[macOS][26.924.22138] Renderer crashes and UI reloads every ~75 seconds in a new chat, even with sidebar collapsed #48848

What version of the Codex App are you using?

Codex/ChatGPT Desktop 26.924.22138, build 11645, production channel. Crashpad records embedded Codex Framework 154.0.8037.57. The in-app update checker reported the app as up to date during the incident.

What platform is your computer?

macOS 26.6.2 (25G83), Apple Silicon (arm64).

What issue are you seeing?

The desktop UI repeatedly refreshes and interrupts normal Codex work. On 28 September 2026, between 09:20:03 and 10:43:40 China Standard Time (UTC+8), the local Crashpad directory accumulated 64 new crash sidecar records: 58 explicitly marked ptype=renderer and six with no process type in the sidecar. Renderer crashes occurred roughly every 70–90 seconds for much of this window.

A read-only process snapshot at 10:25 showed the main ChatGPT application process still running since 09:18, with recently started replacement renderer processes. The evidence supports a renderer/UI crash loop; it does not show the entire app exiting on every refresh.

The problem persists in a new conversation after the user stopped other conversations. Collapsing the sidebar also did not stop it: renderer crash records continued at 10:33:26, 10:34:38, and repeatedly through 10:43:40. The repeated refreshes have materially disrupted the user's work.

What steps can reproduce the bug?

This reproduces consistently in the affected local profile. A clean-profile minimal reproduction has not been established.

1. Launch Codex/ChatGPT Desktop 26.924.22138 on macOS.
2. Use a new Codex conversation with other conversations stopped.
3. Leave the app open. The UI refreshes about every 1–1.5 minutes.
4. Collapse the sidebar. The refresh/crash cycle continues.

What is the expected behavior?

A new conversation should remain open and usable without periodic renderer crashes. Renderer recovery should not repeatedly interrupt work.

Additional information

A representative Crashpad sidecar contains metadata of this shape:

{"capture_kind":"crash","osarch":"arm64","ptype":"renderer","ver":"154.0.8037.57"}

The desktop log also shows frequent successful thread/list responses, typically about 90–160 per minute during the crash loop, including 157 during 10:33 after the sidebar was collapsed. The same log contains intermittent ResizeObserver loop completed with undelivered notifications, post-login Statsig bootstrap timeouts, and unknown-conversation state errors. These are correlations or possible post-restart effects; none has been established as the native crash cause. The available sidecar metadata does not provide a decoded native renderer stack trace.

The user noticed repeated refreshes after updating. A few crash sidecars from the previous embedded framework version (153.0.8010.53) were also present around 08:59–09:01, so this report does not claim that the update alone introduced the defect. It establishes sustained, frequent crashes on build 11645.

Related but not identical: #44787 involves reopening a particular active task on an older build; #48480 reports reloads while a task is running on build 11645; #48654 reports main-process memory growth on build 11645. This report adds a measured renderer crash cycle that continues in a new conversation with other tasks stopped and the sidebar collapsed.

No raw logs, conversation content, account details, local paths, thread IDs, or crash IDs are attached publicly. Sanitized excerpts can be provided to maintainers if needed.