CaseyTso/dsh-analyze-image-tool

CaseyTso★ 0JavaScriptLast synced: 2026-08-15

Open on GitHub

给纯文本 DeepSeek Harness 模型加上识图能力:analyze_image 把图片转发到任意 OpenAI 兼容视觉端点 | Vision bridge for text-only DSH models

README excerpt

analyze-image-tool English 中文 A vision bridge for text-only DeepSeek Harness models: registers an analyze image tool that answers questions about images via any OpenAI-compatible vision/multimodal endpoint . No vendor is hard-coded — bring your own baseURL + apiKey + model . Features - Generic endpoint : One config ( baseURL + apiKey + model ) covers any OpenAI-compatible endpoint — SiliconFlow, DashScope compatible-mode, Zhipu, OpenRouter, Volcano, local Ollama, OpenAI… zero vendor logic in code. - Sandbox-safe reading : Local images优先走宿主沙箱 fs 通道( ctx.fs.readBytes ),遵守会话路径策略;没有该通道的宿主自动回退到 Node 原生读取。也支持 http(s) URL 和 data: URL。 - Schema safety : Tool parameters compiled via defineTool to standard JSON Schema with object root( type: "object" ), avoiding the pitfall of "tool schema root is not object causing whole session 400 crash"(dsh community 297)。 - Robustness : API key auto-scrubbed in error messages; reasoning blocks auto-stripped from thinking models; -only responses treated as "reasoning without answer" with actionable prompts. - Structured return : { text, model, usage } with endpoint model id and token counts. - Coexistence : Tool name analyze image doesn't conflict with c…

View full README on GitHub →
Media / Contentdeepseek-harnessdsh-pluginmultimodalopenai-compatiblevisionvlm

Category