W15: workflow improvements — EXPRESS fast-path, audit→fix closed loop, metadata self-check (W15.1-W15.3)

- W15.1 (杨帆): Add EXPRESS fast-path to §11 state machine (T17/T18, E1-E6 conditions, escalation safety valve)
- W15.2 (王测): Add §14 audit→fix closed loop — findings-registry.md, severity-driven auto-triage, CRITICAL blocking rule
- W15.3 (胡桐): Create scripts/check_agents_metadata.py (5-check: YAML parse, rating range, group/member refs, duplicate IDs)
- Fix YAML orphan bugs in 3 profiles: devops-hu, engineer-sun, security-cao (perf_log entries outside array)
- Pre-fill findings-registry.md with 10 historical findings from W11.1/W11.7 audits

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

agents/WORKFLOW.md

@@ -144,7 +144,7 @@
                     | PROPOSE  |
                     +----+-----+
                          |
-              simple     |     complex
+    simple / EXPRESS    |     complex
            +-------------+--------------+
            |                            |
            |                            v
@@ -183,6 +183,7 @@
 
 - OPTIMIZE --re-vote--> VOTE
 - OPTIMIZE --fundamental rewrite--> PROPOSE
+- EXECUTE --EXPRESS escalation--> VOTE
 - INSPECT fail --fixable--> EXECUTE
 - INSPECT fail --design--> OPTIMIZE
 - INSPECT fail --fatal--> ROLLBACK
@@ -192,6 +193,23 @@
 
 - ANY --CEO abort--> ABORT
 
+### 11.2.1 EXPRESS 快跳路径
+
+快跳（EXPRESS）是将 PROPOSE→VOTE→OPTIMIZE→INTEGRATE 压缩为 PROPOSE→EXECUTE 的合法短路径。快跳适用条件为以下 **全部 6 条** 同时满足：
+
+| # | 条件 | 验证方式 |
+|---|------|----------|
+| E1 | 改动源文件 ≤ 2 个（不含 profile.md） | `git diff --stat HEAD` 统计 changed files |
+| E2 | 不改动 `dstalk-core/include/` 下的任何公共头文件 | `git diff --name-only HEAD` 与 include/ 交集为空 |
+| E3 | 不改动 CMakeLists.txt / cmake/ 目录 / CMakePresets.json | `git diff --name-only HEAD` 与构建文件交集为空 |
+| E4 | 不新增公共 API 面：无新 `dstalk_` 前缀函数声明、无新插件接口结构体 | diff 中公共头文件无新增函数声明 |
+| E5 | 不涉及跨模块依赖变更：改动文件涉及 &le; 2 个顶层目录（如 `dstalk-core/`、单个 `plugins/<name>/`） | `git diff --dirstat HEAD` 目录数 &le; 2 |
+| E6 | CEO 在 WORKFLOW.md &sect;7 任务条目中显式标注 `[EXPRESS]` | 人工核对 &sect;7 |
+
+满足全部 E1-E6 &rarr; CEO 可声明 EXPRESS 快跳，任务直接进入 EXECUTE（跳过 VOTE / OPTIMIZE / INTEGRATE），对应转换规则 **T17**。
+
+**EXPRESS 升级**：若执行者在 EXECUTE 阶段发现任务实际超出 EXPRESS 条件（E1-E5 任一条不再成立），须立即报告 CEO。CEO 核实后移除 `[EXPRESS]` 标签并替换为 `[ESCALATED]`，任务从 EXECUTE 退回 VOTE 走完整治理路径，对应转换规则 **T18**。
+
 ### 11.3 转换条件表
 
 | # | 从 | 到 | 触发条件 | 决策者 |
@@ -212,6 +230,8 @@
 | T14 | INSPECT | ROLLBACK | 验收发现不可逆副作用：文件错误删除或覆盖 / 二进制损坏 / .git 目录状态异常 | CEO |
 | T15 | INSPECT | ABORT | CEO 判定继续修复成本 &gt; 重新执行成本（需改 &gt;5 个文件且涉及多个执行者重新协调） | CEO |
 | T16 | ANY | ABORT | 用户明确指令中止 OR 触发安全红线（凭证泄露、未加密敏感数据落盘） | CEO |
+| T17 | PROPOSE | EXECUTE | EXPRESS 快跳：同时满足 E1-E6（见 &sect;11.2.1 EXPRESS 条件表）AND CEO 在 &sect;7 任务条目标注 `[EXPRESS]` | CEO |
+| T18 | EXECUTE | VOTE | EXPRESS 升级：执行者报告任务实际范围超出 EXPRESS 条件（E1-E5 任一条不再成立）AND CEO 核实后将 `[EXPRESS]` 改为 `[ESCALATED]` | CEO |
 
