dsh-plugin-computer-user
Virtual-desktop computer use for DeepSeek Harness. The agent sees your screen
in the right sidebar and its mouse acts on that snapshot — not on your real
cursor. It plays its own game; you keep your computer.
A fork of jing-hy/computer-user that
adds the half it was missing: a visualization.
The problem
Upstream computer-user gives a text-only model genuine control of a Windows
machine. It works, but you cannot see any of it:
computer_screenshot writes a PNG to a temp directory for the model's eyes
only. You never see what it is looking at.computer_click moves your actual cursor and clicks. While the agent
works, the machine is not yours.- You find out what happened afterwards, from the conversation.
For a long autonomous task that is a bad deal: you cannot tell whether it is
about to click the wrong thing, and you cannot use your own computer meanwhile.
What this fork changes
A live snapshot in the right sidebar. Every capture the agent makes is
published to a dockable panel: the screen, the agent's virtual pointer moving
across it, click ripples, the text it types, the keys it presses. You watch the
work in progress.
Actions land on the snapshot by default. link_mode: virtual means the
agent still screenshots, still "clicks", still "types" — but that pointer lives
on the panel. Nothing reaches Win32 SendInput, so your cursor, your focus, and
your window layout are untouched. The agent plays in the picture; your desktop
is yours.
Flip to real when you mean it. One toolbar button (or the settings card)
switches link_mode to real, and now the Win32 path runs — with the panel
still mirroring what happens, so real mode is visible rather than a hijack.
You can point at things. Click anywhere on the snapshot, optionally with a
note, and that becomes a queued instruction with an exact screen coordinate. The
agent reads them through computer_get_vision, so "click this" stops being a
description problem.
The agent gets numbers, not just a picture. computer_get_vision reports the
captured size, the virtual-screen origin, the capture scale, the current
pointer, the frame's age, and your pending hints. Combined with the panel, the
look-act-verify loop closes on both sides.
How the two halves fit together
| upstream | this fork |
|---|
| capture | PNG to disk for the model | same PNG, published to the sidebar |
| pointer | real cursor via SendInput | virtual pointer on the snapshot (default) |
| typing / keys | real keystrokes | drawn on the snapshot (default) |
| verification | screenshot again, read it | screenshot again, watch it |
| user input | none | click the snapshot to point, with a note |
| real control | always | opt-in, one switch |
Everything upstream did is still here: all nine computer_* tools, the four run
modes, /computer approval, the settings namespace, the output guard that
bounces tool calls written as prose, and the bundled PowerShell that needs no
native module.
Install
pnpm install
pnpm run build # → lib/index.js (host) + lib/client.js (browser)
Then add the package to your DSH profile as a plugin dependency. The bundle patch
in cordis.patch.yml inserts the plugin row and sets the defaults:
- insert:
- id: dsh-computer-user
name: '@asun-labs/dsh-plugin-computer-user'
config:
mode: manual
link_mode: virtual
Settings
Settings → 电脑操作. The two that matter sit at the top:
动作落点 (link_mode)
virtual (default) — mouse and keyboard only affect the sidebar snapshot.
Your desktop and cursor are untouched.real — Win32 SendInput drives this machine. The agent takes the real cursor.
运行模式 (mode)
disabled — refuse every tool call.readonly — screenshot, cursor, vision reads, wait.manual (default) — acting tools need /computer approval in the conversation.auto — the model may call everything.
Advanced: capture interval (how often the panel refreshes), screenshot directory,
typing interval, scroll units, whether the AI may change its own modes, the
output guard, and debug logging.
The panel
| control | what it does |
|---|
| 虚拟快照 / 实机接管 badge | which mode is live, at a glance |
→ real / → virtual | switch link_mode in one click |
− / ＋ / % | zoom; 100% is real 1:1, one panel pixel per snapshot pixel |
| 适应 (Fit) | whole snapshot, centred, aspect ratio preserved |
| 1:1 | original pixels, scrolled if it does not fit |
| 跟随 (Follow) | pan automatically to keep the virtual pointer visible |
⟳ | capture now; otherwise the panel refreshes on the configured interval |
Gestures: left click = point for the agent (with the note field, if you typed
one), Shift+drag = pan when zoomed in, Ctrl/⌘+wheel = zoom.
The snapshot is never stretched. It is scaled by whichever axis fits and centred
on the other, so the shapes you see are the shapes on screen — which is what
makes clicking it meaningful.
Architecture
src/
index.ts plugin body — tools, routes, capture loop, settings, /computer
tools.ts the 12 computer_* tools, the mode gate
session.ts VirtualSession — the board the panel renders (frames, overlays, hints)
ports.ts DesktopPort: real (SendInput) vs virtual (record only)
geometry.ts screen ↔ frame ↔ panel transforms (pure, shared by both halves)
client.tsx the sidebar panel + the settings card
ps.ts bundled PowerShell runner (base64 JSON payload)
capture.ps1 DPI-aware multi-monitor screenshot (upstream, unmodified)
input.ps1 Win32 SendInput mouse + keyboard (upstream, unmodified)
tests/ geometry, session, tools, gate, output guard
The split that makes the fork work is ports.ts: a tool asks a DesktopPort to
click, and whether that means "really click" or "draw a click on the panel" is
one swap. link_mode changes which port is returned; nothing else in the tool
layer knows the difference. It also means the whole tool set is testable without
touching a real desktop — the tests do exactly that. The Host keeps only the
latest 24 generated panel previews and removes its generated screenshots on an
orderly shutdown. A screenshot path supplied explicitly by a tool caller is not
managed by this cleanup.
geometry.ts is shared verbatim by both halves, which is why a click you make on
the panel and a click the agent makes through computer_click resolve to the
same screen pixel instead of merely close to it.
Requirements
Windows. Tested with DSH 0.2.1-alpha.2; the package also declares the earlier
0.1.5 and 0.1.7 lines. Capture uses System.Windows.Forms/System.Drawing and input uses Win32
SendInput, both through the bundled PowerShell — no native module, no build
step, no external API. On other platforms the panel loads and reports that
capture and injection are unavailable.
Attribution
Forked from computer-user by
jing-hy, MIT licensed. The computer_* tool design, the run-mode gate, the
/computer approval flow, the settings namespace, the output guard, and both
PowerShell scripts come from that project; see NOTICE. The
visualization layer, the virtual/real port split, computer_get_vision,
computer_set_link_mode, and the sidebar panel are this fork's additions.
License
MIT — see LICENSE.
