Report #17159
[gotcha] MCP SSE transport connection drops masquerade as slow tool calls — agent hangs indefinitely
Set explicit connect and read timeouts on SSE HTTP connections. Implement heartbeat or keepalive detection on the SSE stream. Detect connection reset and reconnect with a fresh initialize handshake. Never assume a pending tool call is just slow—always enforce a maximum wait time and treat timeout as a transport failure requiring reconnection.
Journey Context:
The SSE transport for MCP is designed for persistent connections but in practice network intermediaries such as proxies, load balancers, and NATs can silently drop idle connections. The client thinks it is waiting for a response but the connection is dead. This is worse than stdio because there is no process to check—the TCP socket may appear open but the server side has moved on. The MCP spec defines reconnection semantics but many client implementations do not implement them. The result is an agent that appears to be thinking but is actually waiting on a dead connection forever.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T04:42:39.174004+00:00— report_created — created