 ### 11.4 状态进入/退出动作
 
@@ -344,4 +364,122 @@
 1. 在 WORKFLOW.md §7 记录：中止原因、时间、影响范围
 2. 决定改动处置：保留（`git stash`）或丢弃（`git checkout`）
 3. 本波 W 编号标记为 ABORTED，下一波使用新编号
-4. 相关执行者 profile.md performance_log 仍追加条目（rating: aborted），保留参与记录
+4. 相关执行者 profile.md performance_log 仍追加条目（rating: aborted），保留参与记录
+
+## 14. 审计→修复闭环机制
+
+审计（audit）产出的发现必须转化为可跟踪、可执行、可验收的修复任务。不允许审计报告写完即归档、发现问题无人跟进。
+
+核心闭环：**Audit report → Finding registration → Severity triage → Fix task (Wave) → CEO verify → Close**
+
+### 14.1 发现登记格式
+
+所有审计发现统一登记在 `agents/audits/findings-registry.md`，分为 Open Findings 和 Closed Findings 两个分区。每条发现包含以下字段：
+
+| 字段 | 说明 | 示例 |
+|------|------|------|
+| ID | `F-<源Wave>-<序号>`，全局唯一 | F-11.7-1 |
+| Severity | CRITICAL / HIGH / MEDIUM / LOW | CRITICAL |
+| Source | 审计报告文件名（相对于 agents/audits/） | W11.7-destructive-test.md |
+| Title | 一句话描述，含关键行号和症状 | `/clear` reports [OK] even when session unavailable — main.cpp:168-172 |
+| Status | 见 §14.2 状态定义 | OPEN |
+| Assigned To | 负责修复的员工 agent-id（OPEN 状态为空） | architect-huang |
+| Fix Wave | 修复所在的 Wave 编号（FIXED 后填写） | W16.1 |
+| Verified By | 验收人 agent-id（VERIFIED 后填写） | qa-wang |
+
+发现数量超过 20 条时，qa-wang 负责将 Closed 分区中超过 30 天的条目归档到 `agents/audits/findings-archive.md`。
+
+### 14.2 发现状态生命周期
+
+| 状态 | 类型 | 含义 | 进入条件 | 退出条件 |
+|------|------|------|----------|----------|
+| OPEN | 活跃 | 发现已登记，待 triage | 审计报告提交后，由审计人或 QA 组长录入 registry | CEO/QA 组长完成 triage 并指定执行者 |
+| ASSIGNED | 活跃 | 已指派修复人，等待执行 | Triage 完成 + CEO 在 PROPOSE 阶段创建对应修复 W 任务 | 执行者提交修复 + 自述 cmake 0 error + ctest 100% pass |
+| FIXED | 活跃 | 修复已提交，等待验证 | 执行者完成修复并更新 registry 状态 | CEO INSPECT 通过（§12）OR QA 组长验证通过 |
+| VERIFIED | 活跃 | 修复已验证，即将关闭 | INSPECT 全部通过 OR 回归测试覆盖通过 | 自动进入 CLOSED |
+| CLOSED | 终止 | 已关闭 | VERIFIED 后自动关闭 | — |
+| WONTFIX | 终止 | 决定不修复 | CEO 明确判定不修复（附理由） | — |
+| BLOCKED | 活跃 | 被阻塞 | 依赖的其他发现未修复 OR 外部条件不满足 | 阻塞解除后回到 ASSIGNED |
+
+状态转换图：
+
+```
+OPEN ──→ ASSIGNED ──→ FIXED ──→ VERIFIED ──→ CLOSED
+  │         │            │           │
+  │         │            │           │
+  ↓         ↓            ↓           ↓
+WONTFIX   BLOCKED      REOPEN      REOPEN
+                        (回到        (回到
+                      ASSIGNED)   ASSIGNED)
+```
+
+回边定义：
+
+- FIXED → ASSIGNED (REOPEN)：CEO INSPECT 发现修复不完整或引入新回归，退回执行者
+- VERIFIED → ASSIGNED (REOPEN)：后续回归测试暴露本发现的修复引入了新问题
+- ASSIGNED → BLOCKED → ASSIGNED：执行者发现依赖未满足，申请阻塞；依赖解除后恢复
+- OPEN → WONTFIX：CEO 判定（如：修复成本远超收益 / 已被后续重构覆盖 / 非 bug 是设计取舍）
