Codex logging bug may write TBs to local SSDs

Original Article ↗ Hacker News Discussion ↗

OpenAI’s Codex coding assistant has been found to generate massive trace logs on users’ machines, in some cases writing tens or hundreds of gigabytes to local SSDs and risking disk wear or system instability. Commenters see it as symptomatic of “vibe-coded” AI-driven development: products shipped quickly with minimal review, weak QA, excessive resource use, and slow or absent fixes for critical issues. Many argue this undercuts claims that AI has “solved” software engineering, and instead highlights the need for stronger testing, safeguards, and human accountability when using AI to build developer tools.

Logging bug and impact

  • Codex’s trace-level logging writes continuously to a local SQLite DB, with WAL files growing to tens of GB and potentially TBs over time.
  • Users report idle sessions still hammering disks; some discover the issue only via tools like iotop or when machines run out of space or hang.
  • Similar heavy logging behavior is noted for other AI tools (e.g., Claude Code, ChatGPT macOS app), sometimes mitigated by symlinking logs to tmpfs.
  • A fix has been committed in the Codex repo and is expected in the next release.

Workarounds and mitigations

  • Suggested mitigations include:
    • SQLite trigger to ignore all inserts into the logs table.
    • Running VACUUM FULL to shrink oversized DB files.
    • Deleting WAL files via scripts or moving logs to a RAM-backed filesystem.
  • Some users simply uninstall Codex or switch tools due to fear of SSD wear.

Vibe coding, slopware, and engineering process

  • Many attribute the bug to “vibe coding”: AI-generated code shipped with minimal human review or design.
  • Commenters highlight that AI often produces plausible-looking but subtly wrong code, making review harder and encouraging over-trust.
  • There’s frustration that flagship AI companies, while claiming to “solve coding,” ship fragile, opaque Electron/Rust apps with severe regressions.

Testing, QA, and accountability

  • Several ask why basic QA or integration tests didn’t catch unbounded logging or resource usage.
  • Others argue such bugs are easy to miss unless you explicitly test for long-running usage and resource ceilings.
  • Strong pushback against developers blaming AI: humans are still responsible for code they ship, regardless of tooling.

User experience and resource usage

  • Broad complaints about Codex, Claude Code, Cursor, and other AI tools: high CPU/GPU use, memory leaks, laggy UIs, background VMs, and unconfigurable behaviors.
  • A minority say these tools work “well enough” for them and prioritize features over polish.

Broader AI and industry implications

  • The incident is used to question claims that AI has made programmers obsolete or that “coding is solved.”
  • Some see this as emblematic of a wider software-quality crisis and hype-driven management pressure to be “10x with AI.”
  • Others note that humans also ship catastrophic bugs; the issue is inadequate process and incentives, not just AI.