apm-issue-autopilot - intake-to-merge issue orchestrator

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 an A11 RECONCILIATION LOOP, MANUALLY INVOKED, over a queue of issues each driven to a terminal state under non-determinism. It generalizes batch-bug-shepherd from bugs-only to ANY issue type, and promotes the full apm-triage-panel rubric to the central front gate. It does NOT re-implement triage, panel review, or the per-PR convergence loop -- it COMPOSES existing skills.

What it composes (do not re-implement)

• apm-triage-panel -- the triage rubric, run per issue in DIRECT (orchestrator-return) mode.
• shepherd-driver -- the per-PR drive- to-merge convergence loop and cross-PR mergeability gate (which in turn composes apm-review-panel).
• pr-description-skill -- authors the anchored, mermaid-validated body of the ONE issue PR opened at Phase 4 acceptance-close (see assets/acceptance-observer.md). The PR body is never hand-rolled.

All three are same-repo LOCAL SIBLINGS. This skill DECLARES each dependency here AND in apm.yml, and PROBES for each at its use-site (Phase 1 triage, Phase 4 PR-open, Phase 5 shepherd) with a tool call -- never an assertion from recall (A9 SUPERVISED EXECUTION).

Hard boundaries

• MANUAL invocation only. No event triggers, no label triggers, no gh-aw. Labels (status/accepted, status/shepherding) are WRITTEN for bookkeeping but are NEVER the trigger.
• Triage is paramount. The autopilot ESCALATES to the maintainer by default; auto-implementation is the narrow exception, reached only for a clear, bounded, high-confidence accept the maintainer approved.
• ONE consolidated triage review for the whole batch, not drop-by- drop. Exactly one human checkpoint (Phase 2).
• Never auto-merge. Mergeability is the terminal state; the human approves the protected merge.
• Escalation NEVER auto-closes or auto-declines an issue. It surfaces the issue to the maintainer running the session and leaves it for human action (inherits batch-bug-shepherd terminal-handling).

Architecture invariants

• Fan-out, not serial. Triage, solution-pipeline (and its Plan lenses + per-wave task children), and shepherd-driver all run as parallel child threads via the runtime task affordance. A single-loop variant is an anti-pattern. Subagent capacity is UNLIMITED and is NEVER a deferral reason.
• Worktree isolation, with one-writer integration. Every solution- pipeline and shepherd-driver child runs in its OWN git worktree (one per issue/PR). Within Phase 4, a pipeline child further spawns ONE task child per task PER WAVE, each in its OWN worktree branched off the issue branch at the wave base; the pipeline child is the SOLE WRITER of the issue branch and integrates task branches via git merge --no-ff at the wave gate. Tasks in a wave are mutually independent and touch disjoint files by construction (planner- enforced), so integration is conflict-free -- a real conflict is a planning error and triggers a re-plan, never a hand-resolve. Do NOT fan out against one shared working tree. Triage and verifier/lens children are read-only and may share a read-only REPO_ROOT.
• One persisted state table. A single plan.md ground-truth table plus a machine-readable proceed_manifest is the canonical session state (B4 PLAN MEMENTO). Reload it at every phase boundary; never keep parallel state in memory. The orchestrator is the SOLE writer (one-writer rule); children return JSON, the parent writes rows.
• Escalate by default. The confidence gate (assets/confidence-gate-rubric.md) routes anything doubtful to the human. Auto-proceed is the exception.
• Triage children are advisory-only. A triage child runs the apm-triage-panel rubric and returns structured JSON. It MUST NOT post a comment, apply a label, touch the working tree, or use any GitHub safe-output channel. A child that emits a comment is a hard retry/fail.
• Fold by default at the PR layer. Inherited from shepherd-driver: every follow-up inside a PR's stated scope is folded; only scope- crossing items defer with a one-line boundary note.
• Deterministic tool bridge. Every consequential write (label, assign, PR open, push, comment, merge probe) and every present-state fact (CI status, mergeable, head sha, duplicate existence, PR-in- flight) goes through a deterministic CLI (gh, git, uv run ruff) wrapped in plan + execute + verify. Never assert these from recall.
• ASCII only. All artifacts (tables, comments, commits, the digest) stay within printable ASCII; status symbols [+] [!] [x] [i] [*] [>].

