After some testing #Haven relay, Claude agent says: ● Bingo! File descriptor leak: ┌───────────┬──────────┬─────┬─────────┐ │ Metric │ Baseline │ Now │ Change │ ├───────────┼──────────┼─────┼─────────┤ │ Open FDs │ 97 │ 105 │ +8 │ ├───────────┼──────────┼─────┼─────────┤ │ TCP conns │ 31 │ 26 │ -5 │ ├───────────┼──────────┼─────┼─────────┤ │ FD max │ - │ 105 │ Growing │ └───────────┴──────────┴─────┴─────────┘ Haven is leaking file descriptors. It opened 8 more FDs but only lost 5 connections, meaning resources aren't being freed properly. This is likely a goroutine leak in haven—when it processes publishes, goroutines spawn but don't exit cleanly. Over time, FDs hit the ulimit and it crashes. Your options: 1. Use the watchdog script I provided earlier to auto-restart haven every ~10 min before it hangs 2. Increase the ulimit for haven (might delay the issue): sudo nano /etc/systemd/system/haven.service 2. Add under [Service]: LimitNOFILE=4096 2. Then restart 3. Report this to haven developers as a goroutine/FD leak after N publishes For now, the watchdog is the practical solution. Want me to help you set that up properly?