#Testing
Two suites, split by language and run by different tools:
cargo test— the Rust workspace: extractor, server, export, CLI. Unit tests live inline in each crate's sources (src/*.rsandcrates/extract/src/*.rs); integration suites live intests/, belong to the root crate, and drive the real binary or an in-processaxumrouter. Both members are in the workspace'sdefault-members, so a barecargo testcovers both — without that, cargo would select only the root package and silently skip the extraction crate's unit tests.bun test— the Svelte SPA and the bun build scripts. Every*.test.tssits beside the module it covers.
See Build & infra for the CI jobs that run them.
#Running
cargo test # Rust; real-nix suites skip when nix is absent
bun test # SPA + build scripts
bun test --coverage # text + lcov reporters, into dist/coverage/#The bun suite
bunfig.toml preloads two setup files for every run and configures coverage (lcov output, test files skipped, web/testing/** ignored):
| Preload | Purpose |
|---|---|
web/testing/happy-dom.ts |
Registers happy-dom as the global DOM and stubs matchMedia / ResizeObserver, which the viewer touches at init time |
web/testing/svelte-loader.ts |
A bun runtime plugin that compiles .svelte files with svelte/compiler (client output, injected CSS, runes) and .svelte.ts modules via compileModule — bun-plugin-svelte can't run under the test runtime because its virtual CSS imports need build-time resolution. It also swaps svelte's index-server.js package entries for their client siblings, since bun test resolves the "default" (server) export condition |
Component tests use the withMount helper in web/testing/helpers.ts: mount into a fresh host element, flushSync(), assert, always unmount.
Tests are co-located: web/lib/indexes.test.ts sits next to indexes.ts, web/components/OptionRow.test.ts next to OptionRow.svelte. Finding a module's tests is a directory listing, not a search.
| Group | Files | Location |
|---|---|---|
| Component tests | 19 | web/components/*.test.ts, plus web/App.test.ts (fixture data injected into the app singleton, components mounted under happy-dom) |
| SPA library tests | 14 | web/lib/*.test.ts — state, indexes, schema, search, segments, colors, URL/hash routing, diffing |
| Build-script tests | 3 | scripts/build-app.test.ts (the </script> escaping invariant), scripts/licenses.test.ts, scripts/release.test.ts |
#The Rust suite
Unit tests live inline beside the code, in src/*.rs for the root crate and crates/extract/src/*.rs for the extractor. The integration suites in tests/ share helpers via tests/common/mod.rs:
| Suite | Covers |
|---|---|
tests/cli.rs |
The binary's flag parsing and help/usage surface, as a subprocess |
tests/serve_http.rs |
The whole route surface against an in-process axum router, using a nix shim on PATH — no real evaluation |
tests/export_html.rs |
End-to-end single-file export, re-parsing the embedded data tags out of the HTML |
tests/degrade.rs |
Per-configuration failure paths — one bad config must not poison the rest |
tests/mini_flake.rs |
The full manifest + option-extraction pipeline against real nix |
tests/determinism.rs |
The two guards on "a blob is a function of the extraction crate and the flake": repeated extractions must be byte-identical (real nix), and the root crate must not grow a new file-writing site (runs everywhere) |
The second suite is the only thing in the repo that fails when the extraction boundary stops holding — see The extraction crate boundary. Its own comment is explicit about what it cannot catch, which is worth reading before relying on it.
#Fixture strategy
fixtures/mini-flake/flake.nix— a real flake evaluated by real nix, but builtins-only (no nixpkgs), so evaluation is cheap and no store downloads happen. It hand-rolls just enough of the module-system option shape (_type = "option",declarations,definitionsWithLocations) for the extractor's structural walk, and includes a nestedpath:input to exercise the Inputs panel.fixtures/broken-flake/flake.nix— a flake whose one configuration throws on evaluation: the attr name is enumerable but forcing the value fails, exercising the per-config error/degradation path without poisoning the healthy fixture.web/testing/fixtures.ts— hand-writtenManifest/ConfigData/OptionEntrybuilders shared by the SPA unit and component tests, with fake store paths for self, inputs, and a patched-input copy.
The nix fixtures live at the repo root rather than under tests/ on purpose: tests/ is in the crate's Nix fileset (see below), so nesting them there would make every fixture edit invalidate the crane dependency layer.
#FLAKE_EXPLORER_REQUIRE_NIX
The real-nix suites skip when nix is not on PATH, which is right for local machines without nix but dangerous in CI — a skipped suite would only show up as a coverage drop. Setting FLAKE_EXPLORER_REQUIRE_NIX=1 makes common::nix_available() panic instead of skipping (tