From Legacy to Reactive: Interpreting Source Code into Requirements for Automated App Generation

Abstract

Legacy systems built with Oracle Forms still support core business operations in many organizations, yet their migration to modern microservice platforms is delayed and hampered by the fact that critical requirements and business rules are embedded in PL/SQL trigger code and form metadata rather than in explicit documentation. Manually recovering these requirements is slow, costly, and error-prone. This dissertation proposes an automated, multi-agent pipeline that reverse-engineers Oracle Forms applications into structured software requirements. Each form's XML export is parsed into an Abstract Syntax Tree that retains only the necessary information, while the database schema and PL/SQL packages are extracted from the accompanying DDL. These artifacts are merged into a Unified Application Behavior Graph, used both to cluster the forms into candidate microservices and to provide context to the LLMs. A deterministic depth-first traversal then collects the relevant context for a Writer agent - a large language model equipped with tools - which generates functional requirements, business rules, integration requirements, and pre/post-conditions in structured JSON, while a Verifier agent checks each requirement for grounding, coverage, and consistency against the source. The approach was applied to a real project within the company. The extracted requirements were externally validated through a questionnaire completed by the application's developers and aligned with the ISO/IEC/IEEE 29148 quality characteristics, using stratified sampling and Krippendorff's α as the inter-rater reliability metric.