dsh-vision-reasoner：在原图像素坐标上标注与逐窗查看
dsh-vision-reasoner 是 LionNatsu 维护的 DeepSeek Harness 插件，通过 camera 与 mark 工具在原图像素坐标上标注、缩放与分窗扫视，返回带坐标网格的视图与可校对的编号记录，并附带 vision-sweep-count skill。文章介绍工具集、安装命令、典型五步用法及会话内存储与 .dsh-vision 预览目录等行为，便于智能体开发者落地看图计数与区域核对类任务。
Read Moredsh-vision-primitives: Adding Pixel-level Visual Reasoning Primitives to DSH Agents
dsh-vision-primitives is a native visual reasoning plugin for DeepSeek Harness (DSH), designed to address issues of imprecise visual localization and the lack of reproducible coordinate benchmarks. It does not require an external MCP server; through an engineering process of "image-to-grid," "coordinate parsing," "local zooming," and "deterministic verification," it achieves precise pixel-level localization. Core functions include SOM grid overlay, pixel-level coordinate parsing, lossless local zooming, and verification tools such as annotation, measurement, OCR, color finding, and frame differencing. The plugin optionally connects to the MiMo V2.5 multimodal backend to enhance understanding capabilities, while also supporting pure local usage. It minimizes OS boundaries, executing screenshots, OCR, and file persistence solely through PowerShell, without including mouse or keyboard control. It features easy installation and supports the DSH plugin system.
Read More