Limitation on updating checkpoints created via update in time travel?

[Arunprakaash]

Hi, I have a question regarding the time travel feature in LangGraph. I understand that we can time travel to a specific checkpoint, update the state there, and resume execution from that point. However, I’ve encountered what seems to be a limitation:

If I update a checkpoint (say, A), it creates a new checkpoint (let's call it A') in the state history with the source set to update. This works as expected.

But if I then try to update A', the update doesn’t seem to work as expected (e.g., no new checkpoint is created, or execution doesn’t resume properly).

So, my question is:

Are there any limitations on which checkpoints can be updated? Specifically, is it not allowed (or not supported) to update a checkpoint that was itself created via an earlier update?

Just trying to understand the design constraints or if this might be a bug. Thanks in advance!

[yf-yang]

#4987

I have the same question too. Ideally people will frequently use this feature if they want to allow some previous message rewritten behavior, but seems nobody cares this bug. If there are other patterns that achieves the same functionality please share with me, thanks.

Currently a workaround: Locate all the messages between the checkpoint and the latest one, then insert RemoveMessages at the end to rewind.