RECODE: Requirements Recovery from CODE for Adoption Review of Autonomous Driving Software

Abstract

Open-source software (OSS) is increasingly adopted in safety-critical autonomous driving (AD) systems, yet adoption is blocked not by functional deficiency but by the absence of adoption evidence: requirements specifications, safety constraints, and traceable linkages to the implementation. This paper frames this gap as requirements debt and proposes RECODE, a hierarchical multi-agent large language model (LLM) pipeline that recovers reviewable requirements artifacts and safety review evidence from under-documented AD repositories. Six specialized agents, organized in three phases (feature recovery, software requirements specification (SRS) construction, and adoption review) decompose the task with iterative verification and inspectable intermediate artifacts. The pipeline is evaluated on two Autoware Universe modules (autonomous emergency braking, AEB; multi-object tracker, MOT) against a monolithic single-agent baseline receiving identical task definitions, using a mutation benchmark covering implementation-type (I-type) defects and omission-type (O-type) gaps alongside practitioner assessment. RECODE detected 30 of 36 injected mutations (Detection Rate 83.3%) compared to 17 for the baseline (47.2%), with the largest gap in O-type detection (68.8% vs. 25.0%); practitioners rated RECODE outputs $2.61 / 3$ vs. $1.92 / 3$ for the baseline. The findings support the position that AD OSS adoption review is fundamentally a reverse requirements engineering problem and provide a practical workflow for addressing it. Supplementary material is available at https://github.com/ailab-hanyang/RECODE.