Windows desktop: browser-process crash loop persists across 26.820.7780 → 26.825.3734

Description

The native Codex Windows desktop app repeatedly terminates without a visible error. Crashpad records each event as a browser/main-process crash, but the desktop log ends abruptly without a fatal exception and no minidump was retained.

Environment

• Windows 11 x64
• Codex package versions observed: 26.820.7780 and 26.825.3734
• Current package: 26.825.3734
• Native Windows/PowerShell mode

Reproduction / timeline

Local timezone: AEST (UTC+10), 28/08/2026.

1. Launch Codex normally.
2. Use the app or leave it running.
3. The entire app closes abruptly.

Crashpad sidecars were created at:

• 14:22:27 — 26.820.7780, about 6 seconds after launch
• 14:28:43 — 26.820.7780, about 2 minutes 30 seconds after launch
• Microsoft Store then updated Codex to 26.825.3734
• 14:48:44 — 26.825.3734, about 18 minutes 39 seconds after launch
• 14:49:02 — 26.825.3734, about 6 seconds after launch
• 14:49:11 — 26.825.3734, about 5 seconds after launch

There were eight earlier browser-process Crashpad events on 24/08 and 26/08. The crash loop therefore persisted across the Store update.

Every retained sidecar contains the equivalent of:

{"capture_kind":"crash","osarch":"x86_64","ptype":"browser"}

Additional diagnostics

During one affected startup, the bundled primary runtime update ran for about 83 seconds and failed during activation:

primary_runtime_install_failed
failureStage=activate_runtime
bundleVersion=26.826.12353
trigger=startup_missing
retryable=false
permission_denied
EPERM while renaming %USERPROFILE%\.cache\codex-runtimes\codex-primary-runtime

Two incomplete codex-runtime-install-* staging directories remained. After updating to 26.825.3734, the log also reported bundled_plugin_status_unknown with The system cannot find the path specified. (os error 3) for bundled components.

Some longer-running crashes ended immediately after three GitHub sidebar warnings saying no git remotes found. However, the same warnings occur in sessions that remain running, so this appears correlated rather than proven causal.

Windows event history showed no corresponding:

• Application Error / WER event
• GPU/display driver reset
• resource-exhaustion or out-of-memory event
• Defender action
• scheduled-task termination

One separate Store deployment did use ForceTargetApplicationShutdownOption; that update termination is distinguishable from the Crashpad-recorded browser-process crashes.

Expected behavior

Codex remains running, or records a usable fatal exception/dump if the main process fails.

Actual behavior

The complete desktop app closes abruptly and only a small Crashpad sidecar remains.

Further evidence

Windows LocalDumps is now enabled for full ChatGPT.exe dumps with a ten-file retention limit. I can attach the next captured dump or provide additional logs if requested.

Start by reproducing the crash on Windows 11 x64 and reviewing the Crashpad sidecars plus startup logs around primary_runtime_install_failed and bundled_plugin_status_unknown. Use the enabled LocalDumps capture to establish a usable failure record; done means the app remains running or produces a usable fatal exception or dump.