arthai-marketplace

Run Autonomous Work

The toolkit has two autonomous loops. Both work without you babysitting them, both auto-continue between phases, and both stop at PR creation — merging is always yours. The difference is the shape of the work.

Which one?

You have… Run
One objective, fuzzy path (“Cut homepage LCP under 2s”) /goal
Multiple issues, ranked queue (“Work through these 8 issues”) /autopilot
One feature with a written plan /implement — not an autonomous loop
One bug with a clear repro /fix

If the work is “I have a destination, figure out the path” — /goal. If it’s “I have a stack of well-scoped tickets, work them” — /autopilot.

/goal — speed-first, single objective

/goal Cut homepage LCP below 2s on mobile

A scout reads your knowledge base and project profile first, then scans the codebase to fill gaps; you answer 3–5 context-aware clarifying questions (or say “go” for defaults) and confirm the plan. Then it loops: pick action → execute → mandatory verify (lint + types + tests) → capture evidence → self-evaluate against each subtask’s done_when clause. It runs inline, pulling in Sonnet agents only when needed, with a budget of 6 agents per goal. Lifecycle: /goal pause, resume, clear, status; state lives in .claude/.goals/current.json.

/autopilot — rigor-first, backlog queue

/autopilot --urgent-first

It ranks open issues P0–P5, then per item: classify → verify repro → plan → implement (with backend/frontend/QA agents as needed) → QA → self-review → PR. The PR body is assembled directly from the captured evidence array (git diffs, test results, lint, types). --dry-run shows what it would do without doing it — useful for trust-building. State lives in .claude/.workflow-state.json.

The safety gates

Both skills ship in the cruise bundle, which requires forge + scalpel + sentinel (see the plugin catalog).