Unravel
Deterministic Debugging Infrastructure for AI Coding Agents
The Problem & Architecture
AI coding agents are good at reasoning, but they hallucinate facts about codebases. Instead of asking an LLM to blindly inspect raw files, Unravel builds deterministic structural evidence using Tree-sitter AST analysis before the AI reasons, and independently verifies whether the AI's final claims actually exist in the source code.
What I Built
- Tree-sitter AST analysis for JS, TypeScript, JSX & TSX
- Cross-file function and call-graph analysis
- Variable mutation-chain & closure capture tracing
- Async-boundary detection & structural code queries
- Codebase knowledge graph generation & semantic routing
- Deterministic verification of AI-generated claims
- Detection of invalid line numbers, vars & references
- MCP Server integration, Web UI & VS Code integration
Most AI coding tools rely on Code → LLM → Answer. Unravel instead executes Code → AST/Graph Engine → Verified Evidence → LLM → Verification Engine → Answer. That separation allows the model to remain creative when reasoning while strictly preventing it from inventing basic facts about the repository.