LangGraph can checkpoint invalid node output and permanently corrupt a thread

Keep your node return values strictly inside the state schema, especially with Pydantic models, because an invalid write can poison the thread beyond repair. A maintainer-proposed fix validates node updates before they reach the checkpointer so the graph fails fast instead of corrupting history; until you confirm it is in your version, treat any unreadable checkpoint as possibly corrupt and start a fresh thread rather than debugging the dead one.

Context: GitHub issue langchain-ai/langgraph#6491 documents a checkpoint corruption trap. LangGraph validated node input but did not validate node output, so a node returning state that violates the Pydantic schema got checkpointed successfully, and the checkpoint could never be retrieved again because validation failed on read. The result was permanently corrupted thread history with no recovery path. A maintainer proposed PR #6512 validating node updates against the schema before they reach the checkpointer, raising immediately during execution, and verified it against the reporter's reproduction script.

Source: GitHub issue langchain-ai/langgraph#6491 documents a checkpoint corruption trap. LangGraph validated node input but did not validate node output, so a node returning state that violates the Pydantic schema got checkpointed successfully, and the checkpoint could never be retrieved again because validation failed on read. The result was permanently corrupted thread history with no recovery path. A maintainer proposed PR #6512 validating node updates against the schema before they reach the checkpointer, raising immediately during execution, and verified it against the reporter's reproduction script.

Verified resolution: Keep your node return values strictly inside the state schema, especially with Pydantic models, because an invalid write can poison the thread beyond repair. A maintainer-proposed fix validates node updates before they reach the checkpointer so the graph fails fast instead of corrupting history; until you confirm it is in your version, treat any unreadable checkpoint as possibly corrupt and start a fresh thread rather than debugging the dead one.

Source URL: https://github.com/langchain-ai/langgraph/issues/6491