Review Code

Read a CODE REFERENCE READ-ONLY and emit a references-first review report to the active thread's implementation/reviews/ folder. This skill runs end-to-end: it reads the code, drafts the report, and writes one record per review run. By default it runs end-to-end without walking findings with the user one at a time, but it honors an invocation that asks it to check in or walk the findings interactively; it does not commit anything.

This is the CODE-QUALITY pass — it evaluates the code on its own merits: quality, safety, idioms, and testability. It is the complement to the implementation-fidelity review: that review asks "does the implementation deliver what the spec's acceptance criteria promised?"; this review asks "is the code any good?". The two stay distinct, but they share one anchor.

Anchor "Right" to the Spec's Acceptance Criteria, Not the Plan

When a spec is available, this code review anchors its sense of "right" to the spec's acceptance criteria — the contract — NOT to the plan. The plan is a disposable compiler-IR the human never needs to read; the spec plus its acceptance criteria are the audited artifact, the definition of what the code is supposed to do. So when a code-quality finding turns on what the code is for — whether an error path matters, whether an edge case is reachable, whether a behavior is the intended one — judge it against the spec's acceptance criteria, never against the plan. Where the code and the spec's acceptance criteria diverge in a way that is a fidelity gap (the code does not deliver what the spec promised) rather than a quality gap, that is the implementation-fidelity review's territory — flag it under ## Next Actions and keep this review on quality.

If no spec is available, this review runs as a pure general-purpose code-quality pass on the code's own merits, and ## Next Actions notes that no spec was on hand to anchor "right".

Inputs

This skill accepts ONE required input: a CODE REFERENCE. Accepted forms: