From Scattered Requirements to Maintainable BDD Specifications: A Multimodal Agentic Pipeline

Replication package for "From Scattered Requirements to Maintainable BDD Specifications: A Multimodal Agentic Pipeline" This repository contains the Gherkin feature corpora generated by a multi-agent pipeline that converts scattered requirements (specification documents, screen-capture videos, and web-crawler output) into maintainable BDD specifications, along with the maintainability measurements, comparison reports, and tooling used to produce and evaluate them. Contents: Feature sets, source materials, and measurements for three projects: Joomla 4.2, Lamassu (IoT), and Spring PetClinic. Feature-selection data and metrics for the ablation study. Cross-project comparison of four LLM models on Gherkin maintainability, plus a detailed per-project breakdown and ablation comparison for Claude Sonnet 4.6. A formal definition of all automated maintainability metrics used (SPC, OAR, BDI, ASL, FMI, SSI). A catalog of good and bad BDD practices used to guide generation. The agent definitions driving the generation pipeline, and the analysis skill used to compute the metrics and produce the Excel/JSON outputs. Note: one of the four evaluated projects (NextPaths) is proprietary to the industrial partner LKS Next; its source materials and feature files are not included, only its aggregate metrics in the comparison documents. The proof-of-concept implementation is available at: https://github.com/lksnext-ai-lab/spec2test