PR Review
With PRs named, check each one. With none, check every open non-draft PR in the
current repository that has no recommend-* label, one subagent per PR.
Check the title, description, and tests against the linked issue and the code at the head commit; they are claims.
Answer five questions in order, each with one line of evidence:
- Meaningful: is the problem real and still present on the base branch?
- Worth adding: does the value justify the scope and ongoing maintenance?
- Right approach: would you solve it this way, at the root cause, with a principled design that still holds as the code grows? A patch or special case that later changes must work around fails. Fail it only by naming an alternative and what it gains: a different design, or a refactor or architecture change that lands first and shrinks or removes this change.
- Correct: does it solve the problem without breaking a caller or edge case, and do the tests it adds for changed behavior fail on the base branch?
- Minimal: is every changed line needed, with nothing that existing code already does?
Take the first failure's verdict:
| Fails | Verdict |
|---|---|
| Meaningful, or Worth adding for a technical reason | recommend-close |
| Worth adding as a product or scope call | recommend-triage |
| Right approach, Correct, or Minimal | recommend-revise |
| Nothing | recommend-merge |
Report to the caller the verdict, the five answers, and for revise the
specific changes needed. The caller decides what reaches the PR.