mirror of
https://github.com/alirezarezvani/claude-skills.git
synced 2026-10-09 03:17:54 +00:00
d1f2396c6f
Adds the agent-launcher/ top-level domain — a plugin re-implementation of Anthropic's launch-your-agent reference skill (Apache-2.0; independent, not a fork) for building Claude Managed Agents (CMA) in the user's own account. Every session starts with a goal (./my-agent/goal.json, surfaced by an opt-in AGENT_LAUNCHER_SESSION=1 SessionStart hook + /cs:goal); loop_compiler.py compiles that goal into a bounded grade->iterate loop (CMA user.define_outcome self-grading, max_iterations 1..20), a recurring POSIX-cron scheduled-deployment loop, or a single-pass interview->stage->launch workflow. - 6 skills: agent-launcher-orchestrator (context: fork goal router) + interview + stage-launch + grade-iterate + run-without-you + wrap-up - 18 stdlib-only deterministic scaffolder tools (NO network/API calls; live launches emitted as BYOK curl that never prints the key); all pass --help/--sample - 4 agents (orchestrator + interviewer + grader + deployer), 8 /cs:* commands - opt-in SessionStart/SessionEnd hooks (exit 0 on any error), 5 shared references, 4 assets (build-sheet schema + overview/next-directions templates + example) - validators enforce CMA limits (<=20 skills/session, <=8 memory stores, depth-1 multiagent, max_iterations <=20, <=1000 deployments/org) - registered in marketplace.json; headline counters trued up via derive_counters.py --check (skills 362->368, domains 18->19, tools 644->664, refs 741->746, agents 102->106, commands 116->124, plugins 88->89) Distinct from engineering/agent-harness (generic bounded loop over any domain) and engineering/write-a-skill (authors Claude Code skills, not CMAs). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012FwXG6TqCXKZQvF4iD69cv
83 lines
3.9 KiB
Markdown
83 lines
3.9 KiB
Markdown
# agent-launcher
|
||
|
||
Build, launch, grade, and schedule **Claude Managed Agents (CMA)** in your own
|
||
Anthropic account — as a Claude Code plugin where **every session starts with a
|
||
goal** and that goal compiles into a **loop or a workflow**.
|
||
|
||
Inspired by Anthropic's reference skill
|
||
[`anthropics/launch-your-agent`](https://github.com/anthropics/launch-your-agent)
|
||
(Apache-2.0). This is an independent re-implementation for the claude-skills
|
||
marketplace — not a fork — that adds agents/sub-agents, an opt-in session-start
|
||
goal, deterministic scaffolders, and explicit loop/workflow compilation.
|
||
|
||
## The four phases
|
||
|
||
| Phase | Skill | Command | Loop/workflow |
|
||
|---|---|---|---|
|
||
| 1 · Interview → Plan | `interview` | `/cs:interview` | single-pass workflow |
|
||
| 2 · Stage → Launch | `stage-launch` | `/cs:stage-launch` | single-pass workflow |
|
||
| 3 · Grade → Iterate | `grade-iterate` | `/cs:grade` | **grade→iterate loop** (bounded by `max_iterations`) |
|
||
| 4 · Run Without You | `run-without-you` | `/cs:run-without-you` | **recurring cron deployment loop** |
|
||
| — · Close out | `wrap-up` | `/cs:wrap-up` | — |
|
||
|
||
`agent-launcher-orchestrator` (`context: fork`) reads the session goal, routes to
|
||
the right phase, and compiles the loop.
|
||
|
||
## Every session starts with a goal
|
||
|
||
- State lives in `./my-agent/goal.json` (your folder — it keeps working after the
|
||
session ends).
|
||
- **Set it:** `/cs:goal set "Launch an agent that triages my inbox every morning"`.
|
||
- **Resume it automatically:** enable the opt-in hook with
|
||
`export AGENT_LAUNCHER_SESSION=1`; the `SessionStart` hook surfaces the current
|
||
goal + phase so you pick up exactly where you left off. Disabled by default — no
|
||
ambient behavior in unrelated repos.
|
||
- **Advance it:** `/cs:goal advance` moves to the next phase.
|
||
|
||
## Loops vs workflows
|
||
|
||
`loop_compiler.py` compiles the goal + phase into exactly one shape:
|
||
|
||
- **single-pass workflow** — interview → plan → stage → launch (Phases 1–2).
|
||
- **grade→iterate loop** — CMA `user.define_outcome` self-grading, **bounded** by
|
||
`max_iterations` (1..20); never unbounded (Phase 3).
|
||
- **recurring deployment loop** — POSIX-cron scheduled deployment that re-runs the
|
||
goal "without you", optionally self-grading each firing (Phase 4).
|
||
|
||
See [`references/loops-and-workflows.md`](references/loops-and-workflows.md).
|
||
|
||
## Safety & hard rules
|
||
|
||
- **Deterministic scaffolders only** — every tool is stdlib-only and makes no API
|
||
calls. Live launches are emitted as **BYOK curl scripts** you run with your own
|
||
`$ANTHROPIC_API_KEY`. The key is never printed, logged, or written.
|
||
- Validators enforce CMA limits (≤20 skills/session, ≤8 memory stores, depth-1
|
||
multiagent, `max_iterations` ≤20, …).
|
||
- The opt-in hook can never break a session (exits 0 on any error).
|
||
|
||
## Quick start
|
||
|
||
```bash
|
||
export AGENT_LAUNCHER_SESSION=1 # optional: auto-surface the goal each session
|
||
/cs:goal set "Nightly repo dependency auditor that writes report.md"
|
||
/cs:launch # runs the orchestrator from the current phase
|
||
```
|
||
|
||
## Layout
|
||
|
||
```
|
||
agent-launcher/
|
||
├── SPEC.md # the build goal (verification target)
|
||
├── skills/ # 6 skills, 3 stdlib tools each
|
||
├── agents/ # 4 agents (orchestrator + interviewer + grader + deployer)
|
||
├── commands/ # 8 /cs:* commands
|
||
├── hooks/ # opt-in SessionStart / SessionEnd
|
||
├── references/ # 5 shared reference docs
|
||
└── assets/ # build-sheet schema, overview + NEXT-DIRECTIONS templates, example
|
||
```
|
||
|
||
## Attribution
|
||
|
||
Inspired by `anthropics/launch-your-agent` (Apache-2.0). CMA primitive semantics
|
||
are drawn from the public [Claude Managed Agents docs](https://platform.claude.com/docs/en/managed-agents/overview).
|
||
No upstream code is copied verbatim. License: MIT (this plugin).
|