vllm - 💡(How to fix) Fix [Bug]: Chat completions emits empty tool_calls arrays after tool results [3 pull requests]

Your current environment

Observed on both:

• vLLM 0.20.2 pinned checkout: bc150f5
• vLLM main/latest checkout tested locally: 7bd738988

🐛 Describe the bug

The OpenAI-compatible chat completions API can serialize a normal assistant response with an empty tool_calls array after the client returns a tool result.

The response is semantically a final assistant message:

• finish_reason is "stop"
• message.content / stream delta.content contains normal text
• there is no tool call to execute

But the JSON payload still contains "tool_calls": []. In the OpenAI Python SDK this becomes a non-None message.tool_calls value, so common client loops treat the response as another tool-call response and then fail when indexing the empty list.

Example client failure:

while assistant_output.tool_calls is not None:
    tool_call = assistant_output.tool_calls[0]

Actual vLLM payload shape after the tool result:

{
  "choices": [
    {
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "...final answer...",
        "tool_calls": []
      }
    }
  ]
}

Expected payload shape:

{
  "choices": [
    {
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "...final answer..."
      }
    }
  ]
}

The same issue appears in streaming chunks as delta.tool_calls: [] alongside normal text deltas.

This is not a DeepSeek V4 tool-call parser failure and not a model accuracy issue. The model returns the expected final natural-language answer after the tool result; the API response serializer emits a misleading empty tool-call field.

Before submitting a new issue...

• I have searched the existing issues and did not find a duplicate.