Contact

← Projects · Sep 2026

repo-ask

Ask a codebase a question and get an answer with file:line citations, or an honest “not found”: every citation is checked against the code it actually read.

repo-ask is a small command-line tool for finding your way around an unfamiliar repository. It splits the code into one chunk per function or class with tree-sitter, retrieves the most relevant chunks with BM25 and ripgrep, and asks Claude to answer from those chunks only. Every citation is then checked against what was actually retrieved; if none survives, it prints “Not found in this repo” instead of guessing.

There is no server and no vector database: it is about 700 lines of Python, and the index is a JSON file in a cache folder, never in the repo. Only the six retrieved chunks are sent to the API, and files that .gitignore excludes, such as .env, are never indexed.

On Soliton, a TypeScript and JavaScript codebase of about 280 files, it answered 8 of 10 held-out questions correctly and declined both questions whose answer wasn’t in the code, with no rejected citations. The question sets are small, so I read these numbers as indicative. I built it alone, starting in September 2026.

A terminal running repo-ask on the Soliton codebase: one question answered with two cited line ranges in runner.mjs and a short explanation, and a second question, about a Kafka consumer the code doesn’t have, answered “Not found in this repo”.
Two questions about Soliton: one answered with its sources, one it refuses to guess.
repo-ask’s pipeline: the repo is split into one chunk per function with tree-sitter and indexed once; each question goes to BM25 and ripgrep, fused into the top six chunks; Claude answers from those only, and a citation check either passes the answer with path:start-end or returns Not found.
How it works: retrieve six chunks, answer from those only, and drop any citation outside them.