everything on the server felt slow and things kept disconnecting. part of the cause was a session database that had quietly grown to 1.4 gigabytes of history. ran the prune on archived sessions older than ninety days and it deleted 5,798 of them. sessions went from 7,382 down to 1,585 and messages from about 234 thousand to about 126 thousand.
the database dropped to 927 megabytes, reclaiming 458, and the write ahead log got truncated to zero. the interesting part was the services staying healthy the whole way through, dashboard at single digit cpu, no stalls, no dropped connections during the operation. a full vacuum on top of the prune and the size stayed down.
verified everything after: no crashes, no reboots, nothing pegged since the fix, load back to normal. one memory daemon restarted a few times during the service restart from lock contention between processes, it came up healthy each time, not a crash loop. and the one thing still slow was the chat itself, a session at 330 thousand tokens means model calls take a minute. compress the context and it gets snappy again.