autopilot v2 — Intent → PR pipeline

Take a natural-language intent and run the full development pipeline to a polished PR. Two gates total: brainstorm approval (Gate 1) + PR review on GitHub (Gate 2 — the existing PR semantics). Multi-agent adversarial review replaces the single-reviewer model from v1.

NOTE: kaijutsu-managed. Re-running jutsu install autopilot overwrites this file. The previous v1 SKILL.md (if any) is preserved at ~/.claude/skills/autopilot/.archived/SKILL.md.

Command Detection

• /autopilot {intent} → Run mode (start new pipeline)
• /autopilot resume → Resume mode (continue interrupted run)
• /autopilot abort → Abort mode (clean up + stop)
• /autopilot status → Status mode (show current phase)

CLI equivalents: jutsu autopilot {init,status,abort,resume,run}. Use the slash command for interactive runs; the cobra group is for scripted use + CI test mode.

Phase Map

/autopilot {intent}
  → Phase 1: BRAINSTORM (interactive — anti-sycophancy via jutsu swarm dream)
  → ★ GATE 1 (configurable: skip | review) ★
  → Phase 2: SPEC
      → write spec to docs/specs/<date>-<slug>.md
      → swarm doc-review on spec [report-only]
      → record approved-spec path in state file
  → Phase 3: PLAN
      → write plan to IMPLEMENTATION_PLAN.md
      → swarm doc-review on plan [report-only]
  → Phase 4: BUILD (stage 1..N — each followed by polish loop)
  → Phase 5: SHIP
      → final swarm pr-review (full lens)
      → reverse-drift gate (informational, tagged on PR)
      → gh pr create  (or test-mode: write planned PR JSON)
  → ★ GATE 2 = PR review on GitHub ★
  → Phase 6: LEARN (state file, findings.db, MEMORY.md)

Cost Cap

Three layers (per spec Decision #3):

1. Soft cap — config default $20 (overridable via .kaijutsu/autopilot.yaml::cost.max_total_usd); at 50% of soft cap, autopilot pauses + asks user to confirm continue.
2. Hard ceiling — MaxAutopilotCostUSD = $100 constant in cli/internal/cli/autopilot.go. Cannot be raised by editing yaml. Per-run override via --max-cost N works only up to ceiling.
3. Per-shell ceiling override — set KAIJUTSU_AUTOPILOT_HARD_CAP_OVERRIDE=N in your shell environment to raise the ceiling for THIS shell. Not stored in repo. Defense against malicious yaml.

Above the ceiling without env var → error. Pre-flight jutsu swarm <preset> --estimate projection runs at Gate 1 before any model dispatch.

Phases (detailed)

Phase 1: BRAINSTORM (interactive)

Interactive conversation with the user about the intent. Anti-sycophancy enabled by default: invoke jutsu swarm dream --mode full --strict --personas brainstorm-creative-claude,architecture-purist-antigravity,claim-auditor-claude against the leading direction. User sees the design ALREADY ARGUED AGAINST before approving Gate 1.

Step 1: Read project context (parallel):

• CLAUDE.md (project rules + constraints)
• ~/.claude/projects/*/memory/MEMORY.md matching this project
• .claude/decisions/*.md with status: active
• git log --oneline -20
• Past autopilot runs from memory (entries named autopilot-run-*)

Step 2: Scan for related code — grep + skill-suggest:

• jutsu suggest "<intent>" to see if installed skills already handle sub-problems
• Grep the codebase for keywords from the intent

Step 3: Anti-sycophancy dream pass (when autopilot.yaml gates.brainstorm: review):

• Run jutsu swarm dream --mode full against the leading design
• Surface load-bearing findings + anti-skill arguments to the user

Step 4: GATE 1 — present design to user with the dream findings already attached:

## Autopilot: Direction Review

**Intent:** "{original intent}"
**Leading approach:** {1-paragraph summary}
**Why this approach:** {2-3 sentences}
**Dream's adversarial findings:** {top 3 load-bearing concerns}
**Estimated cost (pre-flight):** ${X}

Approve, revise, or reject?

Wait for user response. Approve → Phase 2. Revise → re-explore. Reject → abort.

Phase 2: SPEC

Step 1: Write spec to docs/specs/<date>-<slug>.md following the brainstorming spec format.

