The Codex logging bug is a production AI lesson hiding inside a developer-tools issue. A GitHub report on OpenAI Codex said persistent SQLite feedback logs were writing at extreme rates, with one reporter observing 37 TB written to a main SSD after about 21 days of uptime. The issue extrapolated that pattern to roughly 640 TB per year, enough to consume the warranted endurance of some consumer SSDs in under a year.
The mechanics matter. The report identified Codex files such as logs_2.sqlite and its WAL and SHM companions as continuous writers. A later snapshot showed a 1.2 GiB database with about 506,000 retained rows while the autoincrement counter had advanced past 5.5 billion row ids, suggesting massive churn. TRACE-level logging represented about 70.7 percent of retained bytes in one sample, and two PRs merged on June 22 reportedly reduced logs by about 85 percent for the reporter.
Grey Haven’s read: agent observability can become an infrastructure liability when it lacks budgets. Persistent telemetry is not free just because it is local. Long-running coding agents touch files, streams, shells, and networked APIs all day, so their logging path needs the same discipline as any production service: levels, retention, backpressure, privacy review, and resource accounting.
Operators should ask a blunt question before standardizing agentic developer tools: what does this agent write when nobody is watching? Check disk growth, WAL churn, log levels, and telemetry retention on real workloads. If the answer is “unknown,” the rollout is still a pilot.
This becomes more important if agent CLIs become always-on local workers. It becomes less important if vendors expose clear telemetry controls and sane defaults. Source: GitHub issue openai/codex #28224.
