Beyond Smell Detection: Construct-Aware Evaluation of Agentic LLM Pipelines for GitHub Actions Workflow Refactoring

Existing approaches for GitHub Actions configuration smells are mostly rule-based and detection-oriented, leaving developers to interpret findings and write the repairs themselves. We report an empirical study of lightweight agentic LLM pipelines that support both target-smell detection and scope-bounded refactoring. Building on an existing catalog, we operationalize 19 smells with explicit detection criteria, refactoring expectations, exclusion rules, and construct-level safety checks, and compare two configurations: a joint detect-and-refactor agent and a decomposed detect-then-refactor pipeline. Both are instantiated with four LLMs and evaluated on 570 shared annotated (workflow, smell) pairs sampled from a manually curated ground truth of 1,710 pairs. Task decomposition yields modest, model-dependent detection gains (average F1 86.34% → 87.55%, recall 79.12% → 82.37%). End-to-end refactoring is limited mainly by missed detections: 78.2% of assessable positive outputs remove the target smell and 71.9% are clean removals, but conditioned on correct positive detection both rates rise sharply. Negative pairs reveal over-refactoring — 93 unnecessary edits, 54.84% of which touch behavior-relevant workflow constructs. Agentic LLM pipelines can produce useful candidate CI/CD maintenance patches, but practical use requires abstention checks, construct-aware validation, and cost-aware orchestration.