Autopilot PR Merge Worker

autopilot-pr-merge-worker - per-PR drive-to-merge convergence loop

Writes are optional via the activation card (write: on default). autopilot-pr-review-scheduler never spawns this skill. The PR review scheduler never comments, labels, assigns, or requests reviewers.

panel-review requests a review. status/accepted is the human action flag. No accepted, no review. If status/accepted is missing on this PR and every linked issue: remove panel-review if present, do not comment, stop. Never assign. The scheduler and panel stop the same way and leave no comment.

Activation card

activation_card: on. Before any PR read or GitHub write, emit this Enter card with every field filled. Missing field -> stop.

skill: autopilot-pr-merge-worker
skill_path: <resolved directory of this SKILL.md>
mode: run
subject: microsoft/apm#<pr-number>
path: merge
intent: drive one already-selected PR
origin: unattended | actor-session
write: on | off
repo: microsoft/apm
pr: <positive integer>
invocation: agentic-workflow | actor-session
invocation_mode: composed-implementation-review

Rules:

• write defaults to on when the caller omitted it.
• write: off returns the filled template only. Do not comment, add labels, remove labels, push, or request reviewers.
• If write: on, post the one advisory comment via autopilot-comment (source_skill: autopilot-pr-merge-worker) and fold/push inside the PR's stated scope. Never assign. Never request the implementer as a reviewer. Do not call gh pr comment yourself.
• origin fail-closed unknown -> unattended.
• Unattended never assigns and never requests reviewers.
• One PR. Do not nest a scheduler path.
• plan_first defaults to yes. Do not fold, push, or request reviewers until Step 0.P exists. Copilot App spawn uses kickoff_mode: plan. If the session started interactive, still write that plan in-session and do not skip it. Wait for operator approval when a plan-mode gate is present.

After the loop, emit this Exit receipt:

skill: autopilot-pr-merge-worker
subject: microsoft/apm#<pr-number>
path: merge
write: on | off
posted: yes | no
pushed: yes | no
reviewer_requested: no
approved: n/a

This SKILL.md is the natural-language module derived from a genesis design packet; refactors re-run the genesis skill from that packet.

This skill is a BUILDING BLOCK, not the PR-review queue. autopilot-pr-review-scheduler never composes it. Invoke it by name when the caller asked to drive a PR to merge. It COMPOSES the autopilot-pr-review-worker skill for the advisory pass -- it does NOT re-implement panel review.

Boundary (what this skill does and does NOT do)

DOES: take ONE PR that already exists and drive it to a terminal landing-ready state; resolve cross-PR merge conflicts; emit the PR-facing advisory and supersede comments; return a schema-valid completion_return.

Does NOT: triage issues, decide which issues are worth fixing, open the first PR for an issue, choose the batch, or maintain the orchestrator's ground-truth table. Those are the parent orchestrator's responsibility. If you find yourself triaging or opening greenfield PRs, you are in the wrong skill.

How an orchestrator composes this skill

The parent scheduler, for each PR in its batch, spawns ONE autopilot-pr-merge-worker subagent with the spawn body in assets/worker-prompt.md. The orchestrator passes the inputs that prompt declares (PR_NUMBER, ISSUE_NUMBER, AUTHOR, HEAD_REPO, HEAD_BRANCH, MAINTAINER_CAN_MODIFY, REPO_ROOT, ORIGIN, optional PANEL_PRIOR). The subagent owns the whole convergence loop end-to-end and returns a completion_return.

After every driven PR returns ready-to-merge, the orchestrator runs the conflict-resolution phase: probe mergeability and, on DIRTY / BEHIND / CONFLICTING, spawn one conflict-resolution subagent per assets/conflict-resolution-prompt.md. The step-by-step gate procedure is in references/mergeability-gate.md (load it when entering that phase).

Dependency declaration (read before composing)

This skill is a same-repo LOCAL SIBLING. A consuming orchestrator MUST declare the dependency at its own distribution surface and PROBE for this skill before spawning. The probe is a tool call, not an assertion from recall (A9 SUPERVISED EXECUTION; truth #2 CONTEXT EXPLICIT):

test -f ../autopilot-pr-merge-worker/assets/worker-prompt.md \
  && test -f ../autopilot-pr-merge-worker/assets/completion-schema.json \
  && test -f ../autopilot-pr-merge-worker/scripts/owner_touch_gate.py \
  && echo "autopilot-pr-merge-worker present" \
  || echo "MISSING autopilot-pr-merge-worker - stop and ask the operator"

On a probe MISS the orchestrator stops and asks the operator rather than re-implementing the loop inline (avoids HAND-ROLLED HALLUCINATION and PHANTOM DEPENDENCY).

This skill itself COMPOSES autopilot-pr-review-worker. A consuming orchestrator inherits that transitive dependency; the spawned autopilot-pr-merge-worker subagent PROBES for it at preflight (all load-bearing panel assets under $REPO_ROOT/.agents/skills/autopilot-pr-review-worker/) and returns status: blocked ONLY on a genuine asset MISS, before any checkout (see the spawn body Step 0.0). Note: a missing skill TOOL is NOT a miss -- in the normal subagent context the panel is executed INLINE from its on-disk SKILL.md + schemas (Step X.1.1), which is a first-class path, not a fallback.

Convergence loop contract (per PR)

The autopilot-pr-merge-worker subagent runs this loop (full detail in the spawn body):

0. Phase 0.P -- plan before implement (hard gate). Compare work to date with the original task/scope (PR body + linked issue). Then plan: fresh panel, fold in-scope findings and requirement gaps, CI green via agent-merge when available else assets/ci-recovery-checklist.md. Do not checkout, fold, or push until that plan exists (and is approved when plan-mode).
1. Phase X.0 -- fetch + classify copilot-pull-request-reviewer[bot] inline review per assets/copilot-classification-prompt.md.
2. Phase X.1 -- run the autopilot-pr-review-worker against the PR: via the skill tool if present, otherwise (the normal subagent case) execute it INLINE from its on-disk SKILL.md + schemas. Both paths produce the same single recommendation comment.
3. Phase X.2 -- merge follow-ups (LEGIT Copilot + panel recommended_followups) and apply the fold-vs-defer rubric per assets/fold-vs-defer-rubric.md.
4. Phase X.2.5 -- canonical-owner + functional-evidence gate (FAIL CLOSED). Run scripts/owner_touch_gate.py against the exact base/head. It parses the single canonical owner table in .apm/instructions/architecture.instructions.md; no LLM self-classification may override a detected touch. Every touched owner requires executed exact-head functional test IDs/evidence in addition to the boundary lint and any required dual guardrail. Schema-validate and semantically verify the version 2 completion evidence. Missing evidence stays in the loop or returns blocked; it is never deferred.
5. Phase X.3 -- edit code; fold every FOLD item. Run the mutation- break gate on any new regression-trap test.
6. Phase X.4 -- run the lint contract until silent.
7. Phase X.5 -- push (head branch; fall back to a superseding PR that preserves authorship via commit trailers).
8. Phase X.6 -- CI to green. Prefer the agent-merge skill when it is loaded or the prompt contains <agent_merge_state> (do not sleep/watch inside that tick; never merge). Otherwise watch + recover per assets/ci-recovery-checklist.md.
9. Phase X.7 -- decide terminal vs next iteration.
10. Phase X.8 -- at terminal, capture head_sha, mergeable, mergeStateStatus, and CI status for the completion return.