控）
+                    # 旧路径：无 handler 的项目（_general 等）
                    _task_id = task.id
-                    _task_db = Path(project_config["db_path"]) if project_config and "db_path" in project_config else None
+                    _task_db = db_path
                    _dispatcher = self
                    _is_review = action_type == "review"

-                    # #07.2: executor/review 统一 crash 回退
                    ROLLBACK_CURRENT_AGENT_OUTCOMES = frozenset({
                        "crashed", "compact_failed", "process_crash",
                        "session_stuck", "compact_hanging",
                    })

-                    def _task_on_complete(aid, outcome):
+                    def _legacy_on_complete(aid, outcome):
                        try:
-                            # #07.2: 统一 crash 回退——executor 和 review 都回退 current_agent
                            if outcome in ROLLBACK_CURRENT_AGENT_OUTCOMES and _task_db:
-                                _dispatcher._rollback_current_agent(_task_db, _task_id, aid)
-
+                                _dispatcher._rollback_current_agent(
+                                    _task_db, _task_id, aid)
                            if _is_review:
                                if _task_db and outcome in ("completed", "session_revived"):
-                                    # #09: 读 verdict 决定后续动作
-                                    conn = get_connection(_task_db)
+                                    from src.blackboard.blackboard import Blackboard
+                                    # get_connection 已在文件顶部 L22 import
+                                    rconn = get_connection(_task_db)
                                    try:
-                                        review = conn.execute(
-                                            "SELECT verdict FROM reviews WHERE task_id=? ORDER BY created_at DESC LIMIT 1",
-                                            (_task_id,)
-                                        ).fetchone()
+                                        review_row = rconn.execute(
+                                            "SELECT verdict, reviewer, comment FROM reviews "
+                                            "WHERE task_id=? ORDER BY created_at DESC LIMIT 1",
+                                            (_task_id,)).fetchone()
                                    finally:
-                                        conn.close()
+                                        rconn.close()

-                                    if review and review["verdict"] == "approved":
-                                        _dispatcher._mark_task_status(_task_db, _task_id, "done")
-                                        logger.info("Task %s: review approved, marking done", _task_id)
+                                    if review_row and review_row["verdict"] == "approved":
+                                        _dispatcher._mark_task_status(
+                                            _task_db, _task_id, "done")
+                                        logger.info(
+                                            "Legacy %s: review approved, marked done", _task_id)
                                    else:
-                                        # 非 approved → @mention 被审 agent（assignee，非 current_agent）
-                                        verdict_str = review["verdict"] if review else "未知"
-                                        conn2 = get_connection(_task_db)
+                                        verdict_str = review_row["verdict"] if review_row else "未知"
+                                        tconn = get_connection(_task_db)
                                        try:
-                                            task_row = conn2.execute("SELECT assignee FROM tasks WHERE id=?", (_task_id,)).fetchone()
+                                            t_row = tconn.execute(
+                                                "SELECT assignee FROM tasks WHERE id=?",
+                                                (_task_id,)).fetchone()
                                        finally:
-                                            conn2.close()
-
-                                        if task_row and task_row["assignee"]:
-                                            from src.blackboard.blackboard import Blackboard
-                                            bb = Blackboard(_task_db)
-                                            bb.add_comment(_task_id, "daemon",
-                                                           f"@{task_row['assignee']} 审查结论: {verdict_str}，请查看详情并决定接受或反驳",
-                                                           comment_type="review")
-                                        logger.info("Task %s: review verdict=%s, notified assignee=%s",
-                                                    _task_id, verdict_str, task_row["assignee"] if task_row else "?")
-                                        # 不标 done，保持 review 状态
-                                else:
-                                    logger.warning("Task %s: review agent %s (%s), NOT marking done", _task_id, aid, outcome)
+                                            tconn.close()
+                                        if t_row and t_row["assignee"]:
+                                            bb = Blackboard(str(_task_db))
+                                            bb.add_comment(
+                                                _task_id, "daemon",
+                                                f"@{t_row['assignee']} review 未通过 "
+                                                f"(verdict={verdict_str}): "
+                                                f"{review_row['comment'] if review_row else ''}",
+                                                comment_type="review")
+                                        logger.info(
+                                            "Legacy %s: review not approved (%s), "
+                                            "@mentioned assignee",
+                                            _task_id, verdict_str)
                            else:
-                                # executor: 三信号验证 → 标 review
                                _dispatcher._task_auto_complete(_task_id, _task_db)
                        except Exception as e:
-                            logger.error("Task %s: on_complete error: %s", _task_id, e)
-                    on_complete = _task_on_complete
+                            logger.error(
+                                "Legacy %s: on_complete error: %s", _task_id, e)
+                    on_complete = _legacy_on_complete

                session_id = await self.spawner.spawn_full_agent(
                    agent_id=agent_id,
@@ -290,7 +304,8 @@ class Dispatcher:
                    task_id=task.id,
                    on_complete=on_complete,
                    use_main_session=True,  # #02: 统一投递到 main session
-                    task_db_path=Path(project_config["db_path"]) if project_config and "db_path" in project_config else None,
+                    task_db_path=Path(
+                        project_config["db_path"]) if project_config and "db_path" in project_config else None,
                    on_checks_passed=on_checks_passed,
                )

@@ -313,9 +328,14 @@ class Dispatcher:
                else:
                    log_level = logger.debug
                    detail_msg = f"Agent busy: {reason}"
-                log_level("Dispatch skipped %s for task %s: %s", agent_id, task.id, detail_msg)
+                log_level(
+                    "Dispatch skipped %s for task %s: %s",
+                    agent_id,
+                    task.id,
+                    detail_msg)
                # on_checks_passed 未执行（check 失败在它之前），working 未标，无需回退
-                self._record_routing(task, decision, "skipped", detail_msg, _routing_db)
+                self._record_routing(
+                    task, decision, "skipped", detail_msg, _routing_db)
                return {
                    "level": level.value,
                    "agent_id": agent_id,
@@ -325,9 +345,28 @@ class Dispatcher:
                }
            except Exception as e:
                # on_checks_passed 已执行但 subprocess 失败 → 回退 working → pending
-                if _mail_marked_working:
-                    self._mail_revert_to_pending(task.id, db_path)
-                self._record_routing(task, decision, "error", str(e), _routing_db)
+                if handler_marked_working and handler and db_path:
+                    # handler 项目：回退到 pending
+                    try:
+                        conn = get_connection(db_path)
+                        try:
+                            conn.execute("BEGIN IMMEDIATE")
+                            row = conn.execute(
+                                "SELECT status FROM tasks WHERE id=?", (task.id,)).fetchone()
+                            if row and row["status"] == "working":
+                                conn.execute(
+                                    "UPDATE tasks SET status='pending', updated_at=datetime('now') WHERE id=?",
+                                    (task.id,))
+                                conn.commit()
+                                logger.info(
+                                    "Task %s: reverted working → pending (spawn failed)", task.id)
+                        finally:
+                            conn.close()
+                    except Exception as revert_err:
+                        logger.error(
+                            "Task %s: failed to revert to pending: %s", task.id, revert_err)
+                self._record_routing(
+                    task, decision, "error", str(e), _routing_db)
                return {
                    "level": level.value,
                    "agent_id": agent_id,
@@ -386,9 +425,16 @@ class Dispatcher:
    def _build_delegate_prompt(self, task: Task,
                               project_config: Optional[Dict]) -> str:
        """构建 delegate 模式的 prompt（协调员分配任务）"""
-        api_host = getattr(self.spawner, 'api_host', '127.0.0.1') if self.spawner else '127.0.0.1'
-        api_port = getattr(self.spawner, 'api_port', 8083) if self.spawner else 8083
-        project_id = project_config.get("project_id", "") if project_config else ""
+        api_host = getattr(
+            self.spawner,
+            'api_host',
+            '127.0.0.1') if self.spawner else '127.0.0.1'
+        api_port = getattr(
+            self.spawner,
+            'api_port',
+            8083) if self.spawner else 8083
+        project_id = project_config.get(
+            "project_id", "") if project_config else ""

        return f"""你是任务协调员。请分析以下任务，决定最合适的执行者并分配。

@@ -479,7 +525,8 @@ class Dispatcher:

    # ── Legacy 兼容（deprecated） ──

-    def _legacy_decide(self, task: Task, action_type: str = "") -> Dict[str, Any]:
+    def _legacy_decide(
+            self, task: Task, action_type: str = "") -> Dict[str, Any]:
        """旧版三级决策树（兼容过渡用）"""
        LOCAL_ACTIONS = frozenset({
            "L1_guardrail", "format_check",
@@ -519,7 +566,8 @@ class Dispatcher:
                return registered[0]
        return "pangtong-fujunshi"

-    async def _legacy_dispatch(self, task, action_type="", project_config=None):
+    async def _legacy_dispatch(
+            self, task, action_type="", project_config=None):
        """旧版 dispatch（兼容过渡用）

        v2.7.2: counter acquire/release 移到 spawn_full_agent 内部。
@@ -541,16 +589,21 @@ class Dispatcher:
            try:
                # NOTE: _legacy_dispatch 仅在 router=None 时触发，当前配置不会进入。
                # Mail 永远走 dispatch() 主路径（on_checks_passed 方案），不走此路径。
-                # 如果未来 legacy 路径被启用，需同步 on_checks_passed 逻辑。
-                is_mail_legacy = project_config.get("project_id") == "_mail" if project_config else False
-                if is_mail_legacy:
-                    db_path_legacy = Path(project_config["db_path"]) if project_config and "db_path" in project_config else None
-                    if not db_path_legacy or not self._mail_auto_working(task.id, db_path_legacy):
+                # [Step 5] handler 统一：用注册表查 handler
+                project_id_legacy = project_config.get("project_id", "") if project_config else ""
+                handler_legacy = TaskTypeRegistry.get_by_project(project_id_legacy)
+                if handler_legacy:
+                    db_path_legacy = Path(
+                        project_config["db_path"]) if project_config and "db_path" in project_config else None
+                    if db_path_legacy:
+                        handler_legacy.pre_spawn(task.id, db_path_legacy)
+                    else:
                        return {"level": level.value, "agent_id": agent_id,
                                "session_id": None, "status": "error",
-                                "reason": "mail_auto_working_failed"}
+                                "reason": "no db_path for handler"}

-                if hasattr(self.spawner, 'build_spawn_message') and project_config:
+                if hasattr(self.spawner,
+                           'build_spawn_message') and project_config:
                    retry_ctx = self._build_retry_context(task)
                    message = self.spawner.build_spawn_message(
                        task_id=task.id, title=task.title,
@@ -569,32 +622,34 @@ class Dispatcher:

                # v2.7.2: on_complete 只含业务逻辑
                on_complete_legacy = None
-                if is_mail_legacy:
+                if handler_legacy:
                    _t_id = task.id
-                    _m_db = db_path_legacy
-                    _m_mh = task.must_haves or ""
-                    _disp = self
+                    _h_db = db_path_legacy
+                    _h = handler_legacy

-                    def _mail_oc_legacy(aid, outcome):
+                    def _handler_oc_legacy(aid, outcome):
                        try:
-                            _disp._mail_auto_complete(_t_id, aid, _m_db, _m_mh, outcome=outcome)
+                            _h.post_complete(_t_id, aid, outcome, _h_db)
                        except Exception as e:
-                            logger.error("Mail %s: legacy on_complete error: %s", _t_id, e)
-                    on_complete_legacy = _mail_oc_legacy
+                            logger.error(
+                                "Handler %s: legacy on_complete error: %s", _t_id, e)
+                    on_complete_legacy = _handler_oc_legacy

                session_id = await self.spawner.spawn_full_agent(
                    agent_id=agent_id, message=message,
                    task_id=task.id,
                    on_complete=on_complete_legacy,
                    use_main_session=True,  # #02: 统一投递到 main session
-                    task_db_path=Path(project_config["db_path"]) if project_config and "db_path" in project_config else None,
+                    task_db_path=Path(
+                        project_config["db_path"]) if project_config and "db_path" in project_config else None,
                )
                return {"level": level.value, "agent_id": agent_id,
                        "session_id": session_id, "status": "dispatched",
                        "reason": decision["reason"]}
            except AgentBusyError as e:
                reason = getattr(e, 'reason', 'busy')
-                detail_msg = f"Session busy: {reason}" if reason.startswith("session_") else f"Agent busy: {reason}"
+                detail_msg = f"Session busy: {reason}" if reason.startswith(
+                    "session_") else f"Agent busy: {reason}"
                return {"level": level.value, "agent_id": agent_id,
                        "session_id": None, "status": "skipped",
                        "reason": detail_msg}
@@ -608,6 +663,7 @@ class Dispatcher:

    # ── Mail 信封/载荷分离辅助方法 ──

+    # DEPRECATED: Step 5 handler 架构已替代此方法，保留仅供平滑过渡，确认稳定后删除。
    def _mail_auto_working(self, task_id: str, db_path: Path) -> bool:
        """Mail 任务：系统自动标 working（spawn 前）

@@ -619,9 +675,11 @@ class Dispatcher:
            conn = get_connection(db_path)
            try:
                conn.execute("BEGIN IMMEDIATE")
-                row = conn.execute("SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
+                row = conn.execute(
+                    "SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
                if not row:
-                    logger.warning("Mail %s: cannot mark working (task not found)", task_id)
+                    logger.warning(
+                        "Mail %s: cannot mark working (task not found)", task_id)
                    return False
                if row["status"] not in ("pending", "claimed"):
                    logger.warning("Mail %s: cannot mark working (status=%s, expected pending/claimed)",
@@ -632,7 +690,10 @@ class Dispatcher:
                    (task_id,),
                )
                conn.commit()
-                logger.info("Mail %s: auto-marked working (system, was %s)", task_id, row["status"])
+                logger.info(
+                    "Mail %s: auto-marked working (system, was %s)",
+                    task_id,
+                    row["status"])
                return True
            finally:
                conn.close()
@@ -640,27 +701,37 @@ class Dispatcher:
            logger.error("Mail %s: failed to mark working: %s", task_id, e)
            return False

+    # DEPRECATED: Step 5 handler 架构已替代此方法，保留仅供平滑过渡，确认稳定后删除。
    def _mail_revert_to_pending(self, task_id: str, db_path: Path) -> None:
        """Mail spawn 失败时回退 working → pending，避免永久死锁"""
        try:
            conn = get_connection(db_path)
            try:
                conn.execute("BEGIN IMMEDIATE")
-                row = conn.execute("SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
+                row = conn.execute(
+                    "SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
                if row and row["status"] == "working":
                    conn.execute(
                        "UPDATE tasks SET status='pending', updated_at=datetime('now') WHERE id=?",
                        (task_id,),
                    )
                    conn.commit()
-                    logger.info("Mail %s: reverted working → pending (spawn failed)", task_id)
+                    logger.info(
+                        "Mail %s: reverted working → pending (spawn failed)", task_id)
                else:
-                    logger.debug("Mail %s: skip revert (status=%s, expected working)", task_id, row["status"] if row else "not_found")
+                    logger.debug(
+                        "Mail %s: skip revert (status=%s, expected working)",
+                        task_id,
+                        row["status"] if row else "not_found")
            finally:
                conn.close()
        except Exception as e:
-            logger.error("Mail %s: failed to revert to pending: %s", task_id, e)
+            logger.error(
+                "Mail %s: failed to revert to pending: %s",
+                task_id,
+                e)

+    # DEPRECATED: Step 5 handler 架构已替代此方法，保留仅供平滑过渡，确认稳定后删除。
    def _mail_auto_complete(self, task_id: str, agent_id: str,
                            db_path: Path, must_haves: str, outcome=None) -> None:
        """Mail 任务：on_complete 后自动标 done/failed（含幻觉门控）"""
@@ -669,7 +740,9 @@ class Dispatcher:
            performative = "request"
            try:
                meta = json.loads(must_haves) if must_haves else {}
-                performative = meta.get("performative", meta.get("type", "request"))
+                performative = meta.get(
+                    "performative", meta.get(
+                        "type", "request"))
            except Exception:
                pass

@@ -678,13 +751,15 @@ class Dispatcher:
                has_reply = self._mail_check_reply(task_id, db_path)
                if not has_reply:
                    # F3: 立刻标 failed（不等 ticker 30 分钟）
-                    logger.error("Mail %s: no reply found, marking failed (no_reply_found)", task_id)
+                    logger.error(
+                        "Mail %s: no reply found, marking failed (no_reply_found)", task_id)
                    for attempt in range(3):
                        try:
                            conn = get_connection(db_path)
                            try:
                                conn.execute("BEGIN IMMEDIATE")
-                                row = conn.execute("SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
+                                row = conn.execute(
+                                    "SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
                                if not row:
                                    return
                                if row["status"] == "working":
@@ -698,19 +773,24 @@ class Dispatcher:
                                         json.dumps({"reason": "no_reply_found"}, ensure_ascii=False)),
                                    )
                                    conn.commit()
-                                    logger.info("Mail %s: marked failed (no_reply_found)", task_id)
+                                    logger.info(
+                                        "Mail %s: marked failed (no_reply_found)", task_id)
                                    # Mail 失败通知：通知发件人
                                    try:
                                        from src.daemon.mail_notify import notify_mail_failed
-                                        notify_mail_failed(db_path, task_id, "no_reply_found")
+                                        notify_mail_failed(
+                                            db_path, task_id, "no_reply_found")
                                    except Exception as ne:
-                                        logger.warning("Mail %s: failed to send no_reply_found notification: %s", task_id, ne)
+                                        logger.warning(
+                                            "Mail %s: failed to send no_reply_found notification: %s", task_id, ne)
                                return
                            finally:
                                conn.close()
                        except Exception as e:
-                            logger.warning("Mail %s: failed attempt %d: %s", task_id, attempt + 1, e)
-                    logger.error("Mail %s: all 3 failed attempts failed, leaving for ticker", task_id)
+                            logger.warning(
+                                "Mail %s: failed attempt %d: %s", task_id, attempt + 1, e)
+                    logger.error(
+                        "Mail %s: all 3 failed attempts failed, leaving for ticker", task_id)
                    return

            # inform 类型：只对成功 outcome 标 done，失败 outcome 留 working 等 ticker 重投
@@ -718,7 +798,10 @@ class Dispatcher:
            if performative == "inform":
                INFORM_DONE_OUTCOMES = {"completed", "claimed", "no_reply"}
                if outcome not in INFORM_DONE_OUTCOMES:
-                    logger.info("Mail %s: inform outcome=%s, skip auto-done", task_id, outcome)
+                    logger.info(
+                        "Mail %s: inform outcome=%s, skip auto-done",
+                        task_id,
+                        outcome)
                    return

