qwen -d -p hi --output-format json

Output:

{
  "type": "assistant",
  "message": {
    "content": [
      {
        "type": "text",
        "text": "[API Error: Connection error. (cause: fetch failed)]"
      }
    ]
  }
}

Debug log includes:

[ERROR] [OPENAI_ERROR] OpenAI API Error: Connection error.
Error: Connection error.
    at OpenAI.makeRequest (.../qwen-code/cli.js:157505:17)
    at async ContentGenerationPipeline.executeWithErrorHandling (...)

## Expected Behavior

The command should successfully call the configured OpenAI-compatible DashScope endpoint and return
a response.

## What I Verified

The configured API key, endpoint, model, request body, and streaming response are valid.

A direct Node fetch() call to the same endpoint works:

fetch("https://coding.dashscope.aliyuncs.com/v1/chat/completions", {
  method: "POST",
  headers: {
    "content-type": "application/json",
    authorization: "Bearer <redacted>"
  },
  body: JSON.stringify({
    model: "qwen3.6-plus",
    messages: [{ role: "user", content: "hi" }],
    max_tokens: 16,
    stream: true
  })
})

This returns HTTP 200 with text/event-stream.

I also enabled --openai-logging, captured the full request body generated by Qwen Code, and replayed
that exact request body manually with Node fetch(). That also returned HTTP 200 and streamed chunks
correctly.

So the issue does not appear to be caused by:

- invalid API key
- invalid model name
- invalid base URL
- invalid request body
- streaming/SSE support
- DashScope rejecting the request

## Suspected Root Cause

The failure appears related to Qwen Code's runtime fetch options on Node.js 26.

In the bundled code, runtimeFetchOptions.ts creates an Undici dispatcher and passes it to the OpenAI
SDK:

function buildFetchOptionsWithDispatcher(sdkType, proxyUrl) {
  try {
    const dispatcher = getOrCreateSharedDispatcher(proxyUrl);
    return { fetchOptions: { dispatcher } };
  } catch {
    return sdkType === "openai" ? void 0 : {};
  }
}

Where getOrCreateSharedDispatcher() creates either undici.Agent or undici.ProxyAgent.

When I run Qwen Code normally, it fails. When I monkey-patch fetch() to remove init.dispatcher, the
same command succeeds.

Temporary preload used for verification:

const originalFetch = globalThis.fetch;

globalThis.fetch = function qwenFetchWithoutDispatcher(input, init) {
  if (init && typeof init === "object" && "dispatcher" in init) {
    const nextInit = { ...init };
    delete nextInit.dispatcher;
    return originalFetch(input, nextInit);
  }

  return originalFetch(input, init);
};

Run command:

NODE_OPTIONS=--import=/tmp/qwen-strip-dispatcher.mjs qwen -d -p hi --output-format json

This succeeds and returns:

Hi! What can I help you with?

## Workaround

Remove the injected fetchOptions.dispatcher via preload, or run qwen-code with a Node.js version
where this dispatcher path is compatible.

## Request

Could you check whether qwen-code should avoid passing a bundled Undici dispatcher to the OpenAI SDK
under Node.js 26, or gate this behavior behind proxy configuration only?