Do not switch to Anthropic models mid-session

[rxgrant]

Anthropic's hyperactive refusals are paranoid about multi-model sessions. Using something like GLM and then switching to Fable for a particular task (where you really want the full context) triggers a refusal.

For now, refusals are common and fairly benign, but they have a lot of dystopian potential. As a result, JP should (by default) refuse to send "impure" sessions to Anthropic, unless the user's config explicitly overrides this.

[JeanMertz]

I can't say I completely agree with that conclusion. First, it would specialize Anthropic in a way that we haven't done at all, so far, so I want to thread lightly here, especially if/when model providers are removed from JP core and become plugins themselves (of which some will be installed automatically by the main jp binary).

That doesn't mean we can't add that exception to the Anthropic provider code itself, which would self-contain it, but I haven't really seen the issue you described, so I would first like to have some examples so that I can reproduce it, and figure out what the right strategy is to solve the situation, either, as you said, by hard-rejecting it, or, ideally, by solving this in-code so we tune the data sent to Anthropic in such a way that it no longer triggers any issues.

It should be noted that cross-provider work already exists to rewrite how reasoning tokens hare presented from one model to the other, so when we have a relatively stable reproducible case, it could just be a matter of hiding non-anthropic reasoning tokens to Anthropic's models, which is a single toggle in the existing code, and far less invasive than rejecting any model switching when using Anthropic models in a conversation.