About review-pr-comments
A coding-agent skill for handling feedback on an existing pull request. It guides triage, batched fixes, CI validation, and stopping after two rounds without making unauthorized decisions.
Key features
- 01Collect reviews, threads, and failing checks
- 02Classify findings for fixes, replies, or escalation
- 03Validate, commit, push, and monitor CI
- 04Stop pushing after two rounds
Use cases
- 01Address correctness or security feedback on a PR
- 02Respond to style nits without creating a commit
- 03Escalate conflicting guidance or scope decisions
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ hcaiano/ skills --skill review-pr-commentsAfter installing, type/review-pr-commentsin Claude Code to confirm it loaded.
Bundled filesagents/openai.yaml
Details
From the project READMEReview PR Comments
Handle real review feedback on an existing PR. This is an on-demand, exception
path: most PRs have no comments because cloud auto-review is disabled and the
heavy review already happened locally (ship-it's delivery gate). Do not
summon bot reviews (@codex review, @coderabbitai review), and never treat a
bot's silence as something to wait for.
Process, at most TWO rounds (a round = collect → triage → one batched commit → push → required CI green on the new head):
- Collect all open feedback: human reviews, bot comments, unresolved and outdated threads, failing checks.
- Triage each finding against the code:
- Fix in the batch: real correctness, security, or data-integrity issues, and anything a human reviewer explicitly requested.
- Reply, don't push: style nits, defensive hardening for invariants that already hold, severity-inflated or duplicate findings. Answer the thread, resolve it, move on. A nit is never a reason for a push.
- Surface to the user: product/architecture decisions, scope expansions, conflicting reviewer guidance.
- Validate the complete head (original changes plus fixes), commit and push once, and wait for required CI (poll at 60–120 s intervals, never tight loops).
After two rounds, stop pushing: report remaining items with your triage and recommendation. Feedback arriving after your final push does not reopen the loop. Success: required checks green on the final head, human-requested changes addressed, every open thread fixed or answered, and GitHub reports the PR mergeable — bot re-reviews are never a completion requirement.
Do not merge, force-push, modify main, create follow-up PRs (nit-cleanup
follow-up PRs are explicitly banned), or make product and architecture
decisions without authorization.