openclaw - 💡(How to fix) Fix Control UI on Windows resolves media://inbound paths incorrectly and fails image tool [1 pull requests]

Summary

When using OpenClaw's own Control UI on Windows, image upload appears to work initially, but the image tool later fails because the media://inbound/... URI is resolved to an invalid Windows path.

This is separate from the OpenWebUI /v1/chat/completions issue reported in #84918. In this case the image does reach OpenClaw and is saved as a media://inbound/... attachment, but the later local file lookup points to the wrong directory and malformed path.

Environment

• OS: Windows 11
• OpenClaw version: 2026.5.18
• Node runtime in logs: 24.12.0
• Client: OpenClaw Control UI / Gateway web panel
• Gateway URL used locally: ws://127.0.0.1:18789
• Gateway auth mode: token
• Default text model: nvidia/openai/gpt-oss-120b
• Image model: nvidia/nemotron-nano-12b-v2-vl

OpenClaw model status

Relevant output:

Default       : nvidia/openai/gpt-oss-120b
Fallbacks (0) : -
Image model   : nvidia/nemotron-nano-12b-v2-vl
Image fallbacks (0): -
Configured models (5): nvidia/nemotron-3-super-120b-a12b, nvidia/openai/gpt-oss-120b, opencode/glm-5.1, opencode/deepseek-v4-flash, nvidia/nemotron-nano-12b-v2-vl

Steps to reproduce

1. Start OpenClaw Gateway on Windows.

2. Open the local Control UI in the browser:

   http://127.0.0.1:18789/

3. Approve the device pairing request.

4. Start a chat in the Control UI.

5. Upload an image.

6. Ask OpenClaw to describe the image in Spanish.

Example prompt:

Describe esta imagen en español, detallando los objetos, personas, escenarios y cualquier texto visible.

Observed behavior

The Gateway log first shows that the image was offloaded/saved as media:

[gateway] Offloaded image for text-only model. Saved: media://inbound/WhatsApp_Image_2026-05-18_at_09.39.45---6b129134-d7f7-42ea-b617-8145be9a3aa6.jpg

Then image understanding fails:

[media-understanding] image: failed (0/1) reason=Model does not support images

Then the image tool tries to resolve the local file path incorrectly:

[tools] image failed: Local media file not found: C:\Users\tomas\.openclaw\workspace\media:\inbound\WhatsApp_Image_2026-05-18_at_09.39.45---6b129134-d7f7-42ea-b617-8145be9a3aa6.jpg raw_params={"prompt":"Describe esta imagen en español, detallando los objetos, personas, escenarios y cualquier texto visible.","image":"media://inbound/WhatsApp_Image_2026-05-18_at_09.39.45---6b129134-d7f7-42ea-b617-8145be9a3aa6.jpg"}

The assistant then replies that it could not access the image file.

Actual file location

The file does exist, but it is saved under .openclaw\media\inbound, not under .openclaw\workspace\media:\inbound.

Actual path found on disk:

C:\Users\tomas\.openclaw\media\inbound\WhatsApp_Image_2026-05-18_at_09.39.45---6b129134-d7f7-42ea-b617-8145be9a3aa6.jpg

The path OpenClaw attempted to use was:

C:\Users\tomas\.openclaw\workspace\media:\inbound\WhatsApp_Image_2026-05-18_at_09.39.45---6b129134-d7f7-42ea-b617-8145be9a3aa6.jpg

This looks like media://inbound/... is being transformed into a malformed Windows path containing media:\inbound and rooted under the workspace directory.

Expected behavior

When the Control UI uploads an image and the image is saved as:

media://inbound/<filename>.jpg

OpenClaw should resolve that URI to the actual local media store path, for example:

C:\Users\<user>\.openclaw\media\inbound\<filename>.jpg

and pass the image to the configured image model or media-understanding pipeline.

Actual behavior

OpenClaw saves the image successfully, but later fails to resolve the media://inbound/... URI and looks for a malformed local path:

C:\Users\<user>\.openclaw\workspace\media:\inbound\...

The image tool fails with Local media file not found.

Additional note

This appears to be a separate issue from the OpenWebUI image upload path, where the image reached OpenClaw as image:"". In the Control UI case, the image is preserved as media://inbound/..., but the local path resolver appears to be wrong on Windows.