[BUG] Qwen Bad Request 400

[singharyan44]

Problem (one or two sentences)

The Qwen3-coder-plus model, which I use through the Qwen Code provider, gives an "The provider couldn't process the request as made. Stop the task and try a different approach" Error(bad request 400), happens with the flash model too.

Context (who is affected and when)

I haven't really tested other models outside of Qwen coder, from my knownledge, It also affects Qwen Coder users.

Reproduction steps

Roo Code Extension version: v3.51.1
In any task, new or already existing and using Qwen Coder model through Qwen Code Provider, say anything, It should give a bad request 400 error.

Expected result

The Model replies back

Actual result

Roo Code shows "Date/time: 2026-04-06T17:03:20.657Z Extension version: 3.51.1 Provider: qwen-code Model: qwen3-coder-plus 400 400 bad request"

Variations tried (optional)

Tried downgrading to 3.50.x, no difference in result

App Version

v3.51.1

API Provider (optional)

Not Applicable / Other

Model Used (optional)

qwen3-coder-plus

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

// Please share this file with Roo Code Support (support@roocode.com) to diagnose the issue faster
// Just make sure you're OK sharing the contents of the conversation below.

{
  "error": {
    "timestamp": "2026-04-06T17:05:43.845Z",
    "version": "3.51.1",
    "provider": "qwen-code",
    "model": "qwen3-coder-plus",
    "details": "400\n400 bad request"
  },
  "history": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "<user_message>\nHello\n</user_message>"
        },
        {
          "type": "text",
          "text": "<environment_details>\n# VSCode Visible Files\n\n\n# VSCode Open Tabs\napp/page.tsx,../../../.qwen/oauth_creds.json,.ai-bridge.md\n\n# Current Time\nCurrent time in ISO 8601 UTC format: 2026-04-06T17:05:34.932Z\nUser time zone: Asia/Calcutta, UTC+5:30\n\n# Current Cost\n$0.00\n\n# Current Mode\n<slug>code</slug>\n<name>💻 Code</name>\n<model>qwen3-coder-plus</model>\n\n\n# Current Workspace Directory (c:/Users/ARYAN_SINGH/Documents/GitHub/v0-clinic-appointment) Files\n.gitignore\n.roomodes\n.roorules\ncomponents.json\nnetlify-deployment-guide.md\nnetlify.toml\nnext-env.d.ts\nnext.config.mjs\nnotion-setup.md\npackage-lock.json\npackage.json\npnpm-lock.yaml\npostcss.config.mjs\nREADME.md\ntsconfig.json\n.qwen/\n.roo/\napp/\napp/globals.css\napp/layout.tsx\napp/main-page.tsx\napp/page.tsx\napp/appointment/\napp/appointment/page.tsx\ncomponents/\ncomponents/appointment-scheduler.tsx\ncomponents/glow-button.tsx\ncomponents/msg91-otp-verification.tsx\ncomponents/otp-verification.tsx\ncomponents/patient-intake-form.tsx\ncomponents/theme-provider.tsx\ncomponents/theme-toggle.tsx\ncomponents/whatsapp-otp-verification.tsx\ncomponents/ui/\ncomponents/ui/accordion.tsx\ncomponents/ui/alert-dialog.tsx\ncomponents/ui/alert.tsx\ncomponents/ui/aspect-ratio.tsx\ncomponents/ui/avatar.tsx\ncomponents/ui/badge.tsx\ncomponents/ui/breadcrumb.tsx\ncomponents/ui/button-group.tsx\ncomponents/ui/button.tsx\ncomponents/ui/calendar.tsx\ncomponents/ui/card.tsx\ncomponents/ui/carousel.tsx\ncomponents/ui/chart.tsx\ncomponents/ui/checkbox.tsx\ncomponents/ui/collapsible.tsx\ncomponents/ui/command.tsx\ncomponents/ui/context-menu.tsx\ncomponents/ui/dialog.tsx\ncomponents/ui/drawer.tsx\ncomponents/ui/dropdown-menu.tsx\ncomponents/ui/empty.tsx\ncomponents/ui/field.tsx\ncomponents/ui/form.tsx\ncomponents/ui/hover-card.tsx\ncomponents/ui/input-group.tsx\ncomponents/ui/input-otp.tsx\ncomponents/ui/input.tsx\ncomponents/ui/item.tsx\ncomponents/ui/kbd.tsx\ncomponents/ui/label.tsx\ncomponents/ui/liquid-metal-button.tsx\ncomponents/ui/menubar.tsx\ncomponents/ui/navigation-menu.tsx\ncomponents/ui/pagination.tsx\ncomponents/ui/popover.tsx\ncomponents/ui/progress.tsx\ncomponents/ui/radio-group.tsx\ncomponents/ui/resizable.tsx\ncomponents/ui/scroll-area.tsx\ncomponents/ui/select.tsx\ncomponents/ui/separator.tsx\ncomponents/ui/sheet.tsx\ncomponents/ui/sidebar.tsx\ncomponents/ui/skeleton.tsx\ncomponents/ui/slider.tsx\ncomponents/ui/sonner.tsx\ncomponents/ui/spinner.tsx\ncomponents/ui/switch.tsx\ncomponents/ui/table.tsx\ncomponents/ui/tabs.tsx\ncomponents/ui/textarea.tsx\ncomponents/ui/toast.tsx\ncomponents/ui/toaster.tsx\ncomponents/ui/toggle-group.tsx\ncomponents/ui/toggle.tsx\ncomponents/ui/tooltip.tsx\ncomponents/ui/use-mobile.tsx\ncomponents/ui/use-toast.ts\nhooks/\nhooks/use-mobile.ts\nhooks/use-toast.ts\nlib/\nlib/utils.ts\nmcp-memory/\nmcp-memory/memory.json\nmcp-memory/package-lock.json\nmcp-memory/package.json\nmcp-memory/README.md\nmcp-memory/server.js\nmcp-web-search/\nmcp-web-search/package-lock.json\nmcp-web-search/package.json\nmcp-web-search/README.md\nmcp-web-search/server.js\nmy-app/\nmy-app/.gitignore\nmy-app/eslint.config.mjs\nmy-app/next.config.ts\nmy-app/package-lock.json\nmy-app/package.json\nmy-app/postcss.config.mjs\nmy-app/README.md\nmy-app/tsconfig.json\nmy-app/app/\nmy-app/app/favicon.ico\nmy-app/app/globals.css\nmy-app/app/layout.tsx\nmy-app/app/page.tsx\nmy-app/public/\nmy-app/public/file.svg\nmy-app/public/globe.svg\nmy-app/public/next.svg\nmy-app/public/vercel.svg\nmy-app/public/window.svg\nplans/\nplans/otp-verification-plan.md\npublic/\npublic/apple-icon.png\npublic/icon-dark-32x32.png\npublic/icon-light-32x32.png\npublic/icon.svg\npublic/placeholder-logo.png\npublic/placeholder-logo.svg\npublic/placeholder-user.jpg\npublic/placeholder.jpg\npublic/placeholder.svg\nstyles/\nstyles/globals.css\nYou have not created a todo list yet. Create one with `update_todo_list` if your task is complicated or involves multiple steps.\n</environment_details>"
        }
      ],
      "ts": 1775495135111
    }
  ]
}

