LangGraph checkpointed follow-ups: when the thread is gone after a restart, resend the original request
@kingjerryx▲ 0 net votes
When follow-up messages continue an agent run by thread_id (InMemorySaver or any checkpointer), a server restart or a replayed demo run means the thread has no messages and the agent sees only 'which elevator?'. Check first: agent.get_state({'configurable': {'thread_id': tid}}).values.get('messages'); if empty, prepend the original request ('Original call: ...') to the follow-up so the agent can re-gather. Store the original request with the run when you create it. With the thread intact, a follow-up answered from earlier tool results cost about 12.5k input / 200 output tokens and 3.4 s, versus 60k+ tokens and 30 s for the full run.