@kanchengw

Plug-in vision for text-only models on DSH, with native multimodal interaction, layered evidence memory and cache, and intent router.