Pydantic Interrupt Schemas Validate Resume Values
interrupt(value, response_schema=ApprovalModel) surfaces JSON Schema and validates the resumed value against the Pydantic model. An invalid answer raises ValidationError while leaving the same interrupt pending in the SQLite checkpoint, so a valid answer can be supplied after reopening. Passing a raw JSON Schema dict surfaces the schema but does not validate: an out-of-enum value reached the node unchanged. Reproduced three times on LangGraph 1.2.22 / checkpoint-sqlite 3.1.3.
Additional LangGraph 1.2.22 reproduction: with RetryPolicy(max_attempts=3, retry_on=lambda error: True), an invalid typed resume was validated on all three attempts (four total node entries including the initial pause). ValidationError surfaced, the same interrupt ID remained active, and a valid response after SQLite reopen completed. Broad retry classifiers can repeat pre-interrupt work for invalid input.
Retry-policy follow-up on LangGraph 1.2.22: with RetryPolicy(max_attempts=3, retry_on=lambda error: True), an invalid typed resume was validated on all three retry attempts (four node entries including the original pause). ValidationError surfaced, the same interrupt remained active in the checkpoint, and a corrected response after SQLite reopen completed. Broad retry classifiers can repeat pre-interrupt work even though the invalid response does not consume the interrupt.