AI Autopilot

Autopilot v2

Delivers large multi-concern specs and backlog runs autonomously — decomposing into sub-specs (or a backlog DAG), deep-planning with parallel agents, implementing in waves, and shipping via PR with one bounded quality-remediation pass. Use it for end-to-end delivery of specs with ≥3 concerns or ≥10 file changes; use /ai-build for smaller single-concern work.

Purpose

Autonomous execution of large approved specs via a 6-phase pipeline: decompose into N focused sub-specs, deep-plan each with parallel agents, orchestrate a dependency-aware DAG, implement in waves, run one final verify+guard+review pass (with one bounded quality-remediation pass on the full changeset), deliver via PR with a transparency report. One invocation, full disclosure. Done only when sub-spec work converges into a delivery PR against protected main.

Thin orchestrator: phases READ other skills' SKILL.md and EMBED instructions into subagent prompts (no inline implementation), so autopilot inherits their improvements.

When NOT to Use

Triggers (≥3 concerns / ≥10 files, post-/ai-brainstorm approval, --backlog) live in the description + Dispatch threshold. Do NOT use for:

• Need human review between phases — use /ai-build with manual checkpoints.
• Cross-repo changes — coordinate manually.
• Data migrations with destructive DDL — require explicit user approval per step.

Process

Step 0 — Validate: confirm .ai-engineering/specs/spec.md is not a placeholder (else STOP → /ai-brainstorm). On --resume, read .ai-engineering/runtime/autopilot/manifest.md and re-enter at the Resume Protocol. Load stack contexts (manifest providers.stacks + .ai-engineering/overrides/<stack>/conventions.md); pass paths (not content) to subagents. plan.md is not required — Phase 2 agents generate their own. ai-eng host probe is diagnostic/advisory only; ok_to_dispatch is not a standard-flow execution gate and cannot block /ai-autopilot.

Step 1 — DECOMPOSE (handlers/phase-decompose.md): extract N independent concerns; abort if N<3 (recommend /ai-build); write sub-spec dirs + the execution manifest.

Step 2 — DEEP PLAN (handlers/phase-deep-plan.md): dispatch explore+plan agents in parallel; each enriches sub-NNN/spec.md (Exploration) + plan.md (checkbox tasks with exports/imports). Failed agents retry once → mark plan-failed.

Step 3 — ORCHESTRATE (handlers/phase-orchestrate.md): build the file-overlap matrix + import-chain graph from Phase-2 evidence (never from spec text alone); construct the wave-assigned DAG; merge unresolvable conflicts.

Step 4 — IMPLEMENT (handlers/phase-implement.md): per-wave kernel from .claude/skills/_shared/execution-kernel.md. Dispatch build agents per sub-spec in parallel within a wave; each task self-validates via TDD; wave-end guard advisory remains for governance. Collect Self-Reports + per-wave commits. Cascade-block dependents of failed sub-specs.

Step 5 — QUALITY LOOP (handlers/phase-quality.md): read ai-verify / ai-review / ai-governance SKILL.md once at loop entry; dispatch verify+guard+review in parallel on the full changeset; consolidate findings (unified severity). Clean → Phase 6. Blocker/critical/high findings enter Phase 5b only if the one bounded quality-remediation pass is unused and the fixes are finding-scoped. Remaining blocker/critical/high findings after final reassessment → STOP + escalate to user.

Phase 5b — BOUNDED REMEDIATION (handlers/phase-quality.md): persist quality_remediation.max_attempts: 1 in the autopilot manifest, map each finding to sub-NNN, integration, or shared, run cross-platform focal reproducers, return to Step 5 final reassessment. No re-decompose, no re-plan, no second remediation pass.

Step 6 — DELIVER (handlers/phase-deliver.md): build the Integrity Report; follow /ai-pr SKILL.md; cleanup runtime dir; clear spec.md + plan.md; verify cleanup. --resume handles mid-pipeline re-entry.

Flags

Dispatch threshold

Dispatch the ai-autopilot agent when work has ≥3 independent concerns, touches ≥10 files, follows /ai-brainstorm approval (spec.md exists, not a placeholder), or runs any --backlog; smaller scope hands off to /ai-build. The agent handle is .claude/agents/ai-autopilot.md; the procedural contract lives in this SKILL.md.

Governance

DEC-023: invocation is the single approval gate; internal gates (sub-spec validation, DAG verification, quality convergence) are automatic and cannot be bypassed. The consolidation path is mandatory: sub-spec branch/worktree → wave or integration commits → final PR → protected main. State transitions live on disk in .ai-engineering/runtime/autopilot/manifest.md, never in agent memory.

Examples

See references/examples.md for the three canonical invocations (end-to-end delivery, --resume after interruption, --backlog --source github), the failure-recovery rows, the telemetry event taxonomy, and the common-mistakes checklist (never run on draft specs, never cross repos, never carry context across sub-specs, never hand-edit mirrors).

Integration

Called by: user directly post-/ai-brainstorm approval (or with --backlog for backlog runs). Reads: _shared/execution-kernel.md, ai-verify/SKILL.md, ai-review/SKILL.md, ai-governance/SKILL.md, ai-pr/SKILL.md, ai-commit/SKILL.md. Delegates to: ai-explore, ai-build, ai-verify, ai-advise, ai-review agents. Transitions to: /ai-branch-cleanup. See also: /ai-build (smaller scope), /ai-board sync (lifecycle transitions for backlog mode), references/examples.md.

$ARGUMENTS