+
+全局出口：
+
+- ANY → WONTFIX：CEO 在任意状态可强制关闭（需附理由写入 registry Change Log）
+
+### 14.3 自动转化规则
+
+从审计发现到修复任务的转化由严重级别驱动：
+
+| Severity | 转化规则 | 时限 | 触发者 |
+|----------|----------|------|--------|
+| CRITICAL | 下一波 PROPOSE 阶段 MUST 创建对应修复任务（W 编号），优先级最高，阻塞其他任务排期 | 当前 Wave + 1 | CEO |
+| HIGH | 2 个 Wave 内 MUST 安排修复任务 | 当前 Wave + 2 | CEO / QA 组长 |
+| MEDIUM | 每波 PROPOSE 阶段评估，可合并到其他同文件/同模块任务中附带修复 | 不限，但每 5 个 Wave 至少回顾一次 backlog | QA 组长 triage |
+| LOW | 进入 backlog，在相关源文件被其他任务改动时附带修复（opportunistic fix） | 不限 | 执行者自行判断 |
+
+CRITICAL 阻塞规则：
+- 如果进入 EXECUTE 阶段时仍有 OPEN 状态的 CRITICAL 发现，CEO 必须明确决策：(a) 本波优先修 CRITICAL，或 (b) 标记 WONTFIX（附理由），或 (c) 降级为 HIGH（附降级理由）
+- 不允许带着 OPEN CRITICAL 发现进入 SUCCESS
+
+### 14.4 CEO 审查协议（新增验收项）
+
+在 §12 验收清单基础上，INSPECT 阶段追加以下检查项：
+
+| # | 检查项 | 命令/方法 | 通过标准 |
+|---|--------|-----------|----------|
+| A1 | 发现登记完整性 | 检查本波新增审计报告：逐一核对 severity ≥ MEDIUM 的发现是否已录入 findings-registry.md | 无遗漏 |
+| A2 | CRITICAL 发现清零 | `grep "CRITICAL.*OPEN" agents/audits/findings-registry.md` | 输出为空（所有 CRITICAL 已修复或 WONTFIX） |
+| A3 | 修复关联标注 | 检查本波 EXECUTE 子代理报告 | 每个修复任务标注了对应的 Finding ID（格式：`Fixes: F-<Wave>-<N>`） |
+| A4 | 状态同步 | 逐一核对 registry 中本波涉及的发现状态与实际修复结果一致 | FIXED 状态发现的 cmake + ctest 已通过 |
+
+验收结论扩展：
+
+| 失败项 | 处理 |
+|--------|------|
+| A1 失败 | 补充录入 → 重新检查 |
+| A2 失败 | INSPECT → EXECUTE（优先修复 CRITICAL） |
+| A3 失败 | 补标注 → 重新检查 |
+| A4 失败 | 状态回退到 ASSIGNED → EXECUTE |
+
+### 14.5 与 §11 状态机的集成点
+
+| §11 状态 | §14 动作 | 责任人 |
+|----------|----------|--------|
+| PROPOSE | 1. 读取 findings-registry.md Open 分区 2. 将 CRITICAL/HIGH OPEN 发现转为候选 W 任务 3. 评估 MEDIUM backlog | CEO |
+| EXECUTE | 1. 子代理 prompt 中标注 `Fixes: F-<Wave>-<N>` 2. 修复完成后更新 registry 中该发现状态为 FIXED | 执行者 |
+| INSPECT | 1. 执行 §14.4 A1-A4 检查 2. 通过的发现 FIXED → VERIFIED → CLOSED 3. 失败的发现退回 ASSIGNED（REOPEN） | CEO |
+| SUCCESS | 1. 本波 CLOSED 的发现从 Open 分区移到 Closed 分区 2. 记录关闭日期和 Fix Wave | CEO |
+| ABORT | 本波 ASSIGNED 的发现回退到 OPEN（修复未发生） | CEO |
+
+### 14.6 审计人职责
+
+提交审计报告时，审计人必须同时完成以下动作：
+
+1. 在审计报告末尾新增 "## Findings Summary" 小节，列出所有发现的 ID、Severity、Title（与 registry 格式一致）
+2. 将 severity ≥ MEDIUM 的发现录入 `findings-registry.md` Open 分区，状态 OPEN
+3. LOW 发现同样录入（保持完整记录），但可在 triage 时直接标记为 backlog
+
+审计人完成登记后通知 CEO 或 QA 组长进行 triage。
+
+### 14.7 关联文档
+
+- [findings-registry.md](audits/findings-registry.md) — 发现注册表（单一事实来源）
+- [PROMPT_TEMPLATE.md](PROMPT_TEMPLATE.md) — 子代理 prompt 模板（修复任务使用标准模板，Fixes 行添加到交付清单）

