Review loop

/review-loop — drive a PR to a green review

Run the review loop on a pull request in pappcorn/universe: wait for the automated gate, fix or dispute every blocking finding, and loop until the PR is green — then hand it to the human reviewers.

Arguments ($ARGUMENTS, all optional, any order):

• A bare PR number or URL → the PR to run on. Default: the PR for the current branch.
• --max N → max fix/dispute iterations before escalating. Default: 3.
• --dry-run → poll and report the findings, but do not fix, dispute, or push.

---

The gate contract (this repo)

• Check: claude-review, from the workflow "Claude Code Review" (.github/workflows/claude-code-review.yml). It runs on every PR opened / synchronize event, posts a review comment, and fails the check unless the review verdict is PASS.
• Findings: the review is a PR comment containing a ## Findings section. Always read the latest ## Findings comment posted after the current head commit's run started — an older comment describes an older diff.
• Severities:
  • 🔴 Important — blocks merge. The check fails while any exists.
  • 🟡 Nit — does not block. Fix if cheap, otherwise leave a short comment.
  • 🟣 Pre-existing — informational, never blocks.
• Determinism: the gate grades the pushed diff. Same diff → same verdict. The only way to flip a failing check is to change the code and push. A comment triggers a conversation (@claude replies via .github/workflows/claude.yml), never a re-grade.

---

Step 0 — Resolve the PR

• If $ARGUMENTS has a number/URL, use it; else gh pr view on the current branch. No PR → stop and say so (this skill does not open PRs).
• Confirm the PR is OPEN. Record its number, base, and head branch.

Step 1 — Make the branch current and the diff pushed

The gate grades the pushed diff against the base.

1. git fetch origin <base> and check git log --oneline HEAD..origin/<base>. If the base moved ahead, merge it in (git merge origin/<base> or gh pr update-branch). Trivial conflicts only — anything non-obvious stops the loop and goes to a human.
2. Push any unpushed commits with a plain git push. Never force-push.
3. Uncommitted changes in the tree → stop and tell the user; the loop grades committed, pushed code.

Step 2 — Wait for the gate and CI to settle

gh pr checks <PR> --watch --interval 30 --json name,state,bucket,workflow,link

--watch blocks until every check completes and exits non-zero if any failed — that exit code is a signal, not an error. Capture the JSON either way. If the claude-review check never appears after CI completes, surface that instead of treating it as green.

Step 3 — Read the verdict and the findings

• Gate: the claude-review check. Green iff its bucket is pass.
• CI: the CI workflow check (nx affected -t lint test build). For a red CI check, pull the failing detail: gh run view <run-id> --log-failed.
• Findings: fetch the newest qualifying review comment:

gh api "repos/pappcorn/universe/issues/<PR>/comments" --paginate \
  --jq '[.[] | select(.body | contains("## Findings"))] | last | .body'

Verify it postdates the current head commit's review run (compare created_at against the run's start time from gh run list). Split findings by severity.

Red without a verdict — infrastructure failures

claude-review can fail before any review runs. If the check is red and no ## Findings comment postdates the run, read the failure log (gh run view <run-id> --log-failed) and classify instead of looping:

• The PR modifies claude-code-review.yml itself. The action's app-token exchange requires the workflow file on the PR head to be byte-identical to the one on the default branch, and otherwise fails with 401 Workflow validation failed. That is the action's tamper guard, not a review verdict — no push can turn the check green while the workflow change is part of the PR. Post the classification as a PR comment so the human reviewers know the red is expected, then stop; the check heals on the first PR after the workflow change merges.
• Any other pre-review error (bad or expired credentials, a retired model id, runner issues): there is no code finding to fix. Surface the relevant log excerpt on the PR and escalate (Step 6) — never retry-loop hoping for a different outcome, and never weaken the workflow to get past it.

If --dry-run: report verdict + findings and stop here.

Step 4 — Fix or dispute each 🔴 (and red CI)

Default to fix — the reviewer is usually right, and only a code change can flip the gate.

Fix: make the minimal edit for that finding (no drive-by refactors), commit with a conventional-commit message scoped like the repo history (fix(gmail-mcp): …, chore(ci): …), and let Step 5 push. Fix red CI the same way (lint rule, failing test, build error).

Dispute (when the finding is genuinely wrong): reply on the PR tagging the reviewer so the exchange is on the record:

gh pr comment <PR> --body "@claude Re: <finding>. <concrete counter-argument with file:line evidence>. Do you still consider this blocking?"

Know the limit: a comment gets a reply, not a re-grade. If the dispute ends with no code change, that is an escalation (Step 6), never a workaround.

🟡 Nits: fix them when cheap; otherwise leave a one-line comment saying why not. 🟣 Pre-existing: acknowledge if useful; they never block and out-of-scope fixes belong in their own PR.

Step 5 — Re-push and re-grade

If Step 4 produced commits, git push (plain form). That synchronize event re-runs the gate on the new diff. Increment the iteration counter and go back to Step 2 — until green or --max is hit. If Step 4 produced no commits (pure dispute), go straight to Step 6.

Step 6 — Terminate

Green (success)

When claude-review and CI both pass:

1. Post a short summary comment: verdict, iterations, what was fixed.
2. Apply the label:

gh pr edit <PR> --add-label ready-to-merge

In this repo the ready-to-merge label is a signal to humans: it triggers .github/workflows/ready-to-merge.yml, which comments on the PR asking the code owners (@ni500, @lcaloguerea) to review and approve. It does not approve or merge anything — a human code-owner approval is always required by branch protection.

Escalate

When --max iterations are spent with a 🔴 remaining, or a 🔴 is contested and will not be changed in code: post a PR comment tagging @ni500 @lcaloguerea with the unresolved finding (file:line, one line each), a link to any @claude dispute thread, and a recommendation. Then stop — the maintainers break the tie.

---

Hard rules

• Never fake green. No editing the workflow, weakening the check, or merging around it. The only paths to green are a code fix or a maintainer's ruling.
• Never force-push. Plain git push only.
• Never merge. gh pr merge is out of scope in every form. Merging is the human reviewers' call after approval.
• Only 🔴 blocks. Don't chase 🟡/🟣 to green, and don't ignore them silently either — a one-line disposition is enough.
• A comment doesn't re-grade. Only a push re-runs the gate; build the loop around that.
• Scope discipline. Fix the finding, not adjacent code.