review

Deprecated — superseded by /swe:code-review; this skill still works but gets no further updates. Adversarial multi-perspective code review. Spawns independent reviewer agents (correctness, security, maintainability, tests, performance), sharpened by domain lenses where the diff calls for them, then puts every finding through a skeptic verification pass before reporting. Report-only. With no target, reviews the staged changes — or all uncommitted changes if nothing is staged. Small diffs get a reviewer budget scaled to their size, with perspectives beyond it folded into the reviewers that run; --full lifts the budget. User-invoked only — never invoke this skill on your own initiative.

Compatibility

Not fully portable — uses frontmatter extensions beyond the agentskills.io spec. On Claude Code, context: fork runs the whole review in a forked subagent, keeping it out of the main conversation. Hosts that ignore context: fork run the same workflow in the main conversation, with parallel reviewers where a subagent tool exists and sequential passes otherwise. Hosts that ignore disable-model-invocation lose the user-only invocation guarantee.