            # 标 done（重试 3 次）
@@ -727,7 +810,8 @@ class Dispatcher:
                    conn = get_connection(db_path)
                    try:
                        conn.execute("BEGIN IMMEDIATE")
-                        row = conn.execute("SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
+                        row = conn.execute(
+                            "SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
                        if not row:
                            return
                        if row["status"] == "working":
@@ -742,13 +826,20 @@ class Dispatcher:
                    finally:
                        conn.close()
                except Exception as e:
-                    logger.warning("Mail %s: done attempt %d failed: %s", task_id, attempt + 1, e)
+                    logger.warning(
+                        "Mail %s: done attempt %d failed: %s",
+                        task_id,
+                        attempt + 1,
+                        e)
            # 3 次都失败，留 working 等 ticker 超时兜底
-            logger.error("Mail %s: all 3 done attempts failed, leaving for ticker", task_id)
+            logger.error(
+                "Mail %s: all 3 done attempts failed, leaving for ticker",
+                task_id)

        except Exception as e:
            logger.error("Mail %s: auto-complete error: %s", task_id, e)

+    # DEPRECATED: Step 5 handler 架构已替代此方法，保留仅供平滑过渡，确认稳定后删除。
    def _mail_check_reply(self, original_task_id: str, db_path: Path) -> bool:
        """幻觉门控：检查是否有回复邮件（in_reply_to = original_task_id）"""
        try:
@@ -789,7 +880,9 @@ class Dispatcher:
                logger.info("Task %s: verify passed, marking review", task_id)
                self._mark_task_status(db_path, task_id, "review")
            else:
-                logger.info("Task %s: verify not passed (no signal), leaving working", task_id)
+                logger.info(
+                    "Task %s: verify not passed (no signal), leaving working",
+                    task_id)
        except Exception as e:
            logger.error("Task %s: auto-complete error: %s", task_id, e)

@@ -824,7 +917,8 @@ class Dispatcher:
            logger.error("Task %s: verify error: %s", task_id, e)
            return True

-    def _rollback_current_agent(self, db_path: Path, task_id: str, agent_id: str) -> None:
+    def _rollback_current_agent(
+            self, db_path: Path, task_id: str, agent_id: str) -> None:
        """#07.2: crash 后回退 current_agent 到 assignee，避免 exclude_current 卡死"""
        try:
            conn = get_connection(db_path)
@@ -838,11 +932,18 @@ class Dispatcher:
                conn.commit()
            finally:
                conn.close()
-            logger.info("Task %s: rolled back current_agent from %s to assignee", task_id, agent_id)
+            logger.info(
+                "Task %s: rolled back current_agent from %s to assignee",
+                task_id,
+                agent_id)
        except Exception as e:
-            logger.warning("Task %s: failed to rollback current_agent: %s", task_id, e)
+            logger.warning(
+                "Task %s: failed to rollback current_agent: %s",
+                task_id,
+                e)

-    def _mark_task_status(self, db_path: Path, task_id: str, status: str) -> None:
+    def _mark_task_status(self, db_path: Path,
+                          task_id: str, status: str) -> None:
        """更新任务状态 + 写审计事件"""
        try:
            conn = get_connection(db_path)
@@ -858,7 +959,8 @@ class Dispatcher:
                )
                conn.execute(
                    "INSERT INTO events (task_id, agent, event_type, payload) VALUES (?, 'dispatcher', 'status_change', ?)",
-                    (task_id, f'{{"from": "{old_status}", "to": "{status}", "source": "auto_complete"}}'),
+                    (task_id,
+                     f'{{"from": "{old_status}", "to": "{status}", "source": "auto_complete"}}'),
                )
                conn.commit()
            finally:
@@ -867,7 +969,7 @@ class Dispatcher:
            logger.error("Task %s: mark status error: %s", task_id, e)

    @staticmethod
-    def _check_crash_limit(task_id: str, db_path: pathlib.Path, limit: int = 3,
+    def _check_crash_limit(task_id: str, db_path: Path, limit: int = 3,
                           window_minutes: int = 30) -> bool:
        """v2.8.1 Fix-3c: 检查 task 最近 window_minutes 内的 crash 次数是否超限。

@@ -38,7 +38,9 @@ class GuardrailEngine:
            data = yaml.safe_load(f)
        self.rules = data.get("rules", [])
        self.settings = data.get("settings", {"enabled": True})
-        logger.info("Loaded %d guardrail rules from %s", len(self.rules), config_path)
+        logger.info(
+            "Loaded %d guardrail rules from %s", len(
+                self.rules), config_path)

    def check_task(self, task: Any) -> List[GuardrailViolation]:
        """检查 Task 是否触犯安全红线（调度前调用）"""
@@ -95,7 +97,8 @@ class GuardrailEngine:

        return violations

-    def check_token_usage(self, token_count: int) -> Optional[GuardrailViolation]:
+    def check_token_usage(
+            self, token_count: int) -> Optional[GuardrailViolation]:
        """检查 Token 消耗是否超标"""
        if not self.settings.get("enabled", True):
            return None
@@ -103,7 +106,10 @@ class GuardrailEngine:
        for rule in self.rules:
            if rule["id"] != "high_token_usage":
                continue
-            threshold = rule.get("triggers", [{}])[0].get("token_threshold", 100000)
+            threshold = rule.get(
+                "triggers", [
+                    {}])[0].get(
+                "token_threshold", 100000)
            if token_count > threshold:
                return GuardrailViolation(
                    rule_id=rule["id"],
@@ -114,7 +120,8 @@ class GuardrailEngine:
                )
        return None

-    def check_consecutive_failure(self, failure_count: int) -> Optional[GuardrailViolation]:
+    def check_consecutive_failure(
+            self, failure_count: int) -> Optional[GuardrailViolation]:
        """检查连续失败次数"""
        if not self.settings.get("enabled", True):
            return None
@@ -122,7 +129,10 @@ class GuardrailEngine:
        for rule in self.rules:
            if rule["id"] != "consecutive_failure":
                continue
-            threshold = rule.get("triggers", [{}])[0].get("consecutive_failures", 3)
+            threshold = rule.get(
+                "triggers", [
+                    {}])[0].get(
+                "consecutive_failures", 3)
            if failure_count >= threshold:
                return GuardrailViolation(
                    rule_id=rule["id"],
@@ -41,6 +41,7 @@ class HealthChecker:
            {"healthy": bool, "zombie": bool, "stale_ticks": int,
             "alert_written": bool, "resolved": bool}
        """
+        str(db_path)
        result: Dict[str, Any] = {
            "healthy": True,
            "zombie": False,
@@ -57,7 +58,8 @@ class HealthChecker:
        # 用 event count 变化判断是否有真实变更
        conn = queries._conn()
        try:
-            conn.execute("SELECT COUNT(*) FROM events").fetchone()[0]
+            conn.execute(
+                "SELECT COUNT(*) FROM events").fetchone()[0]
            non_tick_events = conn.execute(
                "SELECT COUNT(*) FROM events WHERE event_type != 'daemon_tick' "
                "AND event_type != 'agent_zombie_detected'"
@@ -84,7 +86,8 @@ class HealthChecker:
            self._stale_ticks[project_id] = stale
            result["stale_ticks"] = stale

-            if stale >= self.zombie_threshold and not self._alerted.get(project_id):
+            if stale >= self.zombie_threshold and not self._alerted.get(
+                    project_id):
                # 写告警
                self._write_alert(db_path, project_id, tick_num, stale)
                self._alerted[project_id] = True
@@ -125,7 +128,10 @@ class HealthChecker:
            conn.commit()
        finally:
            conn.close()
-        logger.warning("Zombie detected: %s (stale=%d)", project_id, stale_ticks)
+        logger.warning(
+            "Zombie detected: %s (stale=%d)",
+            project_id,
+            stale_ticks)

    def _write_resolution(self, db_path: Path, project_id: str,
                          tick_num: int) -> None:
@@ -27,7 +27,8 @@ class InboxWatcher:
    def __init__(
        self,
        inbox_path: Path,
-        process_callback: Optional[Callable[[Dict[str, Any]], Coroutine[Any, Any, None]]] = None,
+        process_callback: Optional[Callable[[
+            Dict[str, Any]], Coroutine[Any, Any, None]]] = None,
        watch_interval: float = 1.0,
    ):
        """
@@ -159,7 +160,8 @@ class InboxWatcher:
                                   line_no, type(event).__name__)
                    self._total_errors += 1
            except json.JSONDecodeError:
-                logger.warning("Inbox line %d: invalid JSON, skipping", line_no)
+                logger.warning(
+                    "Inbox line %d: invalid JSON, skipping", line_no)
                self._total_errors += 1

        return events
@@ -0,0 +1,210 @@
+"""mail_handler.py — Mail 任务 handler。
+
+处理 Agent 间通信（飞鸽传书），含 inform 和 request 两种类型。
+"""
+from __future__ import annotations
+
+import json
+import logging
+from pathlib import Path
+
+from src.daemon.base_task_handler import BaseTaskHandler, VerifyResult
+from src.daemon.prompt_composer import PromptComposer, PromptContext
+from src.blackboard.db import get_connection
+
+logger = logging.getLogger("moziplus-v2.handler.mail")
+
+
+class MailHandler(BaseTaskHandler):
+    """Mail 任务 handler。"""
+
+    task_type = "mail"
+    virtual_project = "_mail"
+    display_name = "飞鸽传书"
+
+    def target_success_status(self) -> str:
+        return "done"
+
+    def pre_spawn(self, task_id: str, db_path: Path) -> bool:
+        """auto_working：pending → working"""
+        return self._auto_mark_working(task_id, db_path)
+
+    def build_prompt(self, context: PromptContext) -> str:
+        """通过 PromptComposer 拼装 3 个 section。"""
+        composer = PromptComposer()
+        composer.add_many(self.get_sections())
+        return composer.compose(context)
+
+    def get_sections(self) -> list:
+        return [MailContextSection(), MailApiSection(), MailConstraintsSection()]
+
+    def verify_completion(self, task_id: str, db_path: Path) -> VerifyResult:
+        """Mail 完成验证：区分 inform/request。
+
+        - inform: 始终通过（通知已阅即 done，不需要检查产出）
+        - request: 检查是否已回复
+        """
+        performative = self._parse_performative(task_id, db_path)
+
+        if performative == "inform":
+            return VerifyResult(True, "inform_auto", f"performative={performative}")
+
+        # request: 检查是否已回复
+        has_reply = self._check_reply(task_id, db_path)
+        if has_reply:
+            return VerifyResult(True, "has_reply", f"performative={performative}")
+        return VerifyResult(False, "no_reply", f"performative={performative}")
+
+    # post_complete 由基类 BaseTaskHandler 统一处理（crash→verify→mark→notify）
+    # inform: verify 始终通过 → 基类 mark done ✅
+    # request 有回复: verify 通过 → 基类 mark done ✅
+    # request 无回复: verify 失败 → 基类调 on_failure ✅
+
+    def on_failure(self, task_id: str, agent_id: str,
+                   db_path: Path, verify: VerifyResult) -> None:
+        """request 验证失败 → 标 failed + 通知发件人"""
+        self._mark_task_status(db_path, task_id, "failed")
+        logger.info("Mail %s: request verify failed (%s), marked failed",
+                    task_id, verify.reason)
+
+        # 通知发件人
+        try:
+            from src.daemon.mail_notify import notify_mail_failed
+            notify_mail_failed(db_path, task_id, "no_reply_found")
+        except Exception as e:
+            logger.warning("Mail %s: failed to send notification: %s", task_id, e)
+
+    # === 内部方法 ===
+
+    def _parse_performative(self, task_id: str, db_path: Path) -> str:
+        """解析 mail 类型（inform/request）"""
+        try:
+            conn = get_connection(db_path)
+            try:
+                row = conn.execute(
+                    "SELECT must_haves FROM tasks WHERE id=?", (task_id,)
+                ).fetchone()
+                if row and row["must_haves"]:
+                    meta = json.loads(row["must_haves"])
+                    return meta.get("performative", meta.get("type", "request"))
+            finally:
+                conn.close()
+        except Exception:
+            pass
+        return "request"
+
+    def _check_reply(self, task_id: str, db_path: Path) -> bool:
+        """检查是否已回复（查 tasks 表找 in_reply_to 回复邮件）
+
+        从 dispatcher._mail_check_reply 迁移。
+        Mail 回复机制：创建新 task，must_haves JSON 中包含 in_reply_to = original_task_id。
+        不能查 comments 表——回复邮件是独立的 task，不是 comment。
+        """
+        try:
+            conn = get_connection(db_path)
+            try:
+                row = conn.execute(
+                    "SELECT id FROM tasks WHERE id != ? AND must_haves LIKE ? LIMIT 1",
+                    (task_id, f'%{task_id}%'),
+                ).fetchone()
+                return row is not None
+            finally:
+                conn.close()
+        except Exception as e:
+            logger.error("Mail %s: check reply error: %s", task_id, e)
+            # 查询失败时保守处理：假设有回复（避免误标 failed）
+            return True
+
+    def check_completion(self, task_id: str, db_path: Path) -> bool:
+        """ticker 级别的完成检查：检查是否已回复"""
+        return self._check_reply(task_id, db_path)
+
+
+# ===================================================================
+# Mail PromptSections
+# ===================================================================
+
+class MailContextSection:
+    """邮件上下文段 — 发件人/收件人/主题/内容，区分 inform/request。"""
+
+    name: str = "mail_context"
+    priority: int = 10
+
+    def render(self, context: PromptContext) -> str:
+        if context.mail_type == "inform":
+            return self._render_inform(context)
+        return self._render_request(context)
+
+    def should_include(self, context: PromptContext) -> bool:  # noqa: ARG002
+        return True
+
+    @staticmethod
+    def _render_inform(context: PromptContext) -> str:
+        return (
+            f"你收到一封飞鸽传书(纯通知)。\n\n"
+            f"发件者: {context.from_agent}\n"
+            f"主题: {context.title}\n"
+            f"内容: {context.description}\n\n"
+            f"已阅即可。如需回复,用 in_reply_to 回复发件者(不需要填 to)。\n"
+            f"⚠️ 不要执行任何状态转换命令。"
+        )
+
+    @staticmethod
+    def _render_request(context: PromptContext) -> str:
+        return (
+            f"你收到一封飞鸽传书,需要你处理并回复。\n\n"
+            f"发件者: {context.from_agent}\n"
+            f"主题: {context.title}\n"
+            f"内容: {context.description}\n\n"
+            f"### 如何回复发件者\n\n"
+            f'curl -s -X POST http://localhost:8083/api/mail \\\n'
+            f"  -H 'Content-Type: application/json' \\\n"
+            f'  -d \'{{"from": "{context.agent_id}", '
+            f'"in_reply_to": "{context.task_id}", '
+            f'"title": "回复: {context.title}", '
+            f'"text": "你的回复内容"}}\'\n\n'
+            f"⚠️ 不需要填 \"to\",系统自动回复给发件者。"
+        )
+
+
+class MailApiSection:
+    """Mail API 操作指令段。"""
+
+    name: str = "mail_api"
+    priority: int = 40
+
+    def render(self, context: PromptContext) -> str:
+        return (
+            f"### 如何给其他人发新邮件\n\n"
+            f'curl -s -X POST http://localhost:8083/api/mail \\\n'
+            f"  -H 'Content-Type: application/json' \\\n"
+            f'  -d \'{{"from": "{context.agent_id}", '
+            f'"to": "对方agent-id", '
+            f'"title": "标题", '
+            f'"text": "正文", '
+            f'"type": "inform"}}\'\n\n'
+            f"⚠️ to 必须是有效的 agent id\n"
+            f"⚠️ 纯通知用 type=inform,需要对方回复不填 type(默认 request)"
+        )
+
+    def should_include(self, context: PromptContext) -> bool:
+        return context.mail_type == "request"
+
+
+class MailConstraintsSection:
+    """Mail 硬约束段。"""
+
+    name: str = "mail_constraints"
+    priority: int = 50
+
+    def render(self, context: PromptContext) -> str:  # noqa: ARG002
+        return (
+            "## 硬约束\n\n"
+            "1. ⚠️ 不要执行任何状态转换命令(标 working/done/review/failed 等),系统会自动处理。\n"
+            "2. ⚠️ 不能给自己发邮件\n"
+            "3. ⚠️ 发邮件时 to 必须是有效的 agent id\n"
+            "4. ⚠️ 纯通知用 type=inform,需要对方回复不填 type(默认 request)"
+        )
+
+    def should_include(self, context: PromptContext) -> bool:  # noqa: ARG002
+        return True
@@ -50,7 +50,9 @@ def notify_mail_failed(db_path: Path, original_mail_id: str,
        bb = Blackboard(db_path)
        original = bb.get_task(original_mail_id)
        if not original:
-            logger.warning("notify_mail_failed: original mail %s not found", original_mail_id)
+            logger.warning(
+                "notify_mail_failed: original mail %s not found",
+                original_mail_id)
            return

        # 解析原邮件元数据
@@ -58,7 +60,9 @@ def notify_mail_failed(db_path: Path, original_mail_id: str,

        # 防递归：系统通知邮件失败不再发通知
        if meta.get("system_notify"):
-            logger.info("Mail %s: system notify mail failed, skipping recursive notification", original_mail_id)
+            logger.info(
+                "Mail %s: system notify mail failed, skipping recursive notification",
+                original_mail_id)
            return

        # 获取发件人（优先 assigned_by，fallback must_haves.from）
@@ -67,7 +71,9 @@ def notify_mail_failed(db_path: Path, original_mail_id: str,
        title = original.title or ""

        if not from_agent:
-            logger.warning("notify_mail_failed: cannot determine sender for mail %s", original_mail_id)
+            logger.warning(
+                "notify_mail_failed: cannot determine sender for mail %s",
+                original_mail_id)
            return

