The last 4 days my home server felt cursed. Telegram bots dying almost daily, SSH access randomly blocked (had to VPN through Frankfurt just to reach my own server). On top of that, an older mystery: CPU overload roughly every two weeks that a cleanup would temporarily fix. My AI OS was supposed to be my right hand and instead it kept falling apart. Yesterday I gave Claude Code (Fable 5) SSH access and one instruction: "audit everything, find out why my system keeps dying." What it actually found (none of it was what I thought): 1. My "daily bot crashes" were planned token-refresh restarts that wiped the conversation memory. Fix: resume the last conversation on restart. 2. A second bot had been silently dead for 2 WEEKS: expired OAuth token, no auto-refresh. It looked alive, it just never answered. 3. The "blocked SSH": not the server at all. My hotel and home Wi-Fi block outbound port 22 entirely. tcpdump showed my packets never even arrived. Tailscale solved it permanently, no more VPN. 4. The biweekly CPU overload: orphaned background processes multiplying after every bot restart (this one it found in my own wiki's incident history and confirmed the cleanup cron now handles it). Best part: it documented every finding in my markdown wiki, added self-audit crons that report to Telegram weekly, and wrote "hard rules" into its own config files so future sessions don't repeat the same mistakes. It even caused one incident itself, then found and fixed it 15 minutes later and documented that too. Takeaway: don't ask AI "is my server ok?". Give it real access, ask it to point to evidence for every claim, and make it write everything down where the next session can find it. What's the longest-running "mystery bug" AI has solved for you?