Environment

• qwen-code version: 0.15.10
• Node.js version: v26.1.0
• OS: Linux (Arch/CachyOS)
• Auth type: openai (Standard API Key)
• Endpoint: https://dashscope-intl.aliyuncs.com/compatible-mode/v1 (Singapore/International)

Bug Description

Every prompt fails with [API Error: Connection error. (cause: fetch failed)] when using the DashScope International endpoint (dashscope-intl.aliyuncs.com), even though:

• curl to the same endpoint works (HTTP 200)
• Native Node.js fetch() works
• undici in isolation works
• The API key and base URL are correctly resolved inside the CLI (confirmed via debug patch)

Steps to Reproduce

1. Set DASHSCOPE_BASE_URL=https://dashscope-intl.aliyuncs.com/compatible-mode/v1
2. Configure ~/.qwen/settings.json with selectedType: openai and valid apiKey
3. Run: echo "hi" | qwen --model qwen-plus
4. Result: [API Error: Connection error. (cause: fetch failed)]

Root Cause

DashScopeContentGenerator.buildClient() passes a custom undici Agent dispatcher to the OpenAI SDK via runtimeOptions:

const runtimeOptions = buildRuntimeFetchOptions("openai", this.cliConfig.getProxy());
// runtimeOptions = { fetchOptions: { dispatcher: new undici.Agent({ headersTimeout: 0, bodyTimeout: 0, keepAliveTimeout: 60000 }) } }

return new OpenAI({
  apiKey,
  baseURL: baseUrl,
  ...runtimeOptions || {}  // <-- this dispatcher causes fetch to fail
});

This undici Agent fails silently when connecting to dashscope-intl.aliyuncs.com (Singapore region), while the same request succeeds with native fetch().

Confirmed via debug patch:

DEBUG baseURL: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
runtimeOptions: {"fetchOptions":{"dispatcher":{"_events":{},"_eventsCount":0}}}
→ Connection error. (cause: fetch failed)

The issue is specific to the combination of:

• undici Agent with headersTimeout: 0 / bodyTimeout: 0
• OpenAI SDK using that dispatcher
• dashscope-intl.aliyuncs.com endpoint

Workaround

Remove the undici dispatcher from DashScopeContentGenerator.buildClient() in cli.js:

// Before (broken):
const runtimeOptions = buildRuntimeFetchOptions("openai", this.cliConfig.getProxy());
return new OpenAI({ apiKey, baseURL: baseUrl, timeout, maxRetries, defaultHeaders, ...runtimeOptions || {} });

// After (works):
return new OpenAI({ apiKey, baseURL: baseUrl, timeout, maxRetries, defaultHeaders });

This forces the OpenAI SDK to use Node.js native fetch instead of the undici dispatcher, and the connection succeeds.

Additional Notes

• The same issue affects versions 0.13.0, 0.14.5, and 0.15.10 (all tested)
• The China endpoint (dashscope.aliyuncs.com) may not be affected (not tested)
• QWEN_CODE_DISABLE_PRECONNECT=1 helps avoid a secondary error in the preconnect step but does not fix the main issue
• The undici Agent works fine in isolation with the same endpoint — the failure seems to occur specifically when passed to the OpenAI SDK bundled in the CLI