Feature: Auto-route pasted images to multimodal-looker when main model lacks vision

[itisvincent]

Problem

When a user pastes an image into an opencode session running a non-vision model (e.g. glm-5.2, deepseek-v4, etc.), opencode sends the image directly to the main model API which rejects it with an error like:

ERROR: Cannot read "clipboard" (this model does not support image input). Inform the user.

This happens before any OMO agent logic runs, so the multimodal-looker agent (which exists exactly for this purpose) never gets a chance to handle the image.

Current behavior

1. User pastes image into chat
2. Opencode sends image to main session model (e.g. ollama-cloud/glm-5.2)
3. Model API rejects image (no vision capability)
4. Error shown to user — dead end

The look_at tool exists and routes to multimodal-looker automatically, but only when explicitly invoked. There is no hook that intercepts pasted images and routes them.

Proposed solution

Add a hook (e.g. chat.message.before or session.error) that detects:

1. The user message contains image attachments (clipboard paste)
2. The main session model does not have vision capability (check modalities.input includes image)

When both conditions are met, auto-invoke the look_at tool with the pasted image data, routing it to the multimodal-looker agent — which already has a vision-capable model configured (e.g. ollama-cloud/gemma4:31b).

Alternative: intercept at session.error

When the main model returns an error containing "does not support image input" or similar, automatically retry by routing the image through look_at instead.

Alternative: pre-check model capabilities

Before sending the message, check if the main model supports image input. If not, strip images from the message and invoke look_at separately, then combine the result.

Why this matters

• Users running cost-efficient non-vision models (like glm-5.2) as their main model still need vision occasionally
• Currently the only workaround is manual: save image to file, then ask the agent to use look_at, or switch the entire session to a vision model
• The infrastructure already exists (multimodal-looker agent, look_at tool, vision-capable model cache) — it just needs the auto-routing hook

Environment

• oh-my-openagent: 4.13.0
• opencode with ollama-cloud/glm-5.2 as main model
• multimodal-looker configured with ollama-cloud/gemma4:31b

Happy to help test a prototype if this gets picked up.