fix: set memory step tool call even if code parsing fails

[dechantoine]

Hello huggingface team and thank you for this package, I'm making wonder with it 😄

Long story short, I'm currently working on integrating Google's Gemini API into smolagents using the google-genai package. To achieve this, I've developed a bridge that translates your ChatMessages into the format expected by the Gemini API.

However, I've encountered the following error when sending message lists to Gemini with my agent:

Error in generating model output:
400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Please ensure that the number of function response parts is equal to the number of function call parts of the function call turn.', 'status': 'INVALID_ARGUMENT'}}

This error appears to stem from instances where the List[ChatMessage] contains a tool-response without a preceding tool-call. Notably, this issue arises exclusively with CodeAgent. Specifically, when code parsing fails, the tool_calls attribute for that step isn't set, yet the tool response is still present.

Is this behavior intentional? If so, it might pose challenges in effectively leveraging Gemini with smolagents. 😢

Best regards,
Antoine