LangGraph 1.2.18 fixes error handlers and delta snapshots

LangGraph 1.2.18 was published on October 26, 2026, two days after 1.2.17. Both are patch releases, but they change behaviour you may rely on: node error handlers now actually recover when a graph streams or runs nodes in parallel, delta channels snapshot ten times as often, and an explicit recursion limit equal to the default is no longer ignored.

Release facts

What changed in 1.2.18

• error_handler recovers everywhere. A node's error_handler ran, but the original error was still raised when the node ran on the executor path: streaming messages or custom, subgraphs=True, or next to a parallel node. PR 9408 makes both executors skip failures that were handled. If you saw graphs crash despite a handler, this is the fix.
• Failed task results stream. With only tasks or debug streaming, a failing task's result and its error were never sent; now they are, before the error is raised.
• Delta snapshots every 100 updates. DeltaChannel, including message channels, now takes a full snapshot every 100 updates instead of 1,000 (PR 9420). The release notes say reads on long threads get faster at the cost of more storage. Pass snapshot_frequency=1000 to keep the old behaviour.
• Last checkpoint is streamed. When a run stopped at interrupt_after or the recursion limit, the checkpoint it stopped on never reached the checkpoints or debug stream, so a client resuming from the last checkpoint it saw ran the last node again. The checkpoint is now sent.
• No duplicate tags in subgraphs. Tags passed at invoke time were repeated at every subgraph level (twice one level deep, eight times three levels deep), which showed up in config["tags"], stream metadata and traces.

What changed in 1.2.17

• Explicit recursion limits apply. merge_configs dropped a recursion_limit equal to the default. PR 9417 gives the example of LANGGRAPH_DEFAULT_RECURSION_LIMIT=1000: agent.with_config(recursion_limit=1000) kept the 9999 that create_agent binds, so a looping agent could run to 9999 steps. The PR notes langchain-core has the same check, which it does not change.
• RemoteGraph namespaces. RemoteGraph.stream(subgraphs=True) and astream now keep the caller's namespace prefix on the events they yield.

Questions

Will LangGraph 1.2.18 use more storage? For threads using DeltaChannel, yes: full snapshots are written every 100 updates instead of every 1,000. Set snapshot_frequency=1000 on the channel to keep the previous trade-off.

Do I need to change code for the error_handler fix? No. Handlers you already registered now take effect on the streaming, subgraph and parallel paths too. Code that relied on the error still being raised on those paths will behave differently.

Is this a breaking release? Neither release is labelled breaking, but the snapshot default and the extra checkpoint event on interrupt_after are visible changes to test.

More on the LangGraph hub, the earlier LangGraph 1.2.16 launch and the frameworks category. New today: LangGraph vs Google ADK.