You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
A dropped devpod tunnel kills the Claude session inside the workspace #672
Kill that process and the session inside the container dies with it, mid-turn. Nothing is left to reattach to.
What happened
On 2026-10-02 at 22:25:49 BST, a host command meant to kill one tunnel ran pkill -KILL -f "docker exec -i ". The variable that should have held the container id was empty, so the pattern matched every devpod tunnel on the host. All 9 Claude sessions running in dl workspaces died at once. The containers kept running. The work in their checkouts survived, but every in-flight turn and all session state was lost, and each session had to be restarted and re-briefed.
The command was the bug, and that host now blocks name-pattern kills. But a tunnel can also drop for ordinary reasons: a host-side crash, an OOM kill, a laptop suspend, a devpod restart. Each of those has the same result.
Ask
A dropped tunnel should drop only the connection to the session, not the session.
One way: when dl hands over an interactive shell or starts an agent, run it inside a terminal multiplexer in the container (tmux, or zellij, which the image setup already considered in #391). Attach with new-session -A or the equivalent, so a second dl <ws> re-attaches to the session that is still running instead of starting a new one. dl <ws> -- <cmd> should keep its subprocess contract (exit status, stdin and stdout are the command's), so it probably stays as it is, or gets an opt-in flag.
Activity