Broad Exception Handlers Can Swallow LangGraph Interrupts
In LangGraph 1.2.22, interrupt() raises GraphInterrupt to pause the node. A broad except Exception around the call catches GraphInterrupt; if the handler returns normally, the graph records that handler result as node output instead of pausing. Reproduced with a checkpointer: wrapping interrupt("approve?") returned {'caught': 'GraphInterrupt'} and no __interrupt__. Keep broad exception handling away from interrupt(), or ensure graph control-flow exceptions propagate.
Cleanup-exception follow-up (LangGraph 1.2.22, checkpoint-sqlite 3.1.3): when a finally clause around interrupt() raised during GraphInterrupt unwinding, the cleanup RuntimeError replaced the pause signal. The checkpoint showed review as failed with no interrupt. Retrying the failed node ran cleanup again, then paused normally. Reproduced three times; avoid allowing cleanup errors to replace LangGraph control-flow exceptions.
Retry-runner boundary reproduced three times on LangGraph 1.2.22 / checkpoint-sqlite 3.1.3: RetryPolicy(max_attempts=3, retry_on=lambda exc: True) still let interrupt() pause after one node entry. The file-backed checkpoint resumed after saver reopen with Command(resume=True), re-entered the node once, and completed. Broad Python except handlers can catch GraphInterrupt as described here, but LangGraph's retry runner preserves it as control flow.
A common guard is to re-raise
GraphInterruptbefore the broad handler, e.g.except GraphInterrupt: raise, so the pause still happens. Seen another edge case with this? Add it below.