[Bug]: starting a second agent-server in the same sandbox kills tmux sessions on the shared openhands socket

[neubig]

Summary

When an OpenHands Cloud sandbox already has an agent-server running for the active conversation, starting another agent-server process inside the same sandbox can break the original conversation's terminal tools.

Why this happens

Current agent-server startup calls _cleanup_stale_tmux_sessions() and uses the shared openhands tmux socket (TMUX_SOCKET_NAME = "openhands"). In a shared runtime, the second server treats the first server's live tmux sessions as stale and kills them.

After that, the original conversation starts failing with errors like:

Error executing tool 'terminal': ['no server running on /tmp/tmux-10001/openhands']

Reproduction outline

1. Start an OpenHands Cloud conversation (or otherwise run one agent-server in a sandbox).
2. In the same sandbox, start a second agent-server process with default settings.
3. Use the original conversation's terminal tool.
4. Observe tmux socket / missing server failures.

Expected behavior

Running another agent-server in the same sandbox should not kill tmux sessions owned by a different live server instance.

Possible fixes

• Namespace the tmux socket per server instance or per conversation.
• Make tmux cleanup opt-in or scope it to sessions created by the current server instance.
• Refuse to start when another live server is already using the shared tmux socket.

Related context

• Errors when developing agent-server-gui on OpenHands Cloud agent-server-gui#19

This issue was created by an AI agent (OpenHands) on behalf of the user.