ZhangShurong/dsh-vista

Tags every image the agent sees with a stable id and gives it three tools to revisit, crop and sample pixels long after compaction.

An agent sees a screenshot once and then it is gone — dropped by the next tool call or by compaction, and re-screenshotting is both expensive and lossy because the screen has changed. dsh-vista keeps every image it has seen in an addressable archive and gives the agent three tools to go back: inspect revisits any past image and crops a region enlarged without smoothing, read_pixels samples exact colours from a region as a symbol matrix, and visuals lists what has already been seen. The point the author makes is that looking at the whole image again is not inspection — a cropped, unsmoothed region is what lets a model read a chart axis or a small label.

Vision & Multimodal ★ 1
View on GitHub ↗

Install

dsh plugin --profile web add github:ZhangShurong/dsh-vista#v0.1.3

The README's install section gives the plain npm form (dsh plugin --profile web add dsh-vista) and a pinned form (dsh-vista@0.1.3); this line is the GitHub tag equivalent, pinned to the same version. It suggests verifying with dsh --profile web --dump-config | grep dsh-vista before starting the profile.

Compatibility

Every image that reaches the model is tagged automatically with a stable id in the form <visual id="v12" source="tool:read_image" size="1280x720"></visual>, so the model can reference it later even after the original bytes have left the context window. read_pixels is documented as working even when the model route accepts text only — the region is returned as a compact symbol matrix rather than as an image.

Details

Alternatives

TZHR-invest/dsh-plugins#dsh-vision-tool · Cheng-cheng9669/dsh-deepseek-vision · shinegeer/dsh-vision-window

More plugins in Vision & Multimodal

Browse more in Vision & Multimodal

Guides for Vision & Multimodal plugins