review

Review procedure for task PRs — deterministic tool output first, LLM judgment second, verdict on the board. Use when reviewing a task's pull request or when asked for a task-agent review.

SKILL.md

GitHub

Review

Review one draft PR produced by the task workflow. Output: PR comments plus a single verdict. Reviewers never push fixes — they report.

Order matters: tools before opinions

  1. Gather deterministic signal first (in the PR's worktree or a fresh checkout of the branch):
    • vendor/bin/pint --test — style
    • vendor/bin/phpstan analyse — types/bugs (if installed)
    • php artisan test — behavior
    • gh pr diff <number> — the actual change Any gate failure is an automatic fail verdict; report the output, done.
  2. LLM review only for what tools can't catch, anchored to diff lines:
    • does the change do what the task + approved plan say, nothing more?
    • missing test cases for the changed behavior (not style preferences)
    • security at trust boundaries: validation, authorization, mass assignment, query injection
    • N+1 queries, unbounded queries, missing indexes for new query paths
    • conventions: follows Laravel Boost guidelines, docs/conventions.md, and the code-standards skill (cite CS-n in findings)
  3. Skip entirely: style nits Pint already enforces, hypothetical refactors, scope expansion ("while you're here…"), praise padding.

Report

Post one PR comment via gh pr comment (or reply inline where supported):

task-agent review — TASK-<n>

Gates: pint ✓ / phpstan ✓ / tests ✓
Findings:
1. <file:line> — <problem> — <required fix>   (only real findings; empty is fine)

Verdict: PASS | FAIL — <one-line reason>

Verdict → board

The reviewer never marks a PR ready and never merges — it returns a verdict and the task-workflow skill acts on it.

  • PASS → task-workflow step 9 finishes the task (human merge by default; gated merge only if the developer enabled it).
  • FAIL → a fix round. The findings go back to the same implementation agent in the same worktree, up to MAX_REVIEW_ROUNDS (default 3), then blocked.

Re-review rounds

Round 1 is the full review above. From round 2 on, review the fixes, not the PR afresh:

  1. Verify each finding from the previous round is actually resolved — cite the diff line that resolves it, or repeat the finding.
  2. Look for regressions the fixes introduced.
  3. Anything else you notice goes to workflow/<TASK-id>-<slug>/backlog.md and is named in the summary — not into this PR. A reviewer that finds six new things every round means the loop never converges and the human never gets a merge.

Write each round's report to workflow/<TASK-id>-<slug>/review-<n>.md as well as posting it, so the next round can read what the last one asked for.