[Windows][26.721.11231.0] App exits ~2.4s after launch after two renderer 0xC0000005 crashes

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

• #34923

Powered by Codex Action

I can reproduce what appears to be the same startup failure on a newer Codex Windows build.

Environment

• Codex MSIX: OpenAI.Codex 26.818.5229.0 (x64)
• Windows: 10.0.26200 x64
• Bundled app-server: 0.149.0-alpha.4.1
• Chromium reported by WinDbg: 151.0.7922.170
• AMD Ryzen 7 5800H
• AMD integrated graphics + NVIDIA RTX 3060 Laptop GPU

Behavior

Normal AppX activation succeeds, but the app exits within approximately 1–2 seconds without displaying a usable window.

The desktop log shows successful backend initialization:

[AppServerConnection] initialize_handshake_result ... outcome=success
[AppServerConnection] Codex CLI initialized

followed immediately by:

[window-manager] Failed to load bundle at
errorCode=ERR_FAILED
errorMessage="ERR_FAILED (-2) loading 'app://-/index.html'"

I verified separately with asar list that \webview\index.html is present inside the installed app\resources\app.asar, so this does not appear to be caused by a missing packaged HTML file.

Crashpad

Every reproduced launch creates:

• one browser-process dump (ptype: browser)
• multiple renderer-process dumps (ptype: renderer)

WinDbg analysis of a renderer dump:

ExceptionCode: c0000005 (Access violation)
AV.Type: Execute
Attempt to execute non-executable address

PROCESS_NAME: ChatGPT.exe
MODULE_NAME: chrome
IMAGE_NAME: chrome.dll

FAILURE_BUCKET_ID:
SOFTWARE_NX_FAULT_INVALID_POINTER_EXECUTE_c0000005_chrome.dll!Unloaded

The renderer stack reaches RtlpFlsDataCleanup, LdrShutdownProcess, and RtlExitUserProcess, so this particular crash may occur secondarily during process shutdown.

The browser-process dump shows:

ExceptionCode: 80000003 (Break instruction exception)

PROCESS_NAME: ChatGPT.exe
MODULE_NAME: chrome
IMAGE_NAME: chrome.dll

FAILURE_BUCKET_ID:
BREAKPOINT_80000003_chrome.dll!Unknown

IMAGE_VERSION:
151.0.7922.170

FAILURE_ID_HASH:
{be0d4058-7058-b7fb-5291-62cb1f25d882}

Interestingly, that FAILURE_ID_HASH is identical to the Chromium breakpoint reported in #38310, although the trigger in that issue is different (typing / after the app has successfully launched).

Troubleshooting already attempted

• Terminate / Repair / Reset
• Complete uninstall and reinstall
• Get-AppxPackage reports package status Ok
• Verified app.asar exists (~287 MB)
• Verified webview/index.html exists inside app.asar
• Reproduced through explicit AppX activation:

shell:AppsFolder\OpenAI.Codex_2p2nqsd0c76g0!App

• Tested ELECTRON_DISABLE_GPU=1
• Tested ELECTRON_DISABLE_GPU_COMPOSITING=1
• Neither GPU-related test changed the behavior
• Brave, Discord, Spotify, Steam and a 3D game render normally on the same machine

I also observed the issue around the time of an AMD graphics-driver update, but I currently have no evidence that the AMD driver is the cause. No AMD or NVIDIA driver module appears in the faulting WinDbg stacks, and disabling Electron GPU acceleration did not change the behavior.

I have retained the original Codex logs and Crashpad dumps locally and can provide additional debugger output if useful. I have not attached the raw dumps publicly because they may contain process memory.

One potentially relevant timeline detail: the Codex/ChatGPT Windows app worked normally immediately after I first installed it.

I later installed/updated AMD Adrenalin graphics drivers. The already-running app continued to work before reboot, although I did not explicitly close and relaunch it during that period. After rebooting Windows, the app began consistently failing during startup as described above.

Therefore, the reboot and/or graphics-driver update correlate with the first observed occurrence of the issue, but I cannot establish causation. Other GPU-accelerated applications and a 3D game continue to work normally, the Electron GPU-disable environment-variable tests did not change the Codex behavior, and the analyzed crash stacks do not directly implicate an AMD or NVIDIA driver.

Update: I was able to isolate a strong correlation with the AMD Radeon display driver.

The app originally worked normally on this machine. The startup failure was first observed after an AMD Adrenalin/display-driver update followed by a Windows reboot.

I performed a controlled driver rollback:

Failing configuration

• AMD Radeon(TM) Graphics: 31.0.21925.1001 (2026-05-20)
• NVIDIA RTX 3060 Laptop GPU: 32.0.16.1088
• Codex: 26.818.5229.0
• Windows: 10.0.26200 x64

With this configuration, every Codex launch reproducibly resulted in the previously described ERR_FAILED (-2) loading 'app://-/index.html', followed by the browser/renderer Crashpad failures and application exit.

I then used Device Manager → Roll Back Driver for the AMD Radeon adapter only.

Working configuration

• AMD Radeon(TM) Graphics: 27.20.14032.8 (2020-12-22)
• NVIDIA RTX 3060 Laptop GPU: 32.0.16.1088 (unchanged)
• Codex and Windows versions unchanged

After rebooting with the older AMD driver, Codex opened normally.

I then:

1. closed Codex completely,
2. launched it again successfully,
3. rebooted Windows,
4. launched Codex again successfully.

I am currently using the Codex desktop app successfully from this same installation after the AMD driver rollback.

So, on this machine, changing only the AMD Radeon display driver from 31.0.21925.1001 to 27.20.14032.8 changes the behavior from consistently reproducible startup failure to consistently successful startup.

This does not establish whether the root cause is in the AMD driver itself, Chromium/Electron, or an interaction between them, but it provides a reproducible workaround and strongly suggests the AMD display-driver version is involved.

One potentially interesting comparison: the original report also lists AMD Radeon Graphics (driver 32.0.21037.1004), although it is a different driver version from mine.