Problem

When using streaming responses, some models or API endpoints may return usage fields (prompt_tokens, completion_tokens, total_tokens) as None or omit them entirely. The current implementation passes these values directly to the ResponseUsage Pydantic model, which expects integers. This causes a validation error:

3 validation errors for ResponseUsage
input_tokens
Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType]
output_tokens
Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType]
total_tokens
Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType]

Solution

This PR ensures that all token fields passed to ResponseUsage are always integers. If any of the fields are None or missing, they default to 0. This is achieved by using or 0 and explicit is not None checks for nested fields.

Key changes:

• All input_tokens, output_tokens, total_tokens fields use or 0 fallback.

Impact

• Fixes Pydantic validation errors for streaming responses with missing/None usage fields.
• Improves compatibility with OpenAI and third-party models.
• No breaking changes; only adds robustness.

fixes #1179