GHA-SmellFix: Agentic LLMs for Detecting and Refactoring GitHub Actions Workflow Smells
Story 15, from arXiv
GHA-SmellFix is a new tool that uses agentic LLMs to detect and refactor GitHub Actions workflow smells. It compares two designs—joint and decomposed—and finds that decomposition helps control when patches are generated but doesn't universally improve quality.
Key points
- GHA-SmellFix uses agentic LLM pipelines to detect and refactor GitHub Actions workflow smells.
- Two designs are compared: joint (one agent detects and refactors) and decomposed (separate agents for detection and refactoring).
- Decomposition helps control when patches are generated, reducing overhead for models prone to missing smells.
- The tool operationalizes 19 smells with explicit detection criteria and safety checks, using 1,710 (workflow, smell) pairs for evaluation.
- Results show that detection is the main bottleneck, and false positives can lead to unnecessary edits.
- Four LLMs (GPT-5.4-nano, Gemini 3.5 Flash, Claude Haiku 4.5, and DeepSeek V4 Flash) were tested on 570 sampled pairs.
Why it matters
As GitHub Actions workflows become more complex, automated tools that detect and fix configuration smells are critical for maintaining security, reliability, and maintainability. This work provides a new approach using LLMs that could improve CI/CD pipelines with smarter refactoring.
Counterpoint
False positives can lead to unnecessary changes in execution-relevant constructs, requiring careful validation and abstention mechanisms to avoid destabilizing workflows.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.