A LangGraph stream can reconnect before its first event forever

A connected socket is not the same as a functioning agent response. An open LangGraph SDK issue reports a reconnect loop when a shared SSE connection is established and lost before the first event. A retry counter can look healthy while the user waits indefinitely.

Measure time to first meaningful event

Record stream-open, first event, disconnect, reconnect, and terminal times. Set a no-event deadline that includes reconnect attempts, not just a per-socket timeout. Show the user a recoverable error when that deadline is exceeded.

Make retries bounded and resumable

Limit consecutive reconnects and total elapsed time. Preserve a run identifier or checkpoint where the platform supports it, and make sure replayed events do not duplicate downstream writes. Test a disconnect both before and after the first event.

Prove the failure mode in staging

Use a proxy to accept the SSE connection and drop it before an event. Observe whether the client returns, reconnects, or loops. Track the upstream issue and retest when the SDK changes its retry behavior.

Adovo’s view

A stream must have a progress contract. Connection status alone cannot tell a buyer whether work is happening. Bound the time to evidence of progress, and turn silence into a clear recovery path.