Client-side model selection (first-class)

[jcourson-bg]

What problem are you trying to solve?

There's no way for a client to choose which model a turn uses.

The model is fixed at author time in agent.ts (defineAgent({ model })). Neither HandleMessageRequestBody nor SendTurnPayload carries a model field.

For a chat app (Claude/ChatGPT-style) the user needs a model picker.

Proposed solution

Let the server declare a selectable set and let the client pick from it. Server stays authoritative.

Alternatives considered

No response

[jcourson-bg]

Seems simple but some considerations from harnesses I've built..

Context-window cliffs - If the model selected has a smaller context window than the last model you need to ensure the harness gracefully handles compactions at the right step..

Mid-session model mixing - i think the ai-sdk abstraction is helpful for this but worth considering how one providers message shape converts to another. Also same with supported file type differences.

Prompt-cache invalidation - can't get around this but just worth understanding the tradeoff of allowing this feature mid convo..

Other thoughts: you might want something like default if not specified or "default/last used" option? This might just be a client side/ out of the scope of eve consideration of how to persist the model the user is intending to use but worth thinking about.