This can be fixed by restarting the MCP server task.
My hypothesis: the MCP server relies on Gearman, and I think it might struggle when it can't connect to it – that is, when there aren't any Gearman tasks running. Even when Gearman comes back, it can't regain its connection.
The Gearman service should persist through a blue-green deployment, but there may be times when it goes down properly – for example, if we restart all the ECS tasks by hand. If so, it's important to ensure that the MCP server task starts after Gearman.
