OpenAI-compatible /v1/chat/completions streaming with stream_options.include_usage returns non-standard SSE (empty choices chunks, empty id/model, trailing {"choices":[],"cost":"0"})

Summary

The OpenAI-compatible gateway (opencode.ai/zen/go, /v1/chat/completions) emits non-standard SSE chunks when the client sends stream_options: {"include_usage": true}. This breaks OpenAI-compatible clients that rebuild the streamed response (e.g. LiteLLM crashes with litellm.APIError: Error building chunks for logging/streaming usage calculation).

Without stream_options, the same endpoint streams clean OpenAI-format chunks.

Repro

curl -sN https://opencode.ai/zen/go/v1/chat/completions \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6-luna","messages":[{"role":"user","content":"say hi"}],"stream":true,"max_tokens":20,"stream_options":{"include_usage":true}}'

Observed SSE (with include_usage)

data: {"id":"gen-...","object":"chat.completion.chunk","created":...,"model":"gpt-5.6-luna","choices":[]}
data: {"id":"gen-...","object":"chat.completion.chunk","created":...,"model":"gpt-5.6-luna","choices":[]}
data: {"id":"","object":"chat.completion.chunk","created":...,"model":"","choices":[]}
data: {"id":"","object":"chat.completion.chunk","created":...,"model":"","choices":[]}
data: {"id":"","object":"chat.completion.chunk","created":...,"model":"","choices":[{"index":0,"delta":{"content":"Hi!"},"finish_reason":null}]}
data: {"id":"","object":"chat.completion.chunk","created":...,"model":"","choices":[]}
data: {"id":"","object":"chat.completion.chunk","created":...,"model":"","choices":[]}
data: {"id":"","object":"chat.completion.chunk","created":...,"model":"","choices":[]}
data: {"id":"gen-...","object":"chat.completion.chunk","created":...,"model":"gpt-5.6-luna","choices":[],"usage":{"prompt_tokens":8,"completion_tokens":6,"total_tokens":14}}
data: {"choices":[],"cost":"0"}

Issues vs OpenAI spec

1. Empty choices: [] on the first chunks — OpenAI streams always start with a chunk containing choices[0].delta.role; clients use the first chunk to seed the response object.
2. id and model empty strings on mid-stream chunks — chunks should carry stable id/model.
3. Trailing data: {"choices":[],"cost":"0"} — not a valid OpenAI chunk (no id/object/created/model); the cost field is non-standard.
4. Content chunk's delta has no role (fine in the middle, but the first real chunk should carry it).

Impact

Any strict OpenAI-compatible client that accumulates chunks and rebuilds the full response (logging/usage calculation) crashes. LiteLLM (v1.84.0, and still in v1.97.0) raises Error building chunks for logging/streaming usage calculation because it assumes the first chunk with choices has a non-empty array with a delta.role.

Suggestion: only emit chunks with empty choices as the final usage chunk (like OpenAI does), keep id/model populated on every chunk, and drop the cost trailer (or put it inside the usage chunk).

Start by running the provided curl reproduction against the OpenAI-compatible /v1/chat/completions endpoint with stream_options.include_usage enabled, then locate that endpoint's streaming implementation in the repository. Compare its SSE output with the stated OpenAI requirements and the output without stream_options. Done means valid initial and intermediate chunks with stable id/model fields, followed only by a standard usage chunk without the non-standard cost trailer.