Crash-Looping AI Pipeline Bug Traced to Truncated JSON Line, Not Model Failure

An automated workflow began crash-looping after six hours when its apply stage repeatedly failed to parse a truncated JSON line in a ledger file, causing the supervisor to restart the process in an endless cycle. The root cause was not model degradation, as manual prompt replay returned valid JSON, but a boundary condition where the writer appended raw, token-budget-truncated model output without validation. The apply stage lacked error handling for malformed lines, so every restart re-read the same corrupted entry and exited with a JSONDecodeError. Debugging was carried out by reproducing the loop in an isolated Docker environment, then inspecting line lengths in the JSON Lines ledger file to quickly identify the anomalously short final entry. The incident highlighted two pipeline design gaps: the writer needed output validation before appending, and the apply stage required a guard to handle or skip unparseable lines rather than exiting on failure.