Windows: Codex repeatedly deploys cua_node runtime under non-ASCII user path, causing huge disk writes

Open 💬 6 comments Opened Jun 13, 2026 by young2857
💡 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)?

26.609.4994.0

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

When Codex starts on Windows, it repeatedly writes a large number of files into:

C:\Users\钖涚殦鍏僜\AppData\Local\OpenAI\Codex\runtimes\cua_node

This causes severe system lag, high disk I/O, and rapid C: drive usage growth.

The Windows user profile path appears to contain corrupted/non-ASCII characters:

C:\Users\钖涚殦鍏僜

This looks similar to issue #28079, where Codex repeatedly created cua_node staging/runtime directories under a non-ASCII Windows user profile path.

What steps can reproduce the bug?

Reproduction steps
1.Use Windows 11 with Codex installed from the Microsoft Store.
2.Use Codex version: 26.609.4994.0
3.Use a Windows account whose user profile path contains corrupted/non-ASCII characters, for example:
C:\Users\钖涚殦鍏僜
4.Start Codex.
5.Observe the following directory:
%LOCALAPPDATA%\OpenAI\Codex\runtimes\cua_node
6.Start Codex multiple times.
7.Notice that Codex repeatedly writes a large number of runtime files into this directory.
8.Observe severe system lag and increasing C: drive usage.

What is the expected behavior?

Codex should deploy the bundled cua_node runtime once under:

%LOCALAPPDATA%\OpenAI\Codex\runtimes\cua_node

After successful deployment, Codex should reuse the existing runtime on later launches.

It should not repeatedly recreate or rewrite the runtime files on every startup, and it should not cause excessive disk usage or system-wide lag.

Codex should also correctly handle Windows user profile paths containing Chinese/CJK/non-ASCII characters. If runtime deployment fails, it should fail gracefully with a clear error message instead of repeatedly writing large runtime bundles.

Additional information

Actual behavior

Every time Codex starts, it writes a large number of files under:

%LOCALAPPDATA%\OpenAI\Codex\runtimes\cua_node

The files appear to be a full local Node-based runtime bundle, including items such as:

node.exe
node_modules
Playwright-related packages
@oai/sky
codex-computer-use.exe
native .node files
DLL files
WASM files
PDF/OCR/image-processing dependencies

This does not look like normal small logging or cache activity. It looks like the bundled cua_node runtime deployment is being retried repeatedly instead of completing once.

The result is:

Severe Windows lag after starting Codex
Very high disk activity
Rapid C: drive space consumption
Repeated writing of thousands of runtime files
Codex becoming difficult or impossible to use normally

Evidence / observations

The affected directory is:

C:\Users\钖涚殦鍏僜\AppData\Local\OpenAI\Codex\runtimes\cua_node

A compressed sample of the generated runtime folder contains thousands of files and appears to be the bundled cua_node runtime. It includes Node.js, node_modules, Playwright-related components, @oai/sky, native binaries, WASM files, and codex-computer-use.exe.

This suggests Codex is repeatedly deploying the local Computer Use / CUA Node runtime.

The issue seems related to non-ASCII or incorrectly encoded Windows user profile paths. The user folder name appears as mojibake/corrupted characters:

钖涚殦鍏僜

What I tried
Closed Codex
Deleted the generated cua_node runtime/cache directory
Restarted Codex
Observed that Codex starts writing the same type of runtime files again
Confirmed that the files are not just small logs, but a large runtime bundle

Deleting the directory only temporarily frees disk space. The issue returns when Codex starts again.

Suspected cause

This may be a Windows path encoding issue during bundled runtime deployment.

The runtime deployer appears to write to a path derived from %LOCALAPPDATA%:

C:\Users\钖涚殦鍏僜\AppData\Local\OpenAI\Codex\runtimes\cua_node

If the deployment code does not handle UTF-16 / Unicode Windows paths correctly, or if it misinterprets the user profile path through an ANSI/UTF-8 conversion boundary, the runtime deployment may fail or never be marked as complete.

As a result, Codex may repeatedly attempt to stage/deploy the same cua_node runtime on every launch.

