microsoft/shepherd-driver

shepherd-driver - per-PR drive-to-merge convergence loop

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 COMPOSED BUILDING BLOCK, not a user-facing entrypoint. It was extracted (genesis R3 EXTRACT) from batch-bug-shepherd when a second consumer (apm-issue-autopilot) needed the same per-PR convergence loop. Both orchestrators COMPOSE this skill; neither re-implements the loop. It in turn COMPOSES the apm-review-panel skill for the review 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 orchestrator, for each PR in its batch, spawns ONE shepherd-driver subagent with the spawn body in assets/shepherd-driver-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 shepherded 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 ../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"

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 apm-review-panel. A consuming orchestrator inherits that transitive dependency; the spawned shepherd-driver subagent PROBES for it at preflight (all load-bearing panel assets under $REPO_ROOT/.agents/skills/apm-review-panel/) 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 shepherd-driver subagent runs this loop (full detail in the spawn body):

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 apm-review-panel 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.3 -- edit code; fold every FOLD item. Run the mutation- break gate on any new regression-trap test.
5. Phase X.4 -- run the lint contract until silent.
6. Phase X.5 -- push (head branch; fall back to a superseding PR that preserves authorship via commit trailers).
7. Phase X.6 -- CI watch + recovery per assets/ci-recovery-checklist.md.
8. Phase X.7 -- decide terminal vs next iteration.
9. Phase X.8 -- at terminal, capture head_sha, mergeable, mergeStateStatus, and CI status for the completion return.

Caps (hard limits)

• Outer convergence iterations: 4.
• Copilot classification rounds: 2.
• CI recovery iterations: 3.

When a cap is hit with foldable items still open, return advisory-with-deferred with a scope-boundary note per deferred item. Caps exist to bound non-determinism; they are not targets.

Terminal returns

The subagent returns exactly one schema-valid completion_return matching assets/completion-schema.json:

• ready-to-merge -- clean convergence; CI observed green; lint silent.
• advisory-with-deferred -- iteration cap hit with foldable items remaining (rare); each deferred item carries a scope-boundary note.
• superseded -- push fell back to a superseding PR (records superseded_by).
• blocked -- CI cap hit, panel assets genuinely absent (Step 0.0 probe miss -- NOT merely the skill tool being unavailable), or unresolvable scope conflict (records a one-paragraph blocker).

The orchestrator schema-validates every return. On malformed, re-spawn ONCE; on a second malformed return, mark the row blocked and continue.

Disciplines enforced (inherited by every consumer)

• Fold-by-default: every follow-up inside the PR's stated scope is folded into THIS PR; only scope-crossing items are deferred, each with a one-line boundary note.
• Mutation-break gate: every regression-trap test added is proven by deleting the guard it protects and confirming the test fails without it, then restoring the guard.
• Lint contract: the canonical ruff pair (uv run --extra dev ruff check src/ tests/ and ruff format --check src/ tests/) must be silent before any push.
• CI-observed-green: terminal ready-to-merge requires real CI evidence, not an assumption.
• One advisory comment per PR per terminal pass, rendered from assets/pr-comment-templates.md.
• Authorship preservation: superseding PRs credit the original author via commit trailers and link back.

Consequential writes cross a deterministic tool bridge

Every side effect (push, comment, PR open/close, label, CI read, mergeability probe) goes through a deterministic CLI (gh, git, uv run ruff) wrapped in plan + execute + verify (A9 SUPERVISED EXECUTION). Present-state facts (CI status, mergeable, head sha) are READ from gh/git at terminal, never asserted from recall.

Bundled assets

• assets/shepherd-driver-prompt.md -- the per-PR convergence-loop spawn body. ONE PR per subagent.
• assets/fold-vs-defer-rubric.md -- the fold-by-default decision authority (Phase X.2).
• assets/copilot-classification-prompt.md -- LEGIT/NOT-LEGIT classification of Copilot review (Phase X.0).
• assets/ci-recovery-checklist.md -- post-push CI watch + recovery loop (Phase X.6; cap 3).
• assets/conflict-resolution-prompt.md -- cross-PR rebase + faithful conflict resolution spawn body.
• assets/completion-schema.json -- JSON schema for the completion and conflict-resolution return shapes. Schema-validate every subagent return.
• assets/pr-comment-templates.md -- PR ADVISORY + SUPERSEDE comment shapes (rendered at terminal).
• references/mergeability-gate.md -- load-on-demand step-by-step for the conflict-resolution phase.