Shadow Traffic Replay Catches AI Code Regressions That Unit Tests Miss

AI-suggested code changes can pass all unit tests and static reviews yet still fail under real-world request patterns, such as burst traffic or edge-case data states. A more reliable verification method involves deploying the candidate change to a shadow service and replaying a full day of actual production traffic against it. The approach compares status codes and response bodies between a baseline and candidate service for each replayed request, producing a reproducible mismatch list. A sample Python harness is provided that reads a JSONL request log, skips write methods to avoid side effects, and strips sensitive headers before forwarding requests to both services. Because the shadow server is disposable and does not need to be production-grade, the cost of running frequent comparisons remains low.