agents/architect-yang/profile.md

@@ -26,5 +26,8 @@ performance_log:
   - date: 2026-05-27
     event: "W13.1: 深度审计 anthropic_plugin.cpp (497行)，6个C ABI函数零try/catch (§8违反)，response_body泄漏 + 全局指针竞态，tool_use静默丢弃。综合评级C。报告写入 agents/audits/W13.1-anthropic-audit.md"
     rating: completed
+  - date: 2026-05-27
+    event: "W15.1: 为 WORKFLOW.md §11 协作状态机设计 EXPRESS 快跳路径。定义 E1-E6 六项客观准入条件，新增 T17(快跳入口) + T18(升级回退) 两条转换规则，§11.2 图增加快跳边标注，新增 §11.2.1 完整说明。建议将 EXPRESS 作为正式快跳标签（非新增状态，避免状态爆炸）"
+    rating: completed
 current_groups: []
 ---

agents/audits/findings-registry.md

@@ -0,0 +1,38 @@
+# Audit Findings Registry
+
+> **维护人**: grp-quality-core (王测)
+> **格式定义**: 见 `agents/WORKFLOW.md` §14.2
+> **最后更新**: 2026-05-27 (W15.2 初始化，从 W11.1/W11.7 审计报告提取)
+
+---
+
+## Open Findings
+
+| ID | Severity | Source | Title | Status | Assigned To | Fix Wave | Verified By |
+|----|----------|--------|-------|--------|-------------|----------|-------------|
+| F-11.7-1 | CRITICAL | [W11.7-destructive-test.md](W11.7-destructive-test.md) | `build/bin/dstalk-cli.exe` corrupt copy (MD5 d8e8c92b vs 803ca2ea); all commands treated as AI prompt, exit code always 3 | OPEN | — | — | — |
+| F-11.7-2 | MEDIUM | [W11.7-destructive-test.md](W11.7-destructive-test.md) | `/clear` reports [OK] even when session unavailable (g_session==null) — main.cpp:168-172 | OPEN | — | — | — |
+| F-11.7-3 | LOW | [W11.7-destructive-test.md](W11.7-destructive-test.md) | `/context` silent no-output when session unavailable; no else branch — main.cpp:175-185 | OPEN | — | — | — |
+| F-11.7-4 | LOW | [W11.7-destructive-test.md](W11.7-destructive-test.md) | `/file write` (no args) matched as unknown command instead of usage hint | OPEN | — | — | — |
+| F-11.1-1 | HIGH | [W11.1-context-audit.md](W11.1-context-audit.md) | C++ exception (`std::bad_alloc`)穿越ABI边界，违反plugin-abi §5.3；trim_impl (L114-226) 无try/catch → std::terminate() | OPEN | — | — | — |
+| F-11.1-2 | HIGH | [W11.1-context-audit.md](W11.1-context-audit.md) | strdup返回值未检查，OOM时静默失败+泄漏；L138-141/L219-222 循环内4次strdup无nullptr检查 | OPEN | — | — | — |
+| F-11.1-3 | MEDIUM | [W11.1-context-audit.md](W11.1-context-audit.md) | context_set_max_tokens死API，g_max_tokens从未被读取（L21/L243-244） | OPEN | — | — | — |
+| F-11.1-4 | LOW | [W11.1-context-audit.md](W11.1-context-audit.md) | UTF-8解码无越界保护（L42-64, L96-104），多字节序列假设后续字节有效 | OPEN | — | — | — |
+| F-11.1-5 | LOW | [W11.1-context-audit.md](W11.1-context-audit.md) | token计数逻辑重复（L34-68 vs L91-106 ~90%重复） | OPEN | — | — | — |
+| F-11.1-6 | LOW | [W11.1-context-audit.md](W11.1-context-audit.md) | 0xC0/0xC1过短编码未识别（L52, L100），仅影响token估算计数 | OPEN | — | — | — |
+
+---
+
+## Closed Findings
+
+| ID | Severity | Source | Title | Closed Date | Fix Wave | Verified By |
+|----|----------|--------|-------|-------------|----------|-------------|
+| — | — | — | 暂无已关闭发现 | — | — | — |
+
+---
+
+## Change Log
+
+| Date | Change | Author |
+|------|--------|--------|
+| 2026-05-27 | W15.2 初始化，从 W11.1/W11.7 提取 10 条发现 | 王测 (qa-wang) |

