GitHub Repo
LangChain Proposes Streaming Error Event Fix to Prevent Failed Agent Steps from Appearing to Run Indefinitely
On November 12, LangChain proposed a fix for missing terminal events in `astream_events` v2 when some runs fail. The proposal covers retry and fallback flows, but has not been merged; interfaces must also distinguish actual failures from control-flow interruptions.

On November 12, LangChain submitted PR #41820 to address a gap in the execution lifecycle of astream_events(version="v2"): after a chain, model, or retriever raises an exception, the event stream may contain only a start event, with no corresponding error event. This is a new fix for an existing issue; as of the time of review, the proposal remained open for review and had not been merged. Fix proposal
The original issue was filed in October and included a minimal reproduction that does not require a model service: make RunnableLambda raise a ValueError, then collect the streaming events. The caller still receives the exception, but the event list lacks a terminal signal. The report also covers fallback flows: after the first subtask fails, a subsequent fallback can continue, but there is no event to close out the failed subtask. Original report
This affects agent interfaces that use events to track progress. The official documentation explains that each run has its own run_id, while v2 uses parent_ids to represent parent-child relationships, allowing the frontend to reconstruct the execution tree. By inference from this mechanism, consumers that pair only start and end events may leave steps open even after the overall flow has completed its fallback. API documentation
The proposal says the event handler previously implemented error handling only for tools. The fix will emit error events for other types and remove internal execution records. For example, if a chat model fails, it will emit on_chat_model_error with the exception and input in its data. A retry flow can then appear as “start, error, start again, complete,” giving each attempt a clear outcome. Fix design
Engineering teams should still pay attention to event semantics: the proposal will also route control-flow exceptions such as LangGraph’s GraphInterrupt into error events, so _error should not automatically be labeled as a failure. Next, teams should track the review and official release, and validate interface states against retry, fallback, and manual interruption scenarios. The public evidence does not yet define the full range of affected versions. Review considerations