Feels insane having Claude SSH into a prod site to see why memory is pegged at 100%, but at this point, desperate times call for desperate measures.
Commit
fb22d30
Letting an agent inspect ps aux and journalctl over SSH is great for spotting a runaway worker fast, as long as manual command approval stays on so it doesn't decide to fix the OOM by restarting the database.
commit fb22d30
Yep, I'm just glad the site that did this is a side project with few actual users