arthai-marketplace

/autopilot

Autonomous work loop, picks up issues, implements, QAs, PRs. Stops for merge approval.

Synopsis

/autopilot [issue_number] [--dry-run] [--urgent-first|--least-first]

When to use it

Quickstart

/autopilot --urgent-first

What you’ll see: the open issues and PRs ranked P0–P5, then the loop starts on the most urgent item. It self-paces through implementation and QA, opens a PR, and stops, it never merges. After you merge, it picks up the next item automatically.

Examples

/autopilot                  # rank the backlog, ask which direction to work
/autopilot 141              # work just issue #141
/autopilot --urgent-first   # P0 → P5: critical items first (recommended default)
/autopilot --least-first    # P5 → P0: warm up on small items
/autopilot --dry-run        # show the ranked queue and plan without doing anything

Arguments & flags

Flag Values Default What it does
issue_number : none Skip ranking, work this issue only
--dry-run : off Assess and classify only; report what it would do: good for trust-building
--urgent-first : : Work the queue P0 → P5
--least-first : : Work the queue P5 → P0

What it does

  1. Assess, scans open issues, your PRs, and review comments; ranks everything P0 (critical) to P5 (nice-to-have) using labels and impact rules. If no direction flag was given, it asks you, user-confirmation checkpoint. Vague issues without acceptance criteria are flagged for your triage, never auto-picked.

  2. Classify risk, scores blast radius, reversibility, confidence, and domain sensitivity (0–12). Low risk proceeds autonomously; 6–8 proceeds with notice; 9–10 stops and asks, user-confirmation checkpoint; 11–12 is refused. Migrations and auth/payment code always escalate. With the decisions product on (/arth-setup → decisions), step 2’s four scores come from the Jev decision layer instead of hand scoring. In enforce mode the bands (0–5 / 6–8 / 9–10 / 11–12) and the hard escalation rules are unchanged, the score is recorded with the item’s risk assessment as jev: {score, decision_ids}, and the loop never prompts because of Jev. Should any dimension be answered with low confidence, the item is scored by hand as today. In shadow mode (the default after setup) nothing changes and the decision is only logged.

  3. Verify, reproduces the bug before touching code; stale already-fixed issues get closed instead of “fixed” again.
  4. Plan, defines expected files/layers and objective completion criteria (each with a verifiable done_when clause).
  5. Implement, simple fixes inline; multi-file work via backend/frontend agents plus a QA agent. Every commit and test run is captured in an evidence log.
  6. QA, runs tests, lint, type checks (2 fix attempts max, then it escalates).
  7. Self-review, a separate gate after QA passes: checks scope growth (>1.5x planned stops the loop), hardcoded values, error handling, security, test coverage, and unintended changes; re-assesses risk now that code exists. Unrelated bugs found along the way are filed as issues, not fixed inline.
  8. PR, mandatory stop, opens a PR whose body is assembled from the evidence log, then stops and waits for your merge approval, user-confirmation checkpoint. It never merges. While waiting, it pre-assesses the next item.
  9. Post-merge, checks CI on main (auto-reverts and reopens the issue if its merge broke CI), records learnings, then loops to the next item. Once this completes, a completion-verifier agent checks the item’s evidence against its completion criteria (see below).

Agents spawned

Agent Model tier Role
backend / frontend sonnet Implementation, per the plan’s layers (simple fixes run inline with no spawn)
qa sonnet Reviews and validates alongside implementation agents

architect (opus) is not spawned directly by autopilot, for risk-6+ features, Phase 4 (PLAN) delegates to the /planning skill, which spawns architect itself.

Post-merge verification: after a work item’s full lifecycle completes, autopilot spawns completion-verifier (sonnet) to check the item’s evidence against its completion criteria. It reports PASS, GAPS FOUND, or INCONCLUSIVE, gaps don’t block, you decide whether to rerun.

Output & artifacts

Troubleshooting

Problem Fix
BLOCKED: Can't resolve test failure after 2 attempts The loop escalates instead of thrashing: investigate and give guidance
SCOPE: Started at 3 files, now at 6 Scope guard fired: approve the growth or trim the change
COST: This item has used 6 agent spawns Per-item budget hit: continue, skip, or split the issue into smaller ones
Issue skipped as unrankable It has no acceptance criteria: add them to the issue and re-run
Loop stopped and you want it back /autopilot resumes from .claude/.workflow-state.json; saying “stop” or any unrelated request pauses it