Code Quality Review

Review against the user's base, or the task's base branch plus its relevant staged, unstaged, and untracked changes. Preserve unrelated work. Read nearby source, changed tests, and applicable repository guidance before judging a hunk.

Focus on evidence-backed bugs, error paths, races, state consistency, trust boundaries, and concrete verification gaps. Challenge each finding against the implementation and intended behavior. Apply architecture and performance guidance only where the changed flow makes it relevant; when files or imports moved, run yarn boundaries instead of judging the module boundaries in src/AGENTS.md by eye.

For simplification, use the repository's solution ladder: remove unnecessary work, reuse existing code or platform facilities, and only then introduce new code. Do not trade away correctness, accessibility, validation, error handling, or useful tests. Preserve an unfamiliar guard or workaround until its purpose is understood.

Return actionable findings with file/line evidence, impact, and a practical correction, prioritized by risk. Skip tooling-handled style nits and speculative cleanup; if none remain, say so without a canned approval or publishing instruction.

A review-only request returns findings without edits. During an authorized implementation, apply high-confidence in-scope fixes without asking again for that authorization. This is an advisory pass, not an extra approval gate. Reuse verification evidence for the same final state; consult docs/agent-playbooks/verification.md only when a change or concrete uncertainty calls for more checks.

When a substantive review correction exposes a preventable mistake, use retro to identify useful prevention. Keep review-only requests read-only and reuse existing coverage where sufficient.

For an explicitly scoped semantic second opinion, use Jev diff review via the repository root. Select only relevant changed files and rules; the helper is opt-in, sends selected content only with --live, and returns advisory issues or uncertainty. Keep deterministic lint/tests and independent review authoritative. Do not wire it into automatic edit, commit, or repair hooks.