Applied AI
A LangGraph time-travel fork can inherit an old approval update
A reported checkpoint behavior can replay a prior Command update into a new branch after an interrupt.
November 17, 2026
Time travel is meant to explore a new decision from an earlier checkpoint. LangGraph issue #9491 reports that Command(resume=..., update=...) can leave the update on the checkpoint; a later fork from that point drops the old resume answer but re-applies its state update. The reproduction uses a simple additive log and reports the same behavior with memory, SQLite and Postgres savers.
Treat the fork as a new approval context
Before any resumed branch calls a tool, inspect the selected checkpoint's state and pending writes. Record the new human answer, the branch ID and the state delta separately. LangGraph's time-travel guide says nodes after the checkpoint execute again; that includes API calls and interrupts. A replay must never be assumed to preserve the old decision's safety boundaries.
Add a regression with two opposite answers
Build a small graph that interrupts for approval, appends a branch marker through Command(update=...), then forks the pre-answer checkpoint. The new branch should contain only its own marker. Test with the checkpointer used in production and the pinned LangGraph version. Also assert that a side-effecting tool is not invoked twice when replaying.
Contain production experiments
Until the test passes, avoid time-travel forks for workflows where an inherited state update could authorize an external action. Use a fresh thread or explicitly reconstructed state for experiments, with read-only tools first. Keep an audit log of checkpoint IDs and decisions; do not rely on the visual branch alone to prove isolation.
Adovo’s view
A branch is useful only if its history is honest. Human approval data must belong to the decision that produced it. If an old update can cross into a new path, the safest design is to make branch provenance explicit before giving the agent a write tool.
