Surface model reasoning as delta.reasoning_content in streaming & sync responses

[robsltd]

Summary

• Rework _StreamFilter to capture thinking/reasoning content instead of silently dropping it, emitting it as delta.reasoning_content in streaming SSE chunks and message.reasoning_content in sync responses
• Follows the DeepSeek / OpenAI convention for surfacing chain-of-thought in OpenAI-compatible APIs
• <think> tag split across gRPC chunk boundaries
• Stream ending mid-think (before </think>)
• Multiple <think> blocks mid-response
• Leading \n after <think> tag stripped from reasoning output
• Non-reasoner models never emit reasoning_content
• Existing clients that don't read reasoning_content are unaffected (additive field)

Problem

When using a reasoning model through the proxy, the model's chain-of-thought is completely stripped from responses. Clients that want to display or log reasoning have no way to access it.

To reproduce:

1. Start the proxy with the reasoning model loaded
2. Send a streaming request:

   curl -N http://127.0.0.1:8080/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{"model":"auto","stream":true,"messages":[{"role":"user","content":"What is 2+2?"}]}'

3. Observe: all <think>...</think> content is silently dropped. Only delta.content chunks appear — no reasoning is surfaced anywhere in the streaming response.
4. For non-streaming, reasoning is only available via the non-standard debug.reasoning field and only when --debug is enabled.

Solution

Instead of discarding thinking content in _StreamFilter, capture it and emit it as delta.reasoning_content (streaming) / message.reasoning_content (sync), matching the convention used by DeepSeek and OpenAI for reasoning models.

After this change, the same curl now produces:

data: {"choices":[{"delta":{"reasoning_content":"Okay, the user is asking..."}}]}
data: {"choices":[{"delta":{"reasoning_content":"...simple arithmetic..."}}]}
data: {"choices":[{"delta":{"content":"4"}}]}
data: {"choices":[{"delta":{},"finish_reason":"stop"}]}

Non-streaming responses include both fields on the message:

{
  "message": {
    "role": "assistant",
    "content": "4",
    "reasoning_content": "Okay, the user is asking..."
  }
}

_StreamFilter changes: feed() and finalize() now return (visible, reasoning) tuples instead of a single string. Phase 1 (initial CoT block) and Phase 2 (mid-stream <think> blocks) both capture reasoning instead of discarding it.

Testing

Tested against Qwen3-30B-A3B:

• Streaming: reasoning_content chunks flow, clean transition to content, no tag leaks
• Non-streaming: both fields present and clean
• finish_reason + data: [DONE] emitted correctly