oh-my-musecode delivery harness

# oh-my-musecode

An OMC-shaped delivery harness for
[Meta Muse Code](https://dev.meta.ai/docs/cookbook#building-with-muse-code) (the
`muse` CLI): a gated pipeline from a vague request to verified working code. It
ships a native Muse plugin manifest. On Muse 1.3.0-R3057.1 plugins are enabled
and the npm installer uses the marketplace plugin route, falling back to
`muse skills install` plus a `settings.json` merge on builds without plugin
support (builds through 1.1.1 reported plugins as unavailable). See Install below.

It ports oh-my-claudecode's Tier-0 pipeline —
`deep-interview → deep-dive/trace
→ ralplan → ralph`, plus `team` and `cancel` —
onto `muse`, respecting Muse's approval/sandbox/trust/audit model instead of
working around it. See [`docs/recipe.md`](docs/recipe.md) for the full
walkthrough and acceptance run, and
[`docs/live-probes-1.3.0.md`](docs/live-probes-1.3.0.md) for the committed
live-binary probe evidence (flags, gates, validate shapes) plus the explicit
unresolved-items list.

## The pipeline

```
/deep-interview  → turns a vague request into an approved spec (.omm/specs/<slug>.md)
/deep-dive       → front door when it's unclear whether the problem is a bug or a scope
                    question; runs trace first, then deep-interview if needed
/trace           → root-causes a concrete symptom via parallel competing hypotheses
/ralplan         → turns an approved spec into a consensus-reviewed PRD (.omm/state/prd.json)
/ralph           → runs the PRD to completion: implement each story, verify its
                    acceptance criteria, loop until done or blocked
/team            → N persona subagents in parallel worktrees on independent tasks
/cancel          → stop whatever pipeline stage is active and clean up .omm/ state
```

Every skill here is **explicit-invocation only**. Muse never fires a skill on
its own because a task looks complex or ambiguous — you (or another skill's
prose) has to name it. Each `SKILL.md` says this in its own body, because there
is no `triggers` field in Muse's skill frontmatter to enforce it. Sequencing
between stages works the same way: no `SKILL.md` chains automatically into the
next one — the skill's body tells you, in prose, what to run next
(`deep-interview` → "run `/ralplan`"; `ralplan` → "run `/ralph`").

## Dynamic workflows

Beyond the staged pipeline, the repo ships a QuickJS-sandboxed workflow