App Crashes When Opening Settings

Update: Settings still crashes after reinstall and full local reset

This remains 100% reproducible.

Environment

• macOS 26.5.2
• Mac14,7
• ChatGPT desktop app version 26.707.72221
• Build 5307
• ChatGPT Plus account

Reproduction

1. Launch the ChatGPT desktop app.
2. Open Settings from the menu or press Cmd+,.
3. The entire application immediately crashes.

The rest of the app is currently usable, but Settings cannot be opened.

Troubleshooting already attempted

• Completely uninstalled and reinstalled the app.
• Cleared the ChatGPT/Codex caches and preference files.
• Reset the local ChatGPT/Codex application profile.
• Reset the plugin cache.
• Verified the app’s code signature.
• Verified that the local Codex SQLite database is healthy.
• Updated macOS.
• Tested launching with V8 optimization disabled, although the app did not appear to honor the flag.
• Temporarily isolated individual Claude Code configuration and session files, restoring each afterward.
• Confirmed that Claude Cowork data was not involved or modified.

None of these steps resolved the Settings crash. The full profile reset also temporarily removed my local task history, which I had to restore from a backup.

Crash evidence

The macOS diagnostic reports consistently show:

• Process: ChatGPT
• Exception: EXC_BREAKPOINT (SIGTRAP)
• Crashing thread: main thread, CrBrowserMain
• Recurring frames involving V8 and Node, including:
• v8::internal::OptimizingCompileTaskExecutor::RunCompilationJob
• v8::String::NewFromUtf8
• node::StringDecoder::DecodeData
• node::fs::AfterInteger

Immediately before several crashes, the application log records successful external-agent detection:

[external-agent-import] external_agent_import_provider_stage_finished
providerId=claude-code
stage=detect
status=success
claudeCoworkItemCount=0

Inspection of the packaged application indicates that opening Settings enables external-agent detection automatically. I could not find a supported preference or feature flag to disable that detection.

Current assessment

This appears to be a native desktop-app bug triggered while Settings processes a nonempty Claude Code detection result. It does not appear to be fixed by reinstalling the application or resetting normal ChatGPT/Codex state.

Could the team please provide one of the following?

• A patched build
• A supported way to disable external-agent detection
• A safe workaround that does not require removing or disabling Claude Code/Cowork data

The application is not realistically usable for me while opening Settings consistently terminates the entire process.

I can reproduce this on the same app version (26.707.72221, build 5307) on macOS 26.3 / Apple Silicon. I inspected six .ips reports from July 14–15; all show the same native main-process EXC_BREAKPOINT (SIGTRAP) signature on CrBrowserMain, with recurring frames in v8::String::NewFromUtf8, node::StringDecoder::DecodeData, and node::fs::AfterInteger.

The bundled codex doctor --json reports installation, config, auth, MCP config, local state databases, networking, and updates as healthy.

Immediately before multiple crashes, desktop logs show successful account/read and config/read, followed by successful external-agent-import detection for claude-code; the process dies before Claude Cowork detection reports completion. This machine has a large Cowork session history with many JSONL files, including files over 50 MB, so the Settings page's automatic cross-agent import scan appears correlated with the crash.

Full sanitized diagnostics and suggested bounds/streaming fixes: #33283

I have the same problem in VSCode.

Codex version 26.715.31925.
VSCode version 1.129.1.
OS: MacOS Sequoia 15.6

When I try to open settings via "VSCode Sidebar -> Gear icon -> Codex settings" it waits a bit and then gives me this pop-up:

The window is not responding
You can reopen or close the window or keep waiitng.

Same crash, with additional diagnosis: the OOM is in external-agent-import detection, and there's no way to disable it.

Reproduced deterministically on macOS (Apple Silicon), app build 26.715.31925, Codex CLI 0.145.0-alpha.18. Settings crashes every launch. Running the app from the terminal shows it's a V8 heap OOM, and the last log line before death is always the external-agent import scan:

[external-agent-import] external_agent_import_provider_stage_finished
  providerId=claude-code stage=detect status=success
  claudeCodeItemCount=2 claudeCoworkItemCount=0

<--- Last few GCs --->
[...] Mark-Compact (reduce) 3848.8 -> 3848.8 MB [...] last resort
ERROR:owl/common/node_bindings.cc:109] OOM error in V8:
  CALL_AND_RETRY_LAST Allocation failed - JavaScript heap out of memory
zsh: trace trap  /Applications/ChatGPT.app/Contents/MacOS/ChatGPT

Memory climbs to the ~3.85GB V8 ceiling and aborts, immediately after the claude-code provider's detect stage.

Things I ruled out by testing (crash persists through all of them):

• Removed all large Codex session logs and the oversized ~/.codex rollout files.
• Renamed ~/.codex entirely (blank profile) — still crashed.
• Renamed ~/.claude and ~/.claude.json — still crashed.
• Moved every Cowork/Claude Code session .jsonl over 40MB aside (including 261MB and multiple 50–80MB files) — still crashed.
• Moved the multi-GB Claude VM disk images aside — still crashed.
• Relaunched with --js-flags="--max-old-space-size=8192" — still crashed (just took longer to hit the higher ceiling).

Because it still OOMs even after the large files are gone, the problem isn't one giant file — the detection pass itself is unbounded (appears to read/accumulate matched content into memory rather than streaming or capping it).

Second, separate issue: there is no way to turn this off. The import feature is documented as an opt-in manual action under Settings → General → Import other agent setup (https://developers.openai.com/codex/import), but the detection pass runs automatically on Settings open. There's no documented config.toml key, environment variable, or launch flag to disable external-agent detection — so the only control lives inside the one screen the bug prevents from opening. A documented off-switch (e.g. a config flag to skip external-agent detection) would unblock every affected user immediately, independent of the memory fix.

Suggested fixes: run detection off the main thread; cap/stream file reads during detection instead of loading whole files; and add a config flag to disable external-agent import entirely.

Any update?

I can confirm this is still happening on the latest desktop app.

Environment:

• App: ChatGPT / Codex desktop app (com.openai.codex)
• Version: 26.721.41059 (build 5848)
• macOS: 26.5.2 (25F84)

Reproduction:

1. Launch the app.
2. Open Settings.
3. The app crashes/exits.

The issue persists after updating to the latest available version.

Still reproducible on the latest available build.

Environment:

• ChatGPT desktop: 26.721.81911 (build 5973)
• Bundled framework: 150.0.7871.128
• macOS 26.5.2 (25F84)
• Apple Silicon, Mac15,9

I reproduced four matching crashes across July 29–30. The current two macOS diagnostic reports show the same deterministic signature:

• EXC_BREAKPOINT (SIGTRAP)
• Faulting thread: main thread, CrBrowserMain
• Recurring frames: v8::String::NewFromUtf8 → node::StringDecoder::DecodeData → node::fs::AfterInteger

The app logs end after successful config/read and account/read operations triggered while navigating Settings. There is no preceding configuration parse error.

Additional validation:

• Preference plists are structurally valid.
• Nearby application JSON files parse successfully.
• The bundled Codex CLI reads the same configuration successfully.
• The application’s Apple signature and notarization verify correctly.
• The issue persisted across the update from build 5848 to 5973.
• The built-in update check reports that 5973 is current.

This reinforces that the failure is in the desktop app’s Settings/main-process path rather than damaged local configuration. Raw .ips reports are available privately if maintainers provide a secure upload route.

I know this may seem insanely obvious but I asked ChatGPT Work/Codex to diagnose and repair itself. It found the crash logs and it took a few passes but it found an overlap with a large Claude folder, edited itself and fixed the problem. Give it a shot.