AbdullahElTiby/dsh-vision-bridge

AbdullahElTiby★ 0JavaScript最后同步: 2026-08-16

在 GitHub 打开

Vision bridge for text-only models (DeepSeek, pi-ai routes): Gemini Flash describes attached images at the llm stream seam and provides a describe_image tool.

README 摘要

dsh-vision-bridge npm: dsh-vision-bridge-dsh · source/GitHub: AbdullahElTiby/dsh-vision-bridge (the npm package is named dsh-vision-bridge-dsh because the shorter name was already taken on npm). Host-plane plugin that gives text-only models (DeepSeek and any provider route that does not declare image input) the ability to "see" images, using a Gemini Flash vision model as the eyes. What it does 1. Dispatch interception — the two LLM dispatch entry points are wrapped: llm.prepareCall (the agent loop's prepared-call path, used for main turns and subagents) and llm.stream (session titles, compaction, unprepared loops). Before the adapter stream is built, every image block in the conversation is described by Gemini and replaced with a [Image (mediaType, WxH): …] text block. The session history and UI keep the real image; only the model request is rewritten. This works for every text-only route ( deepseek-official , pi-ai providers such as opencode-go , …). Routes that genuinely declare image input are passed through untouched. Why method patching instead of the llm/stream waterfall: in this harness build the waterfall ignores arguments passed to next() (listeners always receive the ori…

在 GitHub 查看完整 README →
内容/媒体cordisdeepseekdshdsh-pluginvisionagent

分类