        # 发件人不是有效 Agent（如 system）→ 通知庞统代处理，不触发广播
@@ -111,4 +117,7 @@ def notify_mail_failed(db_path: Path, original_mail_id: str,
                    original_mail_id, target_agent, from_agent, reason, notify_id)

    except Exception as e:
-        logger.warning("notify_mail_failed: failed to send notification for mail %s: %s", original_mail_id, e)
+        logger.warning(
+            "notify_mail_failed: failed to send notification for mail %s: %s",
+            original_mail_id,
+            e)
@@ -0,0 +1,127 @@
+"""
+prompt_composer.py — PromptSection Protocol + PromptContext + PromptComposer
+
+拼装器：有序管理 prompt 段落，按优先级排序后合并为最终 prompt。
+"""
+
+import logging
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Protocol, runtime_checkable
+
+logger = logging.getLogger("moziplus-v2.prompt_composer")
+
+# ---------------------------------------------------------------------------
+# Section 优先级范围约定
+# ---------------------------------------------------------------------------
+PRIORITY_CONTEXT = 10        # 任务上下文
+PRIORITY_PRIOR = 20          # 前序信息
+PRIORITY_ROLE = 30           # 角色规范
+PRIORITY_API = 40            # API 操作指令
+PRIORITY_CONSTRAINTS = 50    # 硬约束
+PRIORITY_EXTENSION = 60      # 扩展段
+
+
+# ---------------------------------------------------------------------------
+# PromptSection Protocol
+# ---------------------------------------------------------------------------
+@runtime_checkable
+class PromptSection(Protocol):
+    """一个 prompt 段"""
+
+    name: str                     # 段名（去重用，同名覆盖）
+    priority: int                 # 排序优先级（小数字=靠前）
+
+    def render(self, context: "PromptContext") -> str:
+        """渲染此段的文本内容。返回空字符串表示不注入。"""
+        ...
+
+    def should_include(self, context: "PromptContext") -> bool:
+        """是否注入此段（默认 True，条件段可覆盖）。"""
+        ...
+
+
+# ---------------------------------------------------------------------------
+# PromptContext 数据对象
+# ---------------------------------------------------------------------------
+@dataclass
+class PromptContext:
+    """Prompt 渲染的统一上下文"""
+
+    task_id: str
+    title: str
+    description: str
+    must_haves: str
+    project_id: str
+    agent_id: str
+
+    task: Optional[Dict] = None
+    role: str = "executor"
+    spawn_type: str = "executor"
+
+    # mail 专用
+    from_agent: str = ""
+    mail_type: str = ""          # inform / request
+
+    # toolchain 专用
+    event_type: str = ""         # ci_failure / review_request / ...
+    event_data: Dict = field(default_factory=dict)
+
+    # 前序产出
+    depends_on_outputs: Optional[List] = None
+
+
+# ---------------------------------------------------------------------------
+# PromptComposer 拼装器
+# ---------------------------------------------------------------------------
+class PromptComposer:
+    """有序拼装 prompt sections"""
+
+    SEPARATOR = "\n\n---\n\n"
+    TOKEN_BUDGET_WARN = 800      # token 预算警告阈值
+    CHARS_PER_TOKEN = 3.5        # 估算比率
+
+    def __init__(self) -> None:
+        self._sections: List[PromptSection] = []
+
+    def add(self, section: PromptSection) -> None:
+        """添加一个 section（同名覆盖）"""
+        self._sections = [s for s in self._sections if s.name != section.name]
+        self._sections.append(section)
+
+    def add_many(self, sections: List[PromptSection]) -> None:
+        """批量添加"""
+        for s in sections:
+            self.add(s)
+
+    def compose(self, context: PromptContext) -> str:
+        """拼装最终 prompt
+
+        1. 过滤 should_include=False 的段
+        2. 按 priority 排序
+        3. 逐段 render
+        4. 过滤空段
+        5. 用分隔符连接
+        6. Token 预算警告（不截断）
+        """
+        active = [s for s in self._sections if s.should_include(context)]
+        active.sort(key=lambda s: s.priority)
+
+        parts = [s.render(context) for s in active]
+        parts = [p for p in parts if p.strip()]
+
+        result = self.SEPARATOR.join(parts)
+
+        # Token 估算
+        tokens = max(1, int(len(result) / self.CHARS_PER_TOKEN))
+        logger.debug(
+            "Composed prompt from %d sections, %d tokens",
+            len(parts), tokens,
+        )
+
+        if tokens > self.TOKEN_BUDGET_WARN:
+            logger.warning(
+                "Prompt exceeds %d token budget: %d tokens (task_id=%s)",
+                self.TOKEN_BUDGET_WARN, tokens, context.task_id,
+            )
+
+        return result
@@ -148,12 +148,14 @@ class ReviewPipeline:
    ) -> ReviewResult:
        """Step 2: 格式合规"""
        if not outputs:
-            return ReviewResult("format", ReviewVerdict.FAIL, 0.0, "No outputs")
+            return ReviewResult(
+                "format", ReviewVerdict.FAIL, 0.0, "No outputs")

        issues = []
        for out in outputs:
            # output.md 必须存在且非空
-            if out.get("type") == "markdown" or out.get("path", "").endswith(".md"):
+            if out.get("type") == "markdown" or out.get(
+                    "path", "").endswith(".md"):
                content = out.get("content", "")
                if not content and out.get("path"):
                    try:
@@ -164,7 +166,8 @@ class ReviewPipeline:
                    issues.append(f"Output too short: {out.get('path', '?')}")

            # 结论 JSON 必须有效
-            if out.get("type") == "json" or out.get("path", "").endswith(".json"):
+            if out.get("type") == "json" or out.get(
+                    "path", "").endswith(".json"):
                content = out.get("content", "")
                if not content and out.get("path"):
                    try:
@@ -174,7 +177,8 @@ class ReviewPipeline:
                try:
                    data = json.loads(content)
                    if not isinstance(data, dict):
-                        issues.append(f"JSON not a dict: {out.get('path', '?')}")
+                        issues.append(
+                            f"JSON not a dict: {out.get('path', '?')}")
                except (json.JSONDecodeError, TypeError):
                    issues.append(f"Invalid JSON: {out.get('path', '?')}")

@@ -191,7 +195,8 @@ class ReviewPipeline:
    ) -> ReviewResult:
        """Step 3: 内容质量（自定义检查）"""
        if not outputs:
-            return ReviewResult("quality", ReviewVerdict.FAIL, 0.0, "No outputs")
+            return ReviewResult(
+                "quality", ReviewVerdict.FAIL, 0.0, "No outputs")

        suggestions = []
        total_score = 0.0
@@ -212,7 +217,8 @@ class ReviewPipeline:
            avg = 1.0  # 无自定义检查默认通过

        verdict = ReviewVerdict.PASS if avg >= 0.6 else ReviewVerdict.FAIL
-        return ReviewResult("quality", verdict, round(avg, 2), suggestions=suggestions)
+        return ReviewResult("quality", verdict, round(
+            avg, 2), suggestions=suggestions)

    def _determine_gate(
        self, task: Task, results: List[ReviewResult]
@@ -326,6 +332,7 @@ class RebuttalManager:
            return 0
        try:
            observations = self.bb.get_observations(task_id=task_id)
-            return sum(1 for o in observations if "Rebuttal round" in (o.body or ""))
+            return sum(
+                1 for o in observations if "Rebuttal round" in (o.body or ""))
        except Exception:
            return 0
@@ -107,7 +107,8 @@ class AgentRouter:

        # ── 快速路径 2: retry → 原执行者 ──
        if action_type == "retry":
-            current = task_info.get("current_agent") or task_info.get("assignee")
+            current = task_info.get(
+                "current_agent") or task_info.get("assignee")
            if current and current in self.agent_profiles:
                return RouteDecision(
                    agent_id=current,
@@ -119,7 +120,8 @@ class AgentRouter:
        # ── Mode B: Agent 声明式交接 ──
        next_cap = task_info.get("next_capability")
        if next_cap and self._validate_capability(next_cap):
-            current = task_info.get("current_agent") or task_info.get("assignee")
+            current = task_info.get(
+                "current_agent") or task_info.get("assignee")
            exclude = {current} if current else set()
            matched = self._match_capability(next_cap, exclude)
            if matched:
@@ -129,7 +131,9 @@ class AgentRouter:
                    mode="agent_handoff",
                    latency_ms=int((time.monotonic() - start) * 1000),
                )
-            logger.info("next_capability '%s' no match, delegate to coordinator", next_cap)
+            logger.info(
+                "next_capability '%s' no match, delegate to coordinator",
+                next_cap)

        # ── 快速路径 3: 生命周期流转查表 ──
        lifecycle = self.LIFECYCLE_CAPABILITY.get(action_type)
@@ -140,7 +144,8 @@ class AgentRouter:
            exclude_current = lifecycle.get("exclude_current", False)
            exclude = set()
            if exclude_current:
-                current = task_info.get("current_agent") or task_info.get("assignee")
+                current = task_info.get(
+                    "current_agent") or task_info.get("assignee")
                if current:
                    exclude.add(current)
            matched = self._match_capability(cap, exclude)
@@ -154,7 +159,8 @@ class AgentRouter:

        # ── 快速路径 4: 有 assignee 且非生命周期流转 ──
        assignee = task_info.get("assignee")
-        if assignee and assignee in self.agent_profiles and action_type not in ("review", "escalation"):
+        if assignee and assignee in self.agent_profiles and action_type not in (
+                "review", "escalation"):
            return RouteDecision(
                agent_id=assignee,
                reason=f"Direct assignee: {assignee}",
@@ -7,7 +7,6 @@ Subagent: 占位(实际通过 OpenClaw Gateway API sessions_spawn,F17 完善)
 from __future__ import annotations

 import asyncio
-import pathlib
 import json
 import logging
 import os
@@ -17,6 +16,7 @@ from pathlib import Path
 from typing import Any, Dict, List, Optional

 from src.blackboard.db import get_connection
+from src.daemon.task_type_registry import TaskTypeRegistry

 logger = logging.getLogger("moziplus-v2.spawner")

@@ -164,9 +164,12 @@ class AgentBusyError(Exception):

    #07: reason 字段区分具体原因,便于 dispatcher 层区分处理。
    """
-    def __init__(self, agent_id: str, reason: str = "busy", detail: Optional[dict] = None):
+
+    def __init__(self, agent_id: str, reason: str = "busy",
+                 detail: Optional[dict] = None):
        self.agent_id = agent_id
-        self.reason = reason   # counter_blocked / session_locked / session_running / session_compacting / session_stuck
+        # counter_blocked / session_locked / session_running / session_compacting / session_stuck
+        self.reason = reason
        self.detail = detail or {}
        super().__init__(f"{agent_id}: {reason}")

@@ -276,13 +279,36 @@ class AgentSpawner:
                task_id, title, description, must_haves,
                project_id, agent_id)

-        # mail 任务用精简模板
-        if project_id == "_mail":
-            return self._build_mail_prompt(task_id, title, description, must_haves, agent_id)
+        # handler 路径：Task/Mail/Toolchain 用各自的 PromptSection 构建
+        handler = TaskTypeRegistry.get_by_project(project_id)
+        if handler:
+            from src.daemon.prompt_composer import PromptContext
+            # 从 must_haves 解析 mail 元数据（from / performative）
+            from_agent = ""
+            mail_type = ""
+            try:
+                meta = json.loads(must_haves) if must_haves else {}
+                from_agent = meta.get("from", "")
+                mail_type = meta.get("performative", meta.get("type", ""))
+            except Exception:
+                pass
+            ctx = PromptContext(
+                task_id=task_id, title=title, description=description or "",
+                must_haves=must_haves or "", project_id=project_id,
+                agent_id=agent_id, role=spawn_type,
+                spawn_type=spawn_type,
+                from_agent=from_agent, mail_type=mail_type,
+            )
+            return handler.build_prompt(ctx)
+
+        # 旧路径保留：_general 等非 handler 项目

        # 走 BootstrapBuilder 新路径
        if self.bootstrap_builder and task is not None:
-            role_map = {"executor": "executor", "review": "reviewer", "discussion": "planner"}
+            role_map = {
+                "executor": "executor",
+                "review": "reviewer",
+                "discussion": "planner"}
            role = role_map.get(spawn_type, "executor")
            bootstrap_prompt = self.bootstrap_builder.build_for_task(
                task=task,
@@ -294,7 +320,8 @@ class AgentSpawner:

        # 无 BootstrapBuilder 或无 task 对象 → 最小 fallback
        # 只保留任务上下文 + API 操作指令
-        logger.warning("No BootstrapBuilder or task object, using minimal fallback")
+        logger.warning(
+            "No BootstrapBuilder or task object, using minimal fallback")
        return self._build_minimal_fallback(
            task_id, title, description, must_haves,
            project_id, agent_id)
@@ -314,8 +341,13 @@ class AgentSpawner:
    def _build_api_section(self, project_id: str, task_id: str,
                           agent_id: str) -> str:
        """构建 API 回写操作指令(BootstrapBuilder 模式下补充)"""
-        # mail 任务直接 done,不走 review
-        success_status = '"done"' if project_id == "_mail" else '"review"'
+        # handler 项目（_mail/_toolchain）的 success_status 由 PromptSection 处理
+        # 这里只处理无 handler 的项目（normal task）
+        handler = TaskTypeRegistry.get_by_project(project_id)
+        if handler:
+            success_status = '"done"' if handler.target_success_status == "done" else '"review"'
+        else:
+            success_status = '"review"'
        return f"""## 操作指令

 ### 状态回写
@@ -369,7 +401,8 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
        if not self.guardrails:
            return "无特殊限制"
        try:
-            return "、".join(r.get("name", r.get("rule_id", "")) for r in self.guardrails.rules[:6])
+            return "、".join(r.get("name", r.get("rule_id", ""))
+                            for r in self.guardrails.rules[:6])
        except Exception:
            return "无特殊限制"

@@ -389,7 +422,9 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
        try:
            meta = json.loads(must_haves) if must_haves else {}
            from_agent = meta.get("from", agent_id)
-            performative = meta.get("performative", meta.get("type", "request"))
+            performative = meta.get(
+                "performative", meta.get(
+                    "type", "request"))
        except Exception:
            pass

@@ -472,7 +507,9 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
                self._revive_session(agent_id)
            elif pre_state.get("status") == "running" and not pre_state.get("lock_pid_alive"):
                # status=running 但 lock PID 已死 → 假死,revive
-                logger.warning("Phase 0: %s status=running but lock PID dead, reviving", agent_id)
+                logger.warning(
+                    "Phase 0: %s status=running but lock PID dead, reviving",
+                    agent_id)
                self._revive_session(agent_id)

        # Phase 1: Counter acquire(互斥锁)
@@ -487,12 +524,15 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
        if use_main_session:
            session_state = self._check_session_state(agent_id)
            logger.info("Phase 2 session check for %s: status=%s lock_pid=%s lock_pid_alive=%s compact=%s",
-                        agent_id, session_state.get('status'), session_state.get('lock_pid'),
+                        agent_id, session_state.get(
+                            'status'), session_state.get('lock_pid'),
                        session_state.get('lock_pid_alive'), session_state.get('recent_compact'))

            blockers = []
-            if session_state.get("lock_pid_alive") and not session_state.get("lock_expired"):
-                blockers.append(("session_locked", session_state.get("lock_pid")))
+            if session_state.get(
+                    "lock_pid_alive") and not session_state.get("lock_expired"):
+                blockers.append(
+                    ("session_locked", session_state.get("lock_pid")))
            if session_state.get("status") == "running":
                if session_state.get("lock_pid_alive"):
                    # 真 running:外部进程占用
@@ -515,7 +555,8 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta

            # Phase 2.5: 假死修复(status=running + lock PID 死 → revive → 重检)
            # 此场景应被 Phase 0 提前修复,这里做兜底
-            if session_state.get("status") == "running" and not session_state.get("lock_pid_alive"):
+            if session_state.get("status") == "running" and not session_state.get(
+                    "lock_pid_alive"):
                logger.warning("Phase 2.5: %s status=running + lock dead (should be caught in Phase 0), reviving",
                               agent_id)
                self._revive_session(agent_id)
@@ -538,7 +579,10 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
                raise

        if self.dry_run:
-            logger.info("[DRY RUN] Would spawn agent %s (session=%s)", agent_id, _sid_key)
+            logger.info(
+                "[DRY RUN] Would spawn agent %s (session=%s)",
+                agent_id,
+                _sid_key)
            self._register_session(_sid_key, agent_id, task_id, pid=None)
            return _sid_key

@@ -554,7 +598,8 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
                        if asyncio.iscoroutine(result):
                            await result
                    except Exception:
-                        logger.warning("Business on_complete failed for %s", aid, exc_info=True)
+                        logger.warning(
+                            "Business on_complete failed for %s", aid, exc_info=True)

        cmd = [
            "openclaw", "agent",
@@ -593,7 +638,11 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
            if self.counter:
                self.counter.release(agent_id, _sid_key)
            logger.exception("Failed to spawn agent %s", agent_id)
-            self._record_attempt(task_id, agent_id, "spawn_failed", error=str(e))
+            self._record_attempt(
+                task_id,
+                agent_id,
+                "spawn_failed",
+                error=str(e))
            raise

    async def spawn_subagent(
@@ -609,7 +658,9 @@ curl -X POST http://{self.api_host}:{self.api_port}/api/projects/{project_id}/ta
        session_id = str(uuid.uuid4())

        if self.dry_run:
-            logger.info("[DRY RUN] Would spawn subagent (session=%s)", session_id)
+            logger.info(
+                "[DRY RUN] Would spawn subagent (session=%s)",
+                session_id)
            self._register_session(session_id, "subagent", task_id, pid=None)
            return session_id

@@ -729,10 +780,16 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                    agent_id, session_id, json_result)

        # 查任务实际状态
-        task_status = self._get_task_status(db_path, task_id) if task_id else None
+        task_status = self._get_task_status(
+            db_path, task_id) if task_id else None

        # 分类
-        cls = self._classify_outcome(exit_code, json_result, stderr_text, task_status, stdout_text)
+        cls = self._classify_outcome(
+            exit_code,
+            json_result,
+            stderr_text,
+            task_status,
+            stdout_text)
        outcome = cls["outcome"]

        # 更新 session 状态
@@ -761,17 +818,21 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                    agent_id, session_id, outcome, exit_code, task_status)

