- Shipped
- September 28, 2026 at 3:10 AM UTC
- Author
- Kamo
- Commit
- 39ef6b5
Hugging Face, Google and Cohere text streams carry no usage event, so the assembled response counts 0 input and 0 output tokens. E2 built its stream_options.include_usage chunk from that response and told the caller prompt_tokens 0 / completion_tokens 0, while the executor metered and billed its estimate (found in production by SP04 Task 38: a Hugging Face stream answered usage 0/0 and its ai_usage_records row read 6/5 estimated). The chunk now carries the call's UsageMeasure, the provider's counts when it reported them, else the metered estimate, which is what E10's done event already says.