dsh-luna-vision-bridge

DSH's native input box currently accepts text only, so pasted or dragged images cannot be consumed by downstream LLMs. dsh-luna-vision-bridge is a host-side adapter plugin that registers a custom provider, keeping the existing image interactions (paste, drag-and-drop, thumbnails, delete, send button) fully functional while transcribing image content into text for text-only models.

The workflow is straightforward: images land in DSH's native attachment store, Codex Luna (gpt-5.6-luna) transcribes them into text descriptions, and those descriptions are routed to any configured downstream text provider. Multiple downstream combinations (e.g. DeepSeek V4 Flash + Luna, Pi Coder + Luna) can be listed in the model selector and switched instantly without restarting. All downstream models are configurable through a visual form in the Web settings panel, and a built-in content-addressed cache prevents redundant Luna calls on identical historical images. Temporary image files are deleted immediately after transcription.

This plugin suits users who work in the DSH Web interface, want to keep the native image-paste experience, but whose downstream models are text-only. Keep in mind that this is a temporary compatibility workaround for DSH's current text-only admission rather than a long-term architecture. The transcription side is validated for Codex CLI and gpt-5.6-luna only, and it is not a general-purpose vision framework. Once DSH ships native image reading or vision routing, migrating to the official implementation and retiring this plugin is strongly recommended.