Bug type

Regression (worked before, now fails)

Summary

image tool fails with "Unknown model: dashscope-vision/qwen-vl-max" even though configuration is correct and API works with direct Python calls

Steps to reproduce

1. Configure dashscope-vision provider in openclaw.json with correct baseUrl and apiKey
2. Verify configuration with openclaw models status (shows model is registered)
3. Test API directly with Python/curl (works correctly)
4. Run image photo.jpg --prompt "describe this image"
5. Error: "Unknown model: dashscope-vision/qwen-vl-max"

Expected behavior

The image tool should successfully call the dashscope-vision/qwen-vl-max model and return image recognition results, just like direct API calls do.

Actual behavior

Always fails with error: "Unknown model: dashscope-vision/qwen-vl-max"

Even though:

• Configuration is verified correct (openclaw models status passes)
• API key is valid (curl test succeeds)
• Direct Python calls to阿里云 API work perfectly
• Text model (dashscope/qwen3.5-plus) works fine

OpenClaw version

2026.2.22-2

Operating system

macOS 13.6 (Darwin 22.6.0)

Install method

npm global install

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

Temporary workaround: Use Python script to call 阿里云 API directly. Script location: tools/image_recognize.py in workspace. This works perfectly, proving the API and configuration are correct. The issue is specifically with OpenClaw's image tool implementation for dashscope-vision provider.