Codex issue 28224 writes 37 TB to local SSD in 21 days via unfiltered SQLite TRACE logs

The GitHub report documents retained rows of 681774 in logs_2.sqlite with 70.7 percent TRACE content dominated by codex_api::endpoint::responses_websocket and inotify events. Write amplification from repeated OpenTelemetry mirrors and websocket payload logging exceeds retained size by orders of magnitude, projecting 640 TB annual writes against typical 600 TBW consumer SSD ratings. Filtering TRACE and codex_otel categories would eliminate 96 percent of volume without disabling feedback entirely.

This pattern matches prior logging amplification failures observed in other high-frequency telemetry systems such as early Kubernetes audit backends and Rust tokio tracing defaults before level gating was enforced. Codex relies on the same otel.trace_safe and trace_only paths that bypass production sampling thresholds, turning routine session events into sustained disk I/O. The 21-day measurement on a single machine indicates the defect scales linearly with uptime rather than session count.

Operational impact includes premature SSD wear on developer workstations and potential silent corruption once write endurance is exhausted. OpenAI has not published a targeted patch or configuration flag in the codex repository as of the report date. Users must either disable logging or apply external log rotation to mitigate immediate risk.

Resolution requires explicit level filters on codex_api and codex_otel targets plus WAL checkpoint tuning. Deployment of such a change would be verifiable through reduced TBW metrics on subsequent issue updates.