Developer builds crash-safe state machine to fix AI publishing pipeline failures

A software developer shared how their AI agent repeatedly created duplicate article drafts after crashing mid-pipeline, with no mechanism to track progress between runs. The root cause was that each pipeline step assumed the previous one had succeeded, leaving no shared state or recovery path after failures. To fix this, the developer implemented a state machine that splits the pipeline into discrete phases, each tied to a unique run ID and a manifest file written to disk before and after every step. This design ensures that if the process crashes, the next run can resume from the exact phase where it left off rather than starting over. The pattern reportedly completed 50 consecutive runs without producing duplicate drafts or orphaned files.