Codex writes an alarming volume of data to its local SQLite log database, generating about 640 TB of writes annually.

After 21 days, a single SSD recorded 37 TB of traffic, implying roughly 640 full‑drive writes on a 1 TB SSD. Consumer drives rated at 600 TBW could exhaust their warranty in under a year.

The bulk of the data originates from global TRACE logging, which captures low‑level events and raw websocket payloads. Filtering out non‑essential TRACE targets could cut retained log size by nearly 96 %.

Fixes include disabling the global TRACE default, raising thresholds for noise‑heavy targets, and storing only summarized payload metrics. Implementing a global log‑size cap would prevent future endurance loss.