[BUG] VS Code extension: image/file attachments silently dropped when using custom model backend (ANTHROPIC_BASE_URL)

Resolved 💬 1 comment Opened May 5, 2026 by Muzili919 Closed Jun 2, 2026

Summary

When using Claude Code VS Code extension with a custom model backend (via ANTHROPIC_BASE_URL pointing to a third-party proxy such as z.ai / GLM 5.1), file and image attachments dragged into the chat are silently dropped — the model receives the image URLs as opaque metadata in tool result blocks, but cannot "see" or render them as actual image content.

Repro

  1. Configure ANTHROPIC_BASE_URL to point to a third-party model proxy (e.g., z.ai for GLM 5.1)
  2. Open Claude Code in VS Code extension
  3. Drag a file or image into the chat input
  4. Send a message asking about the attached file

Expected

The model should be able to see and process the attached image/file, same as when using the default Anthropic API backend.

Actual

The image URL appears in the conversation context as a raw URL (e.g., maas-log-prod.cn-wlcb.ufileos.com/...) appended to tool results, but:

  • The model does not receive it as a renderable image content block
  • The model is unaware the user sent anything and responds as if no attachment exists
  • The user has to explicitly tell the model to look at the URL and use analyze_image to read it

Environment

  • Platform: macOS (Darwin 25.4.0)
  • Claude Code: VS Code extension
  • Model backend: GLM 5.1 via z.ai proxy (ANTHROPIC_BASE_URL)
  • Date: 2026-05-05

Notes

This appears to be a difference in how the VS Code extension handles image attachment rendering between the native Anthropic API and third-party proxy backends. With native Claude models, image attachments are presumably converted to proper image content blocks in the API request. With custom backends, they seem to be passed as raw URLs that the model cannot interpret.

Related: #55070 (similar silent attachment drop on Android), #52647 (image paste support request for VS Code)

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