[mywonderland808]

Summary

• Fill prompt_tokens / total_tokens on approach-synthesized and passthrough chat completions (and batch / legacy /v1/completions), instead of omitting them or leaving prompt_tokens: 0.
• Emit a final SSE usage chunk with choices: [] when stream_options.include_usage is set, matching OpenAI / existing dict-stream behavior.
• Prefer real completion_tokens when known; estimate only when missing or zero with non-empty content. Includes focused unit and HTTP wiring tests.

Test plan

• python -m unittest tests.test_usage_reporting -v
• python -m unittest tests.test_completions_endpoint tests.test_proxy_plugin -v
• Manual: non-stream approach response includes usage.prompt_tokens and usage.total_tokens
• Manual: stream with stream_options.include_usage: true ends with a choices: [] usage chunk before [DONE]
• Manual: passthrough/none response missing upstream prompt_tokens gets filled without clobbering complete upstream usage

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[mywonderland808]

Hi! Before signing the CLA, I’d like to clarify one point regarding AI-assisted contributions.

This contribution was developed using Cursor (AI Agent) based on my own planning, architecture decisions, requirements, and instructions. The agent generated most of the implementation, which I subsequently reviewed and tested.

Cursor's terms grant me rights to the generated output.

Does this type of AI-assisted, agent-generated contribution comply with your Contributor License Agreement, particularly the representation that I am the sole author of the Contributions?

I'd appreciate your clarification before I sign the CLA.

Thanks!