langgraph 1.2.18 — langgraph==1.2.18

LangGraph 1.2.18 fixes a case where a node's error handler ran but the run failed anyway. LangGraph is LangChain's library for building agents as graphs of steps, called nodes. A node can have an error_handler that runs when the node fails, so the graph can recover instead of stopping. According to the fix's pull request, the handler ran but the original exception still escaped when the run streamed messages or custom output, when it streamed subgraphs, or when the failing node ran alongside other nodes in parallel, because the code executing tasks re-raised every failure without checking whether it had been handled. The same change makes streams in tasks or debug mode deliver a failing task's result before the exception is raised. Previously that result was queued but never streamed, because a run with a single task raised the exception before the stream was read.

The release also changes a storage default. DeltaChannel stores a channel as a series of updates plus periodic full snapshots, and reading it means replaying the updates since the last snapshot. Snapshots now come every 100 updates instead of every 1,000, which the release notes say gives faster reads on long threads at the cost of more storage. Projects that prefer the old balance can pass snapshot_frequency=1000. Two smaller fixes stream the last checkpoint when a run stops at an interrupt_after point or hits the recursion limit, and stop tags from being duplicated at every subgraph level.