Code Review answers requests like “review this”, “find bugs” or “is this fix safe”. It looks for disagreements between participants that each look reasonable alone, such as caller and callee, producer and consumer, or a value written now and read later, and reports a finding only when it can show a violating execution.
Correctness is the default. Performance and security are optional dimensions that use the same engine, selected with dimensions=performance, dimensions=security or dimensions=all.
What You Get
- A findings report with expectation, trigger, defect, evidence for every participant, impact, refutation, remedy and verification
- A coverage statement that says what was examined and what was not
- Separate unresolved items for contracts or runtime facts that are unknown
- Reference files for a correctness taxonomy of thirteen lenses, and for performance and security sub-cases
- Invocation arguments for dimensions and scope
Common Use Cases
- Reviewing working changes before a push
- Checking whether a bug fix is safe
- Auditing a flow where two components must agree, such as a retry path or a cache
- Running a combined correctness and security review of a repository
What Makes It Unique
The skill reviews agreements between participants, which a file-by-file checklist cannot hold together, and it forces two probes that agents rarely run: authority reconciliation and identity correlation. Every candidate has to survive refutation, and the skill rules that a cache that usually holds or a race that usually resolves correctly does not count as prevention. Zero supported findings is a valid result, and the report never says the code is bug-free.
Under The Hood
The skill is instructions with reference files and no scripts. The agent reads the repository and the working changes, follows dependencies beyond changed lines and, for a large scope, can spawn read-only workers that search and navigate but must not edit the tree. The agent confirms the tree is unchanged when the run ends. It reports and applies fixes only when asked. It does not fetch web pages or call APIs.