PR Autopilot — Open Mercato extension

This file extends the shared om-pr-autopilot skill from open-mercato/skills (installed at .agents/skills/om-pr-autopilot/SKILL.md). Follow the shared skill's full workflow — agentic setup, PR resolution, outer claim, the ten diagnosis signals, the state matrix, chain execution, the summary report, lock release — with the repo-specific rules below layered on top.

The shared skill's own references (references/diagnose.md, references/state-matrix.md, references/report-templates.md, references/claim-pr.md, …) live next to it under .agents/skills/om-pr-autopilot/; they are not committed here. When the shared skill or any companion it dispatches to is missing, run yarn install-skills and re-enter — never improvise a substitute procedure for the claim, worktree, diagnosis, or setup mechanics.

Diff-scope classification (this repo's layers)

The shared references/diagnose.md signal 4 asks for the diff's scope. In this repository, classify against:

• spec-only — only .ai/specs/ or .ai/specs/enterprise/ (the config's paths.specs). A spec-only PR gets the specification review and must never grow into implementation here.
• docs-only — apps/docs/, root markdown, AGENTS.md files, .ai/skills/, .ai/runs/.
• UI-touching — packages/ui/, any module's backend/ or frontend/ pages and components, and the customer-portal surfaces. This is what drives needs-qa.
• migration/schema — migrations/, .snapshot-open-mercato.json, ORM entities under data/entities.ts.
• contract surface — anything enumerated in BACKWARD_COMPATIBILITY.md: auto-discovery files, exported types and signatures, import paths, event IDs, widget spot IDs, API routes, DB schema, DI keys, ACL feature IDs, notification IDs, CLI commands, generated registries. A touch here forces risk-high and deeper review.

Label taxonomy and the QA gate

The label vocabulary, the mutual-exclusion groups, and the priority/risk inference rules are defined in the root AGENTS.md → PR Workflow section; read it rather than restating the taxonomy. Three repo rules bind the chain:

• The QA-approval merge gate is hard. A PR carrying needs-qa without qa-approved MUST NOT merge, even when everything else is green. skip-qa is the explicit opt-out; never combine it with needs-qa/qa-approved.
• Never set the qa pipeline label. It means "manual QA in progress" and is applied by QA reviewers only. This skill and everything it dispatches request QA with needs-qa alone.
• qa-failed, blocked, and do-not-merge are hard blocks — matrix row 0c stops on them.

When the active account lacks triage rights

Triage permission here depends on which account the run is executing as: a maintainer run assigns and labels normally, an outside-contributor run cannot. Never assume either way — perform the full three-signal claim exactly as the shared references/claim-pr.md prescribes, and degrade only in response to a real permission error.

When assign-pr or apply_label actually returns a 403 (replaceActorsForAssignable / addLabelsToLabelable), that is an expected outcome for a contributor account, not an incident: the outer claim degrades to the 🤖 claim comment alone, the release path tolerates the absent in-progress label, and the intended label set is listed in the summary comment addressed to a maintainer. Do not retry the mutation that failed, and never report a label or an assignee as applied when it was not — the reverse holds too: when the mutations succeed, the claim and the label state machine run in full.

Validation

This skill dispatches rather than validating; the delegated skills own the gate. When one of them runs it, the ordered validation.commands list in .ai/agentic.config.json is the CI-mirroring gate, and the Docker-vs-local runner decision in root AGENTS.md § Validation Commands applies. Prefer GitHub PR check results over re-running the gate locally, per .ai/skills/om-auto-review-pr/SKILL.md.