Find Files

Hey! This one helps you find your way around a codebase you don't know yet.

Ask a question in plain words, like "how is login token refresh handled?", and it gives you a short ranked list of the files most likely to answer it, so you open the top two or three instead of searching blind.

How it works

1. A keyword walk of the repo collects up to 100 candidate files. It skips .git, node_modules, virtual environments, build folders, files over 300 KB, and files with secret-looking names like .env or *.pem.
2. Optional model scoring. A small decision model (OpenJev) scores candidates from short excerpts, keeps the top 10, then re-judges those from longer excerpts. The final order blends both passes.
3. No model? No problem. With no key you get the keyword order, with a note on stderr. It needs no key and no network.

What you get back

  100  auth/token.py
   82  tests/test_token.py
   59  auth/session.py
   29  api/routes.py
   24  README.md

One line per file, percent path, best first.

Using the model, and what leaves your machine

Model mode needs a free Codiv OpenJev API key as OPENJEV_API_KEY, or a small local model set up with jev-tools-setup. In model mode, short excerpts of the candidate files (their opening lines and the lines matching your keywords) are sent to api.codiv.ai, or to a loopback address for a local model. Secret-shaped strings are redacted first, and the key is only ever sent to those hosts. For code you would rather not share, just don't set a key. A full run over 100 candidates costs tens of thousands of input tokens.

What you need

Python 3.10+. Standard library only, nothing to install.

Honest limits

• It is a hint, not an oracle. On 8 labelled queries across 8 cloned repositories, plain keyword counting put the defining file in the top 3 for 4 of 8. The model ranking got 6 of 8 on one run and 4 of 8 on an identical rerun. So it was not measurably better than keywords, and repeat runs differ by up to 2 of 8. Open the top two or three files rather than trusting rank 1.
• READMEs and docs can outrank the code that implements the thing. Skip them when you want code.
• "No keyword candidates" means your words appear nowhere in the repo. Rephrase with identifiers likely to be in the code.
• It appends one small line per model call (timestamp, number of questions, milliseconds, tokens, or an error type) to ~/.jev-tools/log.jsonl, or to the path in JEV_LOG. File contents and keys are never logged.

Part of jev-tools

This skill is one of several from the free, MIT-licensed jev-tools project. The full plugin adds optional hooks that this listing does not include: github.com/Rcidshacker/jev-tools. This skill works fine without it.