Codex Desktop for Windows keeps CPU active after first prompt; CPU drops immediately when Codex is killed
Environment
OS: Windows
Codex Desktop version: 26.601.1994.0
Codex app path: C:\Program Files\WindowsApps\OpenAI.Codex_26.601.1994.0_x64__2p2nqsd0c76g0\app\Codex.exe
Workspace: D:\Develop\monima
Shell: PowerShell
Steps to reproduce
- Start Codex Desktop for Windows.
- Ask a simple question/prompt.
- Wait until Codex finishes responding.
- Leave mouse and keyboard idle.
- Observe Task Manager Details tab or PowerShell process counters.
- Kill Codex/Codex helper processes.
Expected behavior
After Codex finishes responding and no task is running, CPU should return to idle. System Idle Process should go back above about 95-99% on an otherwise idle machine.
Actual behavior
After the first prompt, CPU remains elevated and System Idle Process stays below about 90%. The fan stays active. This happens even after a simple answer, not only after coding/build/test tasks.
When I kill Codex processes with:
Stop-Process -Name Codex,codex -ErrorAction SilentlyContinue
CPU drops immediately.
Observed processes
Examples seen while diagnosing:
Codex.exe
C:\Program Files\WindowsApps\OpenAI.Codex_26.601.1994.0_x64__2p2nqsd0c76g0\app\Codex.exe
codex.exe
C:\Program Files\WindowsApps\OpenAI.Codex_26.601.1994.0_x64__2p2nqsd0c76g0\app\resources\codex.exe
codex.exe
C:\Users\borut\AppData\Local\OpenAI\Codex\bin\716dda49c14d31a0\codex.exe
There were also leftover helper/child processes from earlier Codex activity:
node.exe
git.exe
conhost.exe
Killing those old helper processes did not prevent the issue from returning. After restarting Codex and asking the first question, CPU became high again and did not drop.
Additional notes
Windows Defender (MsMpEng.exe) sometimes also shows activity, likely from scanning file activity caused by Codex/Git, but the key signal is that CPU drops immediately when Codex/codex processes are killed.
This appears to be a Codex Desktop process or helper process staying active after the visible task is complete.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