review-loop

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

Input: PR URL or number and any accepted requirements.

1. One coordinator owns the review loop, fixes and pushes. Run [babysit-pr](../babysit-pr/SKILL.md) alongside review: provide the PR and agreed scope, track findings through fixes or reasoned dismissals, and keep monitoring through the final review round. Treat its 15-minute completion window as final only after all planned reviews finish.
2. Review spec correctness, then clean code, using [pr-spec-review](../pr-spec-review/SKILL.md) and [pr-clean-code-review](../pr-clean-code-review/SKILL.md). Use Astra (`gpt-6-astra`) at medium effort for spec review and Opus 5.5 (`claude-opus-5-5`) at high effort for clean-code review, unless the user overrides these choices. Give each reviewer a task-owned read-only worktree at the recorded PR head, the intended base, requirements, and the skill. Request findings with priority, location, scenario, impact and suggested correction; return findings to the coordinator for validation and disposition.
3. After each review, finish disposition of its findings before selecting the next head. Repeat that review only when validated P0/P1 findings required fixes, up to three iterations per review type. Refresh the review worktree after its reader finishes; remove only clean worktrees owned by this run. Preserve unresolved blockers in the handoff.
4. Read and apply [html-communication](../html-communication/SKILL.md), supplying reviewed heads, finding titles, dispositions and brief dismissal reasons. Maintain one report at `./outputs/<task>/review-loop.html` throughout the loop. Finish with the PR status, remaining blockers and report link. Merge only with explicit user authorization.

From AmitRockBear/skills, MIT.

More from AmitRockBear/skills