HITL works unexpected

[liolay]

im using langgraph 1.0+AsyncShallowRedisSaver , but sometimes user need double confirm langgraph Interruption using invoke(Command(resume=...)), and when it is happed, i found interrupt was returned correctly,but when execute agent.get_state(...), interrupt is missing in the checkpoint, after first time confirm, interrupt comeback again and exist in the checkpoint!

[tobalsan]

Can confirm this erratic behavior.
Sometimes my workflows will properly resume after an interrupt, sometimes they won't.

I dug into redis to check values, and sometimes, while a key contains __interrupt__, has_writes is false.