Specification Portability Across LLM Development Agents: Cross-Agent Compatibility in Specification-Driven Software Migration
Specs moved between coding agents can break badly, even when the migration task is controlled.
The paper tests Oracle-to-PostgreSQL migration specs across Amazon Kiro, Google Gemini, GitHub Copilot, Claude Code, and Cursor. In one replicated case, Gemini using a Kiro-origin spec fell to 0.035 Token F1, 2.33% SQL syntax validity, and 0.015 AST mean similarity. Rewriting specs helped Gemini in the tested setup, while compression was not consistently useful. Retrieval-augmented ingestion was the only shared strategy on both Gemini and Copilot Pareto frontiers. Source: ArXiv · AI/CL/LG's note
The paper tests Oracle-to-PostgreSQL migration specs across Amazon Kiro, Google Gemini, GitHub Copilot, Claude Code, and Cursor. In one replicated case, Gemini using a Kiro-origin spec fell to 0.035 Token F1, 2.33% SQL syntax validity, and 0.015 AST mean similarity. Rewriting specs helped Gemini in the tested setup, while compression was not consistently useful. Retrieval-augmented ingestion was the only shared strategy on both Gemini and Copilot Pareto frontiers. Source: ArXiv · AI/CL/LG's note
score 4