Agent skill · alirezarezvani
agent-launcher-orchestrator
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).
What it needs
About 3k tokens when loaded.
What this skill does
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. The through-line: the session goal State lives at ./my-agent/goal.json (the user's folder). Manage it with goalstate.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: 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 loopcompiler.py emits plan.v1: single-pass, grade-iterate (always with a maxiterations cap 1..20), or cron-loop (optionally nesting a self-grading outcome per firing). See ../../references/loops-and-workflows.md. Pre-flight gates (hard refusals) 1. No goal set. If goal.json is missing, run goalstate.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 $ANTHROPICAPIKEY. 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, agentname, outdir (./my-agent), and the compiled plan.v1. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills alirezarezvani/agent-launcher-orchestrator--b73813