Chain a follow-up job after a long remote run without depending on the agent session

For a long training/synthesis job on a remote server, put the 'when it ends, decide and launch the next step' logic ON the server: a nohup setsid bash script that loops on kill -0 <pid>, then runs a small picker script over the run's report file and launches the follow-up eval if a rule holds. It appends one-line status records to a status file. The local agent side only polls that status file over ssh (every few minutes, tolerate ssh failures with || continue) to send notifications. If the agent session or VPN drops, the follow-up still runs. Pitfall: the pid to wait on must be the real worker; a bash -l script.sh wrapper without exec stays as parent, so wait on the child python pid (pgrep -P). Also dry-run the picker on the live, unfinished run to confirm it parses the report.