| name | review-quality |
| description | Use when producing, consuming, triaging, or revalidating code-review findings — review artifacts, P0/P1 findings, fix loops, and "is this finding real or fixed?" decisions, in any repo. Enforces structured evidence, prompt-grounding, a revalidation gate that refuses model-only "fixed", and stable-signature triage/dedup/rerank. This governs findings you already have — it does not discover them or size the review. |
Review Quality
Intent
Make review findings precise, grounded, traceable, and honestly closable. A good finding names a real defect, points at exact evidence in the reviewed code, and is marked fixed only when the defect is gone and the targeted checks pass.
When to use
Use when an agent is:
- producing a code-review artifact (findings on a diff, task, or feature);
- consuming findings to decide what to fix;
- triaging, deduplicating, or prioritizing findings;
- re-checking whether a finding is still open after a fix;
- coordinating a review/fix loop; the Loop workflow carries its own inline copy of this discipline.
Inputs
Required: a review scope — the diff, task, or files under review — and the findings (or a request to produce them).
Optional: the validation commands available for the changed behavior (tests/lint/build).
Core rules
- Structured evidence. Every finding carries
path, startLine/endLine, symbol, and a quote of the offending code. A finding without resolvable evidence is invalid.
- Prompt-grounded. Evidence must reference files actually in the review scope. Drop a finding whose
path/line range does not resolve, but keep valid sibling findings — one bad finding does not sink the batch.
- Scope proof. A review must report what it
inspected (files/symbols covered). An empty finding set is valid only as { findings: [], inspected: [...] }. "No findings" without inspected is not a pass.
- Revalidation gate. A finding is
fixed only when a fresh review no longer finds it and the targeted validation commands (tests/lint/build for the changed behavior) pass; record the command results. If validation cannot run because of missing infrastructure or environment, the status is blocked or uncertain, never fixed. Manual confirmation or model opinion alone does not close a finding.
- Stable signatures. Each finding carries a signature = severity + normalized file path + normalized finding summary + related scenario/task ID. Use it to detect repeats and dedup.
- Triage before acting. Dedup identical signatures; rerank survivors by severity, confidence, reachability, test coverage, and patchability. After two failed fix attempts or three review observations of the same signature, classify it
false_positive, core_issue, or futility rather than looping.
Non-goals
- Do not write or apply fixes; this skill judges findings. The Loop workflow coordinates fixing.
- Do not replace a repo's own review tooling when it is responsible for emitting findings.
- Do not grade spec or test quality; use
spec-quality and test-quality for those.
Output
When reporting on findings, state: the findings with evidence and signatures; the inspected scope; any dropped findings and why; the validation commands and results used for any fixed status; and any blocked/uncertain items with their blockage.
