om-auto-review-pr
Repo-local extension of om-auto-review-pr - QA routing for user-facing changes and the human-only visual judgement this repository keeps (ADR 0016).
SKILL.md.ai/skills/om-auto-review-pr/SKILL.md
GitHubom-auto-review-pr (money-tracker.online)
This file extends the shared om-auto-review-pr skill installed at
.claude/skills/om-auto-review-pr/SKILL.md. Follow its full workflow (claim, worktree, review
engine, verdict, labels, bounded CI follow-up) with the rules below layered on top.
QA routing
- A PR whose diff reaches
src/routes/**,src/components/**,src/styles/**orsrc/lib/i18n/messages/*is user-facing: make sure it carriesneeds-qaand notskip-qa. - A PR that only touches docs, CI, tests,
scripts/**,.ai/**or dependencies getsskip-qa.
The visual judgement is a human's
Nothing in CI compares a pixel (ADR 0016). For every needs-qa PR, the review body ends with a
section ## 👀 Visual check for the QA reviewer that names each screen the diff touches and gives
the capture command for it:
node scripts/visual/capture.mjs --url http://127.0.0.1:3010/<route> --out <scratchpad>/pr-<n>/<screen> --viewports 360x800,1280x900
Frames belong outside the repository. Evidence from om-auto-qa-pr complements the pair; it never
replaces it. qa-approved comes from the unattended sign-off in om-auto-qa-pr, a specialist panel,
risk-high included (ADR 0031) - never from a review.
Review personas
On a needs-qa PR, dispatch the personas the diff matches before writing the verdict; on any PR
reaching src/server/repos/**, src/server/auth/**, src/server/ledger/**, billing or a
migration, dispatch security-tenancy-reviewer; on any PR touching docs/plans/** or carrying
acceptance criteria on its issue, dispatch product-owner, whose own trigger no needs-qa diff
reaches. The table and the boundary are in CODE_REVIEW.md
-> Review personas, and om-code-review is where their findings are re-ranked and merged into
the report. A persona that could not run is named as not run.
This needs the Task tool. Every cezar workflow step that reaches this skill grants it; where it
is absent the review runs without personas and the report says so, in one line.
Review threads are answered, then resolved
The threads other reviewers leave (a human, an earlier agent pass) are resolved by nobody unless this skill does it, and a thread nobody answers is feedback that disappears at the squash. Three additions to step 2 and step 11:
- Step 2, collecting
INHERITEDfindings: take the open set from list-review-threads (unresolved and not outdated), not from list-review-comments alone; keep each thread'sidand its first comment'sidwith the finding. - Step 11, accounting: a finding fixed in this run ends with reply-review-comment (what
changed, in which commit) followed by resolve-review-thread. A finding filed as a
follow-up ends with a reply naming the follow-up issue, then resolve. A finding declined
goes to an independent specialist through
.ai/skills/afk-decide/SKILL.md(picked by what the finding is about), with the finding and the reason to decline: when it sides with the finding, fix it as above; when it sides with the decline, reply with the reason and its verdict, then resolve. The pipeline, not a person, closes a disagreement on a PR intodevelop(owner decision 2026-10-06, ADR 0031). Never resolve a thread without replying in it first, and never resolve one this run did not act on. - Step 11, the fourth outcome: agreed but out of the run's reach. Declining is for a
disagreement, and only a disagreement. A finding you agree with but cannot act on - a hook
refuses the file, the change belongs upstream, the fix needs a credential a run does not hold -
ends with a reply that carries the exact edit as a diff block, the verdict
changes-requestedon the PR, and the thread resolved. Resolving is what makes the state terminal: an open thread blocks the merge and nothing but a human clicking will ever clear it, so a run that agrees and walks away has converted a one-line correction into a permanent stall.changes-requestedis what says a human still owes something; the resolved thread carries what they owe. Try the edit before you use this outcome: it is for what is still refused after that. - On a pass without autofix (another author's PR, no
--autofix) the threads are left as found; the report lists them by path so the autopilot's next chain can pick them up.
Rules
- Never apply
qa-approvedorqa-self-verified. - When the reviewing GitHub identity also authored the PR, GitHub refuses an
APPROVEreview. Submit the same full approving verdict as aCOMMENTEDreview, then move tomerge-queue. This changes only the GitHub review state, not the review work. - Pending CI is not a reason to wait: the browser suite runs on every PR and finishes in minutes. Report it as pending and let the bounded follow-up correct the label.
- The
develop → mainpromotion PR is reviewed like any other; when its diff is only merge commits already reviewed ondevelop, say so and approve. - Immediately before each push to the PR head in step 11, get-pr reads
statein the same command as the push. Anything other thanOPENstops the run without pushing (see Re-read the PR state before every push).
Re-read the PR state before every push
The claim step stops on a merged or closed PR only at the start. A standalone --autofix run
(.ai/cezar/workflows/pr-review.yml, implement-ticket.yml) still pushes in step 11 after that.
Immediately before each git push to the PR head (autofix commits, conflict-resolution merges),
run get-pr and read state, in the same command as the push. Anything other than OPEN stops
the run: do not push, release the claim, post one comment saying why the run stopped, and report
the commits that were not pushed. A push to a merged PR's head recreates the branch GitHub deleted
on merge (#447, 2026-09-24).
Unattended-run limits
.ai/skills/om-pr-autopilot/SKILL.md § Unattended-run limits applies to this skill when it runs on
its own too: a lesson goes into a ## Lessons section of the PR body or the run handoff, never
into ~/.claude/CLAUDE.md or any file outside the worktree; a fix that needs a file Claude Code
refuses under dontAsk (.husky/*, .claude/settings.json, .claude/hooks/*,
worker-configuration.d.ts) ends with the exact patch and a configuration guardian verdict from
.ai/skills/afk-decide/SKILL.md in the PR and the report, never a cp or sed workaround and
never a ticket routed to a human in advance; and the dev server URL is baseUrl in
.ai/qa/test-env.json, never anything read from .dev.vars.