Impact

This is a serious usability issue for affected Windows users:

Codex causes severe system lag
C: drive space is consumed rapidly
The app repeatedly writes thousands of files
Deleting the cache is only temporary
Users with Chinese/CJK/non-ASCII Windows usernames may be affected
The app becomes impractical to use on the affected Windows account

Requested fix

Please make the Windows cua_node runtime deployment path handling fully Unicode-safe for %LOCALAPPDATA% and other user-profile-derived paths.

If runtime deployment fails, Codex should:

Stop retrying endlessly
Avoid repeatedly writing full runtime bundles
Surface a clear visible error message
Log the exact deployment failure
Provide a supported way to relocate the runtime directory to an ASCII-only path, such as an environment variable or configuration option

Attachments

I can provide a compressed sample of the generated cua_node runtime/staging files if useful.

View original on GitHub ↗

6 Comments

github-actions[bot] contributor · 2 months ago

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

  • #28079
  • #27409
  • #28017
  • #27822
  • #27506

Powered by Codex Action

lizezhen13 · 2 months ago

同样遇到了这个问题,请问怎么解决

MasterOK1975 · 2 months ago

I’m seeing the same issue on Windows with a Cyrillic / non-ASCII Windows user profile path.

In-app feedback ID:
019e7d35-863a-7563-87e5-b0353185ca70

My case:

  • %LOCALAPPDATA%\OpenAI\Codex\runtimes\cua_node
  • many .staging-* folders are created repeatedly
  • each staging folder is about 272 MB
  • runtime folder reached 26.5 GB / 27.1 GB on disk
  • 341,600 files and 65,501 folders
  • the growth happens when opening chats that were not yet opened in the current Codex session
  • reopening already loaded chats does not increase disk usage

This looks related to the same non-ASCII / mojibake Windows profile path issue.

Related:
#28258
#28691

Z1WZ05 · 2 months ago

how to solve it?

MasterOK1975 · 2 months ago

For me it seems to be resolved after updating Codex Desktop.

Current version:
26.616.5445.0

Before the update, Codex kept creating many .staging-* folders under:

%LOCALAPPDATA%\OpenAI\Codex\runtimes\cua_node

Each folder was around 272 MB and the runtime directory grew to tens of GB.

After updating, I tested opening/switching multiple chats. Disk usage may temporarily increase for a short time, but then it returns back to normal and .staging-* folders are cleaned up. Current .staging-* count is 0.

So I recommend:

  1. Update Codex Desktop to the latest version.
  2. Fully close and restart Codex.
  3. Delete old leftover .staging-* folders only after Codex is fully closed.
  4. Check again:

Get-ChildItem "$env:LOCALAPPDATA\OpenAI\Codex\runtimes\cua_node" -Force -Directory -Filter ".staging-*" | Measure-Object

If the issue still happens, a temporary workaround is to redirect the broken/mojibake cua_node path with a junction to an ASCII-only path, but in my case the real fix seems to be the Codex update.

zemeng5208 · 13 days ago

Disclosure: I maintain WinBridge Recovery: https://github.com/zemeng5208/winbridge-recovery

This repeated cua_node deployment pattern is closely related to the runtime-recovery cases WinBridge is intended to diagnose. The important distinction is whether Codex has (a) a valid completed runtime plus abandoned/repeated staging trees, or (b) no coherent completed runtime at all. Blindly deleting the whole cua_node directory can make the next launch redeploy everything again.

WinBridge compares the local runtime generation with resources from the officially installed Codex package, backs up the current state, and can restore a coherent user-local runtime when the source is valid. It is also designed to avoid treating every leftover staging directory as the authoritative runtime.

The CJK/non-ASCII path bug itself is upstream: if the Codex updater keeps failing its Unicode path conversion or keeps redeploying on every startup, a local recovery cannot promise that the next app update/reconcile will not trigger it again. A useful WinBridge result here may therefore be 🟡 “runtime recovered / staging normalized, but upstream redeployment recurs,” not necessarily a permanent fix.

If anyone with rapidly growing .staging-* / cua_node disk usage tries it, please report the outcome at https://github.com/zemeng5208/winbridge-recovery/issues/4. That would help distinguish the cases WinBridge can safely recover from the ones that require the updater itself to be fixed.