orbit-run

Route this task through the Orbit loop — do not free-edit the codebase directly.

Task: $ARGUMENTS

1. READ — read CLAUDE.md and .orbit/STATE.md. If there is no .orbit/ system in this repo yet, tell the user to run /orbit first to set it up, then stop.

2. QUEUE — append the task to .orbit/STATE.md's task queue with a clear done-gate.

3. Board FIRST, then RUN the cycle via the sub-agent roster in .claude/agents/: Dispatcher → Product Discovery → Business Analyst → Market Researcher → Planner → Designer on UI work → Engineer → Safety → Reviewer → QA Engineer → CPO → Reporter. For substantial work every bold stage owner is mandatory. Dispatch each actual role and require its completed activity event; mentioning it, privately applying its “lens,” or listing it as available does not count. Lite mode runs the stages sequentially. Approval controls wider concurrency, not whether Orbit may skip its capabilities. Your FIRST action — before spawning any specialist — is to make the board visible: call .orbit/activity.py set_team([...]) to declare who's active + who's queued, set_tasks([...]) to write the checklist, and build the native list with TaskCreate — all up front, so the user sees the plan and who owns each step immediately, not after the work is done. Open with a one-line assignment. Only then dispatch roles with the Task tool. Before a long sub-agent wait, print the inline board (scripts/orbit-status --team) — never leave the user on only "waiting for background agent." Show the checklist two ways (do both): (a) write .orbit/tasks.json via .orbit/activity.py set_tasks/update_task — the guaranteed-visible path that feeds orbit-status; (b) also build the native checklist with TaskCreate/TaskUpdate (the Task* tools — NOT TodoWrite, which is off by default), role-tagged ([data] …, [safety] …), driven by you, the main orchestrator (a subagent's task calls don't surface). Update items to in_progress/completed as roles finish. Don't just narrate [role] lines and skip the files.

   Do NOT run an Orbit task through the native Workflow(...) background runner. It executes a black-box job (Running in background · /workflows to monitor) that bypasses Orbit's entire operating model: the role-tagged checklist, the visible current owner, .orbit/tasks.json, and .orbit/activity.jsonl. Orbit's promise is watch the team work — a task is not "running through Orbit" unless the user can see who owns each step and what's done / in progress. Use the Task tool for sub-agents (each emits start/done via .orbit/activity.py) and drive the checklist yourself. (Workflow(...) is fine for developing Orbit itself; it is banned as the run path for a scaffolded repo's tasks.)

4. COUNTERFACTUAL PREFLIGHT for T2+ work, before implementation: use .orbit/skills/counterfactual-regret.md and .orbit/counterfactual.py to attack the riskiest assumption with one cheap falsification probe. Record Assumption -> Probe -> Evidence -> Decision. A failed probe must route back to its typed phase (discovery, plan, build, or review) and update the checklist before continuing.

5. EVALUATE against CLAUDE.md §3 and the gates in .orbit/loop.config.json. Honor the stop conditions (§8) and approval checkpoints — propose, never auto-perform, anything irreversible, financial, or outward-facing.

6. REPAIR ITERATIVELY when Reviewer, QA, or Safety finds a failure: create a structured repair-<id>.json packet per .orbit/skills/iterative-repair.md, assign a targeted fix, retest the original failure plus a regression check, and backtrack to the right phase. Stop after two attempts on the same failure and escalate; never repeat the same blind repair.

7. UPDATE + REPORT — fold results into .orbit/STATE.md (snapshot, queue, cycle log) and give a short, decision-ready summary of what changed.