| name | monitoring-pr-ai-reviews |
| description | Use when implementation is already complete, a GitHub PR exists or must be opened, and follow-up work is still needed because Copilot or other AI review comments may arrive after the initial push. |
Monitoring PR AI Reviews
Overview
Close the loop after implementation: open the PR, keep AI review visible on a schedule, and handle each suggestion with compatibility-first judgment instead of blind agreement.
Core principle: AI review comments are inputs to evaluate, not orders. Fix real bugs, races, contract gaps, and missing tests. Reject speculative public APIs, overdesign, or premature abstractions.
REQUIRED SUB-SKILLS: github:gh-address-comments, receiving-code-review, verification-before-completion
When to Use
Use this skill when:
- the task code is already implemented and verified
- a PR exists or must be created now
- Copilot or other AI review comments may still arrive after the first push
- the remaining work is PR refresh, comment triage, fixes, replies, and thread resolution
Do not use this skill for:
- initial feature design
- implementing the original task from scratch
- human review that changes product direction and needs a new design decision
The Flow
- Confirm the task is really ready for PR.
- Run fresh verification.
- Open a non-draft PR.
- Keep periodic monitoring enabled.
- Scheduled workflow:
.github/workflows/pr-ai-review-monitor.yml
- Local or CI scan from this skill directory:
python3 ./scripts/pr_ai_review_monitor.py --repo <owner/name> [--pr <number>] [--ai-login <login>]
- Manual workflow run:
gh workflow run pr-ai-review-monitor.yml
- Read review threads, not flat comments.
- Keep thread ids, file anchors, outdated state, and resolution state.
- Evaluate each AI suggestion against repo principles.
- preserve user-visible compatibility first
- prefer the smallest fix that closes the real issue
- keep internal test seams internal unless a real external caller needs them
- reject "future provider" abstractions without concrete current need
- Implement only justified changes.
- Tie each code change to a review thread.
- Run the narrowest proving test first, then broader repo gates.
- Refresh the PR.
- push
- reply with fix or rejection rationale
- resolve threads only after the code or rationale is visible on the PR
- Re-scan until no unresolved AI review remains, or only explicit user-owned decisions remain.
Quick Reference
Commands below are relative to this skill directory.
| Need | Action |
|---|
| Open or refresh PR | create a non-draft PR from the verified task branch |
| Scan all open PRs | python3 ./scripts/pr_ai_review_monitor.py --repo <owner/name> |
| Scan one PR | python3 ./scripts/pr_ai_review_monitor.py --repo <owner/name> --pr <number> |
| Include another AI reviewer login | python3 ./scripts/pr_ai_review_monitor.py --repo <owner/name> --ai-login <login> |
| Resolve a review thread | gh api graphql -f query='mutation($id:ID!){resolveReviewThread(input:{threadId:$id}){thread{isResolved id}}}' -F id=<thread-id> |
| Re-check after push | run the monitor again and confirm zero unresolved AI review threads |
Evaluation Rules
Accept when the comment identifies:
- a real bug, race, broken contract, missing verification, or missing typed error handling
- a concrete startup or runtime failure for actual callers
- spec drift or parity drift
Reject when the comment mainly asks for:
- new exported APIs without current callers
- public hooks whose only consumer is same-package test code
- cross-provider abstractions added "for later"
- refactors that widen task scope after the task is already complete
Common Mistakes
- resolving threads before the fix or rationale is visible in the PR
- treating Copilot suggestions as mandatory
- expanding the public API to satisfy test-only feedback
- re-running only package tests after a repo-wide behavioral change
- forgetting that new AI review may arrive after the first fix push