dsh-free-vision

Vision bridge plugin for DeepSeek Harness (dsh) that gives text-only models image understanding through a bundled luma-mcp vision engine, defaulting to the free Qwen3-VL-Flash model with zero MCP configuration.

• Bundles the luma-mcp vision engine as a package dependency and starts it in-process; no cordis.patch.yml edits and no runtime npx.
• Registers a single general-purpose tool, image_understand (renamable via config.toolName), on ctx.tools.
• Includes free-tier providers: Qwen3-VL-Flash (qwen, default), Doubao vision (volcengine), and DeepSeek-OCR (siliconflow).
• Also supports zhipu (GLM-4.6V), hunyuan (HY-Vision), and custom OpenAI-compatible endpoints.
• Per-provider API base URL override lets built-in providers point to a proxy, gateway, local service, or OpenAI-compatible endpoint.
• Task modes auto | general | ocr | ui | debug | describe, with automatic multi-crop for large images.
• Accepts image_source as a local path, HTTP(S) URL, or data URI (PNG/JPG/WebP/GIF, ≤10MB), plus a required prompt.
• Settings UI form rendered from the plugin schema, saved to ~/.dsh/free-vision.json, effective on the next tool call.