I was tired of "Chat with Code" tools that blindly chunk text and hallucinate context. So I built RepoReaper – an autonomous agent designed to mimic the cognitive process of a Senior Tech Lead.
Instead of dumping the whole repo into context, it works step-by-step:
Thinking Like a Pro: It starts by reading the file tree, selects the top 20 critical files based on architecture, and "thinks" about what to read next to understand the project structure.
AST-Aware Chunking: It uses Python's ast module to split code by Class/Method boundaries, preserving logic instead of breaking it mid-function.
ReAct Loop & JIT Loading: If it encounters a function call defined in a missing file, it pauses, triggers a tool to fetch that specific file from GitHub (Just-In-Time), and resumes analysis.
Visual Output: It generates Mermaid diagrams (Architecture/Sequence) and markdown tables to explain complex logic visually.
Hybrid Search: Combines Vector Search (BGE-M3) for semantics and BM25 for exact keyword/function name matching.
High Performance: Built with AsyncIO for high concurrency and uses Persistent Storage (ChromaDB on disk) to save state across sessions.
The Live Demo (Global & China optimized) is available directly in the README.