Model routing

Routing lets you set up to three tiers, each with up to four model candidates. Only commands change it. The agent cannot edit it.

Since 24 Oct 2026 · Updated 29 Oct 2026

Picking a model for each message

Tiers

small for quick replies, medium for everyday, power for hard tasks.

Regex rules and length thresholds on the message. Triggers for power: long_horizon, max_intelligence_request, low_confidence, complex_architecture.

/routing, /model, /escalate (alias /fable).

Keys are never stored in the routing file. The /setup key serves the /setup provider. Other providers use their .env key.

The power model receives the chat and writes a three-part answer. Handoffs are saved under data/escalations.

Quick, cheap replies for short or trivial messages.

Fallbacks: Up to 3 fallbacks
How a message is sent here: Short messages under 50 characters or 8 words with no tool history. Optional: small-auto setting.
Commands that change it: /routing set small, /model small
When it is not set: Falls back to medium

The everyday model. Default for most messages.

A stronger, slower model for hard tasks.

Fallbacks: Up to 3 fallbacks
How a message is sent here: Triggers: long_horizon, max_intelligence_request, low_confidence, complex_architecture. Or /escalate.
Commands that change it: /routing set power, /escalate, /model power
When it is not set: Empty means nothing to escalate to