paicat1/dsh-screenshot

English | 中文

Zero-dependency screen capture for DeepSeek Harness: Lightweight: zero deps, zero binaries; Stage & shoot: one-click full screen, window layout, hover-snap capture of occluded windows; Agent self-service: path-only delivery; paths are universal, pair with modlens (optional) for one-call structured evidence.

• Lightweight — pure PowerShell, zero dependencies, zero binary payload; capture is maintained independently and never breaks on upstream updates.
• Stage & shoot — one-hotkey full-screen capture, or keep the desktop live and arrange any window before box-selecting a region; hover any window and it glows with a snap outline — one click captures that window's full content, even when occluded (except a standalone PowerShell, see below) — what you stage is what you get.
• Agent self-service — the model can capture the screen on its own; with modlens (optional) installed, capture + read happen in one call, returning structured content (OCR/layout/semantics) that text-only models can consume directly.

Quick start

dsh plugin --profile web add @paicat1/dsh-screenshot
# restart dsh web

• Ctrl+Alt+S — capture: once armed, click the desktop = full screen; hover a window = snap outline, one click captures that window's full content (even when occluded, except a standalone PowerShell); drag = free region (Esc to cancel)
• Ctrl+Shift+Alt+S — full-screen capture: no interaction, captures the whole virtual desktop
• Want the agent to screenshot on its own? Just tell it to use the modlens_screenshot tool.

Video tutorial (Bilibili): Using DeepSeek Harness to build a Dsh screenshot plugin for DeepSeek Harness

Two doors: hotkey for humans, tool for agents

Entry	Best for	Notes
Browser hotkey (human)	"Here's the screen I want to show you"	Full / region capture; the PNG path is auto-inserted into the DSH input box and copied to the clipboard
modlens_screenshot tool (agent)	"Let me look at the screen myself"	Model-invoked: capture + (with modlens present) in-call structured read, returning evidence + the shot path

Manual capture demo	Agent self-capture demo	Window-snap demo

Why pass the path, not the image?

Screenshots are saved to %USERPROFILE%\Downloads\modlens-screenshots\ (a dedicated, easy-to-clean directory), and only the PNG path is handed to the agent — not the image stuffed into a chat box or temp directory. Deliberate design:

• No image litter: many agent frameworks copy pasted images into their own temp/attachment directories, accumulating untrackable junk. A path keeps the image in exactly one place (modlens-screenshots/); cleaning up is deleting one directory.
• Paths are universal: any image-capable agent (native multimodal models, or modlens-style bridges) can read an image from a path — paths are universal, image formats are not.
• Clean context: a path is tens of bytes of text; an image is hundreds of KB of binary. Paths keep the context clean and traceable.

Capture once, reuse the path in any agent, produce zero junk.

Capability split: what's the plugin's, what's modlens's

Layer	Capability	Owned by
Capture	Full / region / window-snap capture / staged-window layout, zero-dep PowerShell	This plugin
Delivery	Path-only, clipboard, dedicated save dir	This plugin
Entry points	Browser hotkeys + agent-callable capture tool	This plugin
Reading	OCR / layout / semantics structured evidence	modlens (optional)
Consumption	Who understands the image	any multimodal model / vision bridge — agnostic

Capture and delivery are fully self-contained and work standalone; reading is an ecosystem combo — install modlens (or hand the path to any image-capable model/bridge) to unlock it. Without modlens, capture still works.

Why can window-snap capture "occluded" windows? On hover the plugin enumerates the on-screen windows and outlines the one under the cursor; on click it uses PrintWindow to ask the window to render itself — the shot is the window's own content, not the on-screen pixels, so being covered by other windows or wrapped in a DWM shadow doesn't matter. It then crops to the DWM content bounds to drop the shadow, for clean edges.

Known exception: a standalone PowerShell window: its console (conhost) implementation does not respond to PrintWindow, and does not render its GDI surface while covered — both see-through-occlusion paths are unavailable, so it falls back to the visible on-screen pixels (including whatever covers it). Bring it to the foreground and click, or just drag-select over it. All other windows (including cmd and other console windows) capture normally.

Configuration

Optional cordis config (all enabled by default):

• route: false — disable the browser capture route
• tool: false — disable the modlens_screenshot tool

MODLENS_DSH_CLI explicitly sets the modlens CLI path (default probe: ~/.dsh/profiles/{web,headless}/node_modules/@liustack/modlens/dist/main.js).

Platform & License

• Platform: Windows (relies on PowerShell System.Drawing.CopyFromScreen)
• License: MIT

Relationship with modlens

Layer	Depends on modlens?	Notes
Capture action	No	Pure PowerShell CopyFromScreen, zero dependencies
Browser hotkeys / path insert	No	Standalone route /dsh-screenshot/screenshot
modlens_screenshot tool reading	Yes (optional)	If the modlens CLI is missing, the tool is not registered; capture still works
Multimodal models	No	After the path is inserted, multimodal models (e.g. go-mimo) can read the image directly, no modlens needed

Background & credits

The capture capability was originally implemented as an enhancement to the dsh plugin of @liustack/modlens (by Leon Liu): the original modlens dsh plugin integrated capture (this repo's fork of the feat/dsh-screenshot branch), and the author marked it as not planned in issue #48. It was split into this standalone plugin to decouple capture from modlens updates.

Many thanks to the original author for giving DSH image-reading ability — modlens lets text-only models (DeepSeek/GLM) "see" images, and this plugin's modlens_screenshot tool reuses the modlens pipeline to combine "capture + read" into one step. The capture half is maintained separately, but the image-reading ability always belongs to the modlens project.