GXX182/dsh-vision-bridge

README excerpt

dsh-vision-bridge 简体中文 dsh-vision-bridge is an installable DeepSeek Harness bundle. It contributes a Cordis plugin that lets a text-only DeepSeek route delegate session or local PNG, JPEG, WebP, or GIF understanding to an external vision API, then returns only bounded text analysis to the active agent. This is a Harness plugin, not an agent skill. The bundle manifest activates the plugin through cordis.patch.yml ; the plugin registers the deepseek-vision-bridge provider route, the vision bridge tool, and its model guidance through Harness services. The current release uses Google Gemini as its vision provider. Support for additional image-understanding providers is planned for future releases. How it works 1. Select the DeepSeek + Vision Bridge provider route and attach images normally in the Harness conversation. 2. Harness validates and stores the images through its attachment service, then records their immutable references in the session. 3. The bridge provider replaces image blocks only in the upstream request copy with controlled text markers; the durable session and transcript keep the original images. 4. DeepSeek calls vision bridge with a focused question. With no image ar…

Category