OpenAI Codex logging bug may burn out SSDs in under a year

What happened

GitHub issue openai/codex#28224 shows that an installed Codex writes ~37 TB over 21 days of runtime — extrapolated to ~640 TB per year. Cause: a SQLite feedback sink wired to a global TRACE log level.

Context and impact

A typical 1TB consumer SSD has ~600 TB endurance. If Codex runs 24/7 it can burn through that in under a year. The issue resonates on HN as another case of "AI tools treat local disk as an infinite resource" — paralleling the Cursor IndexedDB bug earlier this year.

Details

• 37 TB written in 21 days (measured)
• ~640 TB/year extrapolated
• 1TB consumer SSDs typically endure 600 TBW
• Root cause: SQLite feedback sink at global TRACE level
• Workaround until an official fix: lower the log level, or redirect Codex data dir to a non-SSD volume