SmartRefactorer

AI agent that uses GitLab Orbit to analyze code changes, predict their impact, and help developers safely refactor code without breaking things.

Comment

Inspiration Every developer has experienced the anxiety of making a seemingly small code change, only to discover it broke something unexpected in production.This fear often leads to overly cautious development, slow refactoring, and accumulating technical debt. I wanted to solve this problem by giving developers a tool that could predict the impact of their changes before they even write a line of code.

What it does SmartRefactorer is an AI agent that uses GitLab Orbit's structured knowledge graph to analyze the "blast radius" of any code change. When a developer asks about modifying a file, module, or function, SmartRefactorer:

  1. Identifies the target- Clarifies exactly what needs to change
  2. Queries GitLab Orbit - Finds all dependencies, importing files, recent merge requests, and contributors
  3. Analyzes the Blast Radius- Determines which files have the highest risk of breaking
  4. Provides a structured report - Delivers risk level, impacted files, suggested tests, recommended reviewers, and precautionary steps The agent transforms developer intuition into data-driven confidence, making large-scale refactoring safer and more predictable. How we built it

Technologies Used:

  • GitLab Orbit - The core knowledge graph that provides structured, queryable access to the codebase
  • GitLab Duo Agent Platform - The framework for creating and hosting custom AI agents
  • GitLab Duo Agentic Chat - The interface where developers interact with SmartRefactorer
  • Termux (Android) - The development environment used to build and test the agent Development Process:
  • Set up the development environment on an Android device using Termux
  • Installed the GitLab CLI (glab) to access GitLab Orbit via the terminal
  • Authenticated with GitLab and tested Orbit connectivity
  • Created a custom agent in the GitLab AI Catalog with a carefully crafted system prompt
  • Designed the system prompt to guide the agent through a structured analysis workflow
  • Published the agent as Public in the AI Catalog
  • Tested the agent with real queries on the eco-scan project Key Design Decisions: The system prompt was the most critical component. I structured it to:
  • Define the agent's role as an expert software architect
  • Provide clear, step-by-step instructions for analysis
  • Emphasize the use of GitLab Orbit for real data
  • Deliver actionable, structured reports (not just generic advice) Challenges we ran into Technical Challenges:
  • Installing glab on Termux - The package wasn't initially available in the default repositories. I had to use the glab-cli package and ensure proper authentication.
  • Orbit Accessibility - Testing Orbit queries required the right permissions and scopes. I had to generate a personal access token with api and read_api scopes.
  • Agent Response Formatting - Getting the agent to produce structured, consistent output required iterative refinement of the system prompt.
  • Large Project Analysis - When analyzing large codebases like gitlab-org/gitlab, Orbit queries could take significant time to return results. Resource Constraints:
  • Building and testing entirely on an Android device using Termux
  • Limited screen space for development and testing
  • Managing multiple browser tabs for GitLab, Devpost, and documentation Accomplishments that we're proud of ✅ Created a fully functional AI agent*that meaningfully uses GitLab Orbit ✅ Published the agent as Public in the GitLab AI Catalog ✅ Demonstrated real-world impact by analyzing an actual project (eco-scan) ✅ Built the entire project on an Android phone - proving that hackathons are accessible to everyone ✅ Designed a professional system prompt that delivers structured, actionable reports ✅ Successfully integrated with GitLab Orbit - showing the power of knowledge graphs for software development ✅ Created a solution that solves a genuine developer pain point What we learned Technical Takeaways:
  • GitLab Orbit is incredibly powerful for understanding codebase relationships
  • The knowledge graph approach (nodes and edges) makes complex queries intuitive
  • AI agents are only as good as their system prompts - clear instructions yield better results
  • Building on mobile devices is viable with the right tools (Termux + GitLab CLI) Soft Skills:
  • The importance of storytelling in hackathon submissions
  • How to demonstrate value through structured responses
  • The value of focusing on a specific, well-defined problem GitLab Platform Insights:
  • The AI Catalog makes agent discovery and publishing seamless
  • Duo Agentic Chat provides a natural interface for developers to interact with agents
  • Orbit's schema covers all the essential entities for software development analytics What's next for SmartRefactorer Feature Enhancements:
  • Visualization Dashboard- Create a web interface showing dependency graphs and blast radius reports visually
  • CI/CD Integration - Automatically run SmartRefactorer on every merge request to flag high-risk changes
  • Custom Reports - Support for generating specific reports like "Security Impact" or "Test Coverage Impact"
  • Visualization - Integrate with the GitLab UI to highlight impacted files directly in merge requests
  • Team Insights - Provide metrics on code ownership and expertise to suggest reviewers Expansion Plans:
  • Multi-Project Analysis - Extend to analyze cross-project dependencies in GitLab groups
  • Historical Analysis - Learn from past changes to improve risk predictions
  • IDE Integration- Build plugins for VS Code and other IDEs
  • Team Insights - Provide metrics on code ownership and expertise to suggest reviewers Community Goals:
  • Open Source- Share the agent template so other developers can create similar tools
  • Contributor Friendly - Add documentation to help others contribute
  • Hackathon Submissions - Use SmartRefactorer to help other hackathon participants understand their codebases

Built With

Share this project:

Updates