Review Pr

Skill: Review PR Comments

Fetch review comments from AI code review agents on a GitHub PR, categorize them by severity, and suggest fixes.

When to Use

• After CI passes and AI reviewers have posted comments on a PR.
• When you want a consolidated summary of all review feedback before addressing it.

Always check ALL of these reviewers (any combination may be active on a given repo):

• CodeRabbit (coderabbitai[bot])
• Gemini Code Assist (gemini-code-assist[bot])
• Codex (chatgpt-codex-connector[bot])
• Claude (claude[bot])
• Copilot (copilot-pull-request-reviewer[bot], github-actions[bot] running Copilot review actions)

A reviewer being silent or rate-limited is a valid finding — note it as "no comments" or "rate-limited", do not assume absence means nothing to address.

Inputs

The PR number is passed as an argument: /review-pr 53

If no argument is provided, detect the current branch and find the open PR for it.

Workflow

1. Fetch PR comments using gh CLI. Three surfaces — query all three, they hold different content:

   • Inline review comments: gh api repos/{owner}/{repo}/pulls/{number}/comments
   • Review bodies (summary verdicts): gh api repos/{owner}/{repo}/pulls/{number}/reviews
   • Issue-thread comments (Claude posts here, as do CodeRabbit summaries + rate-limit notices): gh api repos/{owner}/{repo}/issues/{number}/comments
   • Detect the repo owner/name from the current git remote.

2. Filter to AI reviewers. Use this jq pattern across all three surfaces:

   --jq '.[] | select(.user.login | test("(?i)^(coderabbit|gemini-code-assist|codex|chatgpt-codex-connector|claude|copilot|github-actions).*\\[bot\\]$"))'
   

   Always enumerate all five reviewers explicitly in the output (coderabbitai[bot], gemini-code-assist[bot], chatgpt-codex-connector[bot], claude[bot], copilot-pull-request-reviewer[bot]) — if one posted nothing, list it as "no comments" or "rate-limited" so the user can see at a glance which reviewers were active. Do not silently drop a reviewer just because it had no findings.

3. Categorize each comment:

   • Critical (must fix): Security issues, bugs, crashes, type errors.
   • Warning (should fix): Missing error handling, edge cases, best practices.
   • Suggestion (consider): Style, naming, refactoring ideas, future improvements.
   • Already addressed: Comments marked with "Addressed in commit" or similar, or whose target code no longer exists at the referenced line.

4. For each actionable comment, provide:

   • File and line reference.
   • What the reviewer flagged.
   • The suggested fix (code diff if provided by the reviewer).
   • Your assessment: agree, disagree with reason, or "style preference — skip".

5. Output format:

## PR #N Review Summary

### Reviewer status
| Reviewer | Findings |
|---|---|
| CodeRabbit | <count or "rate-limited" or "no comments"> |
| Gemini | <count or "no comments"> |
| Codex | <count or "no comments"> |
| Claude | <count, plus Approve/Comment/Request-changes verdict if posted> |
| Copilot | <count or "no comments"> |

### Critical (must fix)
1. **[file:line]** — [description] _(reviewer consensus: A + B)_
   Fix: [what to change]

### Warning (should fix)
1. **[file:line]** — [description] _(reviewer: X)_
   Fix: [what to change]

### Suggestion (consider)
1. **[file:line]** — [description] _(reviewer: X)_
   Assessment: [agree/skip with reason]

### Already Addressed
- [description] — resolved in commit [hash]

### Skipped (style preferences)
- [description] — [why skipping]

Important Notes

• Do NOT apply fixes automatically — present the summary for the user to decide. (If the user explicitly says "address them" / "fix them" / "apply suggestions", then apply.)
• Cross-reference suggestions with the project's CLAUDE.md coding standards.
• If two or more reviewers flag the same issue, surface the consensus in the line item (_(reviewer consensus: Claude + Codex + Gemini)_).
• Flag any suggestions that contradict the project's established patterns.
• Claude reviews live on the issue-thread endpoint, not the pull-request reviews endpoint. Skipping the issue-thread query will silently miss every Claude review.
• CodeRabbit posts its summary on the issue thread but its inline findings on the pulls/comments endpoint. Both must be checked.
• Some reviewers (Copilot, Gemini) post under multiple bot identities depending on action vs app configuration. If gh shows an unfamiliar *[bot] author with a substantive review body, treat it as a real reviewer until proven otherwise.

Handling Multi-Round Reviews

When a PR has been reviewed multiple times (e.g., after pushing fixes):

• Comments from earlier rounds may reference code that no longer exists (stale comments).
• Check if a comment's file path and line still match the current code.
• Mark stale comments as "Already addressed" rather than re-flagging them.
• Focus the summary on genuinely new comments from the latest review round.
• Each new push usually triggers a fresh review pass from Claude and Codex; CodeRabbit may rate-limit between rounds — note when this happens.

Source: hcslomeu/tfl-monitor — distributed by TomeVault.