agents/devops-hu/profile.md

@@ -32,7 +32,6 @@ performance_log:
       顺带修复: tools_plugin.cpp 缺少前向声明、lsp_plugin.cpp 函数签名 mismatch、
       5 个插件缺少 #include <boost/json/src.hpp> (Boost 1.86 不再识别 HEADER_ONLY)。
     rating: done
-current_groups: []
   - date: 2026-05-27
     event: "W12.4 修复 build 产物路径不一致 (BUG-1)"
     detail: >
@@ -43,4 +42,14 @@ current_groups: []
       ${CMAKE_BINARY_DIR}/bin 作为防御性显式声明；删除陈旧 build/dstalk-cli/dstalk-cli.exe。
       验证: clean rebuild 后仅 build/bin/dstalk-cli.exe 存在，ctest 4/4 pass。
     rating: done
+  - date: 2026-05-27
+    event: "W15.3: 设计 agents/ 目录元数据自检机制 (scripts/check_agents_metadata.py)"
+    detail: >
+      修复自身 profile.md YAML 格式错误 (perf_log 条目被误放在 current_groups: [] 之后)。
+      创建 5 项自检: C1 YAML 解析合法性、C2 rating 值范围、C3 current_groups -> group 引用完整性、
+      C4 group members -> agent 引用完整性、C5 重复 ID 检测 + 目录名一致性。
+      首轮运行发现 engineer-sun + security-cao 的 profile.md 存在同类 YAML 错误 (各 2 条目 orphan)。
+      建议集成到 refresh_status.py 作为前置检查，并加入 WORKFLOW.md §5 CEO 自查清单。
+    rating: done
+current_groups: []
 ---

agents/engineer-sun/profile.md

@@ -29,7 +29,6 @@ performance_log:
       修复前：第一行非 Content-Length 时 continue 丢弃该行，导致 header 解析偏移错位。
       修复后：正确遍历所有 header 行，空行后若仍未找到 Content-Length 则记录错误并跳过帧。
       编译通过，smoke test 通过。
-current_groups: []
   - date: 2026-05-27
     event: "W13.2: 深度审计 deepseek_plugin.cpp (486 行) — SSE 解析/ABI 异常安全/堆纪律/重复度"
     rating: completed
@@ -57,4 +56,5 @@ current_groups: []
       构建验证: cmake --build Release 0 error; ctest 4/4 pass。
       L420-471 reader_loop, L481-559 start, L561-603 stop 三件套, L605-630 open, L632-655 close,
       L657-683 diagnostics, L685-730 hover, L730-780 completion, L807-821 on_shutdown.
+current_groups: []
 ---

agents/qa-wang/profile.md

@@ -36,6 +36,9 @@ performance_log:
   - date: 2026-05-27
     event: "W13.3: network_plugin.cpp 深度审计 (322行, 9维度)。发现 TLS 证书验证完全禁用 (F, CVSS 7.4) + DNS 解析无超时 (永久hang) + 缺 catch(...)。RAII/堆纪律/并发 A 级。综合 C 级"
     rating: A
+  - date: 2026-05-27
+    event: "W15.2: 设计审计→修复闭环机制。定义 findings-registry.md 格式 + OPEN→ASSIGNED→FIXED→VERIFIED→CLOSED 状态生命周期 + 4级严重度自动转化规则 + WORKFLOW.md §14 完整草案。从 W11.1/W11.7 提取 10 条历史发现初始化注册表"
+    rating: A
 current_groups:
   - grp-quality-core (组长)
 ---

agents/security-cao/profile.md

@@ -50,18 +50,18 @@ performance_log:
       命令注入: 未发现。路径遍历: tools 确认。
       评级 session:D+ / tools:D。
       报告: agents/audits/W13.5-session-tools-audit.md
-	  - date: 2026-05-27
+  - date: 2026-05-27
-	    event: "W14.3: 修复 W13.5 审计发现 — 路径遍历 + 全局状态加锁 + 9 vtable try/catch"
+    event: "W14.3: 修复 W13.5 审计发现 — 路径遍历 + 全局状态加锁 + 9 vtable try/catch"
-	    rating: done
+    rating: done
-	    detail: |
+    detail: |
-	      修改 session_plugin.cpp (294行) + tools_plugin.cpp (292行)。
+      修改 session_plugin.cpp (294行) + tools_plugin.cpp (292行)。
-	      (1) is_safe_path() 拒绝空路径、绝对路径(/或盘符)、含..段，lexically_normal二次校验；
+      (1) is_safe_path() 拒绝空路径、绝对路径(/或盘符)、含..段，lexically_normal二次校验；
-	      builtin_file_read(L50) 和 builtin_file_write(L85) 入口调用，不安全→log ERROR + 返回错误JSON。
+      builtin_file_read(L50) 和 builtin_file_write(L85) 入口调用，不安全→log ERROR + 返回错误JSON。
-	      (2) 加锁: session g_history/g_cached_history→g_session_mutex; tools g_tools→g_tools_mutex;
+      (2) 加锁: session g_history/g_cached_history→g_session_mutex; tools g_tools→g_tools_mutex;
-	      g_host/g_file_io→std::atomic<T*> load(acquire)/store(release)。
+      g_host/g_file_io→std::atomic<T*> load(acquire)/store(release)。
-	      (3) 9 vtable try/catch 覆盖: session_add/save/load/history (session) +
+      (3) 9 vtable try/catch 覆盖: session_add/save/load/history (session) +
-	      tools_register_tool/unregister_tool/get_tools_json/execute/on_init (tools)。
+      tools_register_tool/unregister_tool/get_tools_json/execute/on_init (tools)。
-	      编译: cmake --build build --config Release → 0 error 0 warning。
+      编译: cmake --build build --config Release → 0 error 0 warning。
-	      ctest -C Release → 4/4 pass。
+      ctest -C Release → 4/4 pass。
 current_groups: []
 ---

