MiniMax M3 Streaming Reasoning

I’m using MiniMaxAI/MiniMax-M3-MXFP8 behind vLLM’s OpenAI-compatible chat completions API. Our incident-response service exposes a runbook search function:

{
  "model": "MiniMaxAI/MiniMax-M3-MXFP8",
  "messages": [
    {
      "role": "user",
      "content": "Checkout API is returning elevated 502s. Search the current incident runbooks for mitigation steps before answering."
    }
  ],
  "stream": true,
  "tool_choice": "auto",
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "search_incident_runbooks",
        "parameters": {
          "type": "object",
          "properties": {
            "service": {"type": "string"},
            "symptom": {"type": "string"}
          },
          "required": ["service", "symptom"]
        }
      }
    }
  ]
}

With the default adaptive thinking mode, reasoning and its markers are emitted as normal content before the structured function call:

data: {"choices":[{"delta":{"reasoning":null,"content":"<mm:think>"}}]}
data: {"choices":[{"delta":{"reasoning":null,"content":"I should search the current checkout incident runbook first."}}]}
data: {"choices":[{"delta":{"reasoning":null,"content":"</mm:think>"}}]}
data: {"choices":[{"delta":{"reasoning":null,"content":null,"tool_calls":[{"type":"function","function":{"name":"search_incident_runbooks","arguments":"{\"service\":\"checkout-api\",\"symptom\":\"elevated 502s\"}"}}]}}]}

I also tried the same request with:

"chat_template_kwargs": {"thinking_mode": "enabled"}

That stream starts in the reasoning field, but it never transitions to the structured function call:

data: {"choices":[{"delta":{"reasoning":"I should search the current checkout incident runbook first."}}]}
data: {"choices":[{"delta":{"reasoning":"</mm:think>"}}]}
data: {"choices":[{"delta":{"reasoning":"]<]minimax[>[<tool_call>\n]<]minimax[>[<invoke name=\"search_incident_runbooks\">]<]minimax[>[<service>checkout-api]<]minimax[>[</service>]<]minimax[>[<symptom>elevated 502s]<]minimax[>[</symptom>]<]minimax[>[</invoke>\n]<]minimax[>[</tool_call>"},"finish_reason":"stop"}]}

For both modes, sending the same request with "stream":false returns clean reasoning, hides the <mm:think> markers, and returns the structured search_incident_runbooks call with the expected arguments.

Fix both streaming failures so adaptive and enabled thinking separate reasoning, visible content, and structured function calls in the same way as the working non-streaming requests. Preserve existing disabled-thinking, atomic-marker, and plain-text streaming behavior. Text that only resembles an incomplete marker, such as ordinary content ending in < or <mm:, must not disappear.