fix(model_runner): request token usage on streaming completions

[CorrectRoadH]

Problem

Streamed runs against OpenAI-compatible providers record zero tokens and zero cost in the tape.

The run event's usage ends up empty, so any consumer that reads usage from the tape (cost reporting, the OTel tapestore's gen_ai.usage.* attributes, etc.) sees 0.

Root cause

ModelRunner.completion_response calls acompletion(..., stream=True) but never sets stream_options.include_usage. Per the OpenAI streaming spec, a streaming chat-completion omits the usage block unless stream_options.include_usage is explicitly requested. As a result Tape._extract_usage(chunk) never finds usage on any chunk and state.usage stays unset.

This is reproducible directly against an OpenAI-compatible endpoint:

stream=True, no stream_options          -> no usage in any chunk
stream=True, include_usage=True         -> final chunk carries {prompt_tokens, completion_tokens, total_tokens}

Fix

Pass stream_options={"include_usage": True} when streaming, gated on the provider base class so only OpenAI-compatible providers receive it:

if stream and isinstance(llm, BaseOpenAIProvider):
    return {"include_usage": True}

• stream_options.include_usage is an OpenAI-completions field; BaseOpenAIProvider subclasses (openai, openrouter, …) accept it, while anthropic/gemini reject it — the gate leaves those paths untouched.
• Non-streaming completions already carry usage in the response body and are unaffected.
• The streaming consumer already extracts usage from chunks (Tape._extract_usage); it simply never had usage to read until now.

Tests

Adds unit tests covering the gate:

• streaming OpenAI / OpenAI-compatible provider → {"include_usage": True}
• non-streaming → None
• non-OpenAI provider (anthropic) → None

[yihong0618]

wow nice and simple can you also add a test?

[CorrectRoadH]

@yihong0618
Earlier, when AI helped me fix this issue, it also wrote a test, but I felt that version wasn’t great. It felt a bit like the test was just asserting its own setup, so I removed it.

After seeing your comment, I asked to look at the style of other tests in this repo and try to build a more e2e-style regression test. The idea was to verify not only that stream_options is passed, but also that streaming usage eventually gets recorded into the tape. That said, this version now feels a bit heavy, and I’m not sure whether this is actually the best practice for this repo.

[CorrectRoadH]

Or did you mean the successful screenshot proof that the bug exists and works after the fix, rather than an actual test file? Or anything🤔

I found this Bub issue in a eval case while building an agent eval framework. On the Bub version with this PR’s fix, it already works successfully.

[yihong0618]

Or did you mean the successful screenshot proof that the bug exists and works after the fix, rather than an actual test file? Or anything🤔

I found this Bub issue in a eval case while building an agent eval framework. On the Bub version with this PR’s fix, it already works successfully.

for my side both ok just simple test also fine~
can you help fix hte linit?

[CorrectRoadH]

I see the error of CI. I will fix it