Agent Skills › areal-project/AReaL › review-pr

review-pr

GitHub

用于只读审查Pull Request，包含风险分析、目标检查表和Codex专家子代理咨询。遵循硬规则不修改环境，通过五个阶段（上下文解析、变更分析、审查规划、专家咨询、最终审查）输出结构化报告，按严重程度排序发现的问题及修复建议。

.agents/skills/review-pr/SKILL.md areal-project/AReaL

触发场景

用户请求对当前分支或特定PR进行代码审查用户指定PR编号并要求评估风险

安装

npx skills add areal-project/AReaL --skill review-pr -g -y

更多选项

非标准路径

npx skills add https://github.com/areal-project/AReaL/tree/main/.agents/skills/review-pr -g -y

不安装直接使用

npx skills use areal-project/AReaL@review-pr

指定 Agent (Claude Code)

npx skills add areal-project/AReaL --skill review-pr -a claude-code -g -y

安装 repo 全部 skill

npx skills add areal-project/AReaL --all -g -y

预览 repo 内 skill

npx skills add areal-project/AReaL --list

SKILL.md

Frontmatter

{
    "name": "review-pr",
    "description": "Read-only pull request review workflow with risk analysis, targeted checklists, and Codex subagent consultation."
}

Review Pull Request

Use this skill when the user asks for a PR review of the current branch or a specific PR.

Inputs

Optional PR number
Optional --quick to stop after the change analysis phase

Hard Rules

Stay read-only.
Do not edit files, commit, push, rebase, or change GitHub state.
Do not run build, install, or test commands that mutate the environment.
Use gh for PR metadata and git diff retrieval.

Reference Files

references/review-pr-domains-and-signals.md
references/review-pr-templates.md

Workflow

Phase 1: Resolve PR context

Use gh pr view to fetch PR title, body, state, draft status, and changed files.
If no PR exists, stop and report that clearly.
If the PR is closed, stop.
Record the branch name and changed file list.

Phase 2: Change analysis

Classify changed files using references/review-pr-domains-and-signals.md.
Determine the highest overall risk level: CRITICAL, HIGH, MEDIUM, or LOW.
Build a CHANGE_ANALYSIS_REPORT that lists:
- detected domains/signals
- risk level
- affected files
- related frameworks
- likely failure modes

If --quick is set, return the change analysis report and stop here.

Phase 3: Review planning

Select the smallest useful set of review passes from references/review-pr-templates.md.
Split by risk area, not by file count.
Always include at least one general logic pass.

Phase 4: Expert consultation

Consult the matching Codex subagents registered in .codex/config.toml when relevant:

archon-expert
fsdp-expert
megatron-expert
algorithm-expert
launcher-expert

If the Codex runtime supports parallel subagent execution, run independent review passes in parallel. Otherwise, execute them serially.

Phase 5: Final review

Produce findings first, ordered by severity:

CRITICAL
HIGH
MEDIUM
LOW

For every finding, include:

file path
line number when available
why it is a bug, regression, or risk
concrete fix direction

What to Ignore

Pure style nits with no correctness impact
Issues outside the changed scope unless the PR makes them worse
Failures that standard linters or formatters would already catch
Speculative concerns with no concrete trigger in the diff

Output Shape

Use this structure:

CHANGE_ANALYSIS_REPORT:
- detected_domains: [...]
- detected_signals: [...]
- risk_level: ...
- affected_files: [...]
- related_frameworks: [...]
- identified_risks: [...]

Findings
1. [severity] Title — path:line
   - Problem: ...
   - Fix: ...

Open Questions
- ...

Residual Risk
- ...