        # 广播反馈追踪(Phase 1 bug fix)
-        if task_id == "broadcast" and hasattr(self, '_ticker') and self._ticker:
+        if task_id == "broadcast" and hasattr(
+                self, '_ticker') and self._ticker:
            # 广播任务:从 session 信息取真实 task_id 列表,逐一回调 tracker
            sess_info = self._sessions.get(session_id or "main", {})
            bt_ids = sess_info.get("broadcast_task_ids") or []
            # 广播场景一律标 no_reply:Agent 只 claim 一个任务,
            # 其余任务的 tracker 不能被 claimed 清除
            for real_task_id in bt_ids:
-                self._ticker.record_broadcast_response(real_task_id, agent_id, "no_reply")
+                self._ticker.record_broadcast_response(
+                    real_task_id, agent_id, "no_reply")
        elif task_id and hasattr(self, '_ticker') and self._ticker:
-            outcome_str = "claimed" if cls.get("status") == "ok" else "no_reply"
-            self._ticker.record_broadcast_response(task_id, agent_id, outcome_str)
+            outcome_str = "claimed" if cls.get(
+                "status") == "ok" else "no_reply"
+            self._ticker.record_broadcast_response(
+                task_id, agent_id, outcome_str)

        if cls["should_retry"]:
            # cooldown: 新增的可恢复场景(A14/A15/A16/A8/A10)
@@ -850,14 +911,24 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            # v2.8.1 Fix-3a: crash 类 outcome 设 cooldown,给 agent session 恢复时间
            if outcome == "crashed" and self.counter:
                self.counter.set_cooldown(agent_id, seconds=60)
-                logger.info("Crash cooldown set for %s: 60s (outcome=%s)", agent_id, outcome)
+                logger.info(
+                    "Crash cooldown set for %s: 60s (outcome=%s)",
+                    agent_id,
+                    outcome)
            elif outcome in ("compact_failed", "process_crash", "session_stuck",
                             "compact_hanging", "agent_error", "compact_interrupted") and self.counter:
                self.counter.set_cooldown(agent_id, seconds=300)  # 5 分钟
-                logger.info("Error cooldown set for %s: 300s (outcome=%s)", agent_id, outcome)
+                logger.info(
+                    "Error cooldown set for %s: 300s (outcome=%s)",
+                    agent_id,
+                    outcome)
            # F1: 不可恢复 outcome → 立刻标 failed + 写黑板
-            if outcome in ("auth_failed", "agent_error") and db_path and task_id:
-                logger.error("Task %s: unrecoverable outcome=%s, marking failed immediately", task_id, outcome)
+            if outcome in ("auth_failed",
+                           "agent_error") and db_path and task_id:
+                logger.error(
+                    "Task %s: unrecoverable outcome=%s, marking failed immediately",
+                    task_id,
+                    outcome)
                self._mark_task(db_path, task_id, "failed", {
                    "reason": outcome,
                    "stderr_preview": (stderr_text or "")[:500],
@@ -881,13 +952,16 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
        except Exception:
            pass

-        _ = b"".join(stderr_chunks).decode("utf-8", errors="replace")
+        # stderr collected but not used in this handler
+        # (kept for potential future diagnostics)
+        b"".join(stderr_chunks).decode("utf-8", errors="replace")

        # 检查 session 状态
        state = self._check_session_state(agent_id)

        # B1: 假死 - 先复活,连续假死 ≥2 次再 failed
-        if state.get("status") == "running" and not state.get("lock_pid_alive", True):
+        if state.get("status") == "running" and not state.get(
+                "lock_pid_alive", True):
            # 假死计数
            stuck_count = self._stuck_counts.get(task_id, 0) + 1
            self._stuck_counts[task_id] = stuck_count
@@ -913,7 +987,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                await self._do_on_complete_async(on_complete, agent_id, "session_revived")
            else:
                # 复活失败 → 标 failed
-                logger.error("Agent %s revive failed, marking failed", agent_id)
+                logger.error(
+                    "Agent %s revive failed, marking failed", agent_id)
                self._mark_task(db_path, task_id, "failed",
                                {"reason": "revive_failed", "stuck_count": stuck_count,
                                 "diagnostics": state})
@@ -994,7 +1069,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                        "SELECT status FROM tasks WHERE id=?", (task_id,)
                    ).fetchone()
                    # Bug-6 fix: pending 不是终态
-                    if row and row["status"] in ("done", "failed", "cancelled", "review"):
+                    if row and row["status"] in (
+                            "done", "failed", "cancelled", "review"):
                        logger.info("Retry skip: task %s already %s (agent=%s)",
                                    task_id, row["status"], agent_id)
                        # on_complete = wrapped_on_complete,会 release counter
@@ -1003,7 +1079,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                finally:
                    conn.close()
            except Exception:
-                logger.warning("Retry status check failed for %s, proceeding", task_id)
+                logger.warning(
+                    "Retry status check failed for %s, proceeding", task_id)

        # 直接读写 tasks 表的 retry_count
        if retry_field == "retry_count" and db_path and task_id:
@@ -1023,7 +1100,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                finally:
                    conn.close()
            except Exception:
-                logger.exception("Failed to update retry_count for task %s", task_id)
+                logger.exception(
+                    "Failed to update retry_count for task %s", task_id)
                count = 1
        else:
            retry_counts = self._get_retry_counts(db_path, task_id)
@@ -1046,9 +1124,10 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
        # 构建续杯 message(Mail 用专用模板,Task 用标准模板)
        task_info = self._get_task_info(db_path, task_id) or {}
        project_id = task_info.get("project_id", "")
-        is_mail = project_id == "_mail"
+        handler = TaskTypeRegistry.get_by_project(project_id)
+        is_handler = handler is not None

-        if is_mail:
+        if is_handler:
            must_haves = task_info.get("must_haves", "{}")
            try:
                meta = json.loads(must_haves) if must_haves else {}
@@ -1107,7 +1186,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
        """
        text = stdout_text.strip()
        if not text:
-            return {"status": None, "summary": None, "fallback_used": False, "fallback_reason": None, "payloads": []}
+            return {"status": None, "summary": None, "fallback_used": False,
+                    "fallback_reason": None, "payloads": []}
        try:
            data = json.loads(text)
        except json.JSONDecodeError:
@@ -1119,7 +1199,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                except json.JSONDecodeError:
                    continue
            else:
-                return {"status": None, "summary": None, "fallback_used": False, "fallback_reason": None, "payloads": []}
+                return {"status": None, "summary": None, "fallback_used": False,
+                        "fallback_reason": None, "payloads": []}

        # 从 data.result.meta.executionTrace 取 fallback 信息
        result = data.get("result", {})
@@ -1135,7 +1216,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
        }

    @staticmethod
-    def _get_task_status(db_path: Optional[Path], task_id: Optional[str]) -> Optional[str]:
+    def _get_task_status(
+            db_path: Optional[Path], task_id: Optional[str]) -> Optional[str]:
        """查任务实际 API 状态"""
        if not db_path or not task_id:
            return None
@@ -1152,7 +1234,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            return None

    @staticmethod
-    def _get_task_info(db_path: Optional[Path], task_id: Optional[str]) -> Optional[dict]:
+    def _get_task_info(db_path: Optional[Path],
+                       task_id: Optional[str]) -> Optional[dict]:
        """查任务基本信息"""
        if not db_path or not task_id:
            return None
@@ -1160,7 +1243,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            conn = get_connection(db_path)
            try:
                row = conn.execute(
-                    "SELECT id, title, status FROM tasks WHERE id=?", (task_id,)
+                    "SELECT id, title, status FROM tasks WHERE id=?", (
+                        task_id,)
                ).fetchone()
                if not row:
                    return None
@@ -1192,7 +1276,9 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            sessions[main_key] = main_session
            with open(sessions_path, "w") as f:
                json.dump(sessions, f, indent=2)
-            logger.info("Revived %s: sessions.json status changed running→idle", agent_id)
+            logger.info(
+                "Revived %s: sessions.json status changed running→idle",
+                agent_id)
            # #07 O4: 同时清理残留 lock 文件
            sf = main_session.get("sessionFile", "")
            if sf:
@@ -1200,7 +1286,10 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                if lock_path.exists():
                    try:
                        lock_path.unlink()
-                        logger.info("Cleaned stale lock for %s: %s", agent_id, lock_path.name)
+                        logger.info(
+                            "Cleaned stale lock for %s: %s",
+                            agent_id,
+                            lock_path.name)
                    except Exception:
                        pass
            return True
@@ -1209,7 +1298,90 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            return False

    @staticmethod
-    def _check_recent_compaction_jsonl(session_file: str, window_seconds: int = 900) -> bool:
+    def _get_recent_gateway_logs() -> list:
+        """获取当天和昨天的 gateway 日志路径。
+
+        日志路径通过 OPENCLAW_LOG_DIR 环境变量配置，默认 /tmp/openclaw。
+        文件名格式：openclaw-{YYYY-MM-DD}.log
+        """
+        from datetime import timedelta
+        log_dir = os.environ.get("OPENCLAW_LOG_DIR", "/tmp/openclaw")
+        now_local = datetime.now()
+        today = now_local.strftime("%Y-%m-%d")
+        yesterday = (now_local - timedelta(days=1)).strftime("%Y-%m-%d")
+        paths = []
+        for d in [today, yesterday]:
+            p = os.path.join(log_dir, f"openclaw-{d}.log")
+            if os.path.exists(p):
+                paths.append(p)
+        return paths
+
+    @staticmethod
+    def _check_compact_in_progress_gateway(
+            session_key: str, window_seconds: int = 120) -> bool:
+        """§24 v3 rotation-only: 检查 gateway 日志，判断指定 session 是否刚完成 compact。
+
+        检测逻辑：读日志尾部 2MB，按目标 sessionKey 过滤，
+        找最后一个 rotation 事件，如果在窗口内 → compact 可能仍在 retry 循环中。
+        """
+        from datetime import datetime as _dt, timezone as _tz, timedelta
+        log_paths = AgentSpawner._get_recent_gateway_logs()
+        if not log_paths:
+            return False
+
+        now = _dt.now(_tz.utc)
+        window_start = now - timedelta(seconds=window_seconds)
+
+        last_rotation_time = None
+
+        for log_path in log_paths:
+            if not os.path.exists(log_path):
+                continue
+            try:
+                with open(log_path, "rb") as f:
+                    f.seek(0, 2)
+                    size = f.tell()
+                    f.seek(max(0, size - 2 * 1024 * 1024))
+                    tail = f.read().decode("utf-8", errors="replace")
+            except Exception:
+                continue
+
+            for line in tail.splitlines():
+                if not line.strip():
+                    continue
+                try:
+                    obj = json.loads(line)
+                except (json.JSONDecodeError, ValueError):
+                    continue
+
+                msg = obj.get("message", "")
+                # 只看包含目标 sessionKey 的事件
+                if session_key not in msg:
+                    continue
+
+                # rotation 事件
+                if "[compaction] rotated active transcript" in msg:
+                    ts_str = obj.get("time", "")
+                    if ts_str:
+                        try:
+                            event_time = _dt.fromisoformat(
+                                ts_str.replace("Z", "+00:00"))
+                            # timezone-aware: normalize to UTC
+                            if event_time.tzinfo is None:
+                                event_time = event_time.replace(tzinfo=_tz.utc)
+                            if last_rotation_time is None or event_time > last_rotation_time:
+                                last_rotation_time = event_time
+                        except (ValueError, TypeError):
+                            continue
+
+        if last_rotation_time is not None:
+            return last_rotation_time >= window_start
+
+        return False
+
+    @staticmethod
+    def _check_recent_compaction_jsonl(
+            session_file: str, window_seconds: int = 900) -> bool:
        """v2.8.2 Fix-2: 读 session jsonl 末尾,检查是否有 window_seconds 内的 compaction 记录。

        比 compactionCheckpoints 更可靠:Gateway 每次完成 compact 必然在 jsonl 末尾追加记录,
@@ -1219,7 +1391,7 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
        实测 50KB 在长对话中不够（compact 记录被推出窗口导致漏检）。
        正常扫描量不变：从尾部往前扫，遇到超过 15min 的 timestamp 即 break。
        """
-        if not session_file or not pathlib.Path(session_file).exists():
+        if not session_file or not Path(session_file).exists():
            return False
        try:
            from datetime import datetime, timezone
@@ -1241,7 +1413,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                    ts = obj.get("timestamp", "")
                    if ts:
                        try:
-                            ct = datetime.fromisoformat(ts.replace("Z", "+00:00"))
+                            ct = datetime.fromisoformat(
+                                ts.replace("Z", "+00:00"))
                            if (now - ct).total_seconds() < window_seconds:
                                return True
                        except (ValueError, TypeError):
@@ -1265,7 +1438,11 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
        v2.8.1: compact 检测改用 session jsonl 末尾扫描(Fix-1),
        替代失效的 compactionCheckpoints 检测。
        """
-        result = {"status": "unknown", "lock_pid": None, "lock_pid_alive": False, "recent_compact": False}
+        result = {
+            "status": "unknown",
+            "lock_pid": None,
+            "lock_pid_alive": False,
+            "recent_compact": False}
        sessions_path = Path(os.environ.get(
            "OPENCLAW_HOME", str(Path.home() / ".openclaw")
        )) / "agents" / agent_id / "sessions" / "sessions.json"
@@ -1304,8 +1481,10 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                                created_at_str = lock_data.get("createdAt", "")
                                if created_at_str:
                                    from datetime import datetime as _dt, timezone as _tz
-                                    created_dt = _dt.fromisoformat(created_at_str.replace("Z", "+00:00"))
-                                    elapsed = (_dt.now(_tz.utc) - created_dt).total_seconds()
+                                    created_dt = _dt.fromisoformat(
+                                        created_at_str.replace("Z", "+00:00"))
+                                    elapsed = (_dt.now(_tz.utc) -
+                                               created_dt).total_seconds()
                                    if elapsed > 1800:  # 30 minutes
                                        result["lock_pid_alive"] = False
                                        result["lock_expired"] = True
@@ -1316,10 +1495,17 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                    except Exception:
                        pass

-            # v2.8.1 Fix-1: compact 检测改用 session jsonl 末尾扫描
-            # 只在 agent 非空闲时才扫描(减少不必要 I/O)
-            if result["status"] not in ("done", "idle", "unknown", None) and sf:
-                result["recent_compact"] = AgentSpawner._check_recent_compaction_jsonl(sf)
+            # §24 v3: compact 检测优先用 gateway 日志 rotation 事件
+            # 旧方法 _check_recent_compaction_jsonl 作为 fallback
+            # 重要：compact 进行中时 status=done，所以不能按 status 过滤
+            # 只跳过 idle/unknown（完全没有活动过的 session）
+            if result["status"] not in ("idle", "unknown", None):
+                session_key = f"agent:{agent_id}:main"
+                result["recent_compact"] = AgentSpawner._check_compact_in_progress_gateway(
+                    session_key)
+                if not result["recent_compact"] and sf:
+                    result["recent_compact"] = AgentSpawner._check_recent_compaction_jsonl(
+                        sf)
        except Exception:
            pass
        return result
@@ -1364,45 +1550,53 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            # A15/A16: stderr 含 network/compact 关键字 → 可恢复
            if stderr_text:
                stderr_lower = stderr_text.lower()
-                if any(kw in stderr_lower for kw in ["econnrefused", "etimedout", "gateway closed", "econnreset"]):
+                if any(kw in stderr_lower for kw in [
+                       "econnrefused", "etimedout", "gateway closed", "econnreset"]):
                    return {"outcome": "gateway_unreachable", "should_retry": True,
                            "retry_field": "retry_count", "cooldown_seconds": 60}
-                if any(kw in stderr_lower for kw in ["compaction-diag", "context-overflow"]):
+                if any(kw in stderr_lower for kw in [
+                       "compaction-diag", "context-overflow"]):
                    return {"outcome": "compact_interrupted", "should_retry": True,
                            "retry_field": "retry_count", "cooldown_seconds": 60}
            # A17: 真正的 crash → 保持 working,ticker 兜底
-            return {"outcome": "crashed", "should_retry": False, "original": "process_crash"}
+            return {"outcome": "crashed", "should_retry": False,
+                    "original": "process_crash"}

-        # stdout 为空但 exit=0:可能是正常完成但 --json 没输出
-        # 查任务状态判断
+        # A13 revised: stdout 为空但 exit=0 → 信任进程退出码，视为正常完成
+        # 实测发现 openclaw session=None + exit=0 是正常场景（inform 通知等）
+        # 旧逻辑按 task_status 区分，非终态判 agent_error → 导致 inform 邮件永不标 done
        if status is None and not stdout_text.strip() and exit_code == 0:
-            terminal_statuses = {"done", "review"}
-            if task_status in terminal_statuses:
-                return {"outcome": "completed", "should_retry": False}
-            return {"outcome": "agent_error", "should_retry": False}
+            return {"outcome": "completed", "should_retry": False}

        # A7-A12: status=error → 不续杯,stderr 辅助分类
        if status == "error":
            stderr_lower = stderr_text.lower()
-            if any(kw in stderr_lower for kw in ["401", "403", "unauthorized", "auth"]):
+            if any(kw in stderr_lower for kw in [
+                   "401", "403", "unauthorized", "auth"]):
                return {"outcome": "auth_failed", "should_retry": False}
-            if any(kw in stderr_lower for kw in ["econnrefused", "etimedout", "gateway closed", "econnreset"]):
+            if any(kw in stderr_lower for kw in [
+                   "econnrefused", "etimedout", "gateway closed", "econnreset"]):
                return {"outcome": "gateway_unreachable", "should_retry": True,
                        "retry_field": "retry_count", "cooldown_seconds": 60}
-            if any(kw in stderr_lower for kw in ["rate_limit", "500", "503", "api error"]):
+            if any(kw in stderr_lower for kw in [
+                   "rate_limit", "500", "503", "api error"]):
                return {"outcome": "api_error", "should_retry": False}
-            if any(kw in stderr_lower for kw in ["compaction-diag", "context-overflow"]):
+            if any(kw in stderr_lower for kw in [
+                   "compaction-diag", "context-overflow"]):
                return {"outcome": "compact_failed", "should_retry": False}
-            if any(kw in stderr_lower for kw in ["lock", "busy", "concurrent", "lane task error"]):
+            if any(kw in stderr_lower for kw in [
+                   "lock", "busy", "concurrent", "lane task error"]):
                return {"outcome": "lock_conflict", "should_retry": True,
                        "retry_field": "retry_count", "cooldown_seconds": 60}
            return {"outcome": "agent_error", "should_retry": False}

        # 兜底:status 未知值
-        return {"outcome": "agent_error", "should_retry": False, "original": "unknown_status"}
+        return {"outcome": "agent_error",
+                "should_retry": False, "original": "unknown_status"}

