Anthropic’s Claude Code tool is experiencing a bizarre and destructive failure mode. A user reported on DEV.to that their ~/.claude configuration directory was wiped four separate times over the course of three weeks. The most recent incident, documented in GitHub issue #93791 filed on September 12, 2026, occurred mid-session while the developer was actively working, leaving the chat window blank and the folder gone.

The Silent Destroyer

The issue highlights a critical vulnerability in how local AI coding assistants manage persistent state. The wipe wasn't just a minor inconvenience; it resulted in the loss of 19 live API keys stored within the configuration directory. For a developer relying on these keys for immediate access to LLM services, this represents a significant security and workflow disruption. The fact that this happened four times in such a short window suggests a reproducible bug rather than a one-off filesystem error.

Mid-Session Catastrophe

What makes this incident particularly alarming is the timing. The fourth wipe happened while the user was actively engaged in a session. The chat window went blank, and the configuration folder vanished without warning. This behavior points to a potential race condition or a flawed cleanup process within the Claude Code agent itself. If the tool is deleting its own configuration directory during active use, it raises questions about the stability of the underlying file handling logic.

Key Takeaways

  • Claude Code wiped the ~/.claude directory four times in three weeks.
  • The incident resulted in the loss of 19 live API keys.
  • GitHub issue #93791, filed September 12, 2026, documents the recurring problem.
  • The fourth wipe occurred mid-session, causing immediate workflow disruption.

The Bottom Line

This isn't just a bug; it's a trust issue. If your AI coding assistant can't keep its own config alive, you shouldn't trust it with your production keys without rigorous backups.