Show output functions span in logfire

[tekumara]

Description

When an output function is called, show this as a "running tool" span in logfire like standard tools.

References

No response

[DouweM]

@tekumara Can you please give a bit more context on why you'd like this? They don't take any time to execute and their output is already rendered in the "agent run" pane. It could also be confusing to people to see "running tool" if they didn't actually register any.

[tekumara]

I think my mental model is: output functions are just a tool (except they end the agent loop and return the final result rather than going back to the LLM). For example, we have an output function that calls another agent and we don't want to mess with that agent's output and rather return it directly to the user.

And so I was expecting a span, with arguments and return values, like there is for a tool. Would a "running output function" span make sense?

But I think you're suggesting output functions are being modelled separately from tools. ie: it's more a property of the agent than it is a tool. And it's true we can get the details from the agent run span (we just have to handle this differently in our metrics, evals and reporting).

[DouweM]

@tekumara Thanks, that's helpful. A new "running output function" span makes sense to me. Output functions can also be used with output modes other than "tool output" (https://ai.pydantic.dev/output/#output-modes), so a new span would be more consistent than depending on an output tool span.

Would you be up for implementing this in a PR?