Dependency probes (run before composing)

Phase 1 (before triage fan-out):

test -f ../apm-triage-panel/SKILL.md \
  && echo "apm-triage-panel present" \
  || echo "MISSING apm-triage-panel - stop and ask the operator"

Phase 5 (before shepherd-driver fan-out):

test -f ../shepherd-driver/assets/shepherd-driver-prompt.md \
  && test -f ../shepherd-driver/assets/completion-schema.json \
  && echo "shepherd-driver present" \
  || echo "MISSING shepherd-driver - stop and ask the operator"

Phase 4 (before a pipeline child opens the issue PR -- the child re-probes this in its own worktree before gh pr create):

test -f ../pr-description-skill/SKILL.md \
  && test -f ../pr-description-skill/assets/pr-body-template.md \
  && echo "pr-description-skill present" \
  || echo "MISSING pr-description-skill - stop and ask the operator"

On a probe MISS, STOP and ask the operator to restore the sibling; do NOT re-implement the composed logic inline.

Phases

Work through the phases in order. Reload the ground-truth table and proceed_manifest at each phase boundary (B8 ATTENTION ANCHOR). Do not skip the consolidated-review gate (Phase 2).

Phase 0 - scope and seed

Take the issue list or queue from the maintainer (explicit issue numbers, a search query, or "the open backlog"). Resolve it to a concrete set via gh issue list. Seed the ground-truth table (assets/ground-truth-table.md) with one row per issue at status: pending-triage. Record the seed source and HEAD sha in plan.md. No labels are written in this phase.

Phase 1 - triage fan-out (read-only)

PROBE for apm-triage-panel (above). Then, for EACH issue, spawn ONE triage child using assets/triage-prompt.md, at PLANNER class (claude-opus-4.8) per model-routing.md (Phase 1 triage binding -- the paramount front gate is front-loaded heavy: a wrong accept burns a whole downstream pipeline). Each child runs the apm-triage-panel rubric in DIRECT mode and returns ONE autopilot-triage-decision JSON matching assets/autopilot-triage-schema.json. Children are read-only and post nothing.

This is "A11 per-item child thread running an existing triage module", not a nested panel-of-panels: each child invokes apm-triage-panel (which itself spawns no sub-agents) once and returns its decision.

On each return: schema-validate, write the row (decision, type, confidence, red_flags), set status: triaged. On a child that posted a comment or mutated the tree, discard and re-spawn ONCE; on a second violation, mark the row blocked and escalate it in Phase 2.

Phase 2 - consolidated triage review + the ONE human checkpoint

This is the heart of the skill and the single human gate. Do it ONCE for the whole batch, not per issue.

1. For every triaged row, apply assets/confidence-gate-rubric.md to compute a gate: auto-proceed | escalate | terminal. Escalate by default; auto-proceed only for a clear, bounded, high- confidence accept whose implementation brief is complete.
2. Write the machine-readable proceed_manifest into plan.md (one row per issue: issue, gate, maintainer_decision, override_reason, implementation_brief_ref, status). maintainer_decision starts pending.
3. Render ONE consolidated digest from assets/triage-digest-template.md: every issue with its decision, type, confidence, gate, red flags, and (for auto-proceed rows) the implementation brief. Present it to the maintainer in-session.
4. Capture the maintainer's decision per row: approved | rejected | overridden (with override_reason). The maintainer may approve an escalated row (override to proceed) or reject an auto-proceed row. Write the result into the proceed_manifest.

All later phases select rows ONLY where gate resolves to proceed AND maintainer_decision in (approved, overridden-to-proceed). Rows the maintainer left escalated/terminal are handled in Phase 7.

