Multi-agent composition with /compose

What /compose does

/compose takes a free-form goal and produces a publish-ready bundle of:

• Reuse entries — existing registry skills that already cover part of the goal.
• Drafts — new skills, agents, or MCP servers the planner says you still need to write.

You walk away with a plan you can edit, sign, and submit — or just read to figure out what's already on the registry.

The phases under the hood

Each compose run goes through these phases. The progress bar in the UI ticks through them live:

1. Classifying — a fast 1-second router looks at the goal and decides whether you need a single skill, a single agent, an MCP server, or a genuinely complex bundle.
2. Planning — for complex bundles, a reasoning model produces the reuse list and the draft outline.
3. Drafting — each draft entity is written by a type-specific drafter (skill / agent / MCP). They run in parallel.
4. Analyzing — every draft gets four scores from a critic model (specificity, completeness, safety, goal alignment). Drafts under threshold get one re-roll with the critic's notes.
5. Persisting — the plan + drafts are saved under a plan_id so you can pick up the bundle wizard from /publish/bundle?plan=<id>.

Bring your own key

If you have an OpenAI / Anthropic / OpenRouter key, switch on the BYOK panel before submitting. Your key is held in memory for the duration of one request and then dropped — never persisted, never logged. The endpoint shape (/v1/intent/byo) is sync, so closing the tab cancels the run cleanly.

When /compose is and isn't useful

• Useful when you're trying to figure out what to publish. The reuse list shows you what's already on the registry; the drafts scaffold what isn't.
• Useful as a kickoff for a bundle wizard — /publish/bundle?plan=<id> pre-fills with everything compose generated.
• Less useful when you already know exactly what you want. If you've written the SKILL.md by hand, skip compose and go straight to /publish.