muse-orchestration

Files (18)

• SKILL.md
• ACCOUNT_RELAY.md · references/
• CONFIG.md · references/
• ORCHESTRATOR.md · references/
• ORCHESTRATOR_CRON.md · references/
• TOOLS.md · references/
• VAULT.md · references/
• licence_stop_guard.py · scripts/
• memory_audit.py · scripts/
• negative_claim_guard.py · scripts/
• orchestration_config.py · scripts/
• recall.py · scripts/
• vault_coverage_check.py · scripts/
• vault_freshness.py · scripts/
• wall_recover.py · scripts/
• HEAVY_LOCK.example.md · templates/
• RELAY_RUNBOOK.example.md · templates/
• orchestration.example.json · templates/

SKILL.md (first lines)

# MUSE multi-lane orchestration

Optional layer on top of `muse-commands` continuity. A single-lane project needs none of
it: `/save`, `/resume` and `/bye` stay the same. Read only the reference for the current
action.

## Configuration

`.muse/orchestration.json` in the project (template `templates/orchestration.example.json`,
keys in `references/CONFIG.md`). It names the orchestrator and the user's own lanes: id,
title, checkpoint file and required reads. MUSE ships no fixed lane names and no
author-machine paths; tools without a configured path report it as unconfigured instead
of guessing. `python3 -I scripts/orchestration_config.py` prints what was resolved.

## Commands

| Command | Workflow | Use |
|---|---|---|
| `/handoff save` · `/handoff <lane id>` · `/handoff status` | `.agent/workflows/handoff.md` | Account relay: save before the switch, continue under the new account |
| `/away` (`/托`) | `.agent/workflows/away.md` | User leaves; decide what can be decided, queue the rest |
| `/back` (`/回`) | `.agent/workflows/back.md` | User returns; deliver the return list, ask once |
| `/boot` | same as `/resume` | Alias for restoring a role or lane |

## Roles

- **Orchestrator** (总管): the user's single entry point; board, dispatch, follow-up,
  reports, account relay; does not write product code by default.
  `references/ORCHESTRATOR.md` (rules), `references/ORCHESTRATOR_CRON.md` (timer prompts).
- **Lanes**: whatever the project defines (for example build, verification, design).
  Each owns its checkpoint, reports to the orchestrator and messages other lanes directly.

## Tools (`scripts/`, Python 3 standard library, read-only unless stated)

| Script | Purpose |
|---|---|
| `recall.py <word> <synonym> …` | Search every memory directory and checkpoint before saying "not in my records" |
| `memory_audit.py [--sync]` | Audit memory indexes, cross-directory visibility, decision backlog, vault freshness, key patterns; `--sync` writes cross-directory blocks |
| `negative_claim_guard.py` | Stop hook: blocks a negative memory claim made without running `recall.py` |
| `licence_stop_guard.py` | Optional Stop hook encoding a policy: a licence is one line of fact, not by itself a reason to skip evaluation |
| `vault_coverage_check.py` | Every owner quote and tracked id in a checkpoint range exists in the vault |
| `vault_freshness.py` | The vault was written recently |
| `wall_recover.py` | After an unsaved wall, list the old session's unfinished subagents |

Usage, exit codes and hook installation: `references/TOOLS.md`. Vault formats:
`references/VAULT.md`. Relay mechanics: `references/ACCOUNT_RELAY.md`.

## Boundaries

Roles, approvals and receipts never widen authority. Deployments, production data
changes, payments, external messages and deletions still need the user's explicit
consent in the current session. Never open or print `.env*` or credential files. Write
times from `date`, not estimates. A tool PASS is not product QA.

From myths-labs/muse, MIT.

More from myths-labs/muse