Back to Details

council-review

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Council Review

委员会评审

Run Codex and Claude's own 
/review
in parallel, then cross-validate and synthesize into one unified report — like a review board where two reviewers examine the code independently, and Claude as lead reviewer delivers the final opinion.
并行运行Codex和Claude自带的
/review
能力，随后交叉验证并合成为一份统一报告——就像评审委员会中两名评审员独立审查代码，再由作为主评审的Claude给出最终意见。

Prerequisites

前置要求

  • Codex CLI: Install with 
    npm i -g @openai/codex
    , authenticate with 
    codex login
If only one CLI is installed, fall back to the available reviewer with a warning — the review still has value with fewer perspectives, so don't fail entirely.
  • Codex CLI: 执行
    npm i -g @openai/codex
    安装，通过
    codex login
    完成认证
如果仅安装了其中一个CLI，回退到可用的评审工具并给出警告——即使视角更少，评审结果仍然有价值，因此不要直接终止运行。

When to Use

适用场景

  • Reviewing uncommitted changes before committing
  • Auditing a branch diff before opening a PR
  • Reviewing a specific commit for regressions
  • 提交前评审未提交的变更
  • 开PR前审计分支diff
  • 评审特定提交是否存在回归问题

When NOT to Use

不适用场景

  • Reviewing documentation, markdown, or non-code files
  • Trivial single-line changes where a full council review would be overkill
  • 评审文档、markdown或非代码文件
  • 无关紧要的单行变更，完整的委员会评审属于大材小用

Workflow

工作流程

Do not read script source code. Run scripts directly and use 
--help
for usage.
不要读取脚本源代码。直接运行脚本，使用
--help
参数查看用法。

Step 1: Determine Review Scope

步骤1：确定评审范围

If the scope is not already clear, use AskUserQuestion to ask:
  • Uncommitted changes (default) — staged, unstaged, and untracked changes
  • Branch diff — compare current branch against a base branch
  • Specific commit — audit a single changeset
如果范围不明确，使用AskUserQuestion询问用户：
  • 未提交变更（默认）——暂存、未暂存和未跟踪的变更
  • 分支diff——将当前分支与基准分支对比
  • 特定提交——审计单个变更集

Step 2: Run Both Reviews in Parallel

步骤2：并行运行两项评审

Both reviewers read the same diff independently — neither depends on the other's output. Launch them both at once in a single message to eliminate sequential wait time.
Scripts are in 
scripts/
relative to this skill's directory and enforce the correct model and read-only mode internally. Run 
<script> --help
for full usage.
两名评审员独立读取同一份diff——互不依赖对方的输出。在同一条消息中同时启动两项任务，消除串行等待时间。
脚本位于本技能目录下的
scripts/
文件夹中，内部已配置正确的模型和只读模式。运行
<script> --help
查看完整用法。

Codex — 
scripts/codex-review.py
(background Bash task)

Codex — 
scripts/codex-review.py
（后台Bash任务）

Launch as a background Bash task (
run_in_background: true
). Codex CLI may take up to 30 minutes. When it completes, use the 
Read
tool on the 
output-file
path from the notification to retrieve the review.
bash
python3 scripts/codex-review.py uncommitted
python3 scripts/codex-review.py branch --base main
python3 scripts/codex-review.py commit <SHA>
作为后台Bash任务启动（
run_in_background: true
）。Codex CLI最多可能需要30分钟运行时间。运行完成后，使用
Read
工具读取通知中给出的
output-file
路径获取评审结果。
bash
python3 scripts/codex-review.py uncommitted
python3 scripts/codex-review.py branch --base main
python3 scripts/codex-review.py commit <SHA>

Claude — 
/review
skill (background Agent)

Claude — 
/review
技能（后台Agent）

Launch a background Agent (
run_in_background: true
) to run 
/review
on the same scope. Prompt the agent to invoke the 
/review
skill (via the Skill tool) and return its complete findings. The agent's output arrives directly in its completion notification.
启动后台Agent（
run_in_background: true
）在相同范围运行
/review
。提示Agent（通过Skill工具）调用
/review
技能并返回完整结果。Agent的输出会直接返回到其完成通知中。

Step 3: Cross-Validate Findings

步骤3：交叉验证评审结果

Once both reviews have returned, cross-validate:
  1. Validate external findings — For each finding from Codex:
    • Confirm — Claude independently agrees the issue exists and is correctly described.
    • Dispute — Claude believes the finding is a false positive or incorrectly categorized. Note the reasoning.
    • Enhance — The issue exists but the explanation or suggested fix can be improved. Provide the improved version.
  2. Add Claude's own findings — Include any issues from 
    /review
    that Codex didn't catch.
两项评审都返回结果后，进行交叉验证：
  1. 验证外部评审结果——针对Codex给出的每一项发现：
    • 确认——Claude也独立认可该问题存在且描述准确
    • 存疑——Claude认为该发现是误报或分类错误，记录原因
    • 优化——问题确实存在，但解释或建议修复方案可以优化，提供优化后的版本
  2. 补充Claude的自有发现——加入
    /review
    发现的、Codex未检出的所有问题

