dsh-luna-vision-bridge

DSH's native input box is text-only today: pasting or dragging an image does nothing useful for downstream LLMs. dsh-luna-vision-bridge is a host-side adapter that registers a custom Luna Vision Bridge provider so the existing image workflow (paste, drag-and-drop, thumbnails, delete, Enter, send button) keeps working exactly as before, while image content is transcribed into text for text-only models.

Under the hood, images are stored in DSH's native attachment store and then passed to Codex Luna (gpt-5.6-luna) for transcription via codex exec --json. The resulting text description is forwarded to whichever downstream text provider you have configured. You can register multiple downstream targets (for example DeepSeek V4 Flash + Luna, Pi Coder + Luna) and switch between them from the model selector in real time. A visual form inside the Web settings panel lets you add or remove downstream combinations without editing any config file, and a content-addressed cache keyed by image and prompt avoids redundant Luna calls for historical images.

If you live in the DSH Web interface, rely on text-only models, and need a quick way to let the native input box accept images, this plugin gives you a working bridge right now. It is explicitly a temporary workaround for DSH's current text-only admission gate, not a general-purpose vision framework or a long-term architecture. Transcription is validated only for Codex CLI and gpt-5.6-luna; other CLIs or remote vision APIs are out of scope. As soon as DSH ships native image reading or a stable vision-routing API, you should migrate to the official implementation and retire this plugin.