Jev Code Finder

Fast semantic code search powered by TypeSafe Jev. Describe a concept in plain English and get the relevant files, line ranges, confidence scores, and source snippets.

$ jev-code-finder "where does user JWT authentication happen?"

src/auth/middleware.rs:81-180 (92%)
pub async fn authenticate(...)

Unlike text search, Jev Code Finder searches for meaning. A query such as where are permissions enforced? can find middleware, guards, policy checks, and configuration without requiring you to know their exact names.

Install

Install with Homebrew:

brew tap Peu77/tap
brew install --HEAD jev-code-finder

Install the Agent Skill for coding-agent workflows:

npx skills add Peu77/JevFind --skill jev-code-finder

How it works

1. Walks the repository while respecting .gitignore and excluding hidden, dependency, cache, and build directories.
2. Asks Jev to score each source path for likely relevance to your query.
3. Opens only paths that meet --file-threshold and divides them into overlapping, size-bounded code windows.
4. Sends those windows in bounded request batches, then prints matches meeting --threshold.

The path filter keeps unnecessary source code out of later requests and makes large searches cheaper. Set --file-threshold 0 when recall matters more than speed.

Transient network errors, HTTP 408/429 responses, server errors, and invalid responses are retried up to three times with exponential backoff: 250 ms, 500 ms, then 1 second. Permanent client errors fail immediately.

Requirements

• A TypeSafe API key
• Rust 1.88 or newer; the included rust-toolchain.toml selects the correct version when using rustup
• Network access to https://api.typesafe.ai

Install

Clone the repository, then build and install the binary:

cargo install --path .

Set your API key in the environment:

export TYPESAFE_API_KEY="your-api-key"

Do not commit the key or pass it as a command-line argument.

Usage

Search the current directory:

jev-code-finder "where does user JWT authentication happen?"

Search a specific directory:

jev-code-finder "where are database transactions committed?" --path ./backend

Require stronger confidence for both file selection and final matches:

jev-code-finder "rate limiting" \
  --file-threshold 0.40 \
  --threshold 0.75 \
  --parallelism 10

Run without installing:

cargo run -- "how are failed payments retried?" --path .

Live debug tree

Use --debug to display a colored file tree that updates in place:

jev-code-finder "authorization checks" --debug

The status symbols are:

Symbol	Meaning
○	Pending or queued
●	Jev is scoring or scanning the file
×	Path relevance was below --file-threshold
·	File was scanned but no window met --threshold
✓	At least one matching window was found
!	File could not be read as UTF-8 text

Completed files show their path-relevance percentage. Matches additionally show the strongest matching-window percentage.

Options

Option	Default	Description
<CONCEPT>	Required	Natural-language concept to locate
-p, --path <PATH>	.	File or directory to search
--file-threshold <0..1>	0.25	Minimum path relevance before opening a file
-t, --threshold <0..1>	0.55	Minimum probability for printing a code window
--model <MODEL>	jev-latest	TypeSafe model alias
--debug	Off	Show the live colored file tree
--hidden	Off	Include hidden files and directories
--parallelism <N>	10	Maximum concurrent file-scanning requests
-h, --help	—	Show command help

Thresholds use decimal probabilities: 0.25 means 25% and 0.80 means 80%.

Files considered

Supported extensions:

rs  js  jsx  ts  tsx  py  go  java  kt  kts  rb  php
c   h   cpp  hpp  cs  swift  scala  sh  sql
yaml  yml  json  toml

The walker honors repository, nested, global, and .git/info/exclude Git ignore rules. Hidden paths are excluded unless --hidden is supplied.

These directories are always skipped:

.git  .next  .venv  __pycache__  build  coverage
dist  node_modules  out  target  vendor  venv

Output

Each match contains the path, one-based line range, probability, and complete matching window:

./src/auth.rs:81-180 (92%)
fn verify_jwt(token: &str) -> Result<User, AuthError> {
    // ...
}

When no window meets the threshold:

No matches found for: where does user JWT authentication happen?

Accuracy and privacy

• File selection is probabilistic. A high --file-threshold is faster but can exclude unexpectedly named files; lower it or use 0 for a thorough scan.
• Results are overlapping code windows, not exact AST or function boundaries, so neighboring results may contain duplicate lines.
• Large and minified files are split across multiple requests instead of being sent as one oversized payload.
• Path-selection batches run sequentially; selected files are scanned concurrently up to --parallelism.
• Matching source code and file paths are sent to the TypeSafe API. Review TypeSafe's data policies before searching private or sensitive repositories.

Development

Run the test suite:

cargo test

Build an optimized binary:

cargo build --release