Codex Desktop on Windows burns particular CPU core while idle if the window is visible.

Open 💬 3 comments Opened Jun 25, 2026 by 07354d88e5

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using (From “About Codex” dialog)?

26.616.10790.0

What subscription do you have?

plus, previously pro

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Total CPU percentage may look modest at first glance because the machine has 22 logical processors.
However, the main Codex.exe process continuously consumes a large fraction of one CPU core while the window is visible and idle.
In the captured log, visible idle usage averaged ~3.06% total CPU, which is about 67% of one logical core on this machine.
p95 was ~4.83% total CPU, about 106% of one logical core, with peaks around ~6.53% total CPU, about 144% of one logical core.

codex-cpu.csv

What steps can reproduce the bug?

This is sustained, not just a short spike. The laptop fans spin up loudly while the Codex window is visible, even when I am only reading existing text and not running any agent/tool work. After minimizing the Codex window, CPU usage drops close to zero and the fans calm down.
The load appears to come from the main Codex.exe process, not from codex app-server, node_repl, renderer, or GPU process.

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 2 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#29281

Powered by Codex Action

alimohajer · 1 month ago

Current-version datapoint from OpenAI.Codex_26.707.9564.0 on Windows 11 Pro for Workstations 10.0.26200 (Intel i9-9900, 16 logical CPUs, 64 GB RAM).

Two process-time samples isolated the visible main ChatGPT.exe window process:

12 seconds: 12.0 CPU-seconds = 100% of one logical core
after pruning 13 idle MCP/REPL helpers, 20 seconds: 24.46 CPU-seconds = 122.3% of one logical core

In the first sample, the GPU child used 16.7% of one core and the renderer used 8.3%; codex app-server, Node MCP servers, and node_repl each used 0 CPU in the sampling window. Removing the idle helpers reduced working set but did not reduce main-window CPU.

Important scope note: these samples were taken while a visible Codex task was active and waiting on local tool execution, not while truly idle or minimized. This is therefore a related current-build datapoint, not an exact idle reproduction. The process attribution is consistent with this issue: the main desktop window, rather than app-server or Node helpers, was the sustained CPU source.

googleoryx · 13 days ago

Still reproducible after restarting the app and installing the pending update to 26.810.6296.0.

Fresh post-update diagnostics:

Windows 11 x64, 22 logical processors.
The main ChatGPT.exe process averaged 1.388 logical cores over 10.02 seconds (6.31% total CPU).
The other eight ChatGPT.exe processes each measured effectively 0 CPU during the same interval.
%USERPROFILE%\.codex\plugins\cache\openai-bundled\chrome\latest is still being continuously deleted/recreated.
Across 40 samples taken every 500 ms (20.5 seconds), the junction had 28 distinct CreationTime values — about 1.37 recreations/second.
Its target remained unchanged throughout: %USERPROFILE%\.codex\plugins\cache\openai-bundled\chrome\26.810.50856.

This is the same signature seen before the update on 26.810.4967.0, when the main process also produced roughly 0.5–1.4 GB/s of read I/O. Restarting and updating therefore did not clear or fix the loop.

The continuous junction recreation is directly observed. A stale-junction or Windows path-normalization check still appears to be a likely cause, but that cause is an inference.