Step 4: Synthesize into Unified Report

步骤4：合成统一报告

After your own review and validation are complete, merge, deduplicate, and rewrite all findings into one coherent report as if written by a single reviewer. Do not copy-paste or concatenate raw outputs.
Load 
references/output-format.md
for the report template. Load 
references/merge-rules.md
for how to reconcile findings across reviewers.
完成自有评审和验证后，合并、去重、重写所有发现，形成一份连贯的报告，就像由单一评审员撰写的一样。不要直接复制粘贴或拼接原始输出。
加载
references/output-format.md
获取报告模板，加载
references/merge-rules.md
了解如何协调不同评审员的发现。

Rules

规则

  • Run both reviewers in parallel — Codex and 
    /review
    are independent reads of the same diff. Running them concurrently instead of sequentially saves the entire 
    /review
    execution time.
  • Use the same review scope for both reviewers — comparing different scopes would make deduplication meaningless.
  • Wait for both reviews to complete before cross-validation — the council's value depends on comparing complete outputs.
  • Write one unified opinion — the report should read as a single reviewer's assessment. Never structure findings by reviewer (no "Codex found..." sections).
  • Sort findings by priority — P0 → P1 → P2 → P3 → P4.
  • Exclude low-confidence findings — If Claude disputes an external finding or evidence is purely circumstantial, omit it from the report. The council's value is cross-validation; findings that fail it are noise.
  • Always use the wrapper script for Codex — do not call 
    codex
    CLI directly, because the script sets the correct model and read-only mode.
  • Suppress intermediate outputs — Do not display raw Codex or 
    /review
    outputs to the user. Running 
    /review
    in a subagent keeps its output out of the main conversation naturally. The only review output the user should see is the final unified report.
  • Never use 
    TaskOutput
    for background tasks     — 
    TaskOutput
    cannot find background Bash task IDs and will fail. Use the 
    Read
    tool on the 
    output-file
    path from the completion notification instead. For background Agents, read the result directly from the completion notification.
  • If a reviewer fails at runtime — stop the review, report the error to the user, and do not produce a council report. A single-reviewer result lacks cross-validation and should not be presented as a council review.
  • Do not poll or probe background tasks — Do not read output files, check process status, resume agents, or run any commands while waiting for background tasks. End your response after launching them. You will be notified automatically when each task completes.
  • 并行运行两个评审工具——Codex和
    /review
    独立读取同一份diff。并行运行而非串行运行可以节省整个
    /review
    的执行时间。
  • 两个评审工具使用相同的评审范围——对比不同范围会让去重失去意义。
  • 交叉验证前等待两项评审全部完成——委员会评审的价值建立在对比完整输出的基础上。
  • 输出统一的评审意见——报告读起来应该是单一评审员的评估结果，不要按评审员区分发现（不要出现「Codex发现...」这类章节）。
  • 按优先级排序发现——P0 → P1 → P2 → P3 → P4。
  • 排除低置信度发现——如果Claude对外部发现存疑，或是证据完全不充分，从报告中删除该内容。委员会评审的价值在于交叉验证，未通过验证的发现属于无效信息。
  • 始终使用Codex的封装脚本——不要直接调用
    codex
    CLI，因为脚本已经设置了正确的模型和只读模式。
  • 隐藏中间输出——不要向用户展示原始的Codex或
    /review
    输出。在子Agent中运行
    /review
    可以自然避免其输出出现在主会话中，用户唯一能看到的评审输出就是最终的统一报告。
  • 永远不要对后台任务使用
    TaskOutput
    ——
    TaskOutput
    无法获取后台Bash任务ID，会运行失败。对于后台Bash任务，使用
    Read
    工具读取完成通知中的
    output-file
    路径；对于后台Agent，直接从完成通知中读取结果。
  • 如果某一个评审工具运行时出错——停止评审，向用户上报错误，不要生成委员会评审报告。单一评审员的结果缺少交叉验证，不应作为委员会评审结果呈现。
  • 不要轮询或探测后台任务——等待后台任务完成期间，不要读取输出文件、检查进程状态、恢复Agent或是运行任何命令。启动任务后直接结束回复，每个任务完成后你会自动收到通知。