macOS: Codex Desktop has high local GPU/WindowServer energy use during SSH-remote tasks

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

Codex Desktop / ChatGPT app: 26.721.41059

What subscription do you have?

ChatGPT Pro (20× usage plan).

What platform is your computer?

Local client:

• macOS 26.5.2 (25F84)
• MacBook Pro (Mac17,9), Apple M5 Pro, 48 GB memory

Remote execution host:

• Mac mini, connected through Codex's SSH remote environment
• The Codex task execution is remote; it continues when the local MacBook app is closed.

What issue are you seeing?

Codex Desktop creates substantial local heat and energy use on the MacBook even when the actual task is running entirely on a remote Mac mini via the built-in SSH connection.

This is not the remote Mac mini doing local work: closing the MacBook's Codex app does not stop the remote task. However, while the app is open on the MacBook, local UI/GPU-related processes remain hot enough for noticeable sustained heating.

I disabled the “Transparent sidebar” appearance setting and minimized the Codex window. The issue persisted.

A two-second sampled power view after minimizing showed the following local contributors:

• WindowServer: approximately 44–45% CPU, power score approximately 47–49
• Codex (Service) GPU process: approximately 21–27% CPU, power score approximately 22–29
• Codex (Renderer): approximately 15–25% CPU, power score approximately 16–25

The Codex app-server process itself was comparatively low at the time of the earlier sampling. This points to local desktop rendering/compositing and GPU activity, rather than the remote task execution, as the primary heat source.

What steps can reproduce the bug?

1. On a MacBook, open Codex Desktop.
2. Connect a Codex task to a remote Mac mini using the built-in SSH remote environment.
3. Run the task remotely and verify it continues when the local Codex app is closed.
4. Reopen Codex Desktop; disable Transparent sidebar and minimize the window.
5. Use Activity Monitor or a process power sample to inspect local processes.
6. Observe sustained elevated CPU and energy impact from WindowServer, Codex GPU service, and Codex renderer, with noticeable MacBook heating.

What is the expected behavior?

When task execution is entirely remote through SSH, the local Codex client should have low idle energy use. In particular, minimizing the app and disabling transparency should substantially reduce local renderer/GPU/WindowServer work and heat.

Related reports

• #32024: visible Codex window drives high WindowServer CPU on macOS
• #24185: sustained macOS CPU/GPU/WindowServer load and heating
• #34415: transparent sidebar can cause excessive GPU use on macOS
• #23072: sustained macOS desktop CPU/overheating while project processes are idle
• #30606: a separate SSH-remote workflow report, on Windows, showing that remote execution can still have client-side effects

This report is not intended as a duplicate of any one of the above. Its distinguishing condition is that the task execution is fully remote through Codex SSH, continues after the local app is closed, and yet the local client has substantial GPU/renderer/WindowServer energy use even after transparency is disabled and the app is minimized.

Update: process-level evidence during a heating episode

A later live sample, after the app had been running for roughly 37 minutes, captured the following local Codex activity while the remote task remained on the SSH host:

• One Codex (Renderer) process: approximately 66% CPU, approximately 1.5 GB resident memory.
• Codex (Service) GPU process: approximately 23% CPU.
• WindowServer: approximately 44% CPU.

A short stack sample of the busy renderer showed its main thread executing inside the Chromium/V8 JavaScript runtime. The GPU service was also active. This is consistent with a local renderer/GPU update loop feeding WindowServer, rather than remote task execution being the source of the MacBook load.

A contemporaneous Chrome workload was also independently high (approximately 84% CPU), so it contributes to total system heat; it does not explain the separately elevated Codex renderer, Codex GPU service, and WindowServer activity.

A local macOS system diagnostic was captured during the investigation and contains process-level power samples. It is not attached publicly because it may include sensitive local metadata, but sanitized excerpts can be provided to maintainers on request.