The Crashpad pending directory on a Mac grew to 4.9 GB and over 54,000 files in a single day, as reported by a GitHub issue.
*On macOS, Codex Desktop is filling users' drives with gigabytes of crash reports each day. The app stays usable, but Crashpad generates pending dump/sidecar pairs without any cleanup. OpenAI has yet to provide a safe upload path or fix.*
On June 2, 2026 a GitHub user identified as Jolg42 opened issue #25921 reporting that Codex Desktop on macOS was silently filling the ~/Library/Application Support/com.openai.codex/web/Crashpad/pending folder with paired dump and sidecar files. In just one day the directory swelled to 4.9 GB and 54,504 files, a rate that translates to at least five gigabytes of data per day. The growth was not a one‑time leftover; file counts rose continuously during investigation, indicating an active generation loop.
The app itself remained responsive. Jolg42 could browse the filesystem, inspect GitHub, and continue coding while the background process churned out data. The core problem is not a crash that prevents launch, but a runaway Crashpad subsystem that lacks any upload, rotation, or cleanup mechanism. OpenAI has not yet provided a safe path for submitting the private dumps, leaving users to manage the bloat themselves.
The GitHub issue filed on June 2, 2026 shows Crashpad/Electron annotations (prod=Codex_Mac, ver=148.0.7778.179) attached to a live process. Two browser_crashpad_handler processes have been running under PPID=1 since May 30, 2022 11:47:50, both pointing at the same Crashpad database in ~/Library/Application Support/com.openai.codex/web/Crashpad. Instead of uploading or rotating reports, the handler repeatedly writes paired .dmp and _sidecar.json files to the pending directory. Expected behavior is bounded reporting; actual behavior is continuous generation while the app remains functional.
On the reporter's machine the pending folder reached 4.9 GB and 54,504 files in a single day. The count rose from 53,444 to 54,064 to 54,504 within hours, confirming active growth. The directory held 27,252 .dmp files and an equal number of sidecar JSONs. Another user, mxggle, logged a 59 GB total size and hit the 65,535 file‑system limit, yet the background process kept writing. The issue title notes growth of at least +5 GB per day.
After the issue was filed, the reporter deleted everything in the pending folder. Disk usage collapsed to 6.4 MB and 70 files (35 .dmp and 35 sidecar). However, new paired files appeared immediately after the purge, proving that Crashpad generation was still active rather than a stale backlog. No minidumps were posted publicly due to potential private data, and no safe upload endpoint was offered by OpenAI.
Unbounded dump creation threatens macOS storage, especially on machines with limited space. Without automatic rotation or upload, users must manually delete files, risking exposure of memory contents if they share dumps. OpenAI has not supplied a secure upload path, leaving the problem in users' hands. The persistence of two crash‑handler processes suggests a systemic bug in the desktop client rather than an isolated crash. Until OpenAI addresses the handler logic, the issue will continue to erode disks.
The pending folder reached 4.9 GB and 54,504 files in one day, a growth rate of at least +5 GB per day as noted in the issue title.
Deletion reduced the folder to 6.4 MB and 70 files, but new dump/sidecar pairs appeared immediately, showing that generation continued.
Yes, two browser_crashpad_handler processes have been observed running under PPID=1 since May 30, 2026, both using the same Crashpad database.
The evidence points to a systemic failure in Codex Desktop's crash‑reporting pipeline. With two crash‑handler processes persisting since late May and no automatic pruning, every active session adds gigabytes to a user's drive. Until OpenAI implements bounded reporting or a secure upload route, the bug will keep eroding storage and exposing private data. Users are forced into a manual cleanup loop, a situation that undermines confidence in the desktop client.
Sources: GitHub issue #25921 (openai/codex)