Codex Desktop sustained high CPU / overheating on macOS while project processes are idle

Summary

Codex Desktop on macOS entered a sustained high-CPU state during normal use. The laptop became hot and fans ramped up. Local project/dev processes were not the CPU source; the load was concentrated in Codex Desktop, its app-server, renderer/GPU helpers, and WindowServer.

This looks similar to existing Codex Desktop high-CPU / renderer-loop reports, but it is still occurring on a newer app build.

Environment

• App: Codex Desktop
• Codex app version: 26.513.31313
• Codex app build: 2867
• OS: macOS 26.5
• Architecture: arm64

Observed behavior

A live process snapshot showed sustained high CPU from Codex-related processes:

Codex.app main process: ~189% CPU
WindowServer: ~35% CPU
codex app-server: ~31% CPU
Codex Helper Renderer: ~9% CPU
Codex Helper GPU: ~8% CPU

A prior snapshot moments earlier showed the Codex main process already above 100% CPU, so this was not just a single instantaneous spike.

At the same time, project/dev processes such as Node/Vite and other tool helpers were essentially idle. Several MCP/tool helper processes existed, but their individual CPU use was low and they did not appear to be the primary source of heat.

Expected behavior

When Codex is not doing heavy local work, CPU should return close to idle. Since model inference is remote, normal chat/tool orchestration should not keep the laptop hot or fans running hard for a sustained period.

Actual behavior

Codex Desktop stayed in a high-energy state, causing the laptop to heat up and fans to run loudly. The CPU pattern suggests a Desktop/Electron renderer/app-server/UI state loop rather than project code, build activity, or local model inference.

Diagnostic notes

• The Codex main process stayed high across repeated snapshots rather than appearing as a one-off spike.
• WindowServer was also elevated while Codex renderer/GPU helpers were active, which points toward rendering/compositing work.
• Local dev tooling was not the primary CPU source during the observation window.
• MCP/tool helper processes were present, but their individual CPU usage was low.
• The pattern is consistent with a Desktop/Electron renderer or app-server loop.

Safe debug data available

I can provide additional redacted diagnostics if useful, such as:

• process-name-only CPU snapshots
• app version/build and OS version
• a redacted macOS sample of the high-CPU Codex process
• a redacted Activity Monitor sample

I am not attaching raw application logs or raw process arguments here because they may contain local paths, session IDs, repository names, or other user-specific context.

Possibly related issues

• #20435
• #18567
• #20735
• #18467
• #21326