AI Framework for Legacy Code Specification Released

Researchers from the Federal University of Goiás in Brazil and other institutions have released Reversa, a framework that converts legacy software into operational contracts that AI coding agents can utilize. This development addresses the challenge where existing systems lack specifications and knowledge is trapped within the code, making it difficult for AI agents to safely modify or extend them.

Reversa generates executable contracts that specify behaviors to preserve, evidence, gaps, and confidence levels, rather than descriptive documents for human readers, enabling agents to evolve systems while preserving existing behavior. Notably, following prior research indicating that incorrect documentation can hinder LLM code understanding, the design adopts a confidence rating for all claims and explicitly marks uncertainties.

Multi-Agent Pipeline

Reversa analyzes and specifies code through a five-stage Discovery pipeline.

• Reconnaissance: Mapping the project surface, including folder structure, languages, and dependencies
• Excavation: Deep analysis of algorithms, control flow, and data structures at the module level
• Interpretation: Extracting business rules, architectural decisions (ADRs), and state machines, and generating C4 diagrams
• Generation: Writing specifications in the form of operational contracts with code traceability
• Review: Verifying contradictions in specifications and identifying gaps

Research Significance and Contributions

The paper positions Reversa at the intersection of classical reverse engineering, LLM-based documentation, requirements generation, and software agent research. While previous studies primarily dealt with documents consumed by humans or existing requirements, Reversa treats the legacy system itself as the primary source to reverse-extract the structured specifications needed by AI agents. Its utility was validated through a case study of migration from COBOL to Go, and it offers portability by avoiding specific tool dependencies and enabling integration with various coding agent engines.