    @staticmethod
-    def _get_retry_counts(db_path: Optional[Path], task_id: Optional[str]) -> dict:
+    def _get_retry_counts(
+            db_path: Optional[Path], task_id: Optional[str]) -> dict:
        """从最新 task_attempt 的 metadata 读计数器"""
        defaults = {"retry_count": 0, "connect_retry_count": 0,
                    "api_retry_count": 0, "lock_retry_count": 0,
@@ -1442,7 +1636,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                    (task_id,)
                ).fetchone()
                if row:
-                    meta = json.loads(row["metadata"]) if row["metadata"] else {}
+                    meta = json.loads(
+                        row["metadata"]) if row["metadata"] else {}
                    meta.update(counts)
                    conn.execute(
                        "UPDATE task_attempts SET metadata=? WHERE rowid=?",
@@ -1452,7 +1647,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            finally:
                conn.close()
        except Exception:
-            logger.exception("Failed to update retry counts for task %s", task_id)
+            logger.exception(
+                "Failed to update retry counts for task %s", task_id)

    def _mark_task(self, db_path: Optional[Path], task_id: Optional[str],
                   status: str, detail: Optional[dict] = None):
@@ -1470,7 +1666,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                if detail:
                    conn.execute(
                        "INSERT INTO events (task_id, agent, event_type, detail) VALUES (?,?,?,?)",
-                        (task_id, "daemon", status, json.dumps(detail, ensure_ascii=False))
+                        (task_id, "daemon", status, json.dumps(
+                            detail, ensure_ascii=False))
                    )
                conn.commit()
            finally:
@@ -1491,7 +1688,10 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
                                             f"@pangtong-fujunshi 任务执行失败: {reason}，请评估是否需要介入",
                                             comment_type="system")
                        bb.record_mentions(cid, task_id, ["pangtong-fujunshi"])
-                        logger.info("Task %s: failure notified pangtong via comment+mention (reason=%s)", task_id, reason)
+                        logger.info(
+                            "Task %s: failure notified pangtong via comment+mention (reason=%s)",
+                            task_id,
+                            reason)
                except Exception as e:
                    logger.warning("Task %s: failed to notify: %s", task_id, e)
        except Exception:
@@ -1520,7 +1720,10 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
            if asyncio.iscoroutine(result):
                await result
        except Exception:
-            logger.warning("on_complete callback failed for %s", agent_id, exc_info=True)
+            logger.warning(
+                "on_complete callback failed for %s",
+                agent_id,
+                exc_info=True)

    def _register_session(
        self,
@@ -1598,7 +1801,8 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
    def get_session_by_agent(self, agent_id: str) -> Optional[Dict[str, Any]]:
        """v2.7.2: 根据 agent_id 获取活跃 session 信息(用于进程存活性检查)"""
        for sid, info in self._sessions.items():
-            if info.get("agent_id") == agent_id and info.get("status") == "running":
+            if info.get("agent_id") == agent_id and info.get(
+                    "status") == "running":
                return info
        return None

@@ -49,7 +49,8 @@ class SSEEvent:
        """格式化为 SSE 协议文本"""
        lines = [f"id: {self.id}"]
        lines.append(f"event: {self.event_type}")
-        lines.append(f"data: {json.dumps(self.data, ensure_ascii=False, default=str)}")
+        lines.append(
+            f"data: {json.dumps(self.data, ensure_ascii=False, default=str)}")
        return "\n".join(lines) + "\n\n"


@@ -0,0 +1,387 @@
+"""task_handler.py — 黑板任务 handler（task_type='task'）。
+
+标准黑板任务：三信号验证 → review 状态。
+"""
+from __future__ import annotations
+
+import logging
+import os
+from pathlib import Path
+from typing import Dict, Optional
+
+from src.daemon.base_task_handler import BaseTaskHandler, VerifyResult
+from src.daemon.prompt_composer import PromptComposer, PromptContext
+from src.blackboard.db import get_connection
+
+logger = logging.getLogger("moziplus-v2.handler")
+
+TERMINAL_STATES = frozenset({"review", "done", "failed", "cancelled"})
+
+# ---------------------------------------------------------------------------
+# Role → Skill 映射（D8 决策：L2 只给索引+引导语，不注全文）
+# ---------------------------------------------------------------------------
+ROLE_SKILL_MAP: Dict[str, str] = {
+    "executor": "blackboard-executor",
+    "reviewer": "blackboard-reviewer",
+    "reviewer-simayi": "blackboard-reviewer-simayi",
+    "reviewer-pangtong": "blackboard-reviewer-pangtong",
+    "planner": "blackboard-planner",
+    "claim": "blackboard-claim",
+}
+
+SKILL_BASE_PATH = os.environ.get(
+    "MOZI_SKILL_PATH",
+    "/Users/chufeng/.sanguo_projects/sanguo_mozi/skills",
+)
+
+
+# ---------------------------------------------------------------------------
+# PromptSection 实现
+# ---------------------------------------------------------------------------
+
+class TaskContextSection:
+    """段 1：任务上下文（title / desc / must_haves / status）。"""
+
+    name: str = "task_context"
+    priority: int = 10
+
+    def render(self, context: PromptContext) -> str:
+        parts = ["## 任务上下文"]
+        if context.task_id:
+            parts.append(f"任务ID: {context.task_id}")
+        if context.title:
+            parts.append(f"标题: {context.title}")
+        if context.description:
+            parts.append(f"描述: {context.description}")
+        if context.must_haves:
+            parts.append(f"必须完成: {context.must_haves}")
+        if context.task and context.task.get("status"):
+            parts.append(f"当前状态: {context.task['status']}")
+        return "\n".join(parts)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return bool(context.task_id or context.title)
+
+
+class PriorOutputsSection:
+    """段 2：前序产出摘要（depends_on 非空时注入）。"""
+
+    name: str = "prior_outputs"
+    priority: int = 20
+
+    def render(self, context: PromptContext) -> str:
+        outputs = context.depends_on_outputs or []
+        parts = ["## 前序产出"]
+        for out in outputs:
+            tid = out.get("task_id", "?")
+            summary = out.get("summary", "无摘要")
+            parts.append(f"- [{tid}] {summary}")
+        return "\n".join(parts)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return bool(context.depends_on_outputs)
+
+
+class RoleSkillSection:
+    """段 3：角色 Skill 全文注入（对齐设计 §2.3 + BootstrapBuilder 行为）。"""
+
+    name: str = "role_skill"
+    priority: int = 30
+
+    def render(self, context: PromptContext) -> str:
+        skill_name = ROLE_SKILL_MAP.get(context.role, "")
+        lines = [
+            "## 角色操作规范",
+            f"你的角色：{context.role}",
+        ]
+        if skill_name:
+            skill_path = os.path.join(SKILL_BASE_PATH, skill_name, "SKILL.md")
+            try:
+                with open(skill_path, encoding="utf-8") as f:
+                    skill_content = f.read()
+                if skill_content:
+                    lines.append(skill_content)
+                else:
+                    lines.append(f"（Skill 文件为空：{skill_name}）")
+            except FileNotFoundError:
+                lines.append(f"（Skill 文件不存在：{skill_name}）")
+        else:
+            lines.append("无对应 Skill 文件，按通用规范执行。")
+        return "\n".join(lines)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return True
+
+
+class TaskApiSection:
+    """段 4：API 操作指令。"""
+
+    name: str = "task_api"
+    priority: int = 40
+
+    API_HOST = "localhost"
+    API_PORT = 8083
+
+    def render(self, context: PromptContext) -> str:
+        pid = context.project_id
+        tid = context.task_id
+        aid = context.agent_id
+        success_status = '"review"'
+        base = f"http://{self.API_HOST}:{self.API_PORT}/api/projects/{pid}/tasks/{tid}"
+        return (
+            "## 操作指令\n"
+            "### 状态回写\n"
+            f"开始工作:\n"
+            f'curl -X POST {base}/status \\\n'
+            f'  -H "Content-Type: application/json" \\\n'
+            f'  -d \'{{"status": "working", "agent": "{aid}"}}\'\n\n'
+            "### 写入产出\n"
+            f'curl -X POST {base}/outputs \\\n'
+            f'  -H "Content-Type: application/json" \\\n'
+            f"  -d '{{\"type\": \"text\", \"content\": \"<your output>\"}}'\n\n"
+            "### 完成后\n"
+            f"成功: status → {success_status} | 失败: status → \"failed\""
+        )
+
+    def should_include(self, context: PromptContext) -> bool:
+        return True
+
+
+class TaskConstraintsSection:
+    """段 5：硬约束。"""
+
+    name: str = "task_constraints"
+    priority: int = 50
+
+    def render(self, context: PromptContext) -> str:
+        constraints = ["## 硬约束"]
+        role = context.role
+        if role == "executor":
+            constraints.extend([
+                "- 完成后必须标 review",
+                "- 产出物不能为空（系统会验证）",
+                "- handoff comment ≥ 50 字符",
+            ])
+        elif role.startswith("reviewer"):
+            constraints.extend([
+                "- 审查结果必须明确 pass/fail",
+                "- 评审意见须附证据（文件:行号）",
+            ])
+        elif role == "planner":
+            constraints.extend([
+                "- 需求不清时提问，不要猜",
+                "- 子任务必须有明确的终态定义",
+            ])
+        else:
+            constraints.append("- 按规范完成 assigned 任务")
+        return "\n".join(constraints)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return True
+
+
+class TaskHandler(BaseTaskHandler):
+    """黑板标准任务 handler。
+
+    - verify: 三信号检查（output / comment / terminal status）
+    - 成功 → review
+    - 失败 → 保持 working，让 ticker 重试
+    - review 完成 → 读取 verdict，approved 则 mark done
+    """
+
+    task_type: str = "task"
+    virtual_project: Optional[str] = None
+    display_name = "黑板任务"
+
+    # === 子类实现 ===
+
+    def post_complete(self, task_id: str, agent_id: str,
+                      outcome: str, db_path: Path) -> None:
+        """Task on_complete：区分 executor 和 review。
+
+        executor: 基类统一流程（crash → verify → mark review）
+        review: handle_review_complete（读 verdict → done/keep review）
+        """
+        # crash 处理（所有类型共用）
+        if outcome in self.CRASH_OUTCOMES:
+            self._rollback_current_agent(db_path, task_id, agent_id)
+            return
+
+        # 检查当前任务状态：如果是 review 状态 → review 完成流程
+        try:
+            conn = get_connection(db_path)
+            try:
+                row = conn.execute(
+                    "SELECT status FROM tasks WHERE id=?", (task_id,)
+                ).fetchone()
+                task_status = row["status"] if row else "unknown"
+            finally:
+                conn.close()
+        except Exception:
+            task_status = "unknown"
+
+        if task_status == "review":
+            # review 完成流程：只处理正常 outcome
+            if outcome in ("completed", "session_revived"):
+                self.handle_review_complete(task_id, db_path)
+            else:
+                logger.warning(
+                    "Task %s: review agent %s abnormal outcome=%s, keeping review",
+                    task_id, agent_id, outcome)
+        else:
+            # executor 完成流程：基类统一 verify → mark
+            result = self.verify_completion(task_id, db_path)
+            if result.passed:
+                self._mark_task_status(db_path, task_id, self.target_success_status())
+                logger.info("Task %s: verify passed (%s), marked %s",
+                            task_id, result.reason, self.target_success_status())
+            else:
+                logger.info(
+                    "Task %s: verify not passed (%s), leaving working",
+                    task_id, result.reason)
+                # NOTE: executor verify 不通过时不标 failed，留 working。
+                # 原因：Agent 可能还在产出中（幻觉门控的后续轮次），
+                # ticker 超时检查会兜底处理。不调 on_failure 避免误判。
+
+    def target_success_status(self) -> str:
+        """task 类型验证通过后进 review。"""
+        return "review"
+
+    def verify_completion(self, task_id: str, db_path: Path) -> VerifyResult:
+        """三信号验证：output / comment / terminal status。"""
+        try:
+            conn = get_connection(db_path)
+            try:
+                # 信号 1：terminal status
+                row = conn.execute(
+                    "SELECT status FROM tasks WHERE id=?", (task_id,)
+                ).fetchone()
+                if not row:
+                    return VerifyResult(False, "not_found", "task not found",
+                                        can_retry=False)
+                status = row["status"]
+                if status in TERMINAL_STATES:
+                    return VerifyResult(
+                        True, "terminal_status",
+                        f"status={status}", can_retry=False
+                    )
+
+                # 信号 2：outputs
+                output_count = conn.execute(
+                    "SELECT COUNT(*) as cnt FROM outputs WHERE task_id=?",
+                    (task_id,)
+                ).fetchone()["cnt"]
+                if output_count > 0:
+                    return VerifyResult(
+                        True, "has_output",
+                        f"output_count={output_count}"
+                    )
+
+                # 信号 3：非 system 且内容 >= 50 字的 comment
+                comment_count = conn.execute(
+                    "SELECT COUNT(*) as cnt FROM comments "
+                    "WHERE task_id=? AND author != 'system' "
+                    "AND LENGTH(content) >= 50",
+                    (task_id,)
+                ).fetchone()["cnt"]
+                if comment_count > 0:
+                    return VerifyResult(
+                        True, "has_comment",
+                        f"comment_count={comment_count}"
+                    )
+
+                # 无信号
+                return VerifyResult(
+                    False, "no_signal",
+                    f"output=0, comment=0, status={status}"
+                )
+            finally:
+                conn.close()
+        except Exception as e:
+            logger.error("Task %s: verify error: %s", task_id, e)
+            return VerifyResult(False, "verify_error", str(e))
+
+    def pre_spawn(self, task_id: str, db_path: Path) -> bool:
+        """task 类型不需要 pre_spawn 逻辑。"""
+        return True
+
+    def get_sections(self) -> list:
+        """返回 5 个 PromptSection 实例。"""
+        return [
+            TaskContextSection(),
+            PriorOutputsSection(),
+            RoleSkillSection(),
+            TaskApiSection(),
+            TaskConstraintsSection(),
+        ]
+
+    def build_prompt(self, context: PromptContext) -> str:
+        """通过 PromptComposer 拼装 prompt sections。"""
+        composer = PromptComposer()
+        composer.add_many(self.get_sections())
+        return composer.compose(context)
+
+    def on_failure(self, task_id: str, agent_id: str,
+                   db_path: Path, verify: VerifyResult) -> None:
+        """验证失败：不标 failed，保持 working 让 ticker 重试。"""
+        logger.info(
+            "Task %s: verify failed (%s, evidence=%s), leaving working for ticker retry",
+            task_id, verify.reason, verify.evidence
+        )
+
+    # === Review 流程 ===
+
+    def handle_review_complete(self, task_id: str, db_path: Path) -> None:
+        """Review 完成后处理：读取 verdict → approved 则 mark done，
+        否则 @mention assignee via blackboard comment。"""
+        try:
+            conn = get_connection(db_path)
+            try:
+                # 读取最新 review
+                review_row = conn.execute(
+                    "SELECT verdict, reviewer, comment FROM reviews "
+                    "WHERE task_id=? ORDER BY created_at DESC LIMIT 1",
+                    (task_id,)
+                ).fetchone()
+
+                if not review_row:
+                    logger.warning("Task %s: no review found", task_id)
+                    return
+
+                verdict = review_row["verdict"]
+                reviewer = review_row["reviewer"]
+                review_comment = review_row["comment"] or ""
+
+                # 获取 assignee
+                task_row = conn.execute(
+                    "SELECT assignee FROM tasks WHERE id=?", (task_id,)
+                ).fetchone()
+                if not task_row:
+                    logger.warning("Task %s: task not found for review", task_id)
+                    return
+                assignee = task_row["assignee"]
+
+                if verdict == "approved":
+                    self._mark_task_status(db_path, task_id, "done")
+                    logger.info("Task %s: review approved by %s, marked done",
+                                task_id, reviewer)
+                else:
+                    # 非 approved：通过 blackboard comment @mention assignee
+                    # 保持 review 状态，让 assignee 自行决定下一步
+                    conn.execute(
+                        "INSERT INTO comments (task_id, author, content, comment_type) "
+                        "VALUES (?, 'system', ?, 'review')",
+                        (task_id,
+                         f"@{assignee} review 未通过 (verdict={verdict}, "
+                         f"reviewer={reviewer}): {review_comment}")
+                    )
+                    conn.commit()
+                    logger.info(
+                        "Task %s: review not approved (%s by %s), "
+                        "@mentioned assignee %s, keeping review status",
+                        task_id, verdict, reviewer, assignee
+                    )
+            finally:
+                conn.close()
+        except Exception as e:
+            logger.error("Task %s: handle_review_complete error: %s", task_id, e)
@@ -0,0 +1,102 @@
+"""
+task_type_registry.py — Task type handler Protocol + Registry.
+
+启动时一次性加载 handler，运行时只读。
+零依赖：不导入项目内其他模块。
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import TYPE_CHECKING, Dict, Optional, Protocol, runtime_checkable
+
+if TYPE_CHECKING:
+    from src.daemon.prompt_composer import PromptContext
+
+logger = logging.getLogger("moziplus-v2.registry")
+
+
+# ---------------------------------------------------------------------------
+# Protocol
+# ---------------------------------------------------------------------------
+
+@runtime_checkable
+class TaskTypeHandler(Protocol):
+    """所有 task type handler 的统一接口。"""
+
+    # 属性（通过 __init__ 设置）
+    task_type: str  # 类型标识：'task' | 'mail' | 'toolchain'
+    virtual_project: Optional[str]  # 虚拟项目 ID，如 '_mail'、'_toolchain'。普通任务为 None
+
+    def build_prompt(self, context: "PromptContext") -> str:
+        """构建 Agent prompt（通过 PromptComposer 拼 section）。"""
+        ...
+
+    def pre_spawn(self, task_id: str, db_path: Path) -> bool:
+        """spawn 前业务准备。默认 True，mail/toolchain override 为 auto_working。"""
+        ...
+
+    def post_complete(
+        self,
+        task_id: str,
+        agent_id: str,
+        outcome: str,
+        db_path: Path,
+    ) -> None:
+        """spawn 完成后的业务处理。统一流程：crash→verify→mark→notify。"""
+        ...
+
+    def check_completion(self, task_id: str, db_path: Path) -> bool:
+        """ticker 级别的完成检查。"""
+        ...
+
+    def get_sections(self) -> list:
+        """返回此 handler 的 prompt section 列表。"""
+        ...
+
+
+# ---------------------------------------------------------------------------
+# Registry
+# ---------------------------------------------------------------------------
+
+class TaskTypeRegistry:
+    """Task type handler 注册表。启动时一次性加载，运行时只读。"""
+
+    _handlers: Dict[str, TaskTypeHandler] = {}
+
+    @classmethod
+    def register(cls, handler: TaskTypeHandler) -> None:
+        """注册一个 handler。启动时调用一次。"""
+        if handler.task_type in cls._handlers:
+            raise ValueError(f"Task type '{handler.task_type}' already registered")
+        cls._handlers[handler.task_type] = handler
+        vp = getattr(handler, "virtual_project", None)
+        logger.info("Registered task type handler: %s (virtual_project=%s)", handler.task_type, vp)
+
+    @classmethod
+    def get_by_project(cls, project_id: str) -> Optional[TaskTypeHandler]:
+        """通过 project_id 查找 handler（匹配 virtual_project）。"""
+        for h in cls._handlers.values():
+            if h.virtual_project == project_id:
+                return h
+        return None
+
+    @classmethod
+    def get(cls, task_type: str) -> Optional[TaskTypeHandler]:
+        """通过 task_type 标识查找 handler。"""
+        return cls._handlers.get(task_type)
+
+    @classmethod
+    def virtual_projects(cls) -> list[str]:
+        """返回所有已注册的虚拟项目 ID（ticker 自动发现用）。"""
+        return [
+            h.virtual_project
+            for h in cls._handlers.values()
+            if h.virtual_project is not None
+        ]
+
+    @classmethod
+    def clear(cls) -> None:
+        """清空注册表（仅测试用）。"""
+        cls._handlers = {}
@@ -19,6 +19,8 @@ from typing import Any, Callable, Coroutine, Dict, List, Optional

 from dataclasses import dataclass, field as dc_field

+from src.daemon.task_type_registry import TaskTypeRegistry
+
 from src.blackboard.operations import Blackboard
 from src.blackboard.db import get_connection
 from src.daemon.spawner import AgentBusyError
@@ -31,7 +33,8 @@ class BroadcastRound:
    """追踪单个任务的广播状态"""
    task_id: str
    notified_agents: set = dc_field(default_factory=set)   # 已 spawn 过的 Agent
-    responded_agents: set = dc_field(default_factory=set)   # 已返回反馈的 Agent（含 NO_REPLY）
+    responded_agents: set = dc_field(
+        default_factory=set)   # 已返回反馈的 Agent（含 NO_REPLY）
    round_number: int = 0                                   # 当前第几轮（0=未开始，1=第1轮）


