January 7, 2026
Cache me outside
Show HN: RepoReaper – AST-aware, JIT-loading code audit agent (Python/AsyncIO)
Dev tool promises to 'chat with code'—crowd cheers, skeptics roast
TLDR: RepoReaper claims to smartly read code and fetch files on demand, not just guess. The crowd is split: fans say it’s like a senior engineer in a box, while skeptics demand proof, worry about privacy, and gripe about demo rate limits—important if you want AI that truly understands large projects.
RepoReaper crashed the party with a skull emoji and a big claim: an AI agent that actually “talks to your code” and hunts for answers in real time. Fans on the Show HN post swear it feels like a staff engineer in a box, saying it builds a smart map of your project, warms up a cache of key files, and when it gets stumped, it fetches the exact file it needs instead of guessing. Skeptics rolled in fast, calling it “just fancy search wearing a cape,” asking for proof it finds real bugs in messy, giant codebases. Privacy alarm bells rang too: do we really want an autonomous bot rummaging through private repos? Supporters shot back that local deploy solves that, while others grumbled about demo rate limits and the “use SEOUL server” note for China—cue memes of the “Repo Grim Reaper” stuck behind a 429 error. The nerdy brawl over acronyms was surprisingly fun: AST (a code map) got love, RAG (a memory cache for the AI) got rebranded as “L2 brain,” and BM25 (keyword match) vs vectors (meaning match) sparked the old “grep vs vibes” showdown. Bottom line: believers say it’s a real step beyond static indexing; critics want benchmarks, not buzzwords.
Key Points
- •RepoReaper is an autonomous code-audit and semantic search agent that analyzes repositories using AST parsing and dynamic retrieval.
- •The system treats the LLM as the CPU and a vector store as a dynamic L2 cache, with cold-start AST mapping, prefetching of key files, and JIT reads on cache misses.
- •AST-aware semantic chunking preserves class/method boundaries and injects parent context to improve retrieval and understanding.
- •An async pipeline using asyncio/httpx and deployment with Gunicorn/Uvicorn enables scalable, non-blocking ingestion with shared ChromaDB for multi-worker service.
- •A ReAct loop rewrites queries to precise English, self-corrects via tool commands to fetch files through the GitHub API, and uses hybrid BM25 plus BAAI/bge-m3 vector embeddings for retrieval.