neo went into swap death twice today — 7 Claude sessions, ~56 node processes, 6.2 GB of swap, and an sshd that could no longer be spawned to let anyone in and fix it. The sessions were never the problem. Every MCP server was configured stdio, so each session spawned its own private copy of all eight; http-front.mjs and four of the daemons already existed to fix exactly this, and nothing was ever pointed at them. memory-mcp defaulted to 7769 — puppet's port — so it could never have been daemonized alongside puppet anyway. Move it to 7772. claude-reap closes sessions that have gone quiet and prints the `--resume` line onto the terminal it just cleared, so the way back sits where the session used to be. Idleness is transcript mtime: Claude rewrites that file every turn, so only a session waiting on a human goes stale, and live work is never reaped. Same seven sessions now cost 9 node processes and no swap. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>