ryan-brosas Agent Code Quality Gate
Before claiming authored implementation or subagent work is done, inspect its diff and the decisive verification output. Match depth to material risk; review real boundaries rather than inventing reuse.
Five checks
- Scope: each authored change serves the request. Split unrelated cleanup without disturbing pre-existing user work.
- Duplication: check likely shared boundaries for copied logic or competing owners. Consolidate only when reuse is real.
- Behavior: use the closest decisive existing test or runtime probe. For a reproducible regression class, add or expand a durable test when its maintenance value is positive; otherwise state the remaining gap.
- Verification: name the command/probe actually run and inspect its result, including any decisive output hidden by truncation. "Should work" is not proof.
- Regressions: explain new failures, removed tests, and skips. Fix blockers; cosmetic nits do not block completion.
Explain legitimate exceptions in completion evidence or the commit: explicitly approved scope, one-off duplication, quarantined flaky tests, or tests replaced by better coverage. Do not hide exceptions or downgrade violated invariants.
Work is not done while an applicable blocking check fails; cosmetic nits are reported without blocking. Report the inspected diff, commands, results, and remaining limitations; do not repeat the checklist as a separate ritual or substitute it for behavioral evidence.