macOS app crashes with SIGTRAP every time Settings opens during Claude Cowork import detection

Summary

The macOS ChatGPT/Codex Work desktop app crashes every time Settings is opened. This is a deterministic native main-process crash, not a renderer error.

Environment

• App: ChatGPT / Codex Work desktop app
• Bundle ID: com.openai.codex
• App version: 26.707.72221
• Build: 5307
• Bundled Codex CLI: 0.144.2
• Framework version: 150.0.7871.115
• macOS: 26.3 (25D125)
• Architecture: Apple Silicon (arm64)
• Hardware model: Macmini9,1

Reproduction

1. Launch the desktop app.
2. Open Settings from the app menu or Settings control.
3. The Settings route begins loading.
4. The entire app exits within roughly 1–9 seconds.

Reproduction rate: 6/6 recent attempts across July 14–15, 2026.

Expected

Settings opens and remains usable.

Actual

The app process terminates with:

• Exception: EXC_BREAKPOINT (SIGTRAP)
• Termination: Trace/BPT trap: 5
• Faulting thread: main thread (CrBrowserMain)
• Latest crash incident: 02BCCB40-A9B4-4DCB-9E4D-E92477946342

Top recurring stack frames across the crash reports:

ares_llist_replace_destructor
v8::ToExternalPointerTag
v8::internal::OptimizingCompileTaskExecutor::RunCompilationJob
v8::String::NewFromUtf8
node::StringDecoder::DecodeData
node::fs::AfterInteger
uv__work_done

All six inspected .ips reports have the same crash signature.

Diagnostic evidence

The bundled codex doctor --json reports these areas healthy:

• installation
• config loading
• authentication
• MCP configuration
• state paths/databases
• rollout/state DB parity
• provider and WebSocket reachability
• update configuration

The only doctor failure is TERM=dumb because the diagnostic was run non-interactively; it is unrelated to the GUI crash.

The desktop log immediately before multiple crashes shows:

method=account/read ... errorCode=null
method=config/read ... errorCode=null
[external-agent-import] external_agent_import_provider_stage_finished
providerId=claude-code stage=detect status=success

The process then dies before the Cowork provider reports completion. This strongly suggests the Settings page's automatic “Import from other AI apps” detection path—especially Claude Cowork session scanning—is triggering the native V8/Node failure. This is a correlation, not a confirmed root cause.

The machine has a large Claude Cowork history with many session/audit JSONL files, including individual files over 50 MB. Settings should bound/constrain this scan and handle large or malformed source files without terminating the main process.

Workaround

Avoid opening Settings and edit ~/.codex/config.toml directly. This does not restore access to settings that are only available in the UI.

Suggested fix

• Run external-agent import discovery off the Electron main thread or in an isolated worker.
• Bound concurrency and file sizes during Cowork discovery.
• Stream large JSONL files rather than reading them as whole UTF-8 strings.
• Catch decoder/read failures and surface an import warning instead of crashing the app.