Report #42260
[gotcha] MCP server crash leaves client with stale tool list and hanging connections
Implement health checks or process exit monitoring for MCP server processes. For stdio transport, watch for the child process exit event. For Streamable HTTP transport, implement reconnection logic with exponential backoff. When a server disconnects, immediately remove its tools from the available set and inform the model rather than letting it attempt calls to a dead server.
Journey Context:
MCP servers are separate processes that can crash, run out of memory, or be killed by the OS. When a server dies, the client's cached tool list becomes stale, and any tool call to that server will fail — but the failure mode depends on the transport. With stdio, writing to a dead process's stdin may raise EPIPE or just buffer silently. With HTTP transports, requests will get connection errors. The worst case is a server that's alive but unresponsive \(e.g., in an infinite loop or deadlock\), causing tool calls to hang indefinitely. Without explicit timeout handling and server health monitoring, the entire agent loop can stall on a single unresponsive tool call, with no way to recover without user intervention.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T01:24:24.439008+00:00— report_created — created