Agent chat model picker doesn't persist across new chats — always reverts to default model

[cameronapak]

Important

Issue created from Macro agent on behalf of Cam (myself). It could be slop, but the issue/problem itself is real. Ask any questions and I'm happy to help bring any clarity
-- Cam Pak

What's happening

The model picker in the AI chat composer doesn't persist a chosen model into new chats. Picking a non-default model (e.g. GPT 5.6) works fine for the current chat, but starting a new chat always resets the picker back to the default model (Sonnet 5), regardless of what was last selected.

This means the model has to be manually reselected every single new chat.

What I expected

Per the docs (docs.macro.com/product/agents, "Choosing a model" section):

A model picker sits in the chat input. You can choose a model per chat depending on how difficult you deem the question. Your choice is remembered for that chat and carried forward when you start a new one from the composer.

So the last-picked model should carry forward as the default for the next new chat started from the composer. Right now it doesn't — it's pinned to Sonnet 5 every time.

Repro steps

1. Open a chat, switch the model picker away from the default (Sonnet 5) to another model.
2. Send a message or two — confirm the chosen model is in effect for that chat.
3. Start a new chat from the composer.
4. Observe: the picker is back on Sonnet 5, not the model chosen in step 1.

Suggested fix (my read of it, could be off)

Sounds like the model choice is likely scoped only to in-memory/session state for the current chat rather than written to a persisted per-user preference (e.g. local storage or a user settings row) that the composer reads when initializing a new chat. If that's roughly right, the fix is probably: on model-picker change, persist the selection as the user's "last used model" default, and have new-chat composer initialization read from that persisted value instead of a hardcoded/default constant.

Happy to be corrected if the actual architecture is different — just flagging as a starting theory.

I'd be glad to take a crack at fixing this if someone can confirm the diagnosis above (or point me at the actual persistence path) — let me know if that's welcome and I'll pick it up. Not opening a PR yet, just flagging + offering.