Skip to content
CommandChain CommandChain CommandChain Docs

Router policy

Settings → AI Configuration → Router records a routing intent: a default objective, a cost ceiling, a latency target, and named routes that map a condition to a model. It needs the manage_router_policy permission.

That is an uncomfortable thing for a settings page to say, and stating it is better than letting a founder cap spend with a control that is not wired to anything.

What decides a model today

Section titled “What decides a model today”

Three things, in this order:

  1. The agent’s runtime chain, on its Runtime tab, decides which provider and which runner. The workspace default is on AI runtime.
  2. The allowlist on the model catalog decides which of that provider’s models may be used. A provider with nothing allowed is skipped.
  3. The app default and its per-area overrides, also on the model catalog, decide the model for work that does not run through an agent.

Cost control comes from budgets, which stop a run rather than steer it.

What the page holds

Section titled “What the page holds”

Recorded so you know what a saved policy contains if you inherit one:

  • Routing objective: balanced, minimize cost, maximize quality, or minimize latency.
  • Max cost per call, in dollars.
  • P95 latency SLO, in milliseconds.
  • Routes: a name, a model id, an optional fallback model id, a priority, and a JSON condition such as {"taskType": "code_generation"}.

What to do instead

Section titled “What to do instead”

If your aim is cheaper work, untick expensive models in the catalog and lower the per-task-type budget for the types that do not need them. Both take effect on the next run.

If your aim is a specific model for a specific agent, set that agent’s runtime chain.

Next steps

Section titled “Next steps”