OpenAI backend: mot._thinking never populated for vLLM thinking models (reasoning vs reasoning_content field name)

[planetf1]

Observed behaviour

When using OpenAIBackend with a vLLM-served thinking model (Qwen3 with --reasoning-parser qwen3), the thinking trace is silently discarded. mot._thinking remains None even when the model has produced reasoning content.

Root cause

In mellea/backends/openai.py, processing() probes for the thinking trace as:

if hasattr(message, "reasoning_content"):
    thinking_chunk = message.reasoning_content

However, vLLM 0.20.2 does not expose this as a Python attribute on the openai SDK message object. The thinking trace is present in the raw response dict under the key "reasoning", only visible via model_dump():

resp.choices[0].model_dump()
# {
#   "message": {
#     "content": null,
#     "reasoning": "2 + 2 equals 4.",   ← here, not 'reasoning_content'
#     ...
#   }
# }

hasattr(message, "reasoning_content") returns False, so mot._thinking is never set. The thinking trace is present in mot._meta["oai_chat_response"] but never surfaced.

Expected behaviour

processing() should also check the raw response dict for a "reasoning" key (vLLM's field name) when reasoning_content is absent, and populate mot._thinking accordingly.

This is independent of the silent-empty-string issue (#1060) — even when a model produces both text and reasoning content, the thinking trace is currently lost for vLLM-served models.

Environment

• Backend: OpenAIBackend (OpenAI-compatible)
• Model: Qwen/Qwen3-Coder-Next-FP8
• Inference server: vLLM 0.20.2, flags: --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser hermes --tensor-parallel-size 2 --quantization fp8 --max-model-len 262144
• Hardware: 2× GPU (tensor parallel), IBM LSF cluster