/qa-review

A review that skips an axis is half a review, and a review that borrows the builder's conclusions is none. This skill judges a ticket's work along two axes at the same time — whether the change honors the repo's own standards, and whether it delivers what the ticket promised — and walks the hard gates no matter what, because gates are the one part of quality that never depended on the ticket's wording. It runs in a fresh session and re-derives everything (ADR 0004): the Dev session's memory of what it built is not evidence here.

Two commitments hold absolutely. Nothing is repaired. The run writes exactly two things — its report and a one-line comment on the reviewed ticket — and never touches a source file, a test, or a config; the moment a reviewer improves what it reviews, its verdict is worthless. And everything is measured, not believed: a checked acceptance box is a claim, a stylesheet rule is an intention, a passing suite is one input. The verdict rests on what the repo, the tracker, and the running behavior actually say.

Re-derive the context first

Before any axis is judged, rebuild what a fresh session owes itself, from the target repo alone:

• docs/agentic-skills/stack.md — the pinned test command is the only runner this run uses; the pinned type-check and build commands stand ready. Absent config means derive both from the repo's manifests and say which evidence stood in.
• docs/agentic-skills/tracker.md — how tickets are read and commented in this repo; the reviewed ticket is read and answered through these operations, never through assumptions about where tickets live.
• docs/agentic-skills/ui-gates.md — the repo's own gate list, paths floor, and gate exceptions; the walk below is judged against this file, which is the truth the repo pinned.
• GLOSSARY.md when present — the vocabulary findings on the standards axis are graded against it. Absent means the vocabulary checks are skipped and the report says so.
• docs/adr/ — every record plausibly touching the reviewed area; a change that quietly contradicts a settled decision is a finding even when every test passes.

Bind the review target

The invocation names a ticket — or a since-scope: "review everything since <ref>", where the ref is a commit, a branch, a tag, or a merge-base. Resolve either through the repo's own history and the tracker's own operations, then establish the scope of work under review: