Muse Code deterministic replay: golden fixtures gate CI offline
At a glance
Muse Code journals sessions to an append-only event log. Author a focused golden fixture (records + expected_model_context), then `muse trace inspect --fixture` replays projection with no model, tools, or network. Gate merges on zero diffs via replay_gate.sh / GitHub Actions—catch harness drift before it ships. This workflow is listed under Meta-Muse, Muse-Code, CI, devtools, agent-ops, testing.
Official Muse Code cookbook: four built-in playbooks load only on the turn you invoke them. /plan grounds a decision-complete plan and stops for approval; /grilling and /grill-with-docs pressure-test designs (optionally writing ADRs); /taste strips AI-slop UI defaults. Explicit-invocation only—won’t auto-plan ordinary implement requests.
Official Muse Code docs: request a workflow to fan out up to 1,000 child tasks (active cap ~16) for audits, migrations, or research→verify→synthesize. `/workflows` control room cancel/pause/skip; save reusable JS under `.agents/workflows/` via `muse workflows save`; settings auto|explicit|off. Note: some Apple silicon builds omit the workflow engine.
Muse Code is Meta’s terminal/CI coding agent: run `muse` interactively or `muse exec "…"` headless for scripts and pipelines—plans, edits, and validates with approvals and an OS sandbox from the first run. Pair with Muse Spark/Wonder via Meta Model API when you build your own harness; use Muse Code when you want the ready-made agent loop.
Muse Code can load skills from built-in, user, project, and plugin sources, and discovers `~/.claude/skills` and Codex skill roots by default. The bundled `/migrate` skill imports Claude Code or Codex memory notes and MCP servers; `muse skills import --from claude|codex` brings skills across. Built-ins include `/plan`, `/grill` (stress-test a plan into a written scope contract), `/taste`, and `/threejs`.