A system internals modeling and annotation language for large language model-driven software engineering

Abstract

Rapid scaling of large language models has significantly disrupted traditional software engineering paradigms with unprecedented capabilities for code understanding, generation, and automated review. However, practical repository-level deployment is constrained by context limits and token cost. While retrieval-augmented generation is widely used to address this limit, it often splits a codebase into disconnected semantic chunks, omitting high-level structural dependencies. In contrast, alternative approaches that attempt to feed the entire repository structure into the context window typically rely on generic formats such as JSON. While widely recognized, these formats introduce redundant syntactic units that significantly influence the token budget, effectively reintroducing the bottleneck. This work introduces SiMAL (System internals Modeling and Annotation Language), a domain-specific language designed specifically for language-model-driven software engineering workflows, with the primary objective of providing a compact, human-readable, error-tolerant, yet highly structured representation of a software system, optimized for iterative interaction with language models. It combines both static and dynamic aspects of a software system, unifying architectural views, component and endpoint definitions, runtime deployment metadata, and other development artifacts into a single textual schema that can be converted to and from a normalized machine representation. This effort includes the definition of the language’s syntax and grammar, an open-source parser, and a visualizer that renders schemas into nested system diagrams. The proposed language is validated through a comprehensive protocol that assesses token efficiency alongside schema validity and semantic faithfulness. This includes deterministic structural checks (parsing integrity and annotation consistency), schema-to-code correspondence analysis, and an “LLM-as-a-judge” evaluation for repository coverage. The results indicate that prompt-efficient schema modeling reduces token overhead without systematically degrading structural usability or quality, making it a practical backbone for scalable autonomous software engineering.