Parse '<think>' tags in streamed text as thinking parts

[DouweM]

Fixes #2244

[github-actions]

Docs Preview

commit:	f72ad2f
Preview URL:	https://da9c15e8-pydantic-ai-previews.pydantic.workers.dev

[DouweM]

@Kludex Can you please review this?

[AlbertoLanaro]

I think that “START_THINK_TAG” and "END_THINK_TAG" should be customizable at the model level. In fact, sometimes some models (e.g Claude 4), even though prompted to use CoT with <think> tags, use different tag content, such as <thinking>, and these are not parsed correctly.

[DouweM]

@AlbertoLanaro Make sense, can you please file a new feature request issue for that? Note that with Claude, we currently expect thoughts to come in as a specific field in the API response, so I'd be curious to see when it comes as tags in the text response instead.

[AlbertoLanaro]

can you please file a new feature request issue for that?

Sure, no problem.

Note that with Claude, we currently expect thoughts to come in as a specific field in the API response, so I'd be curious to see when it comes as tags in the text response instead.

I am using Claude via OpenRouter, so I am using OpenAIModel and OpenAIProvider to access it.