RootCause

"Why did I write this code?" — answered with real evidence

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of## RootCause: Tracking the "Why" Behind Your Code

"Every line of code has a story — RootCause helps you remember it."

Inspiration

As developers increasingly rely on AI assistants like ChatGPT and Claude, a critical problem emerges: we lose the context of why we wrote our code.

Six months later, you look at a function and think: "What was I trying to solve here?" The AI conversation is buried in browser history, disconnected from the code it helped create.

RootCause was born from this frustration — the need to preserve developer intent and create a traceable link between AI prompts and the code they inspire.

What I Learned

Building RootCause taught me:

  • Browser Extension Development — Manifest V3, content scripts, and cross-origin messaging
  • VS Code Extension API — Workspace events, status bar integration, and command registration
  • Local-First Architecture — SQLite with WAL mode for concurrent access
  • Semantic Matching — Using embeddings to find conceptual similarity between text

The most fascinating part? Designing the multi-signal matching algorithm that intelligently links prompts to files.

How I Built It

RootCause is a monorepo with four integrated components:

Component Tech Stack
Backend Express.js, TypeScript, SQLite, Xenova Transformers
Frontend React 18, ReactFlow, TailwindCSS, Vite
VS Code Extension TypeScript, VS Code API
Chrome Extension Manifest V3, Content Scripts

The Matching Algorithm

The core innovation is a weighted multi-signal scoring system:

$$ \text{confidence} = 0.25T + 0.45S + 0.20C + 0.10L $$

Where:

  • \( T \) = Temporal score (prompt must precede save)
  • \( S \) = Semantic similarity (via local AI embeddings)
  • \( C \) = Contextual score (file path matching)
  • \( L \) = Language alignment score

Temporal Decay

Prompts lose relevance over time using exponential decay:

$$ T(t) = e^{-0.0008t} $$

Where \( t \) is seconds since the prompt. This means:

  • At \( t = 0 \): score = 1.0
  • At \( t = 300 \) (5 min): score ≈ 0.78
  • At \( t = 900 \) (15 min): score ≈ 0.49

Challenges Faced

1. Capturing AI Prompts Without Breaking Sites

ChatGPT and Claude have complex, dynamic DOMs. I had to carefully observe DOM mutations and identify the right elements without interfering with the user experience.

2. Cross-Component Communication

Synchronizing session state between VS Code, Chrome, and the backend required a polling-based approach with a dedicated sync endpoint on port 9847.

3. Local Embeddings Performance

Running transformer models locally (via Xenova) was slow initially. I implemented caching and lazy loading to make it practical for real-time use.

4. Privacy by Design

Every architectural decision prioritized privacy: no cloud, no telemetry, no source code capture — only metadata and prompts stay local.

🎯 Key Features

  • Automatic prompt capture from ChatGPT & Claude
  • File save tracking in VS Code
  • Smart matching with semantic understanding
  • Interactive knowledge graph visualization
  • Timeline view of your work progression
  • 100% local — your data never leaves your machine

What's Next

  • [ ] Support for more AI platforms (Gemini, Copilot Chat)
  • [ ] Git commit integration with AI context
  • [ ] Team sharing with encrypted sync
  • [ ] "Explain this code" feature using stored provenance

RootCause — Because understanding why is just as important as understanding how.

What we learned

What's next for RootCause

Built With

Share this project:

Updates