subagent-review

human-only. Start this only when a human asks for it by name. If you arrived here from another skill, stop and get explicit confirmation before running any step.

An independent review of code that is committed but not necessarily on a PR — a branch mid-build, the work /implement just finished, a fan-out's output before any of it is proposed. One subagent per dimension, none of them able to see the others, findings verified by you before they are repeated.

Where the work is already on a PR, /review-pr-in-worktree is the better door. It knows the PR's own context — the head SHA, the existing reviews, what a human already said — and it ends by asking whether to review another round. This skill exists for the case that has none of that: there is no PR, so there is no PR to read.

Nothing here writes. No commits, no fixes, no pushes, no comments — even when the fix is one line and obvious. A review that edits its subject stops being a review. Findings go to the human, who decides what happens next.

How a fan-out is run — isolation, the brief, the prohibitions, and how results are verified — is subagent dispatch, shared with subagent-implement. Read it before dispatching anything.

Process

1. Settle exactly what is under review

The subject is a commit range, never the working tree. Uncommitted changes are not the thing being reviewed and their presence is itself a finding.

Take the range from the human where they named one. Otherwise: