Computer Use runtime not exposed on Windows + /feedback does not open
What version of the Codex App are you using (From “About Codex” dialog)?
Codex Desktop 26.611.8604.0 Session cli_version: 0.140.0-alpha.19
What subscription do you have?
ChatGPT pro
What platform is your computer?
Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
Summary
Computer Use appears installed and enabled, but the visual runtime/tool bridge is not exposed to the model in Codex Desktop on Windows. When I mention @Computer / @Ordinateur, Codex detects the Computer Use plugin/skill, but cannot use visual Computer Use APIs. It falls back to shell commands such as Start-Process, notepad.exe, or firefox instead of controlling the UI visually.
Also, /feedback does not open the feedback dialog in this thread.
Environment
- Windows
- Codex Desktop package visible locally:
OpenAI.Codex_26.611.8604.0_x64__2p2nqsd0c76g0 - Session metadata shows
cli_version: 0.140.0-alpha.19 - Computer Use plugin enabled:
computer-use@openai-bundled - Computer Use plugin version visible locally:
26.611.62324
Expected behavior
When I mention @Computer / @Ordinateur, Codex should expose and use the Computer Use visual runtime so it can list apps, capture windows, click, type, and inspect UI state.
Actual behavior
The Computer Use plugin is detected as a skill, but the expected runtime bridge is not available to the model. The agent reports that the node_repl/js bridge required by computer-use-client.mjs is not exposed. It then falls back to shell commands to open applications.
Examples:
- Asked to open Firefox with
@Ordinateur - Asked to open Notepad and type text
- Asked to use WhatsApp visually
In each case, Codex could only launch apps via shell or refused to continue because it could not visually inspect the app.
Checked
computer-use@openai-bundledis enabled in~/.codex/config.tomlcomputer-use-client.mjsexists locally- No
requirements.tomldisablingcomputer_usewas found - No local Computer Use deny config was found
- Restarting the PC did not fix it
- Reinstalling/reactivating the plugin did not fix it
/feedbacktyped in Codex Desktop does not open the feedback dialogcodex doctor --summary --ascii --no-colorfails withAccès refusé
Impact
Computer Use cannot be used for visual desktop automation on Windows, even though the plugin appears installed and enabled.
What steps can reproduce the bug?
- On Windows, install/enable the Computer Use plugin in Codex Desktop.
- Start a new Codex Desktop thread.
- Mention
@Computer/@Ordinateurand ask it to open or visually control an app, for example:
@Ordinateur ouvre Firefox@Ordinateur ouvre Notepad et écris bonjour
- Observe that Codex reads the Computer Use skill/plugin instructions but cannot access the expected visual runtime bridge.
- Codex falls back to shell commands such as
Start-Process,firefox, ornotepad.exe, or refuses to continue because it cannot visually inspect the target app. - Type
/feedbackin the same Codex Desktop thread. - Observe that the feedback dialog does not open.
- Run
codex doctor --summary --ascii --no-colorfrom PowerShell. - Observe that it fails with
Accès refusé.
What is the expected behavior?
When @Computer / @Ordinateur is mentioned in Codex Desktop, Codex should expose the Computer Use visual runtime tools to the model and use them to operate desktop apps visually.
Expected capabilities include:
- list available apps/windows
- capture target window state
- click/type/scroll via Computer Use
- inspect UI state before acting
- ask for app permissions when needed
Also, /feedback should open the feedback dialog and codex doctor should run or provide an actionable diagnostic error.
Additional information
This appears to be different from a simple app permission issue: the plugin is detected and the Computer Use skill is available, but the underlying runtime/tool bridge is not exposed to the model.
The issue persisted after:
- restarting Windows
- restarting Codex Desktop
- reinstalling/reactivating the Computer Use plugin
The thread can still run shell commands, so the failure seems specific to Computer Use runtime/tool injection rather than general local execution.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