dsh-win-cua

Windows computer-use for DeepSeek Harness: real mouse and keyboard input plus screen capture, with no external driver.

How it works

Capability	Win32 API
Mouse move, click, wheel	user32!SendInput (INPUT_MOUSE)
Keyboard taps and shortcuts	user32!SendInput (INPUT_KEYBOARD)
Type literal text (incl. CJK)	SendInput with KEYEVENTF_UNICODE
Screen capture	gdi32!BitBlt + CAPTUREBLT (layered windows included)
Window list / focus	user32!EnumWindows, ShowWindow, SetForegroundWindow
DPI	SetProcessDpiAwarenessContext(-4) — per-monitor v2

FFI is koffi. Coordinates are physical pixels of the virtual screen; the primary display's top-left is (0, 0). Absolute pointer moves use MOUSEEVENTF_ABSOLUTE | MOUSEEVENTF_VIRTUALDESK, so displays left/right/above the primary one are reachable too.

Tools

Look

• cua_screenshot — capture the whole virtual screen to a PNG
• cua_screenshot_window — capture one window by handle (PrintWindow, screen-blit fallback)
• cua_zoom — capture a region and upscale it x1–x4 for reading small UI text
• cua_windows — visible top-level windows with bounds and flags
• cua_app_list — same list enriched with the process executable name
• cua_cursor — pointer position and virtual-screen geometry

Act

• cua_mouse_move — move the pointer along a human path
• cua_mouse_click — click / double / right / middle, optionally at x,y
• cua_drag — press, travel and release with the button held
• cua_scroll — wheel, vertical or horizontal
• cua_key — one key with optional modifiers
• cua_type — type text (unicode, incl. surrogate pairs)

Window & app

• cua_focus_window — restore and foreground a window
• cua_wait — sleep and/or poll until a window title appears
• cua_app_launch — launch an executable without a shell

Every tool declares output { schema, render } — the DSH tool registry rejects raw definitions without it, so keep that shape when adding tools. Read-only tools are marked readOnly and register as isConcurrencySafe.

Config

Set in this bundle's cordis.patch.yml under config:

Key	Default	Meaning
humanize	true	Pointer travels a smoothstep path, typing and clicking keep a human cadence
keepScreenshots	30	Newest captures kept per directory
moveSteps	16	Max pointer steps for a humanized move
keyDelayMs	[12, 34]	Per-character delay range for humanized typing
clickSettleMs	[60, 140]	Pause between the pointer arriving and the button press
zoomFactor	2	Default cua_zoom upscale factor (hard cap 4)
zoomMaxPixels	24000000	Largest upscaled image allowed
maxWaitMs	10000	Upper bound for cua_wait
allowLaunch	true	Set false to disable cua_app_launch

Pointer targets are clamped into the virtual screen, and astral code points (emoji, CJK extension B) are sent as surrogate pairs so KEYEVENTF_UNICODE types them correctly.

allowLaunch lets the model start any program the host process can run; set it to false when the session is not trusted.

No external driver

Unlike dsh-computer-use (which shells out to the cua-driver binary) this plugin calls Win32 through koffi inside the host process: nothing to install, no daemon, no CUA_DRIVER_BIN. The trade-off is honest: without the driver there is no UIA/accessibility tree, so targeting is pixel-based — use cua_zoom / cua_screenshot_window to read a window precisely before clicking.

Install

From npm:

dsh plugin --profile <profile> add dsh-win-cua

Or from a checkout / local directory:

dsh plugin --profile <profile> add <this directory>

Or in the Desktop app: Plugin Manager → install bundle → pick this directory.

Windows only. SendInput injects into the session whose input queue is active — a window on a locked or disconnected desktop receives nothing.

Changelog

0.2.1

• Fixed: absolute pointer coordinates now carry MOUSEEVENTF_VIRTUALDESK; without it a multi-monitor desktop silently mapped every move/click/drag onto the primary display.
• Fixed: cua_app_launch no longer crashes the host when the target is a directory or the launch is refused — it waits for the OS to confirm the process started and rejects cleanly.
• GetDIBits is called with the bitmap selected out of the DC, as the API requires.
• keepScreenshots now rotates window-* and zoom-* captures too.
• cua_focus_window reports whether the window actually took the foreground instead of always returning ok: true.
• Arrow/navigation keys are sent with KEYEVENTF_EXTENDEDKEY.
• Config values are normalized: keyDelayMs: 20 (a number) works the same as [20, 20].
• cua_type sends one Enter for "\r\n"; cua_zoom coerces string coordinates safely; cua_scroll reports the delta it sent.
• package.json gained icon, repository, bugs, homepage and a >=20.11 engine range.

0.2.0

• New look tools: cua_screenshot_window (PrintWindow with screen-blit fallback), cua_zoom (region capture upscaled x1–x4), cua_app_list (windows with process names).
• New act tools: cua_drag, cua_wait, cua_app_launch (disabled with allowLaunch: false).
• humanize is now a real config knob: smoothstep pointer path, per-character typing delay and click settle range, with moveSteps / keyDelayMs / clickSettleMs tuning.
• PNG encoder, screenshot retention (keepScreenshots) and coordinate clamping tightened.

0.1.0

First public release: screenshot, cursor, window list/focus, mouse move/click/scroll, key press and unicode typing.