[Bug]: separate agent-server processes can attach to the same conversations_path and resume active conversations

[neubig]

Summary

Two agent-server processes that point at the same conversations_path can load the same persisted conversations, which creates split-brain behavior when one of those conversations is already active.

Why this matters

The default Config.conversations_path is workspace/conversations. In shared runtimes (for example an OpenHands Cloud sandbox), a second agent-server that starts with default settings can attach to the same persisted conversations as the server that is already backing the active session.

This is closely related to the scenario covered in #2943, where a second service resumes or attaches to a still-running conversation from the same persistence directory.

Expected behavior

A second agent-server should not be able to attach to or resume a conversation that is already owned by another live server process unless there is an explicit handoff.

Possible fixes

• Acquire an exclusive startup lock for conversations_path.
• Refuse startup when another live server is already using the same persistence directory.
• Enforce conversation leases / ownership checks before resuming persisted active conversations.

Related context

• Errors when developing agent-server-gui on OpenHands Cloud agent-server-gui#19
• test(agent-server): cover split-brain conversation resume #2943

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