BreakScan

Developers break things when they change code and don't know what else is affected. Static analysis tools catch import errors, they don't catch a model trained on 10 features receiving 9 at inference, test mocks that pass but miss production failures, or an API endpoint two hops away from the changed function. BreakScan solves this using IBM Bob's full-repository reasoning. The tool scans the entire codebase using Python AST to build a bidirectional dependency map, reads AGENTS.md architectural rules, and injects full file context into a structured prompt for Bob. Bob reasons across the whole repo and returns a ranked impact report. BreakScan parses the response and renders color-coded cards: direct crashes, indirect behavioral failures, and confirmed safe files. Demonstrated on DemandSense. a production ML forecasting system with XGBoost, Isolation Forest, 60 trained models, and a React frontend. Removing one feature column produced 8 direct breaks, 9 indirect failures, and 20+ confirmed safe files instantly.