@@ -46,7 +49,8 @@ class Ticker:
        registry: ProjectRegistry,
        tick_interval: float = 30.0,
        max_ticks: Optional[int] = None,
-        on_tick_complete: Optional[Callable[[], Coroutine[Any, Any, None]]] = None,
+        on_tick_complete: Optional[Callable[[],
+                                            Coroutine[Any, Any, None]]] = None,
        dispatcher: Optional[Any] = None,
        spawner: Optional[Any] = None,
        max_dispatch_per_tick: int = 3,
@@ -194,7 +198,10 @@ class Ticker:
                pr = await self._tick_project(project_id, project_info)
                results["projects"][project_id] = pr
            except Exception as e:
-                logger.exception("Tick %d project %s error", tick_num, project_id)
+                logger.exception(
+                    "Tick %d project %s error",
+                    tick_num,
+                    project_id)
                results["projects"][project_id] = {"error": str(e)}

        # 虚拟项目 _general：不在 registry 但需要调度
@@ -210,20 +217,26 @@ class Ticker:
                logger.exception("Tick %d _general error", tick_num)
                results["projects"]["_general"] = {"error": str(e)}

-        # 虚拟项目 _mail：飞鸽传书
-        mail_db = Path(self.registry.root) / "_mail" / "blackboard.db"
-        if mail_db.exists() and "_mail" not in active_projects:
-            try:
-                pr = await self._tick_project("_mail", {
-                    "id": "_mail", "name": "飞鸽传书",
-                    "status": "active", "source": "virtual",
-                })
-                results["projects"]["_mail"] = pr
-            except Exception as e:
-                logger.exception("Tick %d _mail error", tick_num)
-                results["projects"]["_mail"] = {"error": str(e)}
+        # 虚拟项目：从注册表自动发现 + _general 硬编码
+        for vp in TaskTypeRegistry.virtual_projects():
+            vp_db = Path(self.registry.root) / vp / "blackboard.db"
+            if vp_db.exists() and vp not in active_projects:
+                try:
+                    vp_handler = TaskTypeRegistry.get_by_project(vp)
+                    vp_name = vp_handler.display_name if vp_handler and vp_handler.display_name else vp
+                    pr = await self._tick_project(vp, {
+                        "id": vp, "name": vp_name,
+                        "status": "active", "source": "virtual",
+                    })
+                    results["projects"][vp] = pr
+                except Exception as e:
+                    logger.exception("Tick %d %s error", tick_num, vp)
+                    results["projects"][vp] = {"error": str(e)}

-        logger.debug("Tick %d complete: %d projects", tick_num, len(active_projects))
+        logger.debug(
+            "Tick %d complete: %d projects",
+            tick_num,
+            len(active_projects))

        if self.on_tick_complete:
            try:
@@ -314,7 +327,8 @@ class Ticker:
        # 8. 健康检查（僵尸检测）
        if self.health_checker:
            try:
-                self.health_checker.check(project_id, db_path, self._tick_count)
+                self.health_checker.check(
+                    project_id, db_path, self._tick_count)
            except Exception as e:
                logger.warning("HealthChecker error for %s: %s", project_id, e)

@@ -335,7 +349,8 @@ class Ticker:
                            task_id=t.id, task_title=t.title, task_type=t.task_type
                        )
            except Exception as e:
-                logger.warning("ExperienceDistiller error for %s: %s", project_id, e)
+                logger.warning(
+                    "ExperienceDistiller error for %s: %s", project_id, e)

        # 10. 扫描后状态
        result["summary_after"] = queries.task_summary()
@@ -375,7 +390,8 @@ class Ticker:
                        (computed, pid),
                    )
                    refreshed.append(pid)
-                    logger.info("Parent %s status aggregated: → %s", pid, computed)
+                    logger.info(
+                        "Parent %s status aggregated: → %s", pid, computed)

            if refreshed:
                conn.commit()
@@ -543,6 +559,7 @@ Parent Task ID: {parent_task.id}
        """
        try:
            agent_id = "pangtong-fujunshi"
+            f"review-{parent_task.id}-r{new_round}"

            # 构造 on_complete 回调：解析庞统结论，更新 parent 状态
            async def _on_review_complete(aid: str, outcome: str):
@@ -554,7 +571,8 @@ Parent Task ID: {parent_task.id}
                        latest_meta = None
                        latest_time = ""
                        for sid, sess in self.spawner._sessions.items():
-                            if sess.get("agent_id") == agent_id and sess.get("meta"):
+                            if sess.get(
+                                    "agent_id") == agent_id and sess.get("meta"):
                                t = sess.get("completed_at", "")
                                if t > latest_time:
                                    latest_time = t
@@ -586,7 +604,9 @@ Parent Task ID: {parent_task.id}
                return True
            return False
        except Exception:
-            logger.exception("Failed to spawn pangtong review for %s", parent_task.id)
+            logger.exception(
+                "Failed to spawn pangtong review for %s",
+                parent_task.id)
            return False

    def _set_parent_reviewing(self, parent_id: str, project_id: str):
@@ -618,7 +638,8 @@ Parent Task ID: {parent_task.id}
        conn = get_connection(db_path)
        try:
            # 解析 GOAL_ACHIEVED
-            is_achieved = bool(review_text and "GOAL_ACHIEVED" in review_text.upper())
+            is_achieved = bool(
+                review_text and "GOAL_ACHIEVED" in review_text.upper())

            if is_achieved:
                # Goal 达成 → parent 最终完成
@@ -648,7 +669,9 @@ Parent Task ID: {parent_task.id}
                    "(round %d, subs=%d)",
                    parent_id, round_num, sub_count)
        except Exception:
-            logger.exception("Failed to handle review conclusion for %s", parent_id)
+            logger.exception(
+                "Failed to handle review conclusion for %s",
+                parent_id)
            # 安全恢复：reviewing → working
            try:
                conn.execute("BEGIN IMMEDIATE")
@@ -686,7 +709,8 @@ Parent Task ID: {parent_task.id}
            return []

        bb = Blackboard(db_path)
-        mentions = bb.get_pending_mentions(max_retries=self.MENTION_MAX_RETRIES)
+        mentions = bb.get_pending_mentions(
+            max_retries=self.MENTION_MAX_RETRIES)
        if not mentions:
            return []

@@ -750,16 +774,19 @@ Parent Task ID: {parent_task.id}

                                if new_review and new_review["verdict"] == "approved":
                                    _ticker._transition_status(
-                                        get_connection(rdb_path), _t_id, "done",
+                                        get_connection(
+                                            rdb_path), _t_id, "done",
                                        agent="daemon",
                                        detail={"reason": "rebuttal_approved"})
-                                    logger.info("Rebuttal: task %s approved after rebuttal", _t_id)
+                                    logger.info(
+                                        "Rebuttal: task %s approved after rebuttal", _t_id)
                                else:
                                    # 仍非 approved → @mention assignee
                                    verdict_str = new_review["verdict"] if new_review else "未知"
                                    rconn2 = get_connection(rdb_path)
                                    try:
-                                        t_row = rconn2.execute("SELECT assignee FROM tasks WHERE id=?", (_t_id,)).fetchone()
+                                        t_row = rconn2.execute(
+                                            "SELECT assignee FROM tasks WHERE id=?", (_t_id,)).fetchone()
                                    finally:
                                        rconn2.close()
                                    if t_row and t_row["assignee"]:
@@ -768,9 +795,11 @@ Parent Task ID: {parent_task.id}
                                        bb2.add_comment(_t_id, "daemon",
                                                        f"@{t_row['assignee']} 审查结论: {verdict_str}，请查看详情并决定接受或反驳",
                                                        comment_type="review")
-                                    logger.info("Rebuttal: task %s still %s after rebuttal", _t_id, verdict_str)
+                                    logger.info(
+                                        "Rebuttal: task %s still %s after rebuttal", _t_id, verdict_str)
                        except Exception:
-                            logger.exception("Rebuttal on_complete failed for task %s", _t_id)
+                            logger.exception(
+                                "Rebuttal on_complete failed for task %s", _t_id)

                    result = await self.spawner.spawn_full_agent(
                        agent_id=agent_id,
@@ -793,22 +822,30 @@ Parent Task ID: {parent_task.id}
                    for item in items:
                        bb.mark_mention_notified(item["id"])
                    processed.append(agent_id)
-                    logger.info("Mention spawn success: %s (%d mentions)", agent_id, len(items))
+                    logger.info(
+                        "Mention spawn success: %s (%d mentions)",
+                        agent_id,
+                        len(items))
                else:
                    # spawn 返回 None（其他原因）→ 递增 retry_count
                    for item in items:
                        bb.mark_mention_retry(item["id"])
-                    logger.warning("Mention spawn failed: %s, retrying next tick", agent_id)
+                    logger.warning(
+                        "Mention spawn failed: %s, retrying next tick", agent_id)

            except AgentBusyError:
                # Agent 忙，不递增 retry_count，等下次 tick 自然重试
-                logger.info("Mention spawn skipped: %s busy, will retry next tick", agent_id)
+                logger.info(
+                    "Mention spawn skipped: %s busy, will retry next tick",
+                    agent_id)

            except Exception:
-                logger.exception("Mention processing error for agent %s", agent_id)
+                logger.exception(
+                    "Mention processing error for agent %s", agent_id)
                for item in items:
                    try:
-                        if item.get("retry_count", 0) >= self.MENTION_MAX_RETRIES - 1:
+                        if item.get("retry_count",
+                                    0) >= self.MENTION_MAX_RETRIES - 1:
                            bb.mark_mention_failed(item["id"])
                        else:
                            bb.mark_mention_retry(item["id"])
@@ -821,8 +858,14 @@ Parent Task ID: {parent_task.id}
                              mention_lines: List[str],
                              project_id: str) -> str:
        """#03: @mention prompt（身份注入）"""
-        api_host = getattr(self.spawner, 'api_host', '127.0.0.1') if self.spawner else '127.0.0.1'
-        api_port = getattr(self.spawner, 'api_port', 8083) if self.spawner else 8083
+        api_host = getattr(
+            self.spawner,
+            'api_host',
+            '127.0.0.1') if self.spawner else '127.0.0.1'
+        api_port = getattr(
+            self.spawner,
+            'api_port',
+            8083) if self.spawner else 8083
        api_base = f"http://{api_host}:{api_port}/api"

        # 获取 Agent 专长
@@ -898,7 +941,8 @@ Parent Task ID: {parent_task.id}
        from datetime import datetime

        conn.execute("BEGIN IMMEDIATE")
-        row = conn.execute("SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
+        row = conn.execute(
+            "SELECT status FROM tasks WHERE id=?", (task_id,)).fetchone()
        if not row:
            return False
        old_status = row["status"]
@@ -909,9 +953,10 @@ Parent Task ID: {parent_task.id}

        now = datetime.utcnow().isoformat()
        # 重置到 pending 时清空 assignee（避免残留导致重复路由到同一 Agent）
-        # 但 Mail 的 assignee 是收件人，永不清空
+        # handler 虚拟项目（_mail 等）的 assignee 是收件人，永不清空
        if new_status == "pending":
-            if self._current_project_id == "_mail":
+            handler = TaskTypeRegistry.get_by_project(self._current_project_id)
+            if handler:
                conn.execute(
                    "UPDATE tasks SET status=?, updated_at=? WHERE id=?",
                    (new_status, now, task_id),
@@ -937,7 +982,8 @@ Parent Task ID: {parent_task.id}
            event_type = "daemon_tick"
        conn.execute(
            "INSERT INTO events (task_id, agent, event_type, detail) VALUES (?,?,?,?)",
-            (task_id, agent, event_type, json.dumps({"from": old_status, "to": new_status, **(detail or {})})),
+            (task_id, agent, event_type, json.dumps(
+                {"from": old_status, "to": new_status, **(detail or {})})),
        )
        conn.commit()
        return True
@@ -977,20 +1023,24 @@ Parent Task ID: {parent_task.id}
            try:
                result = await self.dispatcher.dispatch(
                    task,
-                    project_config={"project_id": project_id, "db_path": db_path},
+                    project_config={
+                        "project_id": project_id,
+                        "db_path": db_path},
                )
-                if result["status"] == "dispatched" and result["level"] in ("full", "escalate"):
+                if result["status"] == "dispatched" and result["level"] in (
+                        "full", "escalate"):
                    conn = get_connection(db_path)
                    try:
-                        # [v2.7.1] Mail 已在 dispatcher 中标 working，跳过 claimed
-                        if project_id == "_mail":
+                        # [Step 5] handler 项目已在 dispatcher 中标 working，跳过 claimed
+                        handler = TaskTypeRegistry.get_by_project(project_id)
+                        if handler:
                            conn.execute(
                                "UPDATE tasks SET current_agent=? WHERE id=?",
                                (result["agent_id"], task.id),
                            )
                            conn.commit()
                            dispatched.append(task.id)
-                            logger.info("Dispatched %s to %s (session=%s, mail auto-working)",
+                            logger.info("Dispatched %s to %s (session=%s, handler auto-working)",
                                        task.id, result["agent_id"],
                                        result.get("session_id"))
                        else:
@@ -1072,7 +1122,8 @@ Parent Task ID: {parent_task.id}
                    detail={"reason": "no_taker_after_3_broadcasts",
                            "round_number": self._broadcast_tracker.get(t.id).round_number if self._broadcast_tracker.get(t.id) else 0},
                )
-                logger.warning("Escalated %s: no taker after 3 broadcast rounds", t.id)
+                logger.warning(
+                    "Escalated %s: no taker after 3 broadcast rounds", t.id)
                self._broadcast_tracker.pop(t.id, None)
            finally:
                conn.close()
@@ -1082,7 +1133,8 @@ Parent Task ID: {parent_task.id}

        idle_agents = self._get_idle_agents()
        if not idle_agents:
-            logger.warning("No idle agents for broadcast, skipping (capacity issue)")
+            logger.warning(
+                "No idle agents for broadcast, skipping (capacity issue)")
            return []

        task_ids = [t.id for t in broadcastable]
@@ -1113,7 +1165,8 @@ Parent Task ID: {parent_task.id}

        spawned = []
        for agent_id in idle_agents:
-            prompt = self._build_claim_prompt(agent_id, broadcastable, project_id)
+            prompt = self._build_claim_prompt(
+                agent_id, broadcastable, project_id)
            try:
                session_id = await self.spawner.spawn_full_agent(
                    agent_id=agent_id,
@@ -1127,7 +1180,8 @@ Parent Task ID: {parent_task.id}
                    spawned.append(session_id)
                    # 记录已通知的 Agent
                    for t in broadcastable:
-                        self._broadcast_tracker[t.id].notified_agents.add(agent_id)
+                        self._broadcast_tracker[t.id].notified_agents.add(
+                            agent_id)
            except AgentBusyError:
                logger.debug("Broadcast skip %s: busy", agent_id)
            except Exception:
@@ -1138,8 +1192,14 @@ Parent Task ID: {parent_task.id}
    def _build_claim_prompt(self, agent_id: str, tasks: list,
                            project_id: str) -> str:
        """#03: 广播认领 prompt（身份+专长注入）"""
-        api_host = getattr(self.spawner, 'api_host', '127.0.0.1') if self.spawner else '127.0.0.1'
-        api_port = getattr(self.spawner, 'api_port', 8083) if self.spawner else 8083
+        api_host = getattr(
+            self.spawner,
+            'api_host',
+            '127.0.0.1') if self.spawner else '127.0.0.1'
+        api_port = getattr(
+            self.spawner,
+            'api_port',
+            8083) if self.spawner else 8083
        api_base = f"http://{api_host}:{api_port}/api"

