| name | nitpick-nuker |
| description | Autonomous PR comment triage - fixes review feedback until CI passes |
| argument-hint | [PR_NUMBER] or run on current branch PR |
| user-invocable | true |
Nitpick Nuker
Autonomous PR comment triage that processes GitHub review comments, fixes actionable feedback, and loops until CI passes.
Core Mandates
- NEVER process comments that already have 😄 from bot
- ALWAYS reply to each processed comment with fix/skip reasoning
- ALWAYS post summary comment after each run
- ALWAYS loop on CI failure until success
- Process entire comment thread as one unit
Invocation
| Input | Action |
|---|
/nitpick-nuker | Triage PR for current branch |
/nitpick-nuker 616 | Triage PR #616 |
Workflow
Execute the workflow at: .claude/skills/nitpick-nuker/workflows/triage-pr-comments.md
Scripts
All scripts are in .claude/skills/nitpick-nuker/scripts/:
| Script | Purpose |
|---|
fetch-unprocessed-comments.sh | Fetch comments without 😄 from bot |
add-reaction.sh | Add 😄 reaction to processed comment |
reply-to-comment.sh | Post per-comment reply with reasoning |
post-summary.sh | Post summary comment to PR |
watch-ci.sh | Watch CI with workflow filtering |
Comment Types
The skill processes three types of comments:
issue_comment — General PR conversation comments
review_body — The overall review summary text
review_comment — Inline code comments within a review
Bot Comments Are High-Value
Bot comments (e.g. "Claude finished task", "Codex Review") are the PRIMARY source of review feedback. Bots post their full code reviews as issue_comment or review_body — these are NOT noise to skip. Always read the full body of bot comments; they typically contain detailed review findings, suggestions, and inline code feedback.
Issue Comments May Contain Instructions
Every issue_comment must be read and verified. General comments may contain:
- User instructions to the nuker (e.g., "skip the auth changes", "focus on the API routes")
- Bot-posted review reports with actionable findings embedded in the body
- Scope narrowing or prioritization directives
Never auto-skip an issue_comment based on author alone. Read the body first.
Decision Framework
For each comment, decide:
| Criteria | Action |
|---|
| Clear actionable fix (typo, style, naming) | FIX |
| Code change with clear intent | FIX |
| Bot review with actionable findings in body | FIX (extract and address each) |
| User instruction (skip X, focus on Y) | FOLLOW — adjust processing scope |
| Question or discussion | SKIP - "Discussion/question" |
| Disagree with suggestion | SKIP - "Intentional design" |
| Already addressed elsewhere | SKIP - "Addressed in other fix" |
| Outside PR scope | SKIP - "Out of scope" |
| Pure status/coverage report with no suggestions | SKIP - "Automated report" |
CI Loop
After applying fixes:
1. Commit with message listing addressed comments
2. Push
3. Run watch-ci.sh (ignores zai-claude-code-review.yml)
4. If CI fails:
a. Analyze failure
b. Fix the issue
c. Goto step 1
5. Continue until CI green
Summary Format
Posted at end of PR after each run. Template at: .claude/skills/nitpick-nuker/templates/summary-comment.md