Pixel-to-text image reading for text-only models

Key features

• Proxy in-situ packagingFor the provider of the selected model, use Proxy to package its adapter into "twins" and replace them in-situ (registerTwinAdapters, uninstall and restore by ctx.effect), without repeated registration.
• Every time I look at the picture, the problem faced by the model is actually the same: "What cost is this picture worth, and which route should be used to understand it?"
• Privacy mode (Privacy) - zero outbound call hard access control
• Intelligent/strict mode only uses external semantic understanding when needed and configured.
• Stream interceptioncapture the image block in the request → export to /.dsh/picturereader-vision/images/ → replace with a text path + local tool chain guidance → forward to the original adapter.