Users of OpenAI's Codex Desktop app on macOS have identified a significant bug where the application's crash reporting mechanism, Crashpad, generates an unbounded number of pending memory dumps. According to a report on GitHub, the issue manifests as the continuous creation of paired `.dmp` and `_sidecar.json` files in the application's support directory. In one documented instance, this process added approximately 4.9GB of data and over 54,000 files to the local disk in a single day, with the count continuing to rise during investigation.

GitHub issue summary describing the Codex Desktop app's crashpad file growth
A GitHub issue report detailing how the Codex Desktop app continuously generated paired .dmp and _sidecar.json files, leading to rapid disk consumption.github.com

The image illustrates the specific bug where the Crashpad mechanism generates an unbounded number of pending memory dumps, as documented in the GitHub report.

Technical analysis of the issue suggests that the `browser_crashpad_handler`—a helper process used to manage crash reports—is entering a crash-and-restart loop. One user observed that the handler was producing a `FATAL` error related to an 'invalid capability' when requesting a Mach port notification. Because the handler crashes and restarts repeatedly, it generates a new crash report for every cycle. This creates a feedback loop where the crash reporter itself is the source of the crashes, leading to rapid disk consumption without the primary application becoming unusable.

While the app remains functional during this process, the rapid accumulation of files can quickly exhaust available storage. Some users reported even more severe instances, with one reporting a directory size of 61GB and a generation rate of roughly one crash report per second. A temporary workaround identified by users involves making the Crashpad database directory read-only and immutable using system flags, which prevents the handler from writing new files. However, this is a local mitigation that prevents the app from successfully submitting any crash reports, rather than a permanent fix for the underlying loop.