Feature: Support for in-session model routing

[Gohlub]

Issue

The model router menu should be accessible throughout a session, not just the beginning.

Context

Currently, the user is unable to switch model providers within an existing session. This is great for session-based provider artifacts, but can be restrictive for multi-model workflows. A potential workaround is the subagent tool call, but that delegates the control to the model and requires an extra prompt.

Consideration:

Changing a provider is a destructive action, as all session-based artifacts will likely need to be dropped. Context is preserved through the message thread, but is nevertheless impacted by a switch.

[Gohlub]

@philz here's the issue referencing the now closed #288. I'd be happy to follow up if you share more info about the specific subtleties you were thinking about. IMO, a longer-term solution is to expose some sort of Ask/Plan/Agent workflow, and be able to configure these roles to different model providers.

[philz]

I want to make sure that the user sees what model is being used at a given point. That means we need to add model names to the messages that generated them, and then include the model name in the timeline when it changes, one way or another. When you switch between models that support images and ones that don't, the conversation will die. That needs to be dealt with. It's generally been possible to know how much a conversation used in tokens; what does that mean if the model changed. Should changing the model be in the UI or just a slash-command or both? Where? (I don't want this prominent; I think it's a bit of a distraction.) Is changing a model an explicit RPC or does the new model come with a "send chat" command. There's other configurability per conversation like the set of tools; should that be mutable too in the same vein? What happens when you fork a conversation seven messages up; what model does it get forked with?

None of these are intractable but they all have to combine coherently.