Step 2: Run jutsu swarm doc-review --personas architecture-purist-antigravity,paranoid-security-claude,claim-auditor-claude <spec-path>.

Step 3: Findings printed to stderr inline + appended to <workspace>/findings.md. Findings are report-only (per spec Decision #2). Autopilot does NOT mutate the spec automatically. User can apply specific findings via re-run after editing.

Step 4: Record approved spec path in .kaijutsu/autopilot-state.md (approved_spec_path key) for the reverse-drift gate to consume.

Phase 3: PLAN

Step 1: Write IMPLEMENTATION_PLAN.md following the writing-plans skill format — bite-sized stages with exact file paths + test commands.

Step 2: Run jutsu swarm doc-review --personas perf-purist-codex,claim-auditor-claude,cross-file-antigravity <plan-path>.

Step 3: Findings report-only (same as spec).

Phase 4: BUILD

For each stage in the plan:

1. Implement the stage — fresh subagent dispatched per task per the subagent-driven-development protocol.
2. Polish loop — invoke /polish skill against the stage's diff. Up to 4 review-fix passes per stage. If findings remain after 4, escalate to user.
3. Run tests — project's test command must pass before moving to next stage.
4. Commit — descriptive message linking to the spec + plan.

Track cost across all stages. Pause at 50% of soft cap; abort at hard ceiling without override.

Phase 5: SHIP

Step 1: Final swarm pr-review (full lens) — jutsu swarm pr-review --diff-from-branch main --strict --personas paranoid-security-claude,architecture-purist-antigravity,perf-purist-codex,claim-auditor-claude --mode full. Apply load-bearing findings; defensible-skip the rest.

Step 2: Reverse-drift gate — jutsu swarm reverse --spec <state.approved_spec_path> --diff origin/main...HEAD. Drift findings appended to PR description as <details> block. PR labeled autopilot-drift if drift surfaced.

Step 3: Open PR — gh pr create with synthesized body. In test mode (KAIJUTSU_AUTOPILOT_TEST_MODE=1), writes planned PR to .kaijutsu/autopilot-pr.json instead.

Phase 6: LEARN

After PR open (or abort):

1. Write run summary to ~/.claude/projects/<project>/memory/autopilot-run-<date>-<slug>.md.
2. Findings recorded into ~/.kaijutsu/findings.db (existing v0.7 schema). Each autopilot run adds rows so per-agent confidence weights tune the next run.
3. Update MEMORY.md if a lesson is broadly applicable.

State File

Path: .kaijutsu/autopilot-state.md. YAML frontmatter, no body.

---
intent: "<original intent>"
phase: brainstorm | spec | plan | build | ship | learn | done
slug: <feature-slug>
approved_spec_path: docs/specs/<date>-<slug>.md
plan_path: IMPLEMENTATION_PLAN.md
branch: feat/<slug>
pr: <url-or-empty>
started_at: <ISO timestamp>
gate_responses: { brainstorm: approve | revise | reject }
cost_so_far_usd: <running total>
---

State file enables resume + status + abort to work after session interruption.

Findings Application Path

Per spec Decision #2: doc-review findings are report-only. Never mutate the artifact silently.

• Findings printed to stderr inline during the swarm run.
• Same findings written to <workspace>/findings.md.
• User reviews findings; if changes are wanted, user edits the spec/plan directly OR re-runs autopilot after editing.
• Autopilot continues to next phase without applying findings unless severity is blocker (per findings.escalate_threshold config) — in which case autopilot pauses + asks user how to proceed.

This preserves the reviewer/generator boundary that dream finding #4 surfaced.

Skill References

See references/:

• config-schema.md — full .kaijutsu/autopilot.yaml reference
• cost-model.md — soft cap + hard ceiling + env override
• persona-defaults.md — which personas run at which phase + how to override

See runbooks/:

• abort-and-resume.md — state file format + resume protocol
• drift-detection.md — reverse-drift gate behavior + interpretation

Hard Rules

• Two gates only. Brainstorm approval + PR review on GitHub. No mid-flight gates by default.
• Findings are report-only. Never auto-apply.
• Cost ceiling is non-negotiable. Hard ceiling lives in skill code, not yaml.
• Reverse-drift is informational, not blocking. Drift surfaces on the PR; user decides.
• Replace, don't extend. v1 autopilot at ~/.claude/skills/autopilot/.archived/SKILL.md (preserved on install).