Allow structured output and text for all ModelResponses except the last one

[nonchris]

history of the issue

When a model is forced to provide a structured output, it will not longer output text messages as described in #2158 (comment).
And it would be helpful to allow non-reasoning models to still "reason" using text output.

idea

Discussing this issue on Slack thread, namely with @DouweM and @dmontagu a new idea arose.:
Allow the model to output structured output or text and only force it to use structured output if it's the last message.

elaboration

So the proposed agent mode would effectively do what the following pseudo-code implementation with current pydatic AI tooling would do:

def validator(_input):
    if isinstance(_input, str):
       rise ModelRetry("please use the structured output tool, not string")
    ....

agent = Agent(output_types=[str, MyOutputType], output_validator=validator)
res = agent.run()
....

But instead of nicely asking the LLM to use the other tool, the last tool call could be reissued with a tool-set that has only the structured output tool at it's disposal. This could be done internally by PydanticAI, subsequently forcing a structured output for the last tool call (if it didnt work without force)

considerations

Issues that I could imagine with this mode:

• what if the message of the AI was some kind of "i give up", which does not fit the format of a structured-"success" output. but on the other hand this issue would also occur if the only available tool to begin with was the "success"-tool...
• it could result in more messages than needed otherwise, raising costs and making the behavior of the agent less understandable for a user....

References

The slack thread this idea originated in: https://pydanticlogfire.slack.com/archives/C083V7PMHHA/p1756933362513579

[DouweM]

@nonchris Thanks for creating the issue.

In this mode, we'd not pass tool_choice=required like we currently do in the default tool output mode (provided that str is not among output types), so we'd allow the model to write just text or write text along with tool calls (including non-output function tool calls). When the model generates a response with only text, without a call to a function or output tool, we'd force the model to call an output tool (as we already do here using a prompt) on the next request by passing tool_choice=[output tools].

I'm not sure yet how this mode should be entered. We can only make this work with tool output mode, as the native output mode (response_format=json_schema) prevents the model from outputting non-structured text entirely. So perhaps it should be an argument on the ToolOutput marker class, even though it doesn't really feel like a tool-output-specific feature.

allow_text_output=True is the name used internally for this, but that could be confusing as "output" in agent context means the final output, not just a single model request's response output, and the final output still couldn't be text unless str is explicitly included.

Another downside of it being on ToolOutput is that output_type can take a list of them, so it would be unclear if it should be added to one or all (realistically, we'd respect it if it's listed on one).

(In retrospect, allowing the model to talk through its actions before calling tools or generating final output would probably have been good default behavior (although reasoning models effectively solve this through special thinking parts), with "always jump straight to a tool call" the opt-in alternative behavior, but it's too late to change that.)

output_type=ToolOutput(MyOutputType, allow_text_before_tool_calls=True) seems like it'd be the least-bad option.

What do you think?