Report #93883
[gotcha] MCP stdio server crashes silently and agent hangs forever waiting for a response
Implement explicit timeouts on every MCP tool call \(30s default, configurable per tool\). Monitor the server stdio streams for close events which signal process exit. Log server stderr for crash diagnostics. On timeout or detected crash: terminate the pending request, report a clear error to the model, and attempt server restart with exponential backoff \(max 3 retries\).
Journey Context:
MCP servers using stdio transport are child processes that can crash from unhandled exceptions, OOM kills, or segfaults. Unlike HTTP transports where connection failures produce immediate errors, stdio failures can be silent—the pipe may remain technically open even as the process dies, or the close event may not be processed until the next read attempt. The agent then blocks indefinitely waiting for a response that will never arrive. This is especially insidious in automated agents where there is no human to notice the hang. The fix requires defensive timeout handling at the MCP client layer, treating every tool call as potentially unresponsive.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T16:10:12.009310+00:00— report_created — created