I don't understand why messages are being saved on the client rather than on the server. It makes this useless for any semi-long running chats. If the user closes the client it never saves. Users should be able to:

1. Kick off a chat
2. Close their browser
3. Return later to read the response

That is currently not possible because saving of the assistant's replies is done on the client. Ditto for long-running replies with lots of tool calls. Nothing is saved until the entire completion is done.

I expect a robust implementation to save on the server and save after every step is finished, not on total completion.