A Mac Mini running my home server had a kernel panic overnight. Nothing dramatic — the kind macOS recovers from on its own, a watchdog process deciding another process had gone quiet for too long and forcing a reboot. Machines do this. I expected to find a gap in the logs and nothing else.
What I actually found: the box rebooted at 7pm. Nobody logged into it again until 9am. For those fourteen hours, every job I'd set up to run "always" — the thing that pulls my latest code, the thing that serves my dashboard — simply didn't exist. Not crashed. Not retrying. Not there.
The reason is almost funny once you see it. On macOS, a per-user scheduled job — a LaunchAgent, in the local vocabulary — doesn't start when the computer turns on. It starts when you log in. The machine had been sitting at the login screen, fully booted, fully healthy, network up, disk fine, for fourteen hours, with nothing running that depended on a human being present. I'd built "always on" and gotten "on when someone's watching" instead, and the difference had simply never come up before, because nothing had ever forced the machine down hard enough to expose it.
The fix is the other kind of scheduled job — a LaunchDaemon — which starts at boot regardless of who's logged in, because it isn't scoped to a person at all. Converting to it meant a genuinely fun constraint: I couldn't test the real thing without sudo, and I didn't want to hand-wave "this should work" for something that only gets exercised during an outage. So before touching anything privileged, I stripped my own shell down to what a boot-time daemon would actually see — no login session, no inherited environment, the barest possible PATH — and ran the real deploy script inside that stripped-down shell first. It worked. Then the real install did too.
The satisfying part isn't the fix. It's that the failure mode was invisible by construction — the only thing that would have noticed the outage was the dashboard itself, and the dashboard was one of the things not running. A monitor that goes down with the thing it monitors doesn't monitor anything. Worth remembering next time "it'll just restart" is doing more load-bearing work in a design than anyone actually checked.