Return last text parts on empty message

[samuelcolvin]

We sometimes get an empty response with anthropic (and perhaps other models) when the model has already returned text alongside tool calls in this scenario, if text responses are allowed, we return text from the most recent model response, if any.

[github-actions]

Docs Preview

commit:	cde9fd8
Preview URL:	https://767168b7-pydantic-ai-previews.pydantic.workers.dev

[Kludex]

Maybe this PR solves #1292

[Kludex]

There's a similar PR I think? #1294

[dmontagu]

This is kind of an opinionated approach to how the graph should behave if you get an empty response, but I think it's a reasonable opinion. I feel like we might want to refactor this in the future but even if we do I suspect any change would be backwards compatible, and this is clearly an improvement over the current behavior, so I think let's go with it.

However, I'll note that I think it would be nice to provide a way for a tool to indicate that its result should not be used to continue the conversation. I have some ideas about this but they are mostly opinionated and probably we don't need to start with that complexity.

For this specific case, I think if you wanted to avoid having the conversation continue just because there was text and tools, you could add an output tool with a single text field (or even do something like OutputTool(str) from the result->output PR), and that would make it possible for the model to decide precisely when it's agentic run was finished (in particular, not needing to receive a request with the result of any non-output tools called alongside an output tool).

Either way, I'm in favor of merging this as is.