hanchn/dsh-vision-router

README excerpt

Vision Router for DeepSeek Harness 简体中文 A zero-config, multi-provider vision tool for DeepSeek Harness. It adds one model-facing tool, inspect image , while keeping model discovery, routing, privacy policy, and provider fallback behind a small Cordis plugin. Why Vision Router? DeepSeek remains the reasoning agent. Vision Router delegates image perception to a local or explicitly configured multimodal model, then returns text evidence to the agent. The default path is private and requires no model name or endpoint configuration. Features - Zero-config discovery of vision-capable Ollama models via model metadata - Native DSH drag-and-drop and clipboard image intake with automatic local analysis - Original thumbnails remain visible in chat; images are archived locally by content hash - Local-first routing with remote fallback disabled by default - Multiple prioritized providers - Ollama and OpenAI-compatible vision APIs - Explicit provider selection when a task needs it - PNG, JPEG, WebP, GIF, and image data URL inputs - Presets for general description, OCR, UI, charts, and visible code/errors - API keys read from environment variables, never stored in the plugin config Requirements -…

Category