rq-review - SKILL.md Agent Skill

name: rq-review description: Review the OTHER agent's open PRs across renquant repos. Invoke when the user says /rq-review, "review codex's PRs", "do my reviews", or asks to run the review workflow. Uses the orchestrator's cross-repo queue; you (Claude) write each review.

rq-review — review the peer agent's PRs

You are Claude acting as the reviewer in the orchestrator-driven multi-agent PR model. You review Codex's open PRs (never your own — GitHub blocks self-approve and the queue excludes self-authored).

Steps

Get the queue from the control plane. Default to the whole manifest; if the user named a repo, pass --repo <name>:
```
PYTHONPATH=/Users/renhao/git/github/renquant-orchestrator/src \
  python3 -m renquant_orchestrator repos agent --as claude --workflow review --repo ${ARG:-all}
```
The JSON repos[].plan.queue lists each PR (number, head_ref, url) that needs your review. Empty queue → tell the user "nothing to review" and stop.
For each queued PR, in its repo:
- Read the diff: gh pr diff <n> --repo <owner/repo>.
- Read the PR body for intent.
- Judge per CLAUDE.md §7 (correctness, data-flow safety, tests, leakage, operational fail-closed). Review the diff, not the codebase.
Post ONE consolidated review with your token, with severity tags (BLOCKER / HIGH / MED / LOW) and a clear verdict. The body MUST contain the visible attribution string reviewed by claude:
```
gh pr review <n> --repo <owner/repo> <STATE> --body "reviewed by claude — ..."
```
Choose STATE by highest severity found:
- --request-changes if any BLOCKER / HIGH / MED.
- --comment if only LOW / nits.
- --approve if zero findings (note: if the PR is under the same account you operate, GitHub blocks self-approve — then post --comment stating "VERDICT: APPROVE").

Rules

One review per PR, consolidated — not a thread of comments.
Run the PR's focused tests when a claim hinges on them; cite the result.
Never review a PR authored by agent:claude (the queue already excludes these; double-check the author_agent field).
After finishing, report a one-line summary per PR (verdict + key finding).