Add support for reasoning_content from DeepSeek R1

[charlieyou]

New DeepSeek reasoning model outputs reasoning_content in messages alongside the regular content. See their API docs here.

Requested feature is to be able to access this reasoning content from the Agent.run() response. According to their docs for multi-turn conversations, reasoning should not be passed back into the messages, so there will need to be consideration of that in the API.

Appreciate your work here, I'm loving the framework so far!

EDIT: OpenRouter supports this content now too, but in a diff format: https://openrouter.ai/announcements/reasoning-tokens-for-thinking-models

[samuelcolvin]

Thanks, should be possible, we'll need to think about how.

[achimstruve]

Can the DeepSeek R1 model use tools?
@samuelcolvin

[Wh1isper]

@achimstruve According to this, not yet

[Wh1isper]

I created a simple project to implement deepclaude-style-agent: https://github.com/Wh1isper/pydantic-ai-deepagent, making models such as claude can use deepseek r1's thinking as a reference for tool use. Check the example.

Hope this helps :D

[Kludex]

Let's use a single issue to track reasoning/thinking.