dsh-desktop-control

A DeepSeek Harness plugin that gives the model a native desktop tool: screenshot the Windows screen, locate controls by their text, drive the mouse/keyboard.

It wraps the battle-tested gui.ps1 script (screenshot + Win32 input + local OCR + UI Automation) as a DSH plugin, so the model calls a typed tool instead of assembling PowerShell commands every time.

Windows only. The package ships its own assets/gui.ps1 (UTF-8 with BOM); it does not depend on a workspace script.

Actions

action	What it does	Main args
find	Local OCR locates on-screen text and returns screen coordinates (optionally clicks). No image.	text / window / click / threshold / screen
elements	UIA control list for a window: name + centre point, no image	window / filter / depth / all
ocr	Local OCR of an image: text + per-line boxes	imagePath / filter
screenshot	Save a PNG (cursor stamped by default) and return its path	window / region / print / grid / cursor / maxWidth
windows	List visible windows (HWND, PID, rect, title)	top
click	Move and click, or click a control by UIA name	x y / uia invoke / button count
move	Move the cursor	x y
drag	Press, move, release	x y x2 y2
scroll	Wheel	x y amount
type	Paste text via the clipboard	text
key	Key combo, e.g. ctrl+s, alt+f4	key
activate	Bring a window to the front	window
pos	Cursor position	—
wait	Sleep	waitMs
sequence	Run several actions in one call (including screenshots)	steps

All coordinates are physical pixels (per-monitor DPI awareness v2 is enabled internally).

Why it is cheap

The real cost is round-trips, not images: every extra tool call replays the whole history. So the order is "no image first, screenshot last".

1. find runs the built-in Windows OCR and returns one line of coordinates:

   # find "Save" | window 1296x828 at 312,110 | 64 lines | matches=1
   356,358,42,12|1.00|Save
   

   Measured 87 characters (~25 tokens), versus 4,000–9,500 characters (1,300–3,100 tokens) of structured evidence when a vision model reads the same screen. ~97% cheaper, no quota used. click:true clicks the hit in the same call.

2. elements gives native controls as name + centre point, 70–120 tokens.

3. screenshot is the last resort: prefer one window or a region, cap the size with maxWidth.

4. sequence collapses click/wait/verify chains; measured 3 actions in 1 call / 535 ms instead of 3 calls / 1650 ms.

find "Plugins" -> click (or find click:true) -> screenshot only to verify

Accuracy

• find coordinates are the OCR box centres, not an eyeballed guess.
• OCR input is upscaled 2x (12px CJK UI text is otherwise unreadable), coordinates divided back.
• Inter-CJK spaces inserted by Windows OCR are stripped.
• Fuzzy matching: exact 1.00, contains 0.92, reverse-contains 0.78, character overlap otherwise; tune with threshold.
• A miss returns the three closest lines so the model can reword instead of re-reading the screen.
• Screenshots stamp the cursor; grid:true overlays labelled coordinate grid lines (the labels are screen pixels) but they get read as text by OCR engines, so the grid is off by default.

Install

Client (Electron): Plugins → + Add plugin, paste dsh-desktop-control, install, then Enable now.

CLI profile:

dsh plugin --profile web add dsh-desktop-control

Local development: remove the npm copy in the plugin page, then install from the directory path instead; source edits then take effect after reloading the plugin.

Configure

- id: dsh-desktop-control
  name: dsh-desktop-control
  config:
    enabled: true
    powershellPath: powershell.exe
    timeoutMs: 30000
    screenshotDir: ''          # empty = $DSH_HOME/desktop-control
    screenshotMaxWidth: 1440   # 0 disables scaling
    screenshotGrid: false      # grid labels are picked up as text by OCR
    screenshotCursor: true
    keepShots: 40
    elementsMaxRows: 140

Option	Default	Meaning
enabled	true	false registers no tool
powershellPath	powershell.exe	PowerShell executable
timeoutMs	30000	Per-call timeout
screenshotDir	$DSH_HOME/desktop-control	Where screenshots are saved
screenshotMaxWidth	1440	Long-edge cap; the reply states the scale
screenshotGrid	false	Stamp the labelled coordinate grid
screenshotCursor	true	Stamp the cursor crosshair
keepShots	40	Prune older screenshots after each shot
elementsMaxRows	140	Truncate elements output (all:true disables)

Why a plugin instead of the pwsh tool

The DSH pwsh tool runs children at Low integrity, so SetCursorPos / SendInput are often refused. This plugin spawns PowerShell from the host process, outside that sandbox, so input usually works without switching the permission preset.

Safety

• Full desktop control: it can click anything and type anything, and screenshots enter the model context once you open them.
• Do not enable it on an unattended machine.
• UAC secure desktop, the lock screen, and some games/anti-cheat/terminal apps cannot be driven.

Self-test

npm run selftest

Pure argument-construction tests: no screenshot, no mouse movement.

Limitations

• find depends on OCR. Measured: 2–4 character menu items like "Insert" or "Workspace" hit at score 1.00; "Slide Show" was misread as "幻灯片囗节" and only matched partially. A miss returns the closest lines so the model can reword.
• Background capture is blank for some apps. Windows that are not in the foreground are grabbed with PrintWindow, which returns an empty bitmap for GPU-rendered apps such as PowerPoint; find then tells you to activate first. Foreground windows use real screen pixels.
• UIA does not expose Chromium/Electron/Qt painted content. Measured here: the DSH client, Edge and Abaqus only yield the window shell, title-bar buttons and menu bar.
• A missing or minimised window raises an error instead of silently falling back to the full screen (which would capture another window and click the wrong place).
• sequence does not wait for animations; insert an explicit wait step.
• Windows only.

License

MIT