        # 获取 Agent 专长
@@ -1194,7 +1254,8 @@ Parent Task ID: {parent_task.id}
    @property
    def counter(self):
        """从 Dispatcher 获取 counter"""
-        return getattr(self.dispatcher, 'counter', None) if self.dispatcher else None
+        return getattr(self.dispatcher, 'counter',
+                       None) if self.dispatcher else None

    @staticmethod
    def _is_pid_alive(pid: int) -> bool:
@@ -1206,7 +1267,8 @@ Parent Task ID: {parent_task.id}
        except (ProcessLookupError, PermissionError):
            return False

-    def record_broadcast_response(self, task_id: str, agent_id: str, outcome: str):
+    def record_broadcast_response(
+            self, task_id: str, agent_id: str, outcome: str):
        """记录 Agent 对广播任务的反馈（Spawner 调用的公共 API）"""
        tracker = self._broadcast_tracker.get(task_id)
        if not tracker:
@@ -1227,7 +1289,8 @@ Parent Task ID: {parent_task.id}

    def _get_all_agent_ids(self) -> List[str]:
        """获取所有配置的 Agent ID"""
-        if self.dispatcher and hasattr(self.dispatcher, 'router') and self.dispatcher.router:
+        if self.dispatcher and hasattr(
+                self.dispatcher, 'router') and self.dispatcher.router:
            return list(self.dispatcher.router.agent_profiles.keys())
        return []

@@ -1236,15 +1299,17 @@ Parent Task ID: {parent_task.id}
        if not self.counter:
            return []
        # agent_profiles 在 Router 初始化时从 config 填充，是完整 Agent 列表
-        all_agents = list(self.dispatcher.router.agent_profiles.keys()) if self.dispatcher else []
+        all_agents = list(
+            self.dispatcher.router.agent_profiles.keys()) if self.dispatcher else []
        active = self.counter.active_agents
        return [aid for aid in all_agents if active.get(aid, 0) == 0]

    async def _dispatch_reviews(self, db_path: Path,
                                project_id: str) -> List[str]:
        """扫描 review 状态任务，检查是否有产出，调度审查 Agent"""
-        # mail 任务不走 review 流程，直接跳过
-        if project_id == "_mail":
+        # handler 项目（_mail/_toolchain）不走 review 流程
+        handler = TaskTypeRegistry.get_by_project(project_id)
+        if handler:
            return []

        queries = Queries(db_path)
@@ -1290,7 +1355,9 @@ Parent Task ID: {parent_task.id}
                result = await self.dispatcher.dispatch(
                    task,
                    action_type="review",
-                    project_config={"project_id": project_id, "db_path": db_path},
+                    project_config={
+                        "project_id": project_id,
+                        "db_path": db_path},
                )
                if result["status"] == "dispatched":
                    dispatched.append(task.id)
@@ -1374,8 +1441,10 @@ Parent Task ID: {parent_task.id}
        working = queries.tasks_by_status("working")
        for task in working:
            # #07.2: crash_limit 统一检查（比超时更严重的信号）
-            if self.dispatcher and hasattr(self.dispatcher, '_check_crash_limit'):
-                if self.dispatcher._check_crash_limit(task.id, db_path, limit=3, window_minutes=30):
+            if self.dispatcher and hasattr(
+                    self.dispatcher, '_check_crash_limit'):
+                if self.dispatcher._check_crash_limit(
+                        task.id, db_path, limit=3, window_minutes=30):
                    conn = get_connection(db_path)
                    try:
                        self._transition_status(
@@ -1387,7 +1456,8 @@ Parent Task ID: {parent_task.id}
                    finally:
                        conn.close()
                    reclaimed.append(task.id)
-                    logger.error("Task %s: executor crash limit (3/30m), marking failed", task.id)
+                    logger.error(
+                        "Task %s: executor crash limit (3/30m), marking failed", task.id)
                    continue

            # #07.3 ACT-1: updated_at fallback 覆盖 mail auto-working（无 started_at/claimed_at）
@@ -1399,7 +1469,8 @@ Parent Task ID: {parent_task.id}
                # per-task timeout: deadline 优先，否则用默认值
                if task.deadline:
                    deadline_time = datetime.fromisoformat(task.deadline)
-                    timeout_minutes = (deadline_time - start_time).total_seconds() / 60.0
+                    timeout_minutes = (
+                        deadline_time - start_time).total_seconds() / 60.0
                    if timeout_minutes < 1:
                        timeout_minutes = self.default_task_timeout_minutes
                else:
@@ -1407,25 +1478,24 @@ Parent Task ID: {parent_task.id}

                elapsed = (now - start_time).total_seconds() / 60.0
                if elapsed > timeout_minutes:
-                    # [v2.7.1] Mail 幻觉门控兜底：有回复 + working → done
-                    if self._current_project_id == "_mail":
-                        has_reply = self._mail_check_reply(task.id, db_path)
-                        if has_reply:
-                            conn = get_connection(db_path)
-                            try:
-                                ok = self._transition_status(
-                                    conn, task.id, "done",
-                                    agent="daemon",
-                                    detail={"reason": "mail_auto_done_recheck",
-                                            "elapsed_minutes": round(elapsed, 1)},
-                                )
-                                if ok:
-                                    reclaimed.append(task.id)
-                                    logger.info("Mail %s: ticker recheck found reply, marked done (%.1fm)",
-                                                task.id, elapsed)
-                            finally:
-                                conn.close()
-                            continue
+                    # [Step 5] handler 幻觉门控兜底：check_completion 通过 + working → done
+                    handler = TaskTypeRegistry.get_by_project(self._current_project_id)
+                    if handler and handler.check_completion(task.id, db_path):
+                        conn = get_connection(db_path)
+                        try:
+                            ok = self._transition_status(
+                                conn, task.id, "done",
+                                agent="daemon",
+                                detail={"reason": "mail_auto_done_recheck",
+                                        "elapsed_minutes": round(elapsed, 1)},
+                            )
+                            if ok:
+                                reclaimed.append(task.id)
+                                logger.info("Mail %s: ticker recheck found reply, marked done (%.1fm)",
+                                            task.id, elapsed)
+                        finally:
+                            conn.close()
+                        continue

                    conn = get_connection(db_path)
                    try:
@@ -1446,8 +1516,10 @@ Parent Task ID: {parent_task.id}
                pass

        # v2.7.2: 进程存活性检查 — counter 占用但进程已死的兜底
-        if self.spawner and self.counter and hasattr(self.counter, "active_agents"):
-            for agent_id in list(self.counter.active_agents.keys()) if hasattr(self.counter, "active_agents") else []:
+        if self.spawner and self.counter and hasattr(
+                self.counter, "active_agents"):
+            for agent_id in list(self.counter.active_agents.keys()) if hasattr(
+                    self.counter, "active_agents") else []:
                session_info = self.spawner.get_session_by_agent(agent_id)
                if not session_info:
                    continue
@@ -1464,20 +1536,24 @@ Parent Task ID: {parent_task.id}
                            conn = get_connection(db_path)
                            try:
                                current_row = conn.execute(
-                                    "SELECT status FROM tasks WHERE id=?", (task_id_check,)
+                                    "SELECT status FROM tasks WHERE id=?", (
+                                        task_id_check,)
                                ).fetchone()
                                if current_row and current_row["status"] == "review":
-                                    logger.info("Task %s in review, keeping status (process dead)", task_id_check)
+                                    logger.info(
+                                        "Task %s in review, keeping status (process dead)", task_id_check)
                                else:
                                    self._transition_status(
                                        conn, task_id_check, "pending",
                                        agent="daemon",
-                                        detail={"reason": "process_dead", "pid": pid},
+                                        detail={
+                                            "reason": "process_dead", "pid": pid},
                                    )
                            finally:
                                conn.close()
                        except Exception:
-                            logger.exception("Failed to handle process dead for task %s", task_id_check)
+                            logger.exception(
+                                "Failed to handle process dead for task %s", task_id_check)

        # #07.2: Fix-3b 已删除。review 超时/crash 统一由 process_dead + _check_timeouts 处理

@@ -1496,7 +1572,10 @@ Parent Task ID: {parent_task.id}
            finally:
                conn.close()
        except Exception as e:
-            logger.error("Mail %s: ticker reply check error: %s", original_task_id, e)
+            logger.error(
+                "Mail %s: ticker reply check error: %s",
+                original_task_id,
+                e)
            return True  # 保守：查询失败假设有回复

    def _check_recent_routing(self, db_path: Path, task_id: str,
@@ -1505,7 +1584,8 @@ Parent Task ID: {parent_task.id}
        try:
            conn = get_connection(db_path)
            try:
-                # 检查是否有 from_status=review 的 dispatched 记录（防止重复 review dispatch）
+                # 检查是否有 from_status=review 的 dispatched 记录（防止重复 review
+                # dispatch）
                if action_type == "review":
                    row = conn.execute(
                        "SELECT COUNT(*) as cnt FROM routing_decisions "
@@ -1536,17 +1616,23 @@ Parent Task ID: {parent_task.id}
        NON_TERMINAL = {"claimed", "working", "review", "reviewing"}

        projects = self.registry.list_projects()
-        recovery_report = {"projects": {}, "total_recovered": 0, "total_noop": 0}
+        recovery_report = {
+            "projects": {},
+            "total_recovered": 0,
+            "total_noop": 0}

        # 收集所有需要扫描的项目（registry + 虚拟项目）
        project_dirs = {}
        for project_id, project_info in projects.items():
            if project_info.get("status") == "active":
-                project_dirs[project_id] = self.registry.root / project_id / "blackboard.db"
+                project_dirs[project_id] = self.registry.root / \
+                    project_id / "blackboard.db"

-        # 虚拟项目
-        for virtual_id in ("_general", "_mail"):
-            virtual_db = Path(self.registry.root) / virtual_id / "blackboard.db"
+        # 虚拟项目：_general + 注册表自动发现
+        virtual_ids = ["_general"] + TaskTypeRegistry.virtual_projects()
+        for virtual_id in virtual_ids:
+            virtual_db = Path(self.registry.root) / \
+                virtual_id / "blackboard.db"
            if virtual_db.exists() and virtual_id not in project_dirs:
                project_dirs[virtual_id] = virtual_db

@@ -1566,13 +1652,15 @@ Parent Task ID: {parent_task.id}
            old_pid = self._current_project_id
            self._current_project_id = project_id
            try:
-                recovered, noop_count = self._recover_project(db_path, NON_TERMINAL)
+                recovered, noop_count = self._recover_project(
+                    db_path, NON_TERMINAL)
                if recovered:
                    recovery_report["projects"][project_id] = recovered
                    recovery_report["total_recovered"] += len(recovered)
                recovery_report["total_noop"] += noop_count
            except Exception:
-                logger.exception("Startup recovery failed for project %s", project_id)
+                logger.exception(
+                    "Startup recovery failed for project %s", project_id)
            finally:
                self._current_project_id = old_pid

@@ -1584,7 +1672,8 @@ Parent Task ID: {parent_task.id}
            logger.info("Startup recovery: %d tasks kept as-is (no recovery needed)",
                        recovery_report["total_noop"])
        else:
-            logger.info("Startup recovery: no non-terminal tasks found, clean start")
+            logger.info(
+                "Startup recovery: no non-terminal tasks found, clean start")

        return recovery_report

@@ -1607,10 +1696,13 @@ Parent Task ID: {parent_task.id}

                for task in rows:
                    try:
-                        action = self._determine_recovery_action(conn, task, status, db_path)
+                        action = self._determine_recovery_action(
+                            conn, task, status, db_path)
                        if action:
-                            self._execute_recovery(conn, task["id"], action, db_path)
-                            recovered.append({"task_id": task["id"], "from": status, "action": action})
+                            self._execute_recovery(
+                                conn, task["id"], action, db_path)
+                            recovered.append(
+                                {"task_id": task["id"], "from": status, "action": action})
                        else:
                            # 审计：保持原状的任务也记录事件
                            noop_count += 1
@@ -1621,7 +1713,8 @@ Parent Task ID: {parent_task.id}
                            )
                            conn.commit()
                    except Exception:
-                        logger.exception("Startup recovery failed for task %s", task["id"])
+                        logger.exception(
+                            "Startup recovery failed for task %s", task["id"])
        finally:
            conn.close()

@@ -1699,7 +1792,8 @@ Parent Task ID: {parent_task.id}
        # 无审查结论 → 保持 review，ticker 自然会 dispatch reviewer
        return None

-    def _execute_recovery(self, conn, task_id: str, action: str, db_path: Path):
+    def _execute_recovery(self, conn, task_id: str,
+                          action: str, db_path: Path):
        """执行恢复动作"""
        # 获取原始状态（用于审计）
        orig_row = conn.execute(
@@ -1711,17 +1805,22 @@ Parent Task ID: {parent_task.id}
            self._transition_status(
                conn, task_id, "pending",
                agent="daemon",
-                detail={"reason": "startup_recovery", "original_status": orig_status},
+                detail={
+                    "reason": "startup_recovery",
+                    "original_status": orig_status},
            )
            # 清空 current_agent（常规推 pending，无特定 agent 接手）
-            conn.execute("UPDATE tasks SET current_agent=NULL WHERE id=?", (task_id,))
+            conn.execute(
+                "UPDATE tasks SET current_agent=NULL WHERE id=?", (task_id,))
            conn.commit()

        elif action == "push_to_pending_keep_agent":
            self._transition_status(
                conn, task_id, "pending",
                agent="daemon",
-                detail={"reason": "startup_recovery", "original_status": orig_status},
+                detail={
+                    "reason": "startup_recovery",
+                    "original_status": orig_status},
            )
            # 保留 current_agent，让同一 agent 重新接手
            conn.commit()
@@ -1730,7 +1829,9 @@ Parent Task ID: {parent_task.id}
            self._transition_status(
                conn, task_id, "review",
                agent="daemon",
-                detail={"reason": "startup_recovery", "original_status": "working"},
+                detail={
+                    "reason": "startup_recovery",
+                    "original_status": "working"},
            )
            conn.commit()

@@ -1738,7 +1839,9 @@ Parent Task ID: {parent_task.id}
            self._transition_status(
                conn, task_id, "done",
                agent="daemon",
-                detail={"reason": "startup_recovery", "original_status": orig_status},
+                detail={
+                    "reason": "startup_recovery",
+                    "original_status": orig_status},
            )
            conn.commit()

@@ -1746,22 +1849,30 @@ Parent Task ID: {parent_task.id}
            self._transition_status(
                conn, task_id, "failed",
                agent="daemon",
-                detail={"reason": "startup_recovery", "original_status": orig_status},
+                detail={
+                    "reason": "startup_recovery",
+                    "original_status": orig_status},
            )
            conn.commit()

        # 记录恢复审计事件
        conn.execute(
            "INSERT INTO events (task_id, agent, event_type, detail) VALUES (?, ?, ?, ?)",
-            (task_id, "daemon", "startup_recovery", json.dumps({"action": action}))
+            (task_id, "daemon", "startup_recovery",
+             json.dumps({"action": action}))
        )
        conn.commit()

-        logger.info("Recovery: task %s → %s (action=%s)", task_id, action, action)
+        logger.info(
+            "Recovery: task %s → %s (action=%s)",
+            task_id,
+            action,
+            action)

