| name | review-changes |
| description | Reviews the diff since a fixed point along two axes run as parallel subagents — Spec (does the change match what was asked?) and Standards (repo conventions plus a code-smell baseline) — and reports them side by side. Use when the user asks to review a branch, work-in-progress changes, a diff, or "review since X". |
Two-axis review of the diff between HEAD and a fixed point the user names.
Fixed point
Whatever the user names — a SHA, branch, tag, main, HEAD~5. Ask if they haven't given one. Before dispatching anything, confirm it resolves (git rev-parse <point>) and the diff is non-empty (git diff <point>...HEAD — three-dot, against the merge-base, so unrelated history on the fixed point's side doesn't show up). A bad ref or empty diff should fail here, not inside two parallel subagents. Also capture the commit list: git log <point>..HEAD --oneline.
The two axes
- Spec — does the change match what was asked?
- Standards — does the change follow this repo's conventions, including the smell baseline below?
Each runs as its own subagent so one axis's reasoning doesn't leak into the other's.
Finding the spec
Look in this order: an agreement already reached in this conversation, a spec file the user points at, or a spec file that matches the branch or feature name. If none of these turn up anything, don't invent a spec — the Spec axis reports "no spec available" and stops there.
Finding the standards
Whatever the repo documents — CONTRIBUTING, CLAUDE.md style notes, lint configs — plus the smell baseline in references/smell-baseline.md. The subagent can't read the plugin directory, so paste that file's contents into its prompt directly.
Two rules bind the baseline:
- documented repo standards override it — where the repo endorses something the baseline would flag, suppress the smell.
- every baseline hit is a judgment call, never a hard violation. Skip anything tooling already enforces (linter, formatter, type checker).
Dispatching
Send both subagents in a single message so they run in parallel. Give each the diff command, the commit list, and a report brief capped at 400 words.
Spec subagent — gets the spec source (or the "no spec available" note) and reports: requirements missing or partial, behavior added that wasn't asked for (scope creep), and requirements that look implemented but wrong. Quote the spec line for each finding.
Standards subagent — gets the standards files found plus the full smell baseline pasted in, and reports: documented-standard violations (citing file and rule) and baseline smells spotted (naming the smell, quoting the hunk). Distinguish the two — documented-standard breaches can be hard violations, baseline smells are always judgment calls.
Aggregating
Present the two reports under ## Spec and ## Standards headings, unmerged and unranked — a change can cleanly pass one axis and fail the other, and combining them would hide that. Close with one line per axis: total findings, and the worst issue within that axis. Don't pick a single winner across axes; that's the reranking the separation exists to prevent.
