Send back thinking parts in a field of request message

[ZeroAurora]

Question

Pydantic AI sends back thinking parts inside model response content wrapped in <think> tags. Interestingly, there are some providers like Moonshot and MiniMax recommending to send back thinking parts directly in a field of request message (like messages[0].reasoning_content or messages[0].reasoning_details. Moonshot specially mentioned that "The model will decide which parts are necessary and forward them for further reasoning", which leaves space for discussion.

Additional Context

N/A

[DouweM]

@ZeroAurora This is because the OpenAI Chat Completions API format doesn't natively has fields for reasoning. We currently read reasoning from non-standard fields:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/openai.py

Lines 560 to 569 in e72452f

	# The `reasoning_content` field is only present in DeepSeek models.
	# https://api-docs.deepseek.com/guides/reasoning_model
	if reasoning_content := getattr(choice.message, 'reasoning_content', None):
	items.append(ThinkingPart(id='reasoning_content', content=reasoning_content, provider_name=self.system))
	# The `reasoning` field is only present in gpt-oss via Ollama and OpenRouter.
	# - https://cookbook.openai.com/articles/gpt-oss/handle-raw-cot#chat-completions-api
	# - https://openrouter.ai/docs/use-cases/reasoning-tokens#basic-usage-with-reasoning-tokens
	if reasoning := getattr(choice.message, 'reasoning', None):
	items.append(ThinkingPart(id='reasoning', content=reasoning, provider_name=self.system))

But we don't currently send thinking back in non-standard fields.

I'd be open to a contribution that does so conditional on a new field on OpenAIModelProfile.

Note also that in #3089 we're introducing an OpenRouterModel subclass of OpenAIChatModel that will implement proper OpenRouter reasoning handling through new hooks that subclasses can override. That approach will be useful for providers that implement a lot of non-standard functionality on top of Chat Completions, which seems to be the case for Moonshot as well as it has features like built-in Web Search: https://platform.moonshot.ai/docs/guide/use-web-search#web_search-execution. So once that PR is merged, a MoonshotModel class would be another option.

[ZeroAurora]

I'd be open to a contribution that does so conditional on a new field on OpenAIModelProfile.

I'm quite interested in implementing such a feature. Is there a general guide to follow?

[DouweM]

@ZeroAurora Basically:

1. Add a field to OpenAIModelProfile like openai_chat_send_back_reasoning_field: Literal['reasoning', 'reasoning_content']
2. Modify OpenAIChatModel._map_messages ThinkingPart handling to read OpenAIModelProfile.from_profile(self.profile).openai_chat_send_back_reasoning_field
3. Modify pydantic_ai.profiles.moonshotai.moonshotai_model_profile to return OpenAIModelProfile with that field set.
4. Then it should work automatically when using Agent('moonshotai:kimi-latest') or OpenAIChatModel with MoonshotAIProvider

[ZeroAurora]

will start working recently