    def _find_pre_reviewing_status(self, conn, task_id: str) -> str:
        """查 events 表找到 reviewing 之前的状态（done 或 failed）"""
-        # _transition_status 写入 event_type=f"task_{new_status}"，detail 用 from/to
+        # _transition_status 写入 event_type=f"task_{new_status}"，detail 用
+        # from/to
        rows = conn.execute(
            """SELECT detail FROM events
               WHERE task_id=? AND event_type='task_reviewing'
@@ -1772,7 +1883,8 @@ Parent Task ID: {parent_task.id}
        for event in rows:
            try:
                detail = json.loads(event["detail"])
-                # _transition_status detail 格式: {"from": old_status, "to": new_status, ...}
+                # _transition_status detail 格式: {"from": old_status, "to":
+                # new_status, ...}
                prev = detail.get("from") or detail.get("old_status")
                if prev in ("done", "failed"):
                    return prev
@@ -0,0 +1,277 @@
+"""toolchain_handler.py — 工具链事件 handler。
+
+处理 Gitea Webhook 事件（CI 失败、Review 请求、Issue 指派等）。
+"""
+from __future__ import annotations
+
+import json
+import logging
+import urllib.request
+from pathlib import Path
+from typing import Dict
+
+from src.daemon.base_task_handler import BaseTaskHandler, VerifyResult
+from src.daemon.prompt_composer import PromptComposer, PromptContext
+from src.daemon.toolchain_templates import render_template, _TEMPLATE_MAP
+from src.blackboard.db import get_connection
+
+logger = logging.getLogger("moziplus-v2.handler.toolchain")
+
+
+# ---------------------------------------------------------------------------
+# Toolchain PromptSections
+# ---------------------------------------------------------------------------
+
+class ToolchainContextSection:
+    """事件类型 + 事件详情（priority=10）"""
+
+    name: str = "toolchain_context"
+    priority: int = 10
+
+    def render(self, context: PromptContext) -> str:
+        event_type = context.event_type
+        event_data: Dict = context.event_data or {}
+
+        if event_type in _TEMPLATE_MAP:
+            # 使用模板引擎渲染已知事件
+            variables = {k: str(v) for k, v in event_data.items()}
+            return render_template(event_type, variables)
+
+        # fallback：通用事件描述
+        lines = ["## 工具链事件", ""]
+        lines.append(f"- **事件类型**: {event_type or '未知'}")
+        if event_data:
+            lines.append("- **事件详情**:")
+            for key, value in event_data.items():
+                lines.append(f"  - {key}: {value}")
+        lines.append("")
+        return "\n".join(lines)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return True
+
+
+class ToolchainApiSection:
+    """API 操作指令（priority=40），success_status=done"""
+
+    name: str = "toolchain_api"
+    priority: int = 40
+
+    API_HOST = "localhost:8083"
+
+    def render(self, context: PromptContext) -> str:
+        lines = [
+            "## API 操作指令",
+            "",
+            f"项目 ID: `{context.project_id}`",
+            f"任务 ID: `{context.task_id}`",
+            "",
+            "### 完成后必须更新任务状态",
+            "完成后务必通过以下命令将任务标记为 **done**：",
+            "```bash",
+            f'curl -s -X POST "http://{self.API_HOST}/api/projects/{context.project_id}/tasks/{context.task_id}/status" \\',
+            '  -H "Content-Type: application/json" \\',
+            '  -d \'{"status": "done"}\'',
+            "```",
+            "",
+            "### 提交产出",
+            "如有产出（如 review 结果、修复方案），提交到任务 outputs：",
+            "```bash",
+            f'curl -s -X POST "http://{self.API_HOST}/api/projects/{context.project_id}/tasks/{context.task_id}/outputs" \\',
+            '  -H "Content-Type: application/json" \\',
+            '  -d \'{"content": "<你的产出内容>", "type": "text"}\'',
+            "```",
+            "",
+        ]
+        return "\n".join(lines)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return True
+
+
+class ToolchainConstraintsSection:
+    """硬约束（priority=50）"""
+
+    name: str = "toolchain_constraints"
+    priority: int = 50
+
+    def render(self, context: PromptContext) -> str:
+        lines = [
+            "## 硬约束",
+            "",
+            "1. **必须标 done**：处理完成后必须通过 API 将任务状态更新为 `done`，否则视为未完成",
+            "2. **产出不能为空**：必须提交有意义的产出（output 或 comment），不能只改状态",
+            "3. **单一职责**：只处理本次事件相关的操作，不要越界执行无关任务",
+            "4. **出错即报告**：如果无法处理（如权限不足、资源不存在），在 comment 中说明原因并标 done",
+            "5. **不要创建新任务**：工具链事件只处理当前事件，不衍生新任务",
+            "",
+        ]
+        return "\n".join(lines)
+
+    def should_include(self, context: PromptContext) -> bool:
+        return True
+
+
+# ---------------------------------------------------------------------------
+# ToolchainHandler
+# ---------------------------------------------------------------------------
+
+class ToolchainHandler(BaseTaskHandler):
+    """工具链事件 handler。"""
+
+    task_type = "toolchain"
+    virtual_project = "_toolchain"
+    display_name = "工具链事件"
+
+    def target_success_status(self) -> str:
+        return "done"
+
+    def pre_spawn(self, task_id: str, db_path: Path) -> bool:
+        """auto_working：pending → working"""
+        return self._auto_mark_working(task_id, db_path)
+
+    def get_sections(self) -> list:
+        """返回 3 个 Toolchain PromptSection 实例"""
+        return [
+            ToolchainContextSection(),
+            ToolchainApiSection(),
+            ToolchainConstraintsSection(),
+        ]
+
+    def build_prompt(self, context: PromptContext) -> str:
+        """通过 PromptComposer 拼装 sections 为最终 prompt"""
+        composer = PromptComposer()
+        composer.add_many(self.get_sections())
+        return composer.compose(context)
+
+    def verify_completion(self, task_id: str, db_path: Path) -> VerifyResult:
+        """检查行动输出（output 或 comment 有实质内容）"""
+        try:
+            conn = get_connection(db_path)
+            try:
+                # 检查 output
+                output_count = conn.execute(
+                    "SELECT COUNT(*) FROM outputs WHERE task_id=?", (task_id,)
+                ).fetchone()[0]
+                if output_count > 0:
+                    return VerifyResult(True, "has_output", f"output_count={output_count}")
+
+                # 检查 comment（非系统、有实质内容）
+                comment_count = conn.execute(
+                    "SELECT COUNT(*) FROM comments WHERE task_id=? "
+                    "AND author != 'system' AND LENGTH(content) >= 20",
+                    (task_id,)
+                ).fetchone()[0]
+                if comment_count > 0:
+                    return VerifyResult(True, "has_comment", f"comment_count={comment_count}")
+
+                return VerifyResult(False, "no_action", "output=0, comment=0")
+            finally:
+                conn.close()
+        except Exception as e:
+            logger.error("Toolchain %s: verify error: %s", task_id, e)
+            return VerifyResult(False, "verify_error", str(e))
+
+    def on_failure(self, task_id: str, agent_id: str,
+                   db_path: Path, verify: VerifyResult) -> None:
+        """验证失败 → 标 failed + Mail API 通知主公"""
+        self._mark_task_status(db_path, task_id, "failed")
+        logger.info("Toolchain %s: verify failed (%s), marked failed", task_id, verify.reason)
+
+        # 从 db 读取事件上下文
+        event_type = ""
+        event_data: Dict = {}
+        try:
+            conn = get_connection(db_path)
+            row = conn.execute(
+                "SELECT must_haves FROM tasks WHERE id=?", (task_id,)
+            ).fetchone()
+            if row and row["must_haves"]:
+                meta = json.loads(row["must_haves"])
+                event_type = meta.get("event_type", "")
+                raw = meta.get("event_data", "{}")
+                event_data = json.loads(raw) if isinstance(raw, str) else raw
+            conn.close()
+        except Exception:
+            pass
+
+        self._notify_via_mail_api(
+            task_id, verify.reason, verify.evidence,
+            event_type, event_data,
+        )
+
+    def _build_gitea_links(self, event_type: str, event_data: dict) -> str:
+        """根据事件类型构建 Gitea 链接。"""
+        links = []
+        repo = event_data.get("repo", "")
+        base_url = "http://192.168.2.154:3000"
+
+        if "pr_number" in event_data:
+            links.append(f"PR: {base_url}/{repo}/pulls/{event_data['pr_number']}")
+        if "issue_number" in event_data:
+            links.append(f"Issue: {base_url}/{repo}/issues/{event_data['issue_number']}")
+        if "commit" in event_data:
+            links.append(f"Commit: {base_url}/{repo}/commit/{event_data['commit']}")
+        if "branch" in event_data and "commit" not in event_data:
+            links.append(f"分支: {event_data['branch']}")
+
+        return "\n".join(links) if links else "（无法提取链接，请检查黑板任务详情）"
+
+    def _notify_via_mail_api(
+        self,
+        task_id: str,
+        reason: str,
+        evidence: str,
+        event_type: str,
+        event_data: Dict,
+    ) -> None:
+        """通过 Mail API 发送丰富的失败通知给主公。"""
+        # 构建行动指引
+        action_hint = "请检查黑板任务并手动处理。"
+        et_lower = event_type.lower()
+        if "ci" in et_lower or "deploy" in et_lower:
+            action_hint = "建议创建任务派给 jiangwei-infra 检查 CI/部署问题。"
+        elif "review" in et_lower:
+            action_hint = "建议查看 PR review 状态，必要时通知相关开发者。"
+        elif "issue" in et_lower:
+            action_hint = "建议创建任务派给对应开发者处理 Issue。"
+
+        # 构建事件详情
+        event_details = ""
+        if event_data:
+            event_details = "\n".join(
+                f"  - {k}: {v}" for k, v in event_data.items()
+            )
+
+        # 构建 Gitea 链接
+        gitea_links = self._build_gitea_links(event_type, event_data)
+
+        title = f"[toolchain-handler] 工具链事件处理失败: {task_id}"
+        text = (
+            f"任务 {task_id} 验证失败\n\n"
+            f"事件类型: {event_type or '未知'}\n"
+            f"事件详情:\n{event_details or '  (无)'}\n\n"
+            f"失败原因: {reason}\n"
+            f"证据: {evidence}\n\n"
+            f"{gitea_links}\n\n"
+            f"行动指引: {action_hint}"
+        )
+
+        payload = json.dumps({
+            "from": "daemon",
+            "to": "pangtong-fujunshi",
+            "title": title,
+            "text": text,
+            "type": "inform",
+        }, ensure_ascii=False).encode("utf-8")
+
+        try:
+            req = urllib.request.Request(
+                "http://localhost:8083/api/mail",
+                data=payload,
+                headers={"Content-Type": "application/json"},
+            )
+            urllib.request.urlopen(req, timeout=5)
+            logger.info("Toolchain %s: sent failure notification via Mail API", task_id)
+        except Exception as e:
+            logger.warning("Toolchain %s: failed to notify via Mail API: %s", task_id, e)
@@ -21,6 +21,8 @@ _TEMPLATE_MAP: Dict[str, str] = {
    "issue_assigned": "issue_assigned.md",
    "ci_failure": "ci_failure.md",
    "deploy_failure": "deploy_failure.md",
+    "review_updated": "review_updated.md",
+    "review_comment": "review_comment.md",
 }

 # 模板缓存
@@ -426,7 +426,7 @@ export default function TaskModal() {
            {/* 状态操作 */}
            <div style={{ marginBottom: 16 }}>
              <SectionLabel icon="🔄" title="状态操作" />
-              <StatusButtons status={task.status} taskId={task.id} resumedFrom={task.resumed_from} />
+              <StatusButtons status={task.status} taskId={task.id} resumedFrom={task.resumed_from ?? undefined} />
            </div>

            {/* v2.7: 子 Task 进度 + 列表 */}
@@ -57,6 +57,7 @@ export interface V2Task {
  estimated_duration_minutes: number | null;
  escalated: number;
  archived: number;  // v2.8: 归档标记
+  resumed_from: string | null;  // v2.8: 续杯来源
  // API 聚合字段
  comments_count?: number;
  outputs_count?: number;
@@ -1,6 +1,13 @@
 """v2.6 主入口 - FastAPI + Daemon ticker 共享 asyncio event loop"""

 from __future__ import annotations
+from src.api.toolchain_routes import router as toolchain_router
+from src.api.mail_routes import router as mail_router
+from src.api.sse_routes import router as sse_router
+from src.api.project_routes import router as project_router
+from src.api.daemon_routes import router as daemon_router
+from src.api.checkpoint_routes import router as checkpoint_router
+from src.api.blackboard_routes import router as blackboard_router

 import logging
 from contextlib import asynccontextmanager
@@ -14,6 +21,10 @@ from fastapi.staticfiles import StaticFiles

 from src.blackboard.registry import ProjectRegistry
 from src.daemon.ticker import Ticker
+from src.daemon.task_handler import TaskHandler
+from src.daemon.mail_handler import MailHandler
+from src.daemon.toolchain_handler import ToolchainHandler
+from src.daemon.task_type_registry import TaskTypeRegistry
 from src.daemon.spawner import AgentSpawner
 from src.daemon.bootstrap import BootstrapBuilder
 from src.daemon.dispatcher import Dispatcher
@@ -131,7 +142,8 @@ async def lifespan(app: FastAPI):
    counter = ActiveAgentCounter(
        max_global=daemon_config.get("max_global_agents", 5),
        max_per_session=daemon_config.get("max_per_session", 1),
-        max_concurrent_sessions=daemon_config.get("max_concurrent_sessions", 3),
+        max_concurrent_sessions=daemon_config.get(
+            "max_concurrent_sessions", 3),
        default_cooldown_seconds=daemon_config.get("cooldown_seconds", 120),
    )
    # BootstrapBuilder（L2 四段式引擎注入层，v2.1）
@@ -181,7 +193,10 @@ async def lifespan(app: FastAPI):
        spawner=spawner,
        counter=counter,
        db_path=default_db_path,
-        guardrails=GuardrailEngine(config_path=Path(__file__).parent.parent / "config" / "guardrails.yaml"),
+        guardrails=GuardrailEngine(
+            config_path=Path(__file__).parent.parent /
+            "config" /
+            "guardrails.yaml"),
    )

    # ── 集成模块 ──
@@ -191,6 +206,7 @@ async def lifespan(app: FastAPI):
    )

    # ExperienceDistiller（经验自动蒸馏）
+    config.get("experience", {})
    experience_distiller = ExperienceDistiller(
        store=ExperienceStore(store_path=DATA_ROOT / "experiences.jsonl"),
    )
@@ -202,6 +218,11 @@ async def lifespan(app: FastAPI):
        watch_interval=inbox_config.get("watch_interval", 1.0),
    )

+    # [Step 5] 注册 TaskType handler（必须在 ticker 启动前）
+    TaskTypeRegistry.register(TaskHandler())
+    TaskTypeRegistry.register(MailHandler())
+    TaskTypeRegistry.register(ToolchainHandler())
+
    ticker = Ticker(
        registry=registry,
        tick_interval=tick_interval,
@@ -251,13 +272,6 @@ app.add_middleware(
 # API 路由注册
 # ---------------------------------------------------------------------------

-from src.api.blackboard_routes import router as blackboard_router  # noqa: E402
-from src.api.checkpoint_routes import router as checkpoint_router  # noqa: E402
-from src.api.daemon_routes import router as daemon_router  # noqa: E402
-from src.api.project_routes import router as project_router  # noqa: E402
-from src.api.sse_routes import router as sse_router  # noqa: E402
-from src.api.mail_routes import router as mail_router  # noqa: E402
-from src.api.toolchain_routes import router as toolchain_router  # noqa: E402

 app.include_router(blackboard_router)
 app.include_router(checkpoint_router)
@@ -303,6 +317,7 @@ if DIST_DIR.exists():

    class CachedStaticFiles:
        """包装 StaticFiles，添加 Cache-Control 头"""
+
        def __init__(self, app):
            self._app = app

@@ -0,0 +1,9 @@
+Review 评论
+
+PR: http://192.168.2.154:3000/{repo}/pulls/{pr_number}
+标题: {pr_title}
+评论者: {reviewer}
+
+{comment_body}
+
+请查看评论并回复或修改代码。
@@ -0,0 +1,16 @@
+PR 有新提交，请重新 Review
+
+PR: http://192.168.2.154:3000/{repo}/pulls/{pr_number}
+标题: {pr_title}
+作者: {pr_author}
+新 commit: {new_sha}
+
+您之前已审查过此 PR，作者已根据反馈更新了代码。请重新 Review。
+
+流程：
+1. 读取 PR diff（Gitea API: GET /repos/{repo}/pulls/{pr_number}.diff）
+2. 重点检查上次 Review 意见的修改部分
+3. 提交 Review（Gitea API: POST /repos/{repo}/pulls/{pr_number}/reviews）
+4. 提交后改动者会自动收到通知
+
+完成后回复此 Mail 确认。
@@ -1,12 +1,12 @@
 import pytest

-pytestmark = pytest.mark.e2e
-
 skip_no_integration = pytest.mark.skipif(
    not __import__("os").environ.get("RUN_INTEGRATION"),
    reason="Set RUN_INTEGRATION=1 to run E2E tests against real daemon",
 )

+pytestmark = [pytest.mark.e2e, skip_no_integration]
+
 """v2.7 端到端测试 — 全链路真实环境

 覆盖：项目管理 → Task CRUD → SubTask → Stage进度 → 状态聚合 → 依赖链 → 超时 → Mail → 真实Agent调度
@@ -0,0 +1,92 @@
+"""单元测试：§24 v3 rotation-only compact 检测
+
+测试 _get_recent_gateway_logs 和 _check_compact_in_progress_gateway。
+用 tmp_path 构造 mock gateway 日志文件。
+"""
+
+import json
+import os
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+
+from src.daemon.spawner import AgentSpawner
+
+
+# ── helpers ──
+
+_SESSION_KEY = "agent:pangtong-fujunshi:main"
+_TODAY_STR = datetime.now().strftime("%Y-%m-%d")
+
+
+def _make_rotation_event(session_key: str, ts: datetime) -> dict:
+    """构造一条 rotation 日志事件"""
+    return {
+        "time": ts.isoformat(),
+        "message": f"[compaction] rotated active transcript after compaction (sessionKey={session_key})",
+    }
+
+
+def _make_other_event(session_key: str, ts: datetime, msg: str = "something else") -> dict:
+    """构造一条普通日志事件"""
+    return {
+        "time": ts.isoformat(),
+        "message": f"{msg} (sessionKey={session_key})",
+    }
+
+
+def _write_log(tmp_path: Path, date_str: str, lines: list[dict]):
+    """写 mock 日志文件"""
+    log_file = tmp_path / f"openclaw-{date_str}.log"
+    with open(log_file, "w") as f:
+        for obj in lines:
+            f.write(json.dumps(obj, ensure_ascii=False) + "\n")
+
+
+@pytest.fixture(autouse=True)
+def _set_log_dir(tmp_path, monkeypatch):
+    """每个测试自动设置 OPENCLAW_LOG_DIR 到 tmp_path"""
+    monkeypatch.setenv("OPENCLAW_LOG_DIR", str(tmp_path))
+
+
+# ── 测试用例 ──
+
+
+class TestCheckCompactInProgress:
+    """§24 v3: _check_compact_in_progress_gateway 单元测试"""
+
+    def test_rotation_within_window_returns_true(self, tmp_path):
+        """TC1: rotation 事件在窗口内 → True"""
+        now = datetime.now(timezone.utc)
+        recent = now - timedelta(seconds=30)
+        _write_log(tmp_path, _TODAY_STR, [_make_rotation_event(_SESSION_KEY, recent)])
+        assert AgentSpawner._check_compact_in_progress_gateway(_SESSION_KEY) is True
+
+    def test_rotation_outside_window_returns_false(self, tmp_path):
+        """TC2: rotation 事件超出窗口 → False"""
+        now = datetime.now(timezone.utc)
+        old = now - timedelta(seconds=200)
+        _write_log(tmp_path, _TODAY_STR, [_make_rotation_event(_SESSION_KEY, old)])
+        assert AgentSpawner._check_compact_in_progress_gateway(_SESSION_KEY) is False
+
+    def test_no_rotation_event_returns_false(self, tmp_path):
+        """TC3: 无 rotation 事件 → False"""
+        now = datetime.now(timezone.utc)
+        _write_log(tmp_path, _TODAY_STR, [
+            _make_other_event(_SESSION_KEY, now, "model.completed"),
+        ])
+        assert AgentSpawner._check_compact_in_progress_gateway(_SESSION_KEY) is False
+
+    def test_log_file_not_exists_returns_false(self, tmp_path):
+        """TC4: 日志文件不存在 → False"""
+        # tmp_path 为空目录，无日志文件
+        assert AgentSpawner._check_compact_in_progress_gateway(_SESSION_KEY) is False
+
+    def test_session_key_mismatch_returns_false(self, tmp_path):
+        """TC5: sessionKey 不匹配 → False"""
+        now = datetime.now(timezone.utc)
+        recent = now - timedelta(seconds=10)
+        other_key = "agent:simayi-challenger:main"
+        _write_log(tmp_path, _TODAY_STR, [_make_rotation_event(other_key, recent)])
+        assert AgentSpawner._check_compact_in_progress_gateway(_SESSION_KEY) is False
@@ -123,7 +123,7 @@ class TestClassifyNoJsonExit0:

    def test_task_status_pending(self):
        result = Spawner._classify_outcome(0, {}, "", "pending", "")
-        assert result["outcome"] == "agent_error"
+        assert result["outcome"] == "completed"
        assert result["should_retry"] is False