scripts/check_agents_metadata.py

@@ -0,0 +1,421 @@
+#!/usr/bin/env python3
+"""
+agents/ metadata self-check: profile.md YAML validity, rating range,
+group cross-references, member cross-references.
+
+Usage:
+  python scripts/check_agents_metadata.py
+  python scripts/check_agents_metadata.py --strict   # treat warnings as errors
+  python scripts/check_agents_metadata.py --json       # machine-readable output
+
+Exit code: 0 = all checks pass, 1 = errors found, 2 = warnings only (--strict).
+
+Checks:
+  C1  YAML parse      - every profile.md + grp-*.md front matter parses legally
+  C2  rating range    - every performance_log entry uses a known rating token
+  C3  group ref       - every current_groups entry points to an existing grp-*.md
+  C4  member ref      - every group members entry points to an existing agent dir
+
+Requirements: Python 3.8+, PyYAML (pip install pyyaml).
+"""
+
+import sys
+import re
+import argparse
+import json
+from pathlib import Path
+
+# Enforce UTF-8 I/O on Windows
+for _stream in (sys.stdout, sys.stderr):
+    try:
+        _stream.reconfigure(encoding='utf-8')
+    except Exception:
+        pass
+
+try:
+    import yaml
+except ImportError:
+    print("FATAL: PyYAML not installed. Run: pip install pyyaml", file=sys.stderr)
+    sys.exit(1)
+
+
+# =============================================================================
+# Constants
+# =============================================================================
+
+# Allowed rating tokens (union of PROMPT_TEMPLATE.md spec + observed usage)
+ALLOWED_RATINGS = frozenset({
+    'ongoing',           # task in progress
+    'done',              # DevOps shorthand
+    'completed',         # standard completion
+    'success',           # engineer-chen style
+    'good',              # engineer-zhou / qa-xu style
+    'A', 'A+', 'A-',     # top grade
+    'B', 'B+', 'B-',     # mid grade
+    'C', 'C+', 'C-',     # low grade (spec says up to C)
+    'aborted',           # WORKFLOW.md §13.7
+})
+
+# Valid roles (for optional C5 check, not enforced by default)
+KNOWN_ROLES = frozenset({
+    '架构师', '工程师', '质量工程师', 'DevOps 工程师',
+    'UX/CLI 设计师', '安全工程师', '技术作家',
+})
+
+
+# =============================================================================
+# Path helpers
+# =============================================================================
+
+def _repo_root():
+    return Path(__file__).resolve().parent.parent
+
+
+def _agents_dir():
+    return _repo_root() / 'agents'
+
+
+# =============================================================================
+# YAML front matter extraction
+# =============================================================================
+
+def _extract_front_matter(filepath):
+    """Return (parsed_dict, error_string).
+    On success: (dict, None).  On failure: (None, 'reason string')."""
+    try:
+        text = filepath.read_text(encoding='utf-8')
+    except (OSError, UnicodeDecodeError) as e:
+        return None, f"read error: {e}"
+
+    m = re.match(r'^---\s*\n(.*?)\n---', text, re.DOTALL)
+    if not m:
+        return None, "no YAML front matter (missing --- delimiters)"
+
+    raw = m.group(1)
+    try:
+        parsed = yaml.safe_load(raw)
+    except yaml.YAMLError as e:
+        return None, f"YAML parse error: {e}"
+
+    if parsed is None:
+        return None, "YAML front matter is empty"
+
+    if not isinstance(parsed, dict):
+        return None, f"YAML front matter is not a mapping (got {type(parsed).__name__})"
+
+    return parsed, None
+
+
+# =============================================================================
+# Check C1: YAML parse
+# =============================================================================
+
+def check_yaml_parse(agents_dir):
+    """Return list of (severity, file, msg) tuples."""
+    findings = []
+
+    # Profile files
+    for child in sorted(agents_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith('.') or child.name == 'groups':
+            continue
+        pf = child / 'profile.md'
+        if not pf.is_file():
+            findings.append(('warn', str(pf), 'profile.md not found'))
+            continue
+        result, err = _extract_front_matter(pf)
+        if result is None:
+            findings.append(('error', str(pf), err))
+        else:
+            required = ['agent_id', 'name', 'role']
+            for key in required:
+                if key not in result:
+                    findings.append(('error', str(pf), f"missing required field '{key}'"))
+            if 'performance_log' not in result or result['performance_log'] is None:
+                findings.append(('warn', str(pf), "missing performance_log"))
+
+    # Group files
+    groups_dir = agents_dir / 'groups'
+    if groups_dir.is_dir():
+        for gf in sorted(groups_dir.glob('grp-*.md')):
+            result, err = _extract_front_matter(gf)
+            if result is None:
