gh-ai-review-triage

name	gh-ai-review-triage
description	Triage and optionally address AI-generated GitHub PR review feedback from Copilot, CodeRabbit, or similar bots. Use when the user asks to check AI review comments, decide whether comments require action, ignore non-actionable suggestions, fix required/should-fix/worth-fixing feedback, or summarize bot review findings on a pull request.

GitHub AI Review Triage

Goal

Separate AI review noise from useful feedback. Inspect review threads, classify each comment, fix comments that are required, should be addressed, or are clearly worth addressing, and leave optional/nitpick comments alone unless the user asks to address them.

Workflow

1. Resolve the target PR.

   • Use a PR number or URL from the user when provided.
   • Otherwise infer the PR from the current branch and repository.
   • Prefer thread-aware data over flat comments when available, because resolution and outdated state matter.

2. Collect review context.

   • Fetch inline review threads with resolved/outdated state.
   • Fetch top-level PR comments and review submissions to catch bot summaries, approvals, and nitpick-only reviews.
   • Note the author for each item. Treat Copilot, CodeRabbit, and similar bot comments as advisory until verified.
   • Extract binding constraints from the user request, explicit maintainer corrections, and accepted canonical decisions. Use them as the classification boundary rather than treating them as advisory inputs equal to bot feedback.

3. Use the best available retrieval method.

   • If GitHub connector tools are available, prefer _list_pull_request_review_threads for inline threads and _fetch_pr_comments for the merged PR timeline.
   • If the GitHub MCP server is available, use get_pull_request_comments and get_pull_request_reviews as a fallback; note that flat comments may not fully preserve thread resolution state.
   • If only gh is available, use gh pr view to resolve the PR, then gh api graphql for reviewThreads and REST endpoints for /pulls/{number}/comments, /pulls/{number}/reviews, and /issues/{number}/comments.
   • If gh auth or network access fails, report the blocker and ask for re-authentication or permission instead of guessing from incomplete data.

4. Classify each comment.

   • Required: correctness bug, build failure, clippy/lint warning that reproduces, test failure, security/soundness issue, or behavior that conflicts with the user request. Note: a comment may only be classified Required after verification (see Step 5); if verification is not feasible, classify as Should Fix with a 'needs verification' tag instead.
   • Should Fix: non-blocking but clearly correct feedback that removes misleading code, unrealistic tests, stale comments, fragile behavior, or likely reviewer follow-up.
   • Worth Fixing: small low-risk change that improves clarity, test realism, performance in a hot path, user-visible behavior, or future maintainability.
   • Optional: nitpick, style preference, docstring/coverage suggestion outside project requirements, or UX improvement with fallback already implemented.
   • Ignore: outdated, resolved, duplicate, incorrect after local verification, unrelated to this PR, or conflicts with a binding maintainer constraint without evidence that the decision must be reopened.

5. Verify before editing.

   • Check the proposed edit against the binding constraints before judging only its local technical correctness. A plausible implementation can still be the wrong product behavior or scope.
   • Reproduce claimed CI-impacting issues locally when feasible.
   • Inspect the referenced code, not only the bot wording.
   • If a bot says "warnings-as-errors" or similar, run the relevant check before treating it as required.
   • Only mark Required if the issue is verifiable with reproduction/test evidence or deterministically verifiable from available data; if verification isn't possible, downgrade to Should Fix with a 'needs verification' tag (or Ignore if clearly stale/duplicate).
   • If a comment conflicts with a binding constraint, preserve the constraint and report the conflict. If the comment exposes new correctness, safety, or consistency evidence, return that evidence to the maintainer instead of silently overriding the direction or implementing a compromise that weakens it.

6. Decide action.

   • Treat Required, Should Fix, and Worth Fixing as fix targets by default. Note: fix targets assume verification feasibility has been assessed per Step 5.
   • If the user explicitly says "必須だけ" or "only required", implement only Required items and report the skipped Should Fix / Worth Fixing items.
   • Do not address Optional or Ignore items unless the user explicitly asks.
   • Do not post replies, resolve threads, or submit reviews on GitHub unless the user explicitly requests that write action.

7. Implement and validate when changes are needed.

   • Keep fixes narrowly scoped to the review comment.
   • Run the smallest relevant tests/checks first, then broader checks if the fix touches shared behavior.
   • Commit/push only when the user asks for that.

Output Format

When only triaging, summarize in Japanese if the user is using Japanese:

確認しました。修正対象は N 件です。

- Required: ...
- Should Fix: ...
- Worth Fixing: ...
- Optional/Ignore: ...

対応する/しない理由:
...

When changes are made, include:

• which comments were addressed
• which comments were intentionally ignored and why
• validation commands and results
• whether GitHub replies/resolution were left untouched

Heuristics

• AI approvals and bot overview comments usually do not require code changes.
• CodeRabbit "Nitpick" sections are optional by default, but promote them to Worth Fixing when the change is clearly safe and improves the PR directly.
• "Consider ..." wording is usually optional unless it points to a real bug, misleading test/comment, measurable hot-path issue, or low-risk improvement with clear value.
• A stale or incorrect bot claim should be called out briefly and left unchanged.
• For Rust PRs, local cargo clippy success is stronger evidence than an AI claim about unused imports.