image-reader-relay

Opencode plugin that lets text-only models read pasted images via a vision-capable model.

When an image appears in a session whose main model cannot see images (pasted by the user or produced by a tool), the plugin:

1. Stashes pasted images in memory (never writes to disk). They stay visible in the chat exactly as you pasted them.
2. Injects a note into the model's context only — the model sees [An image was pasted... Use the read_image tool...] while you still see the image and your text in the UI. Notes are injected for image parts in any message (user pastes and tool/assistant results).
3. Registers a read_image tool the main model can call with its own question — for pasted images (imageIndex) or files on disk (filePath, e.g. tool screenshots).
4. Relays the image to a configured vision model in a throwaway session, cleans up, and returns the answer as a tool result.

For models that support images natively, the plugin does nothing — the image stays intact.

Install

Add the plugin to your config (opencode.json in a project, or global ~/.config/opencode/opencode.json):

{
  "plugin": [
    ["massiveits/image-reader-relay", { "model": "clinepass/cline-pass/mimo-v2.5" }]
  ]
}

Options

Option	Type	Default	Description
model	string	clinepass/cline-pass/mimo-v2.5	Vision model used to read images. provider/modelID format.
timeoutMs	number	60000	Max time a relay call may take.

The model can also be set with the environment variable IMAGE_READER_MODEL (takes priority over the model option).

Changing the model later is just editing the model option — no plugin file changes.

How it works

• chat.message hook detects pasted image parts and stashes them per session (last 10 images, 30-minute expiry; vision-capable models are skipped).
• experimental.chat.messages.transform swaps image parts for a note in the model's context only — for user messages and tool/assistant results — while the stored message (and the UI) keeps the images untouched.
• read_image tool: question + optional imageIndex (0 = most recent) for pasted images, or filePath to read an image file from disk (png/jpg/webp/gif/bmp/svg/avif, max 20MB, resolved against the session directory). It sends the image with a system prompt to the configured vision model and returns its answer.
• Everything is in-memory: the relayed image never touches the filesystem.
• A throwaway session is created per call and deleted afterwards.

Requirements

• A main model that supports tool calls.
• Access to the configured vision model through one of your providers.