Testing

make                 # fmt check, clippy, unit and end-to-end tests
make real-world      # scenarios against registry.npmjs.org and GitHub
docker/windows/test.sh   # the whole suite built for Windows, run under Wine
make bench           # race craft against npm, pnpm and Yarn

Unit tests

Each crate has unit tests next to its code. They never touch the network or the real home directory: everything runs in temporary directories with an explicit Context.

End-to-end tests

apps/craft/tests/e2e/ runs the real craft binary against MockRegistry (crates/testkit), an npm-compatible registry on 127.0.0.1 that serves packuments and tarballs built in memory from PackageFixtures:

#[test]
fn installs_dependencies_and_their_subtree() {
    let registry = MockRegistry::start();
    registry
        .publish(PackageFixture::new("ms", "2.1.3"))
        .publish(PackageFixture::new("debug", "4.4.0").dependency("ms", "^2.1.3"));
    let project = Project::new(&registry, json!({ "dependencies": { "debug": "^4.0.0" } }));

    project.craft().arg("install").assert().success();

    assert_eq!(project.required_version("debug"), Ok("4.4.0".to_string()));
}

• Each test gets its own project directory and craft home, so tests run in parallel and leave nothing behind.
• Fixtures have real integrity hashes, and their default index.js requires their dependencies, so project.require(…) under Node proves the tree resolves.
• The mock registry can require tokens, fail requests with chosen statuses and Retry-After, send max-age, serve over TLS with a self-signed certificate, act as an HTTP proxy, and record request paths, headers and concurrency.
• Scripts in fixtures use node -e rather than shell commands, so the same tests run on Windows.

Real-world tests

apps/craft/tests/real_world/ installs pinned real packages from the npm registry, JSR and GitHub: an Express server, React server rendering with Redux and React Query, a craft dlx create-vite project that builds, TypeScript, native packages, git and tarball dependencies, and more. They are ignored by default; make real-world runs them, and CI runs them in the registry-tests job.

Windows

CI runs the full suite on windows-latest. Locally, docker/windows/test.sh cross-compiles every test executable for x86_64-pc-windows-gnu and runs them under Wine with the Windows builds of Node.js and Git:

docker/windows/test.sh               # everything
docker/windows/test.sh run::         # arguments go to each test executable

Wine is not Windows: CI's windows-latest job has the final word. On Apple Silicon, Docker's Rosetta emulation can crash Wine processes; switching Docker Desktop to QEMU emulation avoids that.

Benchmarks

make bench times craft install against npm, pnpm and Yarn on three real projects, from a cold machine to an already installed project, with hyperfine. The Benchmark workflow runs the same on a GitHub-hosted runner. See bench/README.md for the fixtures, the scenarios and how each package manager is kept on equal terms.