mirror of https://github.com/alirezarezvani/claude-skills.git synced 2026-08-28 04:24:58 +00:00

feat(agent-launcher): new domain plugin for building Claude Managed Agents

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

2026-08-17 02:39:34 +00:00

2.8 KiB

Raw Blame History

CLAUDE.md — agent-launcher domain

Guidance for working inside agent-launcher/. See the root CLAUDE.md for repo-wide rules.

What this domain is

A plugin that scaffolds and launches Claude Managed Agents (CMA) in the user's own Anthropic account, organized around a per-session goal that compiles into a loop or workflow. Inspired by anthropics/launch-your-agent (Apache-2.0); independent re-implementation, not a fork.

Non-negotiable rules (enforced by SPEC.md + validators)

Deterministic scaffolders only. Every script under skills/*/scripts/ is stdlib-only and makes no network/API calls. Live launches are emitted as BYOK curl. Do not add requests, anthropic, or any network client.
Never surface the API key. Launch scripts read $ANTHROPIC_API_KEY; no tool echoes/logs/writes a key. payload_validator.py and reviewers check this.
CMA limits are law. Keep references/cma-primitives.md as the source of truth for ceilings; validators must match it.
Bounded loops only. loop_compiler.py never emits a grade→iterate loop without a max_iterations cap (1..20).
The hook is opt-in and crash-proof. Gated by AGENT_LAUNCHER_SESSION=1; exits 0 on any error. Never make it fire unconditionally.
The folder is the user's. All artifacts go under ./my-agent/; scripts accept --out-dir and default there. Never write into the plugin folder.

Structure

skills/agent-launcher-orchestrator/ — context: fork goal router (goal_router.py, goal_state.py, loop_compiler.py).
skills/{interview,stage-launch,grade-iterate,run-without-you,wrap-up}/ — one phase each, 3 tools each.
agents/ — cs-agent-launcher-orchestrator + 3 phase specialists.
commands/ — 8 /cs:* commands.
hooks/ — opt-in session_start.py / session_end.py + hooks.json.
references/ — 5 shared docs. assets/ — schema + templates + example.

Tool conventions

Every tool: argparse with real --help, a --sample that runs a deterministic demo and exits 0, and JSON output via --json where a machine reads it.
Default output dir ./my-agent/; never assume network access.
Import shared logic via relative sys.path insert (see how the orchestrator tools import goal_state).

Forcing-question discipline

Every SKILL.md ships a "Forcing-question library" (Matt Pocock grill-with-docs): walk one question at a time, recommend an answer, cite the reference. The /cs:grill-agent-launcher command surfaces them.

When editing

Changing a CMA limit → update references/cma-primitives.md and every validator in lockstep.
Adding a loop shape → update loop_compiler.py and references/loops-and-workflows.md.
Keep SPEC.md authoritative; if you ship something different, record it in the delivery report.

2.8 KiB Raw Blame History