Expected Behavior
When replay changes a run_in_child_context operation from non-virtual to ChildConfig(is_virtual=True), the SDK should reject the existing container checkpoint with NonDeterministicExecutionError before returning its cached result or emitting replay hooks.
Virtual child contexts never write their own container checkpoint. Therefore, any existing checkpoint at the virtual child operation ID proves that the prior execution used non-virtual checkpointing.
Actual Behavior
is_virtual is not represented in OperationIdentifier. A successful CONTEXT checkpoint from a previous non-virtual child passes type, subtype, name, and parent validation. ChildOperationExecutor.check_result_status() then returns the cached result without executing the current virtual child body, so changes to the inner hierarchy are never validated.
This reproduces on both main and PR #698, so it is pre-existing and separate from #692, whose scope is type/subtype/name checkpoint identity.
Suggested Fix
- Reject any existing child-container checkpoint when the current child is virtual.
- Perform the rejection before terminal checkpoint handling and replay/update plugin hooks.
- Add a unit regression for
ChildOperationExecutor and a composed durable-execution test covering non-virtual to virtual replay.
Context
Found while reviewing #698. Related review threads:
Expected Behavior
When replay changes a
run_in_child_contextoperation from non-virtual toChildConfig(is_virtual=True), the SDK should reject the existing container checkpoint withNonDeterministicExecutionErrorbefore returning its cached result or emitting replay hooks.Virtual child contexts never write their own container checkpoint. Therefore, any existing checkpoint at the virtual child operation ID proves that the prior execution used non-virtual checkpointing.
Actual Behavior
is_virtualis not represented inOperationIdentifier. A successfulCONTEXTcheckpoint from a previous non-virtual child passes type, subtype, name, and parent validation.ChildOperationExecutor.check_result_status()then returns the cached result without executing the current virtual child body, so changes to the inner hierarchy are never validated.This reproduces on both
mainand PR #698, so it is pre-existing and separate from #692, whose scope is type/subtype/name checkpoint identity.Suggested Fix
ChildOperationExecutorand a composed durable-execution test covering non-virtual to virtual replay.Context
Found while reviewing #698. Related review threads: