Routing
Each request goes to one backend model:
| Request | Model |
|---|---|
| Haiku-tier model name | --small-model |
| Carries an image (anywhere in the history) | --vision-model |
| Anything else | -m/--model |
A conversation keeps routing to --vision-model while an image stays in its history. Set it
when the main model is text-only.
--vision-mode caption instead has --vision-model describe each image as text and keeps
the main model. Use it for vision models without tool calling (e.g. pixtral-12b-2409).
Captions are cached per image.