orchestratelisted

# Orchestrate Runs a **single structured multi-step task** end-to-end — "Claude with a plan": picks the right execution surface (background task, durable workflow, or direct subagent), runs the steps, hands back a result or a trackable `run_id`. `/orchestrate <goal>` — the text after `/orchestrate` IS the goal, use it verbatim, never re-ask it. Invoked bare (no argument) → ask ONE free-text `AskUserQuestion`: `"What should I orchestrate?"` Whatever the phrasing — a task, or a question about orchestrate itself — it's the goal to run; never substitute a hand-written explanation for actually picking a surface and running it. ## Operating loop 1. Ground: goal (from the argument, or the answer above), expected deliverable, acceptance signal. 2. Pick the narrowest execution surface (`AskUserQuestion` only when the user's intent doesn't decide it): - durable/resumable run → the **`workflow`** MCP tool: smallest valid spec, `workflow` with `op="run"` - **`isolated`** (detached/background) → the host's background-task surface - otherwise → a direct child subagent 3. Return the `run_id` / task handle / child-run handle + how to inspect progress. ## Questions to gather `AskUserQuestion` for what's still missing after the goal is set — unknowns batched into a single call (up to 4 questions). Gather until clear: - deliverable / acceptance signal, if not already clear from the goal - launch mode: durable workflow, isolated/background, or direct subagent - workflow shape, if