Feature Request: Auto-switch to vision model when image attachment is detected

Summary

When a user drags/drops or pastes an image into the chat, automatically switch to a configured vision-capable model instead of silently dropping the image or requiring manual intervention.

Motivation

Currently (per issue #49518), when a user attaches an image but the active model does not support image input:

1. The image is silently dropped by the gateway
2. The assistant replies as if no image was provided
3. No warning is shown to the user

This creates a poor user experience where:

• The image appears to be attached in the UI
• Sending succeeds
• But the assistant cannot see the image
• The user must manually switch models or use a command like /model vision-model

Proposed Solution

Add a new configuration option visionFallbackModel that:

1. Specifies a vision-capable model to use when images are attached
2. When an image attachment is detected and the current model does not support images:
   • Automatically routes the request to the configured vision model
   • Processes the image successfully
   • (Optionally) Returns to the original model after

Example configuration:

{
  "agents": {
    "defaults": {
      "visionFallbackModel": "llama-3.2-vision"
    }
  }
}

User Experience After Implementation

Before: User drags image → Image dropped silently → Assistant: "I cannot see the image" After: User drags image → Automatically switches to vision model → Image processed successfully

Alternative Solutions Considered

1. User-facing warning at send time: Show a warning before sending if the model doesn't support images. But this still requires manual action.

2. Always use vision model when available: Could be confusing if user wants text-only responses for text conversations.

3. Per-conversation model setting: Allow users to set model per conversation. Adds complexity.

The auto-switch approach is the most seamless for the user.

Implementation Notes

Based on code review of parseMessageWithAttachments in src/gateway/server-methods/chat.ts:

1. The function already checks opts?.supportsImages to determine if the model supports images
2. When supportsImages === false and attachments exist, it currently just logs a warning and returns empty images
3. The fix would add logic to:
   • Check if visionFallbackModel is configured
   • If yes, temporarily use that model for the request
   • Continue normal processing with the vision-capable model

Key files to modify:

• src/gateway/server-methods/chat.ts — add visionFallbackModel logic in parseMessageWithAttachments or its callers
• Configuration schema to accept the new visionFallbackModel option

Backward Compatibility

This is purely additive — if visionFallbackModel is not configured, behavior remains unchanged (images are dropped with warning for text-only models).

Priority

Medium-High — this is a significant UX gap that causes confusion and makes the image attachment feature effectively unusable for users with text-only primary models.

Related: #49518 (same underlying issue, this is a specific solution proposal)