Phase 3 - ownership signaling + PR-in-flight xref

For each proceed row: cross-reference open PRs that already address the issue (gh pr list --search). Record pr and pr_in_flight. Apply status/accepted to the issue and, on the issue (and the PR if one exists), assign @me and add status/shepherding. Record every label THIS run adds in the row's labels_added column so Phase 7 (and Phase 5 teardown) strip ONLY those and never touch pre-existing labels.

Phase 4 - solution pipeline (Ideate -> Plan -> Implement waves)

For each proceed row WITHOUT an in-flight PR, spawn ONE solution- pipeline child in its OWN git worktree on the issue branch (provision with git worktree add at HEAD; record its slug in the row's worktree column so Phase 7 tears down only worktrees this run created). Spawn it at IMPLEMENTER class (claude-sonnet-4.6); it and every child it spawns route models per assets/model-routing.md (B12 MODEL ROUTER). The child runs assets/solution-pipeline-prompt.md, a four-stage per-issue pipeline (A2 PIPELINE), and returns the opened PR. It is the SOLE WRITER of the issue branch:

1. Ideate (assets/ideate-prompt.md, devx-ux-expert) -- frame the brief and derive a testable acceptance_shape (the B5 contract). Spawn at PLANNER class (claude-opus-4.8) per model-routing.md -- front-loaded heavy because this contract is the verification spine for every wave.
2. Plan (assets/plan-panel-prompt.md, python-architect lead + conditional performance / test-coverage / supply-chain / auth lenses) -- emit a persisted task DAG matching assets/plan-schema.json: per-task staffing, interdependencies, ordering, and WAVES with checkpoints (B4 PLAN MEMENTO). Trivial issue -> ONE task in ONE wave; skip the full gate ceremony (scale-down).
3. Implement (A5 WAVE EXECUTION) -- per wave, spawn ONE task child (assets/task-implement-prompt.md) per task, each in its OWN worktree off the issue branch at the wave base; the pipeline integrates the task branches into the issue branch, then runs the inter-wave checkpoint (assets/wave-gate-rubric.md, plan- guardian + ideator verifiers). PASS advances; FAIL re-plans from the failed wave (cap 2 re-plans).
4. Acceptance close (assets/acceptance-observer.md, B5) -- verify every acceptance_shape condition deterministically, then open ONE PR (Closes #N) and return its number.

Each task child writes the TYPED coverage gate first (bug: failing regression trap + mutation-break; feature: failing acceptance test; docs: docs build/link check; refactor/perf: behavior-preserving test + benchmark) for its task type, never opens a PR, and never spawns children. The orchestrator then applies the Phase 3 ownership signaling to the new PR. Rows WITH an in-flight PR skip Phase 4 and go straight to Phase 5. On a status: escalate|blocked return, write the reason to the row and surface it in Phase 7 (no PR opened). On a pr-opened return, also record the child's routing_receipts array in the row's notes (B12 cost audit) so the Ideate=opus / architect=opus front-load is auditable from plan.md alone, never from a child transcript.

Phase 5 - shepherd-driver fan-out (drive to merge)

PROBE for shepherd-driver (above). For each PR (own-implemented or in- flight), spawn ONE shepherd-driver subagent using ../shepherd-driver/assets/shepherd-driver-prompt.md. It owns the convergence loop (Copilot classification, apm-review-panel, fold-vs- defer, push, CI watch) and returns a completion_return matching ../shepherd-driver/assets/completion-schema.json. Caps: 4 outer iterations, 2 Copilot rounds, 3 CI recovery iterations.

On each terminal return: schema-validate, then write head_sha and the mergeable/merge_state_status/ci_status projection into the row's head_sha and merge_state columns (the crash-survivable A11 stop evidence), and remove ONLY the status/shepherding labels listed in the row's labels_added column (assignment stays). Also record the return's panel_execution (skill-tool|inline), panel_personas, and routing_receipt in the row's notes -- the inline panel path is EXPECTED in subagent context, so panel_execution: inline is a normal healthy value, not a degradation.

