PR review loop

Two rounds, then a human. Each round: wait, read, act, update the PR. Never a third round.

The profile

Read <main checkout>/.claude/ship-profile.json first (the main checkout: dirname of git rev-parse --path-format=absolute --git-common-dir). It gives:

• repo: owner/name. base: the branch AI review runs on PRs into.
• review.workflow: the AI review workflow (a reviewer posts findings, a verifier keeps or dismisses each, a bot posts one comment; cdr's pr-review skill is one). review.checks: the gates workflow (it may be the same workflow). Both run when a PR into base is opened and on every push to it, on review.runners, so the push that ends round 1 is what starts round 2's review.
• gates: the commands that must pass before you push (<wt> is the checkout you work in).
• review.flakes: known flaky failures, which get a rerun, not a code change.

No profile: say so in one line and stop.

Start

1. gh pr view <N> -R <repo> --json baseRefName,headRefName,createdAt,state. The base must be base: no AI review runs for a PR into another branch, so for one, skip the loop and say so in one line.
2. Work in the checkout or worktree that holds the PR's branch (the one it was created from).
3. Tell the user one line: "Watching PR #N: review round 1 of 2."

Each round (r = 1, then 2)

1. Wait. Run REPO=<repo> REVIEW_WORKFLOW=<review.workflow> CHECKS_WORKFLOW=<review.checks> ${CLAUDE_PLUGIN_ROOT}/bin/wait-review <N> <since> with run_in_background: true and timeout: 7200000. <since> is the PR's createdAt for round 1 and the time you recorded just before round 1's push for round 2. Don't poll: the session is re-invoked when it exits. The runners are shared, so a round can queue behind other PRs.
   • Exit 0: both runs finished; read on.
   • Exit 3: no review run appeared for the head (the runners are down, or the workflow trigger changed). Act on the comments it printed if any, then hand off, and tell the user the review never ran.
   • Exit 4: timed out. Hand off, saying so.
2. Read its output: the run results and every comment, inline comment and review since <since>. The AI review comment lists findings with the verifier's verdicts. A pr-review comment starts with <!-- cdr:pr-review verdict=<v> head=<sha> -->: request-changes means fix its Required changes, accept-with-suggestions means triage its Suggestions and Uncertain lines, accept means nothing from the review (checks skipped after a request-changes verdict are not failures). A failed checks run: gh run view <id> -R <repo> --log-failed | tail -200. If the PR merged while you waited (gh pr view <N> --json state), fixes can't go onto it. Open a follow-up PR against base from the same branch; the hook starts that PR's own loop. Post the hand-off on the merged PR, pointing to the follow-up.
3. Triage every finding the verifier kept and every comment that asks for a change. Check each against the code yourself: a verdict is not proof. Real: fix it. Not real: decline it, with file:line evidence. If nothing is actionable (no kept findings, no change requests, checks green), stop here and hand off.
4. Fix the way the repo does: a regression test shown failing without its fix, one finding per commit, the profile's gates.typecheck and gates.test green (and gates.smoke when gates.smokeWhen holds), no Claude co-author trailer. A checks failure that matches one of review.flakes gets gh run rerun <id> -R <repo> --failed, not a code change.
5. Update the PR. Record date -u +%Y-%m-%dT%H:%M:%SZ (round 2's <since>), then push. Post one PR comment that starts with <!-- pr-review-loop round r -->: what was fixed (with commit) and what was declined (with the reason). Edit the PR body if the fixes change what it says. No Claude attribution in either.
6. After round 1, go back to step 1 for round 2. After round 2, hand off: don't wait for or act on the review your round-2 push starts.

Hand off

Post a PR comment that starts with <!-- pr-review-loop done --> and leads with Needs human review. Give the rounds run, what was fixed, what was declined and why, anything still open, and, when round 2 pushed, that a review of that push is still coming. Then tell the user in chat: "PR #N needs human review", one line on where it stands, and the link. Never merge.