MUSE continuity commands

Read only the workflow needed for the current action. The installed shared runtime is .agent/skills/muse-commands, its workflows are .agent/workflows, and project configuration is .muse/config.json. Paths in this guide are relative to the installed project; use the absolute resolved script path when the shell cwd differs.

User interaction

The user may say continue, /resume strategy Lane A, /save, or /bye. The agent derives the goal, acceptance checks, constraints and next action from the request, shared role record and relevant checkpoint. Do not require the user to fill a goal/acceptance/constraints form. Do not change role based on cwd, provider or project. A project may have Strategy implement work directly; respect its policy. Do not introduce a Strategy-to-Build handoff unless actually requested.

Keep the same Lane by default. Use a new logical Lane only when independent scope or continuation warrants it within user authorization; changing provider or reaching a context threshold does not create one. A new app task is separate and requires user intent. Do not silently reopen a CLOSED Lane; first read its explicit reopen conditions. A missing checkpoint does not prove an old task is new.

Dispatch

Use bash .agent/skills/muse-commands/scripts/muse-doctor.sh as the helper prefix. It never calls a model or executes product tasks by itself.

1. On /resume, run parse-resume '<exact command>' '<workspace>' and lock the returned role_home/role/Lane. Run resume-workflow --command '<exact command>' --workspace '<workspace>', then read only the returned daily or legacy workflow.
2. On ordinary save, read .agent/workflows/save.md; use save-daily, not full Bye.
3. On explicit Bye, select bye-workflow --command '<locked resume>' --workspace '<workspace>' and follow that workflow. A valid unchanged closeout receipt can be checked; changed requirements/evidence or failing gates must be addressed.
4. Other commands use resolve /<command> '<workspace>' and the returned workflow.

State and identity

memory/lanes/<role>-lane-<name>.md under the configured role home is canonical. Schema1 is retained. The provider/session is a writer identity, not the role. Use helpers to update state; do not hand-rewrite checkpoints or bypass markers. Before writing, verify the expected hash, native writer and current workspace. New writers review and explicitly claim/adopt; old writers are rejected. Do not reuse another session ID, fabricate a native ID, or clear environment guards.

Read references/DAILY_V1.md for save/claim fields and references/ONBOARDING_V1.md for reviewed initialization and legacy adoption. Read references/SOURCES_V1.md for native source import/review and references/CLOSEOUT_V1.md before formal closeout.

Context and evidence

Save meaningful deltas and at least every ten interaction rounds if there are unsaved changes. At observed 80% pressure, save, compact if available, verify the current Lane and continue. If occupancy is unavailable, say UNKNOWN; cumulative tokens and model specifications do not establish current occupancy. After compaction, load the checkpoint and necessary evidence rather than full Boot.

Source coverage is not guaranteed. Missing native messages remain unknown; a partial source note does not prove complete transcript coverage. A hook exit0, hash, successful save or successful compilation is not product QA. Verify the actual user flows relevant to the changed scope before claiming completion.

Use the installed doctor for configuration/files preflight, not as a QA verdict.

Before a large save or after TOO_LARGE, follow checkpoint capacity and evidence storage. Keep detailed evidence outside the checkpoint and preserve the guarded save limits.

For interpreter or optional image-decoder failures, follow runtime dependency preflight. Check the actual selected interpreter, then rerun the affected workflow. Ordinary MUSE continuity does not require Pillow; preflight PASS is not workflow acceptance.