Takes a skill PR from "opened" to "ready to merge": review, fix, verify, summarise. Maintainer
tooling, not shipped in any plugin. It composes the
other authoring skills rather than repeating them: the bar is
umbraco-skill-author's conformance checklist,
runtime proof comes from the dotnet test gate (umbraco-reference-instance)
and the repo-root scripts/boot-check.mjs, and evals from umbraco-skill-evaluator.
Ground rules (decided by the maintainers)
- Reviews and updates run on their own. Fix it yourself and push to the PR branch (or, when you can't push, open the fixes as a pull request into the author's branch, as fix-and-push.md describes). Post the summary comment and set the labels without asking first. Don't leave suggestions: they bounce back and forth.
- Evals run only when something looks off. The red flags are in verify.md.
- The labels are decorative, so swap them freely:
for-ai-review,ai-review-doneandai-review-pending-edits. - Never approve, request changes, or merge. A human merges.
- Never claim a check you didn't run. Not in the comment, not in your report: the comment is what a human uses to decide whether to merge, so an unrun check that reads as run makes a red PR look ready.
The stages
Open each reference as you reach its stage. The traps they list are why this skill exists.
| Stage | Reference | Output |
|---|---|---|
| 1. Review the skill | review.md, in a worktree set up as fix-and-push.md describes | A written findings list (format at the end of review.md) |
| 2. Fix on the PR branch | fix-and-push.md | Commits on the PR branch, in the author's voice |
| 3. Verify | verify.md | Validation, gate and boot results; evals only on a red flag |
| 4. Push and summarise | fix-and-push.md, with scripts/push-review.mjs |
Pushed commits, one summary comment, labels swapped |
For more than one PR, start from batch.md: it lists the queue with
scripts/review-queue.mjs, sets up the worktrees with scripts/setup-review-worktree.mjs, and runs
one reviewer per PR, then finishes PRs reviewed earlier (--follow-ups).
Report back
End with a table, one row per PR: the PR, the skill, what you fixed, what you verified, and anything you left for a human (a design question, a fixture you couldn't repair, a check you couldn't run, and why).