From Bytes to Semantics: LLM-Agent-Driven Commit Message Generation for Non-Code Assets

Abstract

Modern software projects increasingly version binary assets alongside source code—files that encode semantic logic authored by engineers yet are stored as opaque blobs by version control systems. Such assets appear across diverse domains: gameplay logic in game engines, circuit designs in EDA tools, and scene hierarchies in 3D pipelines. Existing commit message generation (CMG) tools rely exclusively on text diffs and produce no output for these commits, leaving a pervasive blind spot in CMG research and practice. We propose B2S (Bytes-to-Semantics), a general framework that bridges native tool diff APIs with LLM-driven CMG via an Asset Type Registry and a standardized structured differ interface. We instantiate B2S for Unreal Engine Blueprint assets: BlueprintDiff, a custom UE5 plugin exposing native diff output as an MCP server, provides typed structured diffs; a hybrid semantic routing strategy then enriches execution-flow changes with targeted BlueprintSL subgraph context while passing self-contained changes as structured diff entries alone. Evaluated on 300 real-world Blueprint commits across four LLM backends, B2S+Claude achieves BERTScore 0.893 and BLEU-4 9.47, outperforming the filename-only baseline by 0.061 BERTScore. Ablation confirms independent contributions from subgraph context (+0.032 BERTScore) and semantic routing (+0.021). Human evaluation yields Accuracy 4.03, Informativeness 3.57, and Fluency 4.53 (Cohen’s κ = 0.74). The architecture admits another asset type when its native headless differ can map evidence to the shared interface.