| name | review-agent |
| description | Findings-first review agent for code changes. Use when the user asks to review a branch, PR, diff, commit range, or recent changes and wants prioritized, actionable issues instead of implementation work. |
Review Agent
Use this skill as a broad code-review agent for change sets.
Use When
- The user asks to review a branch, diff, PR, or recent commits.
- The user wants prioritized findings, not a patch.
- The work is about regressions, risk, correctness, or missing tests introduced by a change.
Do Not Use When
- The request is plan or architecture review before implementation. Use
$plan-exit-review.
- The request starts from a symptom like logs, crash output, or a failing test. Use
$diagnostics-agent.
- The user only wants one narrow trace or test failure interpreted. Use the relevant triage skill directly.
Inputs
- Diff, branch, merge base, commit range, or explicit files.
- Surrounding code, touched call sites, and relevant tests.
- Optional docs if the change affects behavior contracts.
Workflow
- Identify the review scope and inspect the change set.
- Read surrounding code, not just the patch.
- Apply the severity and coverage rules in
references/review-rubric.md.
- Prefer provable regressions, behavior changes, and missing tests over style comments.
- Emit findings first, ordered by severity.
- If no findings are discovered, say so explicitly and note residual risks / testing gaps.
Output Contract
Findings: prioritized, actionable, file/line specific.
Open questions / assumptions: only when needed.
Residual risks: what was not fully validated.
Rules
- Do not drift into plan review unless the user is reviewing a plan.
- Do not implement fixes while reviewing.
- Only flag issues the original author would likely fix if shown clear evidence.
- Prefer a small number of high-confidence findings over a noisy list.
- If runtime symptoms appear during review, optionally recommend
$diagnostics-agent for deeper symptom-first analysis.