Support Thinking part

[Kludex]

• Closes How to include thoughts from the Gemini model in Pydantic output? #793
• Closes Reasoning response support #907
• Closes Disable thinking mode in Gemini 2.5 flash #1555

Checklist

• Streaming
• Documentation
• Add id in the ThinkingPart

[github-actions]

Docs Preview

commit:	19c275e
Preview URL:	https://66df8787-pydantic-ai-previews.pydantic.workers.dev

[Kludex]

I would like to support OpenAI responses API before working on this PR.

[sam-cobo]

Any progress on this or any release plan?

[Kludex]

I'll release this in some days.

[Wh1isper]

Very nice PR, I was wondering if we could split it into two parts and release the ThinkingPart first so developers can implement it for their own models first.

For example, deepseek-r1 uses reasoning_content to show the full content. this doesn't seem to be in the current primary support.

https://api-docs.deepseek.com/api/create-chat-completion

[Kludex]

Why GitHub keeps breaking this PR?

[jerry-reevo]

I know this is work in progress, but I tested the current PR branch using streaming with Claude 3.7 Extended thinking and it generally works, but I did run into this one strange error.

Hope this is helpful. Excited for this PR to get merged!

This occurred after successfully streaming a full phase of reasoning tokens, followed by a tool call. This occurred on the first part following the tool call.

...site-packages/pydantic_ai/models/anthropic.py", line 346, in _map_message
    assert response_part.signature is not None, 'Thinking part must have a signature'
           │             └ None
           └ ThinkingPart(content=None, signature=None, part_kind='thinking')

AssertionError: Thinking part must have a signature

[mwildehahn]

I was testing this and noticed that when making requests, it will sometimes include this message to the responses API (i think from a tool call?):

{'id': '', 'summary': [{'text': '', 'type': 'summary_text'}], 'type': 'reasoning'}

which throws an error:

openai.BadRequestError: Error code: 400 - {'error': {'message': "Invalid 'input[7].id': empty string. Expected a string with minimum length 1, but got an empty string instead.", 'type': 'invalid_request_error', 'param': 'input[7].id', 'code': 'empty_string'}}

[mwildehahn]

it seems like here:

                if thinking_parts:
                    openai_messages.append(
                        responses.ResponseReasoningItemParam(
                            id=thinking_parts[0].signature or '',
                            summary=[Summary(text=item.content, type='summary_text') for item in thinking_parts],
                            type='reasoning',
                        )
                    )

we should not include if it has an empty id?

[mwildehahn]

Some of this was fixed with: openai/openai-python#2311 but now seeing:

• UserWarning: Handling of this event type is not yet implemented. Please report on our GitHub: ResponseReasoningSummaryTextDeltaEvent(delta='...', item_id='rs_680b0e42dd748191baf3259afcab72b40792f36bcb0feba3', output_index=0, summary_index=2, type='response.reasoning_summary_text.delta')
• UserWarning: Handling of this event type is not yet implemented. Please report on our GitHub: ResponseReasoningSummaryPartDoneEvent(item_id='rs_680b0e42dd748191baf3259afcab72b40792f36bcb0feba3', output_index=0, part=Part(text='...', type='summary_text'), summary_index=2, type='response.reasoning_summary_part.done')

[mwildehahn]

a couple things i found:

• the response tool call from openai is ResponseFunctionToolCall(arguments='', call_id='call_IstYYmboPGwj3IZpQBm1KOUr', name='...', type='function_call', id='fc_680b21b9319c8191b215aa9e4a73538004237e9105833371', status='in_progress')

specifically, they have a:

• call_id
• id

Right now both ToolCallPart and ToolReturnPart just have a single tool_call_id where we need to pass back both id and call_id.

I added an id property to ToolCallPart and ToolReturnPart so that I could capture both id and call_id and it works.

I'm not sure if that should just be called vendor_id or what -- I'll open a PR against this branch to show you the changes.

[mwildehahn]

Changes I had to get this to work with openai: #1587

[Kludex]

I think signature should only be available for Anthropic. I think it's confusing to convert id to signature.

[hyperlint-ai]

PR Change Summary

Introduced support for the 'Thinking' feature, enhancing the model's reasoning capabilities.

• Added a new section on 'Thinking' to the documentation
• Closed multiple related issues regarding the 'Thinking' feature

Added Files

• docs/thinking.md

---

How can I customize these reviews?

Check out the Hyperlint AI Reviewer docs for more information on how to customize the review.

If you just want to ignore it on this PR, you can add the hyperlint-ignore label to the PR. Future changes won't trigger a Hyperlint review.

Note specifically for link checks, we only check the first 30 links in a file and we cache the results for several hours (for instance, if you just added a page, you might experience this). Our recommendation is to add hyperlint-ignore to the PR to ignore the link check for this PR.

[aristideubertas]

Does this mean that models won't be able to see their previous thinking if we try to pass previous messages in?

[Kludex]

I think we need to use the <think> tags here.

[aristideubertas]

@Kludex is this PR ready to be merged or is more work to be done?