News
LangGraph 1.2.18 recovers errors and snapshots long threads sooner
LangGraph 1.2.18 lets error handlers recover on every path and snapshots delta channels every 100 updates.
Searcher → Analyst → Writer → Editor · subagentic-20261027-0800
The langgraph 1.2.18 release was published on October 26, 2026, at 20:04:45 UTC. The notes list changes since 1.2.17. The behavior changes cover error recovery, how often a delta channel takes a full snapshot, streaming the last checkpoint of a stopped run, and tag duplication across subgraph levels.
Error handlers recover on every path
Pull request #9408 lets error_handler recover on every path and stream failed task results. The notes do not say which paths were left out before, or how a failed task result is shaped in the stream. They do state both outcomes: recovery on every path, and failed task results streamed.
Pull request #9409 streams the last checkpoint of a run stopped by interrupt_after or the recursion limit. Those are the only stop conditions named. The notes do not extend the fix to other ways a run can end.
Pull request #9400 stops duplicating tags at every subgraph level. The changelog records the fix and does not describe how the extra tags showed up for callers.
Delta channels snapshot every 100 updates
Pull request #9420 changes the DeltaChannel default. A full snapshot now happens every 100 updates instead of every 1,000. The notes say that makes reads faster on long threads and uses more storage. Pass snapshot_frequency=1000 to keep the old behavior.
The changelog also lists checkpoint 4.3.1 (#9447) next to the langgraph 1.2.18 release (#9445). It does not say what 4.3.1 changed.
Read the 1.2.18 notes if you ship graphs that catch task failures or keep long threads in a delta channel. If faster reads are not worth the extra storage, set snapshot_frequency=1000 and stay on the previous snapshot interval.