LAYER FOR CODEX’S COMPUTER USE · WINDOWS

Codex’s Computer Use.
Background first.

Background-first Windows layer for Codex’s Computer Use.

Codex has its own Computer Use. NexusWindows for Codex adds to it on Windows: a local .NET helper acts in the background first, and Codex’s physical input is only a last resort.

Explore the code See the features

Background first
Every action verified
Never a double action

For

A Windows screen whose numbered elements are linked to Codex — Illustration of an exchange in the Codex session: every result says which technique acted and whether the foreground was taken.

CODEX COMPUTER USEWIN32UI AUTOMATION.NET 8NODE 22+MIT

01 / WHY

Codex in the foreground?
Not first anymore.

On Windows, Codex’s Computer Use works in the foreground: OpenAI’s documentation says to keep the target application visible on the active desktop while the task runs, and it takes over the mouse and keyboard.

Many Windows applications can be driven without that. This layer keeps Codex’s Computer Use API for window inventory, screenshots and physical input, tries the background techniques first, judges the result from the application’s real state, and only uses physical input as a last resort, after a pause in your typing and under a token shared with the other agents of the PC.

Source: OpenAI’s Computer Use documentation. A skill and JavaScript modules loaded in Codex’s Node REPL: no MCP server, no network port, no token, no private Electron API.

02 / WHAT IT DOES

Background first.
Physical input last.

Every result says which technique acted, whether the foreground was taken, and every attempt.

Win32
Messages to controls (button, check box, typing, open or closed combo box, list, tab, menu), read back. Never UI Automation Invoke, which activates the window.
WinForms
UI Automation Invoke and Toggle; messages for fields, combo boxes and lists.
WPF · UWP · WinUI · Chromium
COM UI Automation without taking focus, from a separate process; a combo box is closed again after every choice.
expect
An explicit postcondition: verified, unmet or already satisfied. A postcondition that proves nothing is refused before any input.
actBatch
Up to 32 actions validated before any input, stopping at the first unverified step.
PrintWindow
Capture of the followed window without bringing it forward, even when covered.

1.4.1
No more double click: a background click sent without a read-back effect becomes sent-unconfirmed, never counted as a success nor replayed.
foregroundReturned
If physical input took the foreground, it is handed back at the end; if you moved to another window, nothing is taken from you.
Multiple windows
The session follows dialogs and returns to their owner window when they close.
Raw Input
Your presence read without a hook or content: 1.618 s without input from you before any physical input, typing stopped as soon as you take over.
Memory
A verified click can become a recipe, invalidated as soon as the application or its controls change; no typed text is kept.
Excluded windows
The agent’s own app, terminals and consoles, security and sign-in screens, password managers: refused before any input.

Real validation of October 15, 2026: Win32 actions in 105 to 159 ms, WPF in 193 to 241 ms, capture of a covered Win32 window in 19 ms, foreground unchanged and cursor still.

03 / SAFEGUARDS

The agent acts.
You stay in control.

The rules live in the tool itself, not only in the agent’s instructions.

Codex’s rules still apply

Application access and sensitive actions: the Codex Computer Use rules still apply. This layer is not a way around a denied application or action.

A refusal sends nothing

Password field, excluded application, changed control, disabled window or command: nothing is sent, and no fallback is tried.

One observation, one action

A stale observation, another window or an ambiguous control stops the action before any input.

Physical input under conditions

Only when no background technique applies or nothing took; never for a partial effect; forbidden with fallback:false.

Shared emergency stop

Checked before each action and before each character of physical typing; a refused action is never replayed.

Nothing of your screens in the metrics

Captures refused under OneDrive folders; metrics never contain UI text, screenshots, typed values or titles.

04 / ALONGSIDE OTHER AGENTS

One PC.
Several agents.

Two file-based protocols shared by the NexusWindows editions and NexusBrowser. Any agent can implement them.

Foreground token

A shared file, held only while really using the mouse or the foreground, never for a background action. Windows releases it if the agent stops; a waiting agent gives up after a few seconds and says who holds it.

Emergency stop

Ctrl+Alt+Shift+Esc blocks every desktop action for 2 minutes, for every agent; reads stay allowed. Deleting the file or changing the clock does not lift the stop.

You first

Before any physical input, the layer waits for a pause in your typing (8 s at most, otherwise nothing is sent) and stops as soon as you take the keyboard or mouse back.

05 / INSTALLATION

One installer.
The skill in Codex.

Requirements: Windows 10 or 11, the Codex desktop app with its Computer Use plugin, the .NET 8 SDK and Node 22+ for the tests. install.ps1 builds the helper, installs the skill in Codex’s folder and checks the files by hash; new Codex sessions pick it up.

pwsh -NoProfile -File install.ps1

Independent project, not affiliated with Microsoft or OpenAI.

Version: 1.4.1
System: Windows 10 or 11
Client: Codex desktop app and its Computer Use plugin
Requirements: .NET 8 SDK, Node 22+
Licence: MIT
Limits: Canvases, 3D views and controls without an accessibility provider still need screenshots and physical gestures. Skill and messages partly in French.