vllm - 💡(How to fix) Fix [Bug]: Not support role tool of image [1 pull requests]

Your current environment

Title

Kimi K2.5/K2.6 vision model rejects image_url content in role: "tool" messages, although image input works in user messages

Description

When serving Kimi K2.5/K2.6 with vLLM OpenAI-compatible Chat Completions API, I found that multimodal image_url content is accepted in role: "user" messages, but rejected when the same image_url content is returned from a tool message using role: "tool".

The error says that tool message content only supports text content.

I would like to clarify whether this is expected behavior, and if so, whether vLLM plans to support multimodal tool results for vision-capable models such as Kimi K2.5/K2.6.

Reproduction

A tool call flow where the assistant calls a tool, and the tool returns an image:

{
  "role": "tool",
  "tool_call_id": "functions.get_image:0",
  "content": [
    {
      "type": "image_url",
      "image_url": {
        "url": "https://example.com/image.png"
      }
    }
  ]
}

🐛 Describe the bug

{ "role": "tool", "tool_call_id": "functions.get_image:0", "content": [ { "type": "image_url", "image_url": { "url": "https://example.com/image.png" } } ] }

Before submitting a new issue...