read tool returns metadata-only for image files, breaks multimodal model image input

[yyk9137]

What happened?

Problem

When the AFT plugin is active, the read tool returns only a metadata string for image files (e.g., Image read successfully (PNG, 59KB). File: /path/to/image.png). The actual image content is never returned to the model.

Expected behavior

For models configured with attachment: true and modalities: { input: ["text", "image"] }, the read tool should return image content so the model can process it with its vision capabilities — matching the behavior of OpenCode's native read tool.

Observed behavior

AFT's read tool detects image file extensions and returns a metadata-only string, regardless of whether the calling model supports image input. The model receives only the string Image read successfully (PNG, 59KB). File: ... and cannot see the actual image.

Steps to reproduce

1. Install AFT plugin (@cortexkit/aft-opencode@latest)
2. Configure a model with attachment: true and modalities: { input: ["text", "image"] }
3. Have the model call read on any .png or .jpg file
4. Observe: only metadata string is returned, no image content

Impact

Any workflow that uses read to load image files for multimodal models is broken. For example, OMOslim's observer agent (using a multimodal model) cannot read images when delegated via task() — the observer calls read on the image file but only receives metadata, so it cannot perform visual analysis.

Diagnostics

AFT binary version: 0.39.4
AFT plugin version: 0.39.4
Platform: Windows 11 (win32-x64)
OpenCode version: 1.17.11

Plugin version

0.39.4

AFT binary version

0.39.4

Platform

Windows 11 (win32-x64)

Log output (optional)

N/A — relevant logs are in OpenCode's main log, not in `/tmp/aft-plugin.log`. The OMOslim image-hook log shows the flow: image is saved to disk, observer session is created, `read` tool is called on the saved image, but only metadata is returned.