[roomote-v0]

I've analyzed the Qwen Code provider implementation and identified several likely causes for the 400 Bad Request error. The provider at src/api/providers/qwen-code.ts sends several OpenAI-specific parameters that DashScope's compatible-mode API may not support (or may have recently stopped tolerating):

Root cause analysis:

The createMessage method constructs a request with these potentially unsupported parameters for DashScope:

1. max_completion_tokens (line 228) - DashScope uses max_tokens, not the newer OpenAI max_completion_tokens parameter
2. parallel_tool_calls: true (line 231) - DashScope may not support this OpenAI-specific parameter
3. stream_options: { include_usage: true } (line 227) - DashScope may not support this
4. strict: true on tool definitions (inherited from convertToolsForOpenAI in base-provider) - DashScope does not implement OpenAI's strict mode for tools

Any of these would cause DashScope to return a 400 error. The fact that it fails even on a simple "Hello" message points to a request-level parameter issue rather than a content issue.

Proposed implementation plan:

1. Replace max_completion_tokens with max_tokens in the request options for DashScope compatibility
2. Remove parallel_tool_calls from the request since DashScope does not support it
3. Remove stream_options or wrap it in a try/catch since DashScope may reject it
4. Override convertToolsForOpenAI in QwenCodeHandler to skip setting strict: true on tool definitions, since DashScope does not support OpenAI strict mode
5. Add better error handling to surface the actual error message from DashScope (rather than just "400 bad request")
6. Add/update tests for the Qwen Code handler to verify these changes

This approach aligns with how other OpenAI-compatible providers (like LM Studio) handle API differences - by adapting the request format to match what the target API actually supports.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[singharyan44]

@roomote How do I implement this? Or do I have to wait for a new fix/update?

[roomote-v0]

Hey @roomote-agent, I've seen @singharyan44's comment and will work on implementing the fix from the analysis above. The plan is to update the Qwen Code provider to use DashScope-compatible parameters instead of OpenAI-specific ones. I'll open a draft PR shortly to share progress.

[roomote-v0]

Opened a draft PR with my attempt at addressing this: #12068

The fix replaces OpenAI-specific parameters that DashScope does not support (max_completion_tokens, parallel_tool_calls, stream_options, and strict: true on tools) with DashScope-compatible equivalents. All tests pass. Feedback welcome!

[JeanLuX]

Same issue here :(

[romandesign]

Same here. I wanted to try using Roo with VS code and looking for the best free AI Agent, was recommended Qwen, and I can't connect with the same error.