/plan
Ground a plan before editing.
Meta says the skill inspects the real project, records decisions, and stops for approval without changing code.
Runtime map · based on Meta documentation
How Muse Code keeps long work moving.
The launch architecture centers on persistent specialist agents, isolated git worktrees, a local append-only event log, and skills that shape a turn only when explicitly invoked.
01 · Main loop
A simple loop with specialists in the background.
Meta describes a main agent loop augmented by specialized background agents that remain active for the session. They carry out next steps and decide when to report back, reducing repeated information gathering.
Receive the goal
The main agent interprets the task, project state, and current session record.
Plan and delegate
Background agents can research or execute follow-on work without being recreated for every step.
Use tools and worktrees
Fan-out children can work on separate branches and isolated git worktrees, according to Meta’s cookbook.
Validate and report
Results return to the parent for review, merging, and the next decision.
02 · Event log
The session is a sequence of recorded events.
Every model call, tool run, approval, and edit is appended to a local event log, Meta says. That record is the documented basis for auditing and rebuilding state after a crash.
01 session.start
02 goal.accepted
03 agent.spawned
04 tool.requested
05 approval.recorded
06 edit.applied
07 validation.completed
08 session.checkpoint
Note: this sequence explains the concept.
It is not a verbatim Muse Code schema.
03 · Bundled skills
/grill
Stress-test a weak plan.
The research announcement describes an interview that challenges the plan until the decisions hold up.
/goal
Persist toward an objective.
Meta describes goal tracking that keeps a long-running agent aligned with the requested outcome.
Meta’s developer article names additional bundled skills. The set may evolve during beta.
04 · Resume
Recovery is only as trustworthy as the record.
Meta says muse resume reads the same session log and continues from the last recorded step. A practical evaluation should interrupt a real task, verify the reconstructed state, and inspect whether pending approvals or tool results are handled correctly.
Test this boundary
Interrupt after an edit but before validation. Resume the session, confirm the working tree and agent roster, then verify that validation runs exactly once and no approval is silently replayed.