+                findings.append(('error', str(gf), err))
+            else:
+                required = ['group_id', 'name', 'lead', 'mission']
+                for key in required:
+                    if key not in result or result[key] is None:
+                        findings.append(('error', str(gf), f"missing required field '{key}'"))
+
+    return findings
+
+
+# =============================================================================
+# Check C2: rating range
+# =============================================================================
+
+def check_rating_range(agents_dir):
+    """Return list of (severity, file, msg) tuples."""
+    findings = []
+
+    for child in sorted(agents_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith('.') or child.name == 'groups':
+            continue
+        pf = child / 'profile.md'
+        if not pf.is_file():
+            continue
+        result, err = _extract_front_matter(pf)
+        if result is None or not isinstance(result, dict):
+            continue
+
+        perf_log = result.get('performance_log', [])
+        if not perf_log:
+            continue
+
+        for i, entry in enumerate(perf_log):
+            if not isinstance(entry, dict):
+                findings.append(('error', str(pf), f'perf_log[{i}] is not a mapping'))
+                continue
+            rating = entry.get('rating')
+            if rating is None:
+                findings.append(('error', str(pf), f'perf_log[{i}] missing rating'))
+            elif str(rating).strip() not in ALLOWED_RATINGS:
+                findings.append(
+                    ('warn', str(pf),
+                     f'perf_log[{i}] rating="{rating}" not in allowed set'))
+
+    return findings
+
+
+# =============================================================================
+# Check C3: current_groups -> groups/*.md
+# =============================================================================
+
+def check_group_refs(agents_dir):
+    """Return list of (severity, file, msg) tuples."""
+    findings = []
+    groups_dir = agents_dir / 'groups'
+
+    # Collect valid group_ids
+    valid_groups = set()
+    if groups_dir.is_dir():
+        for gf in sorted(groups_dir.glob('grp-*.md')):
+            result, err = _extract_front_matter(gf)
+            if result is not None and isinstance(result, dict):
+                gid = result.get('group_id')
+                if gid:
+                    valid_groups.add(str(gid).strip())
+
+    for child in sorted(agents_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith('.') or child.name == 'groups':
+            continue
+        pf = child / 'profile.md'
+        if not pf.is_file():
+            continue
+        result, err = _extract_front_matter(pf)
+        if result is None or not isinstance(result, dict):
+            continue
+
+        current_groups = result.get('current_groups', [])
+        if not current_groups:
+            continue
+
+        for g in current_groups:
+            gid = str(g).strip()
+            # Strip parenthetical annotations like "grp-xxx (inactive)"
+            gid_clean = re.sub(r'\s*\(.*\)', '', gid).strip()
+            if gid_clean and gid_clean not in valid_groups:
+                findings.append(
+                    ('error', str(pf),
+                     f'current_groups references unknown group "{gid_clean}"'))
+
+    return findings
+
+
+# =============================================================================
+# Check C4: group members -> agents/*/
+# =============================================================================
+
+def check_member_refs(agents_dir):
+    """Return list of (severity, file, msg) tuples."""
+    findings = []
+    groups_dir = agents_dir / 'groups'
+
+    # Collect valid agent_ids
+    valid_agents = set()
+    for child in sorted(agents_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith('.') or child.name == 'groups':
+            continue
+        if (child / 'profile.md').is_file():
+            valid_agents.add(child.name)
+
+    if not groups_dir.is_dir():
+        return findings
+
+    for gf in sorted(groups_dir.glob('grp-*.md')):
+        result, err = _extract_front_matter(gf)
+        if result is None or not isinstance(result, dict):
+            continue
+
+        members = result.get('members', [])
+        lead = result.get('lead')
+
+        # Check lead
+        if lead and str(lead).strip() not in valid_agents:
+            findings.append(
+                ('error', str(gf),
+                 f'lead "{lead}" is not a valid agent_id'))
+
+        # Check members
+        for m in (members or []):
+            mid = str(m).strip()
+            if mid and mid not in valid_agents:
+                findings.append(
+                    ('error', str(gf),
+                     f'member "{mid}" is not a valid agent_id'))
+
+    return findings
+
+
