[Bug]: OpenAI-compatible streaming sends finish_reason=stop on intermediate thought/tool chunks

[IlusiveAtlas]

Summary

OpenAI-compatible streaming responses (/antigravity/v1/chat/completions, and likely /v1/chat/completions) send "finish_reason": "stop" on every SSE chunk, including intermediate thought/reasoning chunks and tool-call chunks. OpenAI-compatible clients (e.g. DeepSeek Harness / DSH) treat the first non-null finish_reason as the end of the stream, so they stop immediately after the first reasoning_content or tool_calls chunk and never consume the actual content / continue the tool loop.

Environment

• gcli2api latest upstream: 2b5d78f / 69638c8 (2026-08-24)
• Route: POST /antigravity/v1/chat/completions
• Model: gemini-3.7-flash-high
• Auth: Bearer token
• Client: DeepSeek Harness (OpenAI-compatible streaming), also reproducible with curl

Steps to reproduce

1. Start gcli2api.
2. Send a streaming request that triggers thinking:

curl -N http://127.0.0.1:7861/antigravity/v1/chat/completions \
  -H 'Authorization: Bearer <password>' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gemini-3.7-flash-high",
    "messages": [{"role": "user", "content": "请一步步推理：17*23+45=?"}],
    "stream": true
  }'

3. Observe the first SSE chunk:

{"choices":[{"index":0,"delta":{"reasoning_content":"..."},"finish_reason":"stop"}]}

Expected OpenAI-compatible behavior: the first chunk should have "finish_reason": null; only the final chunk should have "finish_reason": "stop".

Actual behavior

Every streamed chunk contains "finish_reason": "stop" because:

• The upstream Antigravity/Gemini SSE stream only sets finishReason on the final chunk (verified via the native /antigravity/v1/models/{model}:streamGenerateContent endpoint).
• src/converter/openai2gemini.py -> convert_gemini_to_openai_stream() calls _map_finish_reason(gemini_finish_reason) even when gemini_finish_reason is None.
• _map_finish_reason() defaults any None/unknown value to "stop".

This causes OpenAI clients that honor finish_reason to stop after the first thought/tool chunk.

Impact

• Streaming chat responses that include reasoning_content appear to “think then stop”.
• Streaming tool-call responses appear to return only the first tool call and never continue the agent loop.
• This breaks agentic use-cases through DSH and likely other strict OpenAI-compatible clients.

Suggested fix

In convert_gemini_to_openai_stream(), only map an explicit upstream finishReason; use null when it is absent:

gemini_finish_reason = candidate.get("finishReason")
if gemini_finish_reason:
    finish_reason = _map_finish_reason(gemini_finish_reason)
else:
    finish_reason = None

I verified this locally: intermediate chunks now have "finish_reason": null, and only the final empty chunk has "finish_reason": "stop". The tool-call stream also correctly emits null on the tool-call chunk and stop on the terminal chunk.