Command Code v0.33.4 Windows image attachment bug

首先，在 Windows 上使用 @./screenshots/image-1.png 和絕對路徑，在列出的各個 shell 和模型中重現該問題。追蹤 CLI 在路徑偵測後將 PNG 路由到何處；完成的標準是影像以多模態視覺輸入的形式到達模型，而不是觸發 READ 和純文字回應。

Summary

Command Code v0.33.4 Windows image attachment bug

OS:
Windows

Version:
v0.33.4

Model:
kimi-k2.5 / kimi-k2.6

Input tried:
@./screenshots/image-1.png
C:/Users/UserA/Desktop/Projects/ProjectA/screenshots/image-1.png

Expected:
Image should attach as multimodal vision input.

Actual:
CLI treats the image as a file path, calls READ on the PNG, then the agent says it cannot see pixels and only has file/shell/search tools.

Observed output:
READ [screenshots\image-1.png]
“I can't view image content directly.”

Conclusion:
The image is routed through file reading, not encoded into model context as a vision attachment.

Expected Behavior

Image should attach as multimodal vision input.

Actual Behavior

CLI treats the image as a file path, calls READ on the PNG, then the agent says it cannot see pixels and only has file/shell/search tools.

Steps to reproduce the issue

OS:
Windows

Version:
v0.33.4

Model:
kimi-k2.5 / kimi-k2.6

Input tried:
@./screenshots/image-1.png
C:/Users/UserA/Desktop/Projects/ProjectA/screenshots/image-1.png

Command Code Version

0.33.4

Operating System

Windows

Terminal/IDE

vscode,external

Shell

git bash, cmd.exe, powershell

Additional context

Observed output:
READ [screenshots\image-1.png]
“I can't view image content directly.”

CLI treats the image as a file path, calls READ on the PNG, then the agent says it cannot see pixels and only has file/shell/search tools.