How AI Saved My Unbootable Linux System After 67 Hours of ZFS Hell

# How AI Saved My Unbootable Linux System After 67 Hours of ZFS Hell Last week, my Linux Mint workstation greeted me with every sysadmin's nightmare: a BusyBox prompt instead of my desktop. The diagnosis was grim: ZFS pools marked as "previously in use from another system," GRUB completely blind to my filesystems, and initramfs refusing to cooperate. My 2TB boot drive with years of work was inaccessible. ## Enter the AI Assistants I turned to two AI coding tools: **OpenAI's Codex CLI** and **Anthropic's Claude Code**. What followed was a 67-hour marathon debugging session unlike anything I'd experienced. ### What They Did Well - **Script Generation**: Both tools produced sophisticated bash scripts with proper error handling, logging, and checkpoints. When one approach failed, they'd iterate with numbered versions (`fix-zfs-repair-001.sh` through `fix-zfs-repair-006.sh`). - **Documentation**: Every failure was logged, every error catalogued. By session's end, I had 100+ documented issues with resolutions. - **Persistent Problem-Solving**: When `grub-probe` kept failing with "unknown filesystem," the AIs explored compression settings, pool imports, chroot environments, and GRUB module configurations systematically. ### The Recovery After dozens of iterations, the system finally booted. But the AI assistance didn't stop there—I also used Claude Code to: - Fix Docker's storage driver (ZFS → overlay2) - Restore five containerized services - Debug HiDPI scaling issues - Configure encrypted drive mounts ## The Takeaway AI coding assistants aren't magic. They hallucinate. They go in circles. They need human judgment to stay on track. But for complex, iterative debugging where you need to try dozens of approaches while maintaining meticulous documentation? They're invaluable partners. My system is fully recovered. My data is safe. And I have a newfound appreciation for both ZFS's resilience and AI's potential as a debugging companion. --- *🤖 Note: At my direction and review, I had Claude Code write this. It used my project directory with notes and scritps to understand the context. Human directed, machine written, human reviewed & approved.*