wenjunduan /athena-review — Review stage (v9.6.4)
触发
impl stage 完成 (checklist.yaml 全 completed + tests passing) → 主 agent 进 review.
或 Stop hook prompt 自动检测 + 推进:
stage = impl AND all checklist done AND tests passing → suggest next_stage = review
6 维度 (vs v9.6.2 的 5 维度)
| # | 维度 | 由谁负责 |
|---|---|---|
| 1 | Correctness (代码正确) | reviewer subagent |
| 2 | Security (安全) | reviewer |
| 3 | Test (测试) | reviewer |
| 4 | Design 一致 (代码符合 design) | reviewer |
| 5 | Quality (质量) | reviewer |
| 6 | Spec Compliance (文档覆盖, v9.6.4 新) | spec-compliance subagent |
工作流 (3 subagent 并行)
Step 1: 并行 spawn 3 subagent
# 主 agent 用 Task tool 并行调用
spawn_parallel([
Task(subagent_type="reviewer", prompt="..."), # 1-5 维度
Task(subagent_type="spec-compliance", prompt="..."), # 6 维度 (新)
])
# 等两个完成后, spawn evaluator (后跑, 综合两家)
Task(subagent_type="evaluator", prompt="...")
Step 2: reviewer 输出
写入 sprints/{slug}/reviews/pass1.md, 段:
## Reviewer (代码层 findings)- 按 P0/P1/P2 列 finding
Step 3: spec-compliance 输出 (借 CodeStable cs-feat-accept + OpenSpec /opsx:verify)
读 sprints/{slug}/design.md 提取 ## 验收标准 + ## 实现要点 + ## File Structure Plan.
Bash git diff main...HEAD --stat + --name-only.
逐项对比, 写 sprints/{slug}/reviews/pass1.md 段:
## Spec Compliance (spec-compliance, {timestamp})### MISSING (做少了)### EXTRA (做多了, 分合理 refactor / scope creep)### DEVIATED (做偏了)### 总评 (PASS | REWORK)
Step 4: evaluator 综合 VERDICT
evaluator 读 reviews/pass1.md 的两段 (reviewer + spec-compliance), 综合给 VERDICT:
| VERDICT | 含义 | 下一步 |
|---|---|---|
| PASS | 全部维度合规 | next_action = polish (Refactor/System) 或 ship |
| CONCERNS | 有 P1/P2 但可推后 | next_action = polish (Refactor/System) 或 ship |
| REWORK | 有 P0 或 MISSING/DEVIATED | next_action = rework_impl |
| FAIL | 严重违规 (e.g. 安全漏洞) | next_action = rework_impl, 标 _index.major_issue |
evaluator 写 _index.next_action.
Stop hook prompt 类型联动
stage=review 时, Stop hook prompt 自动检查:
- 若
pass1.md缺## Spec Compliance段 → block + 提示 "spawn spec-compliance subagent" - 若 spec-compliance + reviewer 都跑了但 evaluator 没跑 → block + 提示 "spawn evaluator"
多轮 review (REWORK 后)
REWORK → 主 agent 回 impl 修, 再进 review:
- 新建
reviews/pass2.md(不覆盖 pass1.md) - evaluator 比 pass1 与 pass2 看是否改善
- ≥3 轮 REWORK 仍不过 → 主 agent 必须征求用户介入
不要做
- ❌ 主 agent 自己写 review.md (3 subagent 各自写)
- ❌ 跳过 spec-compliance (Stop hook 会 block)
- ❌ evaluator 早于 reviewer + spec-compliance 跑 (顺序错)
- ❌ 用同一个 reviewer 跑多次 (要换 subagent_type, 借 OMO Metis 多 critic 思想)
例外
- 路径 = Hotfix: 跳过 review (没时间)
- 路径 = Bugfix / Quick: 仅跑 reviewer, 不跑 spec-compliance (无 design.md)
- 路径 = Feature / Refactor / System: 全跑 (3 subagent)