Phase 6 - conflict-resolution

For every PR that returned ready-to-merge, probe mergeability (gh pr view --json mergeable,mergeStateStatus). On DIRTY / BEHIND / CONFLICTING, spawn one conflict-resolution subagent per ../shepherd-driver/assets/conflict-resolution-prompt.md (step-by- step in ../shepherd-driver/references/mergeability-gate.md).

Phase 7 - final report

Read the table and proceed_manifest one last time. Render assets/final-report-template.md to the maintainer: per-issue decision, gate, maintainer decision, PR link, terminal status, ready-to-merge PRs, advisory-with-deferred PRs, blockers (with the responsible child's session ref), and every ESCALATED / terminal row still awaiting human action. Tear down only the solution-pipeline/shepherd worktrees recorded in the worktree column (git worktree remove); leave branches on origin for open PRs. Never auto-close an escalated issue.

Bundled assets

• assets/triage-prompt.md -- any-type triage child spawn body (runs apm-triage-panel in DIRECT mode).
• assets/autopilot-triage-schema.json -- JSON schema for the autopilot-triage-decision return.
• assets/confidence-gate-rubric.md -- escalate-by-default gate policy (Phase 2).
• assets/triage-digest-template.md -- the ONE consolidated review presented to the maintainer.
• assets/solution-pipeline-prompt.md -- A2 PIPELINE: per-issue Ideate -> Plan -> Implement(waves) -> Acceptance close (Phase 4 child; sole writer of the issue branch).
• assets/ideate-prompt.md -- devx-ux-expert Ideate child: design_brief + testable acceptance_shape (read-only).
• assets/plan-panel-prompt.md -- Plan stage: lens-selection rubric + python-architect synthesis emitting the task DAG.
• assets/plan-schema.json -- JSON schema for the persisted issue-solution-plan (tasks, deps, waves, checkpoints).
• assets/model-routing.md -- B12 MODEL ROUTER: authoritative role-class -> concrete-model table + per-spawn bindings + verifier escalation; the pipeline resolves every Phase 4 spawn's model here. Also records the B14b CAVEMAN BRIEF layer (lens advisors + wave-gate verifiers ship compressed, fixed-schema briefs), the B14c audience-boundary PER-SPAWN DECLARATION TABLE, the B13 cache-aware- prefix discipline, and the B15/B16 status for this harness.
• assets/task-implement-prompt.md -- ONE task per child in its own worktree; loads the typed coverage gate by task type; no PR, no further fan-out.
• assets/implement-bug.md, assets/implement-feature.md, assets/implement-docs.md, assets/implement-refactor.md -- per-type coverage-gate references loaded per task by task-implement-prompt.md.
• assets/wave-gate-rubric.md -- inter-wave checkpoint: integrate + plan-guardian/ideator verify + re-plan policy.
• assets/acceptance-observer.md -- B5 close: verify acceptance_shape, open the ONE issue PR.
• assets/ground-truth-table.md -- canonical table template with A11 columns + proceed_manifest.
• assets/final-report-template.md -- end-of-session report to the maintainer.

Composed siblings (declared dependencies)

• apm-triage-panel -- triage rubric (Phase 1; probed before use).
• shepherd-driver -- per-PR drive-to- merge loop + mergeability gate (Phases 5-6; probed before use). Transitively composes apm-review-panel.
• pr-description-skill -- authors the Phase 4 issue-PR body (probed before the PR-open; never hand-rolled). Transitively also used by shepherd-driver for any superseding PR.

Operating contract for the orchestrator thread

• Before each phase: re-read plan.md (table + proceed_manifest). Do NOT rely on recall.
• The orchestrator is the SOLE writer of the ground-truth table and of the GitHub state it explicitly owns (issue labels, assignment, tracking issues). PR-side writes -- pushes and PR comments during the drive and mergeability phases -- are owned by shepherd-driver, never by the orchestrator.
• Every consequential write is plan + execute + verify through a CLI.
• One consolidated review, one human checkpoint, escalate by default.