Expose the instruction generated by PromptedOutput on `ModelRequest`

[xcpky]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

The instruction generated by PromptedOutput is not saved in message_history. Since other instructions are saved, I don't think this is the expected behavior.

I checked the code:

        if instructions := self._get_instructions(messages, model_request_parameters):
            openai_messages.insert(0, chat.ChatCompletionSystemMessageParam(content=instructions, role='system'))

it's inserted into openai_messages ad hoc but not saved in list[ModelMessage].

Example Code

class Step(BaseModel):
    explanation: str
    output: str

class MathReasoning(BaseModel):
    steps: list[Step]
    final_answer: str

  result = await agent.run("how can I solve 8x + 7 = -23", output_type=MathReasoning)
  print(result.all_messages_json())

Python, Pydantic AI & LLM client version

Python 3.13.9
pydantic_ai 1.20.0

[DouweM]

@xcpky This was an intentional change because the output mode that's going to be used (and whether it requires instructions to the model) is a property of the agent (run) rather than the message history.

The old behavior where it was included would cause incorrect behavior if a message history from a run that used prompted output mode was transferred to a new run that uses tool output, native output, or a different output type: the old instructions from the first run would unintentionally affect the output of the new run.

Is this causing issues for you?

[xcpky]

Is this causing issues for you?

I assumed the message history contains all the complete history, so that I can for example retrieve the history to train my model.
When I use PromptedOutput, the only thing I know without looking into the code is that it will be merged into the instruction and sent to the model. Then I sometimes(for example when training) want to know "what exactly are sent to the model".

[DouweM]

@xcpky To know exactly what's sent to the model, you need the combination of message history, model settings, and model request parameters (which hold things like the tool definitions and output schema, including output mode and instructions):

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

Lines 377 to 384 in eae558b

	async def request(
	self,
	messages: list[ModelMessage],
	model_settings: ModelSettings | None,
	model_request_parameters: ModelRequestParameters,
	) -> ModelResponse:
	"""Make a request to the model."""
	raise NotImplementedError()

There's a PR to record that information on the ModelRequest: #3405

Would that solve your problem?

[xcpky]

There's a PR to record that information on the ModelRequest: #3405

Would that solve your problem?

That seems good. But how ModelRequestParameters are used is an internal implementation detail, so we still need a way to reconstruct the real request sent from that ModelRequestParameter as a method.

[DouweM]

@xcpky That's true, but of course how exactly list[ModelRequest | ModelResponse] is mapped to an API request is also a model class implementation detail.

If you want to see the real request, could you use https://ai.pydantic.dev/logfire/? Otherwise the messages + setting + request prams combination should allow you to issue the request again via Model.request.