AskUserQuestion is not wrapping a question in an array

[jcordes73]

What happened?

With Qwen 0.12.1 and the following snippet on a Skill markdown file

Ask:

```
Please provide the path to your integration project. This can be:
- A project directory
- A single configuration file (e.g. a Blueprint XML or Spring XML file)
- A ZIP archive

Path:
```

the following error is shown:

AskUserQuestion {"question":"Please provide the path to your integration project. This can be:\n- A project directory\n- A single configuration file (e.g. a Blueprint XML or Spring XML file)\n- A ZIP… 
 
Parameter "questions" must be an array. 

What did you expect to happen?

The user is prompted for a path

Client information

Qwen Code 0.12.1 (150521d) │
│ Runtime Node.js v22.22.0 / npm 10.9.4 │
│ OS linux x64 (6.18.7-200.fc43.x86_64) │
│ │
│ Auth API Key - openai │
│ Base URL https://litellm-prod.apps.maas.redhatworkshops.io/v1 │
│ Model qwen3-14b │
│ Session ID b2507d1e-9219-4fee-b704-51a4b4832d14 │
│ Sandbox no sandbox │
│ Proxy no proxy │
│ Memory Usage 268.0 M

Login information

No response

Anything else we need to know?

No response

[DragonnZhang]

Thank you for the detailed report and for including the repro, environment information, and error message. That makes it much easier to understand what is happening.

At the moment, AskUserQuestion expects the questions parameter to be an array, so a payload like {"question": "..."} will fail validation with Parameter "questions" must be an array.

From the client information, I also notice you're using qwen3-14b. My suspicion is that this model may not be adhering to the function/tool-calling schema very reliably, and smaller models generally tend to be less consistent for this kind of structured output. It may be worth trying the same prompt a few more times, or switching to a larger model with stronger function-call support.

Thanks again for taking the time to report this. The details you shared were very helpful.

[jcordes73]

Thanks for looking into it, I found Qwen Code really useful. What would be a recommendation for a Qwen3 model that is more reliable? Something like Qwen 3 Coder Next (which has 80b). We from Red Hat provide a 4-bit quantized model (https://huggingface.co/RedHatAI/Qwen3-Coder-Next-NVFP4) but I need to see where I can evaluate it. My colleagues didn't had these kind of issues with Claude, but I explictely want to have something "on-premise".

At least once it seems this particular tool call was working (maybe out of 20 times I tried).

[DragonnZhang]

Thanks for the follow-up, and glad to hear you're finding Qwen Code useful!

If you're looking for a more reliable coding-oriented model, Qwen3-Coder-Next is a very strong option and one of the best Qwen models specifically optimized for programming workloads. You might also want to try the Qwen-3.5 series, which has shown strong performance across reasoning, coding, and agent-style tasks.

You can find more details about these models here:

Qwen 3.5: https://qwen.ai/blog?id=qwen3.5

Qwen3-Coder-Next: https://qwen.ai/blog?id=qwen3-coder-next

If you're running on-premise, evaluating Qwen3-Coder-Next (e.g., the larger variants) or a Qwen-3.5 model may give you more consistent tool-calling behavior than smaller models like qwen3-14b.

Appreciate you taking the time to test and report this!