Towards LLM-Based Context Engineering for Model-Driven Vibe-DevOps-Engineering of Cross-Platform CI/CD Pipeline Support

Abstract

Existing model-driven DevOps automation either depends on handcrafted transformations that do not scale to new CI/CD platforms or on prompt-based LLM generation that lacks determinism and reliability for production. This study introduces a proof of concept for vibe DevOps engineering, where LLM-based context engineering synthesizes model-driven transformation chains for CI/CD pipelines from existing MDE artifacts and platform documentation. Metamodels, ATL rules, Acceleo templates, and CI/CD specifications are decomposed and organized to prepare the ground for a future Graph-RAG and multi-agent architecture that will derive platformspecific DevOps artifacts from a platform-independent DevOps metamodel. In the current phase, the roles of the envisioned agents are emulated by a human researcher executing the chain of M2M2M2C steps and Eclipse-based runtime testing. Crossplatform constraint transfer reduces total refinement rounds from 11 (GitLab CI/CD) to 6 (Atlassian Bamboo) to 5 (Azure DevOps) across successive runs, and assess syntactic correctness, semantic alignment, and executability, indicating potential and current limitations of vibe DevOps engineering.