muse-handoff

Muse handoff (sender side)

Muse is Meta's personal agent, running on its own cloud computer with its own inference, connectors (Zapier, custom API connectors) and a browser sub-agent. It reads the muse queue directly through its Fermi MCP connection. Hand it the WHOLE ask.

1. Enqueue one task per ask

task_enqueue({
  channel: "cloud", queue: "muse", chat_id: "muse",
  sender: "cc:<host>/<session>",
  payload: JSON.stringify({ v: 1, ask, context, repo, expected_output, deadline_min: 30 })
})

• Queue is muse, never main (the Mac drain workers claim main).
• Do NOT pre-split a job into many muse tasks. Give full context and one clear expected_output. Separate unrelated asks may be separate tasks.
• repo is optional (owner/name). Raise deadline_min for big jobs.
• Good fits: work in the owner's Meta-connected apps or Muse's connectors, web research, browsing sites with no API.

2. Wait

Loop task_wait({ id, timeout_seconds: 120 }) until done/failed or deadline_min passes. timeout only means "not yet": call again.

3. Read the result

If the result starts with fs:, the full output is a file: fs_read({ path }) with the rest of the string (e.g. fs:handoff/<id>.md → handoff/<id>.md).

4. On deadline

Report "Muse hasn't finished " to your user and stop. Do not re-enqueue; task_list({ queue: "muse" }) shows whether it was claimed.

Rules

• No secrets in payloads; reference a Fermi secret by name if needed.
• Muse drafts anything outward-facing (email, posts) for the owner's approval instead of sending it.