Codex/ChatGPT VS Code extension triggers fan ramp-up and software GPU/WebGL fallback on Linux with GPT-5.5

[OmerGolcuk]

What version of the IDE extension are you using?

openai.chatgpt@26.5422.21459

What subscription do you have?

ChatGPT Plus

Which IDE are you using?

Visual Studio Code 1.117.0

What platform is your computer?

Linux x86_64 Ubuntu 24.04.4 LTS Kernel: 6.8.0-110-generic Machine: Lenovo Legion 5 15ACH6H CPU: AMD Ryzen 7 5800H GPU: NVIDIA RTX 3060 Mobile / Max-Q VS Code install: Snap package, path includes /snap/code/235/

What issue are you seeing?

The OpenAI Codex/ChatGPT VS Code extension causes sustained fan ramp-up / high CPU behavior on Linux when the extension panel is opened with GPT-5.5.

I performed a controlled experiment:

• VS Code open, Codex/ChatGPT extension panel closed: fans stop / system returns to normal.
• Codex/ChatGPT extension panel opened: fans immediately ramp up and remain high.
• Closing the Codex/ChatGPT panel again: fans stop.
• Running VS Code with extensions disabled avoids the issue.

This appears specific to the OpenAI VS Code extension path, not normal VS Code editing.

code --status also shows that VS Code appears to be falling back to software GPU/WebGL rendering:

• GPU0: Google ANGLE / SwiftShader
• gpu_compositing: disabled_software
• rasterization: disabled_software
• webgl: unavailable_software
• webgl2: unavailable_software

The GPU log contains repeated EGL/OpenGL initialization errors and WebGL/software-rendering warnings, including:

• Could not create a backing OpenGL context.
• eglInitialize OpenGL failed with error EGL_NOT_INITIALIZED.
• Initialization of all EGL display types failed.
• Automatic fallback to software WebGL has been deprecated.
• GPU stall due to ReadPixels.

The OpenAI extension process is visible in code --status as:

openai.chatgpt-26.5422.21459-linux-x64/.../codex app-server --analytics-default-enabled

This may indicate that opening the Codex/ChatGPT extension UI triggers a renderer/WebGL/software-GPU path that causes sustained CPU/fan usage on Linux.

What steps can reproduce the bug?

1. Start VS Code on Linux.
2. Open a normal workspace.
3. Open the OpenAI Codex/ChatGPT VS Code extension panel.
4. Use/select GPT-5.5.
5. Observe that fans immediately ramp up and remain high.
6. Close the Codex/ChatGPT extension panel.
7. Observe that fans stop / CPU behavior returns to normal.
8. Run VS Code with extensions disabled.
9. Observe that the issue is avoided.

Controlled experiments already tried:

• Closed Codex/ChatGPT panel: fans stop.
• Opened Codex/ChatGPT panel: fans ramp up again.
• Ran VS Code with extensions disabled: issue avoided.
• Cleared VS Code Electron/GPU caches.
• Temporarily moved/restored ~/.codex.
• Temporarily moved/restored VS Code workspaceStorage.
• Confirmed the issue correlates with the extension being open.

What is the expected behavior?

Opening the Codex/ChatGPT VS Code extension panel should not cause sustained fan noise, high CPU usage, or continuous GPU/software-rendering activity when the model is not actively generating.

When idle, the extension should remain mostly inactive. CPU/GPU usage should return to normal after a request finishes, and simply keeping the Codex/ChatGPT panel open should not make the fans ramp up continuously.

Additional information

I am consistent codex user, I want it to be solved ASAP. I am experiencing this issue all of the versions after 0.5.80 for codex extension of VS code.

[github-actions]

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

• [VS Code extension][Windows] High CPU regression after 26.325.31654 #19746
• High idle CPU (25%) in Codex IDE extension on Positron: Electron window renderer consumes CPU after opening Codex panel #19731
• High CPU in second VS Code project window when Codex panel is opened, while first project window is already open #19091
• Renderer process stuck in infinite microtask loop, consuming 100–180% CPU and disconnecting Remote SSH sessions #19097

Powered by Codex Action

[igorkuznecov]

I put together a small public workaround repo for this stable-metadata non-git loop:

https://github.com/igorkuznecov/patch-codex-extension

What it does:

• patches the extension host bundle so the non-git stable-metadata path returns null instead of a hard error
• patches the webview asset so failed stable-metadata requests resolve to null instead of feeding the loop
• creates .stable-metadata.bak backups
• includes a restore script
• uses MIT license

Current version of the script patches all installed openai.chatgpt-* versions it finds under ~/.cursor/extensions or ~/.vscode/extensions, which turned out to matter when multiple extension versions are present.

Tested locally on Cursor extension builds:

• openai.chatgpt-26.422.62136-darwin-arm64
• openai.chatgpt-26.422.71525-darwin-arm64

The patch script is here:

• patch-codex-stable-metadata.sh

Usage:

bash patch-codex-stable-metadata.sh ~/.cursor/extensions
# or
bash patch-codex-stable-metadata.sh ~/.vscode/extensions

Then reload the editor window.

This is only a local workaround, not an upstream fix, but it stopped the renderer/log loop in my repro.