agent-launcher-orchestrator — скилл для Claude Code

---
name: agent-launcher-orchestrator
description: Use when a user wants to build, launch, grade, or schedule a Claude Managed Agent (CMA) in their own Anthropic account — "build me an agent", "launch this as a managed agent", "run this on a schedule", "grade my agent against a rubric", "set up a nightly worker". Reads the per-session goal (./my-agent/goal.json), routes deterministically to one of five phase sub-skills (interview → stage-launch → grade-iterate → run-without-you → wrap-up) via goal_router.py, and compiles the goal+phase into an execution shape (single-pass workflow / bounded grade→iterate loop / recurring cron deployment loop) via loop_compiler.py. Forks context so heavy intake (build sheets, payloads, eval cases) stays out of the parent thread. All launches are emitted as BYOK curl the user runs with their own key; no tool makes API calls. Inspired by anthropics/launch-your-agent (Apache-2.0). Distinct from engineering/agent-harness (generic domain loop) and engineering/write-a-skill (authors Claude Code skills, not CMAs).
context: fork
license: MIT
metadata:
  version: 2.11.2
  author: Alireza Rezvani
  tags: [claude-managed-agents, cma, agent, launch, orchestrator, session-goal, loop, workflow, cron, outcome, byok]
  compatible_tools: [claude-code, codex-cli, cursor, antigravity, opencode, gemini-cli]
---

# agent-launcher — Domain Orchestrator

Every session starts with a **goal** — one sentence for one CMA. This orchestrator
reads that goal, routes to the right phase, and compiles the goal into a **loop or
a workflow**. Heavy intake stays in the forked context; the parent gets a digest.

Inspired by Anthropic's `launch-your-agent` reference skill (Apache-2.0). This is
an independent re-implementation; CMA semantics come from
[`../../references/cma-primitives.md`](../../references/cma-primitives.md).

## The through-line: the session goal

State lives at `./my-agent/goal.json` (the user's folder). Manage it with
`goal_state.py` (init / set / status / advance) — it also backs the `/cs-goal`
command and the opt-in `SessionStart` hook. The goal's `phase` selects the lane;
the phase + recurrence selects the loop shape.

## Routing (deterministic)

Run the router, then act on its exit code:

```bash
python3 scripts/goal_router.py --out-dir ./my-agent
# exit 0 ROUTE  -> fork to the named phase sub-skill
# exit 3 ASK    -> ask the one printed forcing question, then re-route
# exit 4 REFUSE -> goal too vague; get one sentence, then re-route
```

| Lane (phase) | Sub-skill | Loop/workflow |
|---|---|---|
| interview | `interview` | single-pass workflow |
| stage-launch | `stage-launch` | single-pass workflow |
| grade-iterate | `grade-iterate` | **bounded grade→iterate loop** |
| run-without-you | `run-without-you` | **recurring cron deployment loop** |
| wrap-up | `wrap-up` | — |

## Compile the loop

```bash
python3 scripts/loop_compiler.py \
  --out-dir ./my-agent --max-iterations 5 --cron "0 9 * * *" --timezone Europe/Berlin --nest-outcome
```

`loop_compiler.py` emits `plan.v1`: `single-pass`, `grade-iterate` (always with a
`max_iterations` cap 1..20), or `cron-loop` (optionally nesting a self-grading
outcome per firing). See [`../../references/loops-and-workflows.md`](../../references/loops-and-workflows.md).

## Pre-flight gates (hard refusals)

1. **No goal set.** If `goal.json` is missing, run
   `goal_state.py init --goal "..."` first. The orchestrator does not guess a goal.
2. **Goal too vague.** Router exit 4 — get one sentence naming the one job before
   routing. Never route on under-3-word goals.
3. **Never make API calls.** Emit BYOK curl; the user runs it with their own
   `$ANTHROPIC_API_KEY`. No script in this plugin touches the network.
4. **Never print the key.** Launch scripts read the key from the environment.

## Hand-off contract

After routing, fork to the sub-skill with: the goal string, `agent_name`,
`out_dir` (`./my-agent`), and the compiled `plan.v1`. When the sub-skill returns,
`goal_state.py advance` moves the phase and the parent gets a ≤100-word digest
(phase done, artifact paths, loop shape, one next step).

## Forcing-question library (walk one at a time; recommend + cite)

1. **"What one job should this agent do end-to-end?"** — *Recommend:* the single
   most repeated task. *Cite:* interview-to-config.md (six intake slots). Refuse to
   route a two-job goal; split into two `./my-agent-*/` folders.
2. **"What kicks it off — you ask it, an event, or a schedule?"** — *Recommend:*
   on-demand for v0, schedule as the Phase-4 upgrade. *Cite:* loops-and-workflows.md.
3. **"How would you grade a good run?"** — *Recommend:* 3–5 rubric lines grounded
   in the output. *Cite:* cma-primitives.md (outcomes; rubric required).
4. **"Is a real integration ready, or do we mock it in v0?"** — *Recommend:* mock
   with a schema-true custom tool; wire the MCP server as v1. *Cite:* interview-to-config.md.
5. **"Should run #10 be smarter than run #1?"** — *Recommend:* attach a memory
   store only if yes; else skip it. *Cite:* cma-primitives.md (memory limits + injection risk).

## Tools

- `scripts/goal_state.py` — own `goal.json` (init/set/status/advance).
- `scripts/goal_router.py` — goal → lane (exit 0 route / 3 ask / 4 refuse).
- `scripts/loop_compiler.py` — goal+phase → `plan.v1` execution shape.