RippleCP: Measuring Counterfactual Checkpoint Advantage in Coding Agents

arXiv:2610.09088v1 Announce Type: new Abstract: Agent checkpoint systems decide what state is recovery-relevant, how to snapshot it, and whether rollback is admissible. None decides which of the safe boundaries they expose are worth materializing. We formulate this as counterfactual checkpoint advantage, the reduction in future recovery cost obtained by checkpointing a candidate rather than skipping it, and measure i…