mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-09-17 23:27:56 +02:00
Elior Nataf LackritzandGitHub
448a763779
Adds an optional keyword-only `response_schema` to `interrupt()`, carried on `Interrupt.response_schema` as JSON Schema so clients (LangGraph Studio first) can render a typed form for the resume value instead of a free-form JSON box. Default `None`, so nothing changes for existing graphs. - Accepts a JSON Schema dict (passed through, not validated) or a Pydantic model / `TypedDict` / dataclass, converted with `TypeAdapter.json_schema()`. For those, the resume value is validated and the validated object is what `interrupt()` returns; an invalid value raises from the node before it is committed, so the next resume works. - Additive on the `stream_mode="debug"` payload: interrupt dicts gain `response_schema` (it is a dataclass field, serialized with `asdict`). - `libs/sdk-py`: mirrors the field on the `Interrupt` TypedDict as `NotRequired`; the server omits the key when no schema was given. JS counterpart: langchain-ai/langgraphjs#2824. Verified: new tests parametrized across every checkpointer backend (schema kinds, coercion, invalid resume) plus the full `libs/langgraph` suite and downstream `prebuilt`, `cli`, `sdk-py`; round-tripped end to end through the agent server with this branch installed.