Describe the bug
The _transform_output method uses simple conditional to build AssistantMessage. If any tool calls are made, the text content of the message, if there is one, is dropped.

At the same time, the AssistantMessage is constructed using either content or reasoning_content field of Delta, in this order of precedence. This means that if both are defined, the reasoning content is ignored.

AssistantMessage(update.content or update.reasoning_content or "", id=chunk.id)

However, both fields can be populated at the same time. [1] In fact, it is possible to have tool_calls, content and reasoning_content set on the same Delta object.

But the initialization of AssistantMessage introduces a third problem. If the Delta object from litellm is initialized without any reasoning content, the attribute reasoning_content is deleted.

        if reasoning_content is not None:
            self.reasoning_content = reasoning_content
        else:
            # ensure default response matches OpenAI spec
            del self.reasoning_content

This is a problem introduced by litellm. But it will possibly manifest here, if the content field doesn't evaluate to true.
In that case the the attempt to access reasoning_content will fail with AttributeError.

In principle, that situation is rather unlikely, as it would require the inference provider not to set reasoning, or regular content.
But it can still happen. I have seen it happen with some Gemini models running in GCP.

To Reproduce

1. Create agent with LiteLLMChatModel
2. Start the agent and compare raw inference output to responses handled by agent.

Expected behavior
All content received from inference should be processed into AssistantMessage objects.

Set-up:

• Bee version: 0.1.80
• Model provider Google Gemini

Additional context
Litellm source:

[1] https://github.com/BerriAI/litellm/blob/aa587bd9d3c1f21decd590a68fd07427f1fa3653/litellm/types/utils.py#L1263