[Feature]: If I want gpt-oss to be able to call custom tools, how should I set the --tool-call-parser parameter during deployment?

[LZAndy]

🚀 The feature, motivation and pitch

Currently, gpt-oss supports built-in tools. But does it also support Agent tool calls? If so, do I still need to configure the --tool-call-parser parameter?

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[tdoudech]

Hello,

There's already a reasonning parser used : https://github.com/vllm-project/vllm/blob/a47e6ffe9366516ea5ca28e27fc87367a869e854/vllm/reasoning/gptoss_reasoning_parser.py

Imo You can init it with --reasoning-parser GptOss

[LZAndy]

Hello,

There's already a reasonning parser used : https://github.com/vllm-project/vllm/blob/a47e6ffe9366516ea5ca28e27fc87367a869e854/vllm/reasoning/gptoss_reasoning_parser.py

Imo You can init it with --reasoning-parser GptOss

Thanks, But what I'm looking for is tool_parsers rather than reason_parsers. I'm not sure if setting up reason_parsers will be helpful for me, but I'll give it a try.

[tdoudech]

Oh mb sorry, I Answred the wrong issue.

Yes it seems that there's not a tool parser for Harmony tools for the moment; If you want to dig more, gpt oss was trained on this response format https://github.com/openai/harmony.

[gaby]

@tdoudech Harmony support was added in #22342 we need a new vLLM release

[Yokesh-RS]

So to make use of this , should we add --tool-call-parser harmony ?

[bbrowning]

You need to use --tool-call-parser openai which was added in #22386 and should be in the v0.10.2 released version.

[bbrowning]

@yeqcharlotte I think we could close this one, since we now document the correct tool call parser at https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html and I answered in this issue as well. Or, is there another way we want to educate users on the correct tool call parser options for gpt-oss before closing this?

[athewsey]

I'm actually not seeing great results with --enable-auto-tool-choice --tool-call-parser openai on vLLM v0.10.2 to be honest, but I'm using the 20b model and haven't had much chance yet to dig in to whether VLLM could handle it better or it's just that the model is bad / poorly-aligned with the prompting framework.

I'm finding that the tool call often doesn't get picked up and just gets output as reasoning/text. For example (using Strands Agents framework and their built-in calculator tool):

[
  {
    "role": "user",
    "content": [
      {
        "text": "What's the square root of 5329, plus 1? Use the calculator to be sure"
      }
    ]
  },
  {
    "role": "assistant",
    "content": [
      {
        "reasoningContent": {
          "reasoningText": {
            "text": "User wants sqrt(5329)+1. Use calculator.{\n  \"expression\": \"sqrt(5329)+1\",\n  \"mode\": \"evaluate\",\n  \"precision\": 10\n}"
          }
        }
      }
    ]
  }
]

Was initially worrying that the reasoning parser might be interfering with it, but tried with & without --reasoning-parser openai_gptoss and it didn't seem to have much impact.

I'm seeing notably worse success rate extracting a couple of MCP-sourced tools we have with 3 consecutive underscores in the name also, like aws___read_documentation.

[bbrowning]

If you can get the entire Chat Competions request out of Strands Agent that it sends, it would be easier to reproduce and fix any issues in vLLM here. I did a bit of digging, and will assume you're using the Strands OpenAIModel class. If so, that defaults to "stream": True, which does streaming Harmony tool calls and actually that code path today doesn't even go through this openai tool call parser. We've had other reports of streaming tool calls in gpt-oss not working well as non-streaming, so I'll take a look at this code today and see what I can find. If you happen to come across a simple reproducer of the type of requests Strands is sending that isn't working well, I can incorporate that directly into my own testing.

I was going to suggest you try Strands with streaming set to false, but looking at their code I don't think it handles the non-streaming case.

[bbrowning]

@athewsey I think you're hitting the bug fixed in #24768 , which specifically impacts streaming tool call response handling of Chat Completions requests to gpt-oss models. I spent some time digging into this today in another bug report, and what was happening in that case (and likely in your case) is that the model was outputting tool calls to the wrong Harmony channel because we are not correctly activating the commentary channel in the system prompt we generate.