agent-launcher-orchestrator

1# agent-launcher — Domain Orchestrator
2 
3Every session starts with a **goal** — one sentence for one CMA. This orchestrator
4reads that goal, routes to the right phase, and compiles the goal into a **loop or
5a workflow**. Heavy intake stays in the forked context; the parent gets a digest.
6 
7Inspired by Anthropic's `launch-your-agent` reference skill (Apache-2.0). This is
8an independent re-implementation; CMA semantics come from
9[`../../references/cma-primitives.md`](../../references/cma-primitives.md).
10 
11## The through-line: the session goal
12 
13State lives at `./my-agent/goal.json` (the user's folder). Manage it with
14`goal_state.py` (init / set / status / advance) — it also backs the `/cs-goal`
15command and the opt-in `SessionStart` hook. The goal's `phase` selects the lane;
16the phase + recurrence selects the loop shape.
17 
18## Routing (deterministic)
19 
20Run the router, then act on its exit code:
21 
22```bash
23python3 scripts/goal_router.py --out-dir ./my-agent
24# exit 0 ROUTE  -> fork to the named phase sub-skill
25# exit 3 ASK    -> ask the one printed forcing question, then re-route
26# exit 4 REFUSE -> goal too vague; get one sentence, then re-route
27```
28 
29| Lane (phase) | Sub-skill | Loop/workflow |
30|---|---|---|
31| interview | `interview` | single-pass workflow |
32| stage-launch | `stage-launch` | single-pass workflow |
33| grade-iterate | `grade-iterate` | **bounded grade→iterate loop** |
34| run-without-you | `run-without-you` | **recurring cron deployment loop** |
35| wrap-up | `wrap-up` | — |
36…