| name | model-orchestrator |
| description | Run local multi-model agent workflows through the model-orchestrator CLI. Use when a task should combine stronger and cheaper models through Advisor, Orchestrator, or Hybrid delegation; when the user asks to delegate independent work across models; or when an interrupted orchestration run must be inspected, resumed, or cancelled. |
Model Orchestrator
Use the model-orchestrator CLI as the only execution interface. Do not reimplement orchestration in the host agent.
Workflow
- Check availability with
command -v model-orchestrator and model-orchestrator doctor.
- Choose the smallest fitting mode:
single: one model can complete the task.advisor: a normal executor should ask a stronger model only for consequential uncertainty.orchestrator: a stronger model should plan, delegate independent tasks, and synthesize results.hybrid: delegated workers may also escalate bounded advice requests.
- Default to
--permission read. Use --permission write only when the user requested mutations.
- Run from the target working directory and request JSON output:
model-orchestrator --output json run \
--mode orchestrator \
--permission read \
--cwd "$PWD" \
"<objective>"
- Read
run_id and result.output from the JSON response. Report the final output, not raw provider events.
- If execution stops, inspect before retrying:
model-orchestrator --output json status <run-id>
model-orchestrator --output json inspect <run-id>
model-orchestrator --output json resume <run-id>
Use model-orchestrator cancel <run-id> only for an active run.
Boundaries
- Never read, copy, or persist provider credentials or native session files.
- Never pass secrets in the objective or tracked config.
- Keep independent research tasks read-only and parallelizable.
- Keep shared-state mutations in the exclusive write lane.
- Do not start a second orchestration for work already owned by an active run.
- Use model aliases and profiles from the local config instead of hardcoding provider-specific behavior in the skill.
