CodeGraph

At UW, student developers constantly inherit codebases, from hackathon projects and research labs to club infrastructures and open-source tools. Yet onboarding onto any new remains painful. New contributors spend hours searching through unfamiliar files, guessing how the logic works, fearing unintended breakage, or just not knowing how anything works. We asked, what if onboarding onto a codebase didn't have to be so painful?

We introduce Codegraph, a Forge-powered platform that takes in any GitHub repository and generates an interactive Directed Graph (DiGraph) representing the true semantic structure of the codebase inspired by systems often used in Personal Knowledge Management and Second Brain Systems. .

Our system uses Gemini models to do the following:

• Parse and classify functions, classes, and architectural “blocks”
• Detect inter-file and intra-file dependencies and logical groupings
• Surface high-level flows (e.g., routing → controllers → utilities → I/O)
• These are rendered into a navigable knowledge graph, essentially like an intelligent Jira layer on top of real code.

We learned the following:

• How to leverage Atlassian Forge beyond UI extensions by integrating numerous API's and finding new ways to visualize data.
• The complexity of program analysis with LLM agents, especially differentiating syntax from actual semantics.
• Beyond task management, it is very important in collaboration to help people understand the systems they’re working on.

Challenges:

• Semantic Code Parsing: LLM agents can hallucinate dependencies unless guided with strict prompting and grounding, and the larger the model context as with the case of larger repositories, they more likely this would happen.
• Cross-File Linking: Reconstructing architecture across files required building our own DAG abstraction layer, and required significant prompt engineering to achieve the proper categorizations.
• Performance on Large Repos: We had to batch/analyze repos incrementally to avoid timeouts and token crashes as well as staying with Model Context windows.
• Visuals UX: Turning raw graph data into something usable and palatable requires careful thinking about design
• Rovo Agent Integration: Having data is really cool and all, but its important especially with Jira to have said data lead to actions that can actually improve

We believe this tool can transform how students collaborate, turning every codebase at UW into a live and collaborative map rather than a black box of jargon. From onboarding new club members to explaining CS capstone architecture, it makes it far easier for others to easily understand your code and join you.

In the future, we would like to implement direct Jira and Rovo integration with no need for user intervention, especially when Forge allows custom UI to invoke Rovo actions. We also want to be able to implement a scoring mechanism for code that measures contribution rather than just basic metrics like commits and lines.

Loom video link: https://www.loom.com/share/cfcd3444ca2c49068d64a291f4db4d08?sid=5b5b294f-6850-492d-8a46-df883181a596

Built With

• forge
• gemini-api
• github-api
• javascript
• node.js
• react