Benchmark-Driven LLM Translation of Production Codex CLI: 648K LOC Rust to 41K LOC Python

arXiv·high signal

Wang and Sengupta present a methodology for LLM-assisted continuous cross-language migration, translating OpenAI's production Codex CLI agent from Rust (648K LOC, 65 crates) to Python (41K LOC, 28 modules) using public agent benchmarks as the objective function. The paper demonstrates that benchmark-driven iterative refinement produces a Python superset that passes the same agent evaluations as the Rust original — the largest documented LLM-assisted production codebase translation to date.