opencode

vision.ts — OpenCode plugin

Gives a text-only model eyes inside OpenCode: the experimental.chat.messages.transform hook replaces image file parts in the outgoing history with focus-hinted descriptions written by your configured vision model.

Install

Copy this one file into a plugin directory:

# per-project
mkdir -p .opencode/plugin && cp vision.ts .opencode/plugin/

# global
mkdir -p ~/.config/opencode/plugin && cp vision.ts ~/.config/opencode/plugin/

Configure

Same env chain as the proxy and CLI tools in this repository — typically ~/.config/agent-vision-toolkit/env (permissions 0600):

VISION_API_KEY=...
VISION_BASE_URL=https://your-vision-endpoint/v1
VISION_MODEL=your-vision-model
# optional: LANG=zh  (or en) to pin the description language

$VISION_ENV_FILE, %LOCALAPPDATA%/agent-vision-toolkit/env and a .env in the working directory are also read; later files override earlier ones.

Behavior

• Attached/pasted images (data URLs, http URLs, or local paths — local files are inlined) are described under their own message's text.
• Descriptions are cached per (image, prompt) for the process lifetime.
• A failed description is replaced with an explicit failure note — the raw image is never forwarded and failures are never silent.

Limitations

• The transform hook does not expose the active model, so the plugin cannot detect a multimodal primary by itself. Set VISION_REWRITE=off in the environment to disable rewriting when you run a vision-capable model.
• Only user-attached image parts are handled in this version; images returned inside tool results are not yet rewritten.