v2 SDK drops `reasoning_content` from model run responses (Message + StreamChunk)

[aix-ahmet]

Summary

The v2 SDK silently drops the reasoning_content field that the platform returns for reasoning LLMs, in both the non-streaming and streaming model-run paths. Consumers (e.g. the agents engine's AiXplainLLM provider) can never see the model's reasoning text — and for some models the entire response is in reasoning_content, so the run appears to have produced empty output.

Evidence

Direct model run on dev (POST {MODELS_RUN_URL}/{model_id} with {"data": "<essay prompt>", "max_tokens": 80}), model Qwen 3.6 35B A3B (69e5f93df018f8050346c1de, alibaba-cloud/qwen-3.6-35b-a3b/deepinfra):

{
  "details": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "",
        "reasoning_content": "Here's a thinking sequence\n\n1.  **Deconstruct the prompt:** ...",
        "name": null,
        "tool_calls": []
      },
      "finish_reason": "length",
      "logprobs": null
    }
  ],
  "status": "SUCCESS",
  "completed": true,
  "data": "",
  "usage": { "prompt_tokens": "26", "completion_tokens": "80", "total_tokens": 106 }
}

All 80 completion tokens went to reasoning_content; content/data are empty. After SDK deserialization, reasoning_content is gone: the Message dataclass in aixplain/v2/model.py only declares role, content, tool_calls, refusal, annotations, and dataclass_json discards unknown keys.

The streaming parser has the same gap: ModelResponseStreamer.__next__ (OpenAI-style branch in aixplain/v2/model.py) reads only delta.content and delta.tool_calls; a delta.reasoning_content key is dropped.

Fix

1. Message dataclass (aixplain/v2/model.py): add

   reasoning_content: Optional[str] = None

   so details[0].message.reasoning_content survives deserialization.

2. StreamChunk (aixplain/v2/model.py): add a reasoning_content: Optional[str] = None field, and in the OpenAI-style chunk branch of ModelResponseStreamer.__next__ populate it from delta.get("reasoning_content") (string-typed only, same defensive coercion as content). A chunk that carries only reasoning_content (empty content) must still be yielded, not skipped.

3. Tests: unit tests deserializing the payload above (non-streaming) and an SSE fixture with {"choices":[{"delta":{"reasoning_content":"..."}}]} chunks (streaming), asserting the field is preserved and that content-only behavior is unchanged.

Notes

• This mirrors the LiteLLM convention (message.reasoning_content / delta.reasoning_content), which is also what the platform emits, so no renaming — keep the field name as-is.
• Backwards compatible: field is optional with None default.
• Downstream consumer change: aixplain/aixplain-agents#222 (the agents engine reads the field once this ships).