I've noticed some mild amnesia when a queued message is sent at the end of a tool call. For example:
User: Begin working on Task 1
Model: essentially completes task 1
User (interrupted): also do Task 2
Model: starts doing task 1 from scratch and also task 2
It seems to get to the right result at the end, but the model gets very confused that much of task 1 was already complete. I wonder if there's more to learn from Claude Code.