+# =============================================================================
+# Check C5: duplicate IDs (bonus safety net)
+# =============================================================================
+
+def check_duplicate_ids(agents_dir):
+    """Check for duplicate agent_id / group_id across files."""
+    findings = []
+
+    agent_ids = {}
+    for child in sorted(agents_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith('.') or child.name == 'groups':
+            continue
+        pf = child / 'profile.md'
+        if not pf.is_file():
+            continue
+        result, err = _extract_front_matter(pf)
+        if result is None or not isinstance(result, dict):
+            continue
+        aid = result.get('agent_id')
+        if aid:
+            aid = str(aid).strip()
+            if aid in agent_ids:
+                findings.append(
+                    ('error', str(pf),
+                     f'duplicate agent_id "{aid}" (also in {agent_ids[aid]})'))
+            else:
+                agent_ids[aid] = str(pf)
+
+    # Also verify dir name matches agent_id
+    for child in sorted(agents_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith('.') or child.name == 'groups':
+            continue
+        pf = child / 'profile.md'
+        if not pf.is_file():
+            continue
+        result, err = _extract_front_matter(pf)
+        if result is None or not isinstance(result, dict):
+            continue
+        aid = result.get('agent_id')
+        if aid and str(aid).strip() != child.name:
+            findings.append(
+                ('warn', str(pf),
+                 f'directory name "{child.name}" != agent_id "{str(aid).strip()}"'))
+
+    # Group ID duplicates
+    groups_dir = agents_dir / 'groups'
+    group_ids = {}
+    if groups_dir.is_dir():
+        for gf in sorted(groups_dir.glob('grp-*.md')):
+            result, err = _extract_front_matter(gf)
+            if result is None or not isinstance(result, dict):
+                continue
+            gid = result.get('group_id')
+            if gid:
+                gid = str(gid).strip()
+                if gid in group_ids:
+                    findings.append(
+                        ('error', str(gf),
+                         f'duplicate group_id "{gid}" (also in {group_ids[gid]})'))
+                else:
+                    group_ids[gid] = str(gf)
+
+    return findings
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='Check agents/ metadata integrity (profile.md + groups/*.md).'
+    )
+    parser.add_argument(
+        '--strict', action='store_true',
+        help='Treat warnings as errors (exit 2 -> exit 1).'
+    )
+    parser.add_argument(
+        '--json', action='store_true',
+        help='Machine-readable JSON output.'
+    )
+    args = parser.parse_args()
+
+    agents_dir = _agents_dir()
+    if not agents_dir.is_dir():
+        print(f'ERROR: agents/ not found at {agents_dir}', file=sys.stderr)
+        sys.exit(1)
+
+    check_suites = [
+        ('C1', 'YAML parse',         check_yaml_parse),
+        ('C2', 'rating range',       check_rating_range),
+        ('C3', 'group refs',         check_group_refs),
+        ('C4', 'member refs',        check_member_refs),
+        ('C5', 'duplicate IDs',      check_duplicate_ids),
+    ]
+
+    all_findings = []
+    for code, label, fn in check_suites:
+        findings = fn(agents_dir)
+        all_findings.extend((code, label, f) for f in findings)
+
+    errors = [f for f in all_findings if f[2][0] == 'error']
+    warnings = [f for f in all_findings if f[2][0] == 'warn']
+
+    if args.json:
+        output = {
+            'passed': len(errors) == 0 and (not args.strict or len(warnings) == 0),
+            'errors': [
+                {'check': f[0], 'suite': f[1], 'file': f[2][1], 'message': f[2][2]}
+                for f in errors
+            ],
+            'warnings': [
+                {'check': f[0], 'suite': f[1], 'file': f[2][1], 'message': f[2][2]}
+                for f in warnings
+            ],
+            'summary': {
+                'total_errors': len(errors),
+                'total_warnings': len(warnings),
+                'checks_ran': 5,
+            }
+        }
+        print(json.dumps(output, ensure_ascii=False, indent=2))
+    else:
+        if not all_findings:
+            print('OK: All 5 metadata checks passed.', file=sys.stderr)
+        else:
+            for code, label, (sev, filepath, msg) in all_findings:
+                tag = 'ERROR' if sev == 'error' else 'WARN'
+                print(f'[{code}] {tag}: {filepath}: {msg}', file=sys.stderr)
+            print(
+                f'\nSummary: {len(errors)} errors, {len(warnings)} warnings',
+                file=sys.stderr
+            )
+
+    if errors:
+        sys.exit(1)
+    if args.strict and warnings:
+        sys.exit(2)
+    sys.exit(0)
+
+
+if __name__ == '__main__':
+    main()