OpenAI's Codex Desktop has a crash-handler loop that silently fills user disks with crash dump files at a rate of at least 5GB per day — and possibly much faster. One user documented 323GB consumed across 4.04 million files, representing 35% of their 926GB disk. Free space fell to 1.4GB before the problem was discovered manually. The application provided no warning. The root cause is specific and reproducible: the browsercrashpadhandler process SIGTRAPs within 2 milliseconds of launch, macOS's launchd respawns it, and each cycle deposits a new dump file in the Crashpad pending/ directory. This creates an infinite loop of crash-generate-respawn-crash with no size cap, no count limit, and no automatic pruning. The app itself continues functioning normally throughout, which is why users don't notice until their disk is full. The bug has been tracked across multiple GitHub issues — #36982, #39875, #51739, #52194, #51361 — spanning months from initial reports through the latest confirmation on October 17, 2026. The reporter measured the scale between October 15-17, finding 180GB in one Crashpad directory and 143GB in another. Build versions confirm this persists through ChatGPT.app 26.1007.71326 with Codex Framework (Chromium) 155.0.8059.40 on macOS 15.7.4. The failure mode is textbook negligence in resource management. Three basic engineering safeguards are missing: a size or count cap on the pending directory, detection and circuit-breaking of the crash loop itself, and user-facing alerts when application data exceeds reasonable thresholds. Any one of these would prevent the problem. Their combined absence means the software treats user disk space as an unlimited resource it has no obligation to monitor. The user's workaround — renaming and deleting the Crashpad folders while the app runs — recovered 50+ GB within minutes, though the deletion of millions of small files continued running. This is not a fix. Without the underlying crash loop resolved, the directories will refill immediately. Users who haven't manually audited their Library folders may be losing disk space right now without knowing it. This pattern reveals something about OpenAI's desktop software maturity. Codex Desktop inherits Chromium's Crashpad infrastructure but apparently lacks the operational discipline to monitor its own crash-reporting pipeline. A crash handler that itself crashes every 2ms is a monitoring failure. An unbounded write loop with no disk-space check is a design failure. Months of open issues with no shipped fix is a prioritization failure. The asks are reasonable and minimal: cap the directory, break the loop, warn the user. These are not novel engineering challenges. They are table-stakes reliability practices that ship with most desktop software by default. The gap between OpenAI's AI ambitions and its desktop engineering fundamentals is on full display.
