PR Review Recycle Loop

Use this skill when you need to process automated PR review feedback (Copilot, Greptile, Codex) and iterate until:

• All automated review threads are addressed (fixed, explained, or explicitly deferred).
• PR checks are green.
• The PR is ready for a human to merge.

Canonical workflow: .codex/prompts/automated-review-cycle.md. This file is a Claude Code mirror of .opencode/skills/pr-review-recycle/SKILL.md and exists for harness-specific standing rules. Precedence: if the canonical workflow conflicts with this wrapper (for example, around committing/pushing), this wrapper wins.

Standing Rules

• Prefix all GitHub prose (PR comments, issue comments, PR descriptions) with [AGENT] (repo rule, see AGENTS.md).
• Do not create commits unless explicitly asked by the user.
• Do not force push.
• Do not amend commits unless explicitly asked.
• Do not bypass git hooks (for example, do not use --no-verify).
• Do not commit secrets or .env files.

Inputs You Need

• PR number (or PR URL).
• Repo owner/name.

Practical Command Set

Snapshot state:

gh pr view <PR> --json url,title,headRefName,baseRefName,state,mergeable,updatedAt
gh pr view <PR> --json statusCheckRollup
gh pr diff <PR>

Collect review threads (paginate until hasNextPage=false):

gh api graphql -f query='query($owner:String!, $name:String!, $number:Int!, $after:String){ repository(owner:$owner,name:$name){ pullRequest(number:$number){ reviewThreads(first:100, after:$after){ totalCount pageInfo{hasNextPage endCursor} nodes{ id isResolved isOutdated comments(first:50){ nodes{ id author{login} body } } } } } } }' \
  -F owner=<OWNER> -F name=<REPO> -F number=<PR> -F after=<CURSOR_OR_null>

Resolve a thread after replying:

gh api graphql -f query='mutation($threadId:ID!){ resolveReviewThread(input:{threadId:$threadId}){ thread{ id isResolved } } }' \
  -F threadId=<THREAD_ID>

Minimal Loop Checklist

Repeat until stable:

• Snapshot: gh pr view + gh pr diff.
• Collect:
  • Review threads via GraphQL (paginate until hasNextPage=false).
  • PR conversation issue comments (Greptile sometimes posts follow-ups there).
• Triage: decide must-fix, should-fix, nit, or defer (with a tracking issue).
• Implement smallest safe fix, add/adjust tests.
• Run the smallest proving checks, then (if needed) the broader gate:

yarn lint:check
yarn test
yarn build:all

• If UI or visual tests changed:

yarn test:e2e
yarn test:visual

• Refresh review threads and issue comments again before declaring done.
• Reply + resolve all AI bot review threads (even if rejecting the suggestion).
• Wait for checks to go green, then wait at least 5 minutes after the latest push and do one last refresh to catch late bot updates.

Repo Notes (Chronote)

• Visual regression is prone to flake if snapshots are updated casually, only run yarn test:visual:update after confirming the UI change is correct.
• If you changed frontend UI, follow AGENTS.md expectations for Storybook + screenshot capture.
• When checking for unresolved AI threads, always paginate reviewThreads, do not assume first: 100 is enough.

Permissions Note

Inline replies and resolveReviewThread can fail due to token permissions.

Fallback behavior:

• Reply where possible.
• Add a PR-level [AGENT] comment listing what you addressed, and ask a maintainer to resolve remaining threads.