Multiple tool calls not supported in streamed response

[Finndersen]

It appears that currently only one tool call is supported for streaming responses. Is this a limitation of streamed responses vs standard complete responses?
Does this potentially mean that some tool calls will be ignored/discarded?

I'm not sure if this PR #468 addresses this issue or not

[samuelcolvin]

We need to support Structured Outputs, I think that's the only way to solve this, see #582.

[Finndersen]

I'm not too familiar with the streamed response behaviour, but why is structured outputs relevant to this?

[samuelcolvin]

because structured outputs make sure you only get one response, where as it's impossible to prevent the LLM from returning multiple tool calls, while only one can returned as the final result.

[Finndersen]

I think the issue I was referring to is now resolved by the streaming refactor. I was referring to handling of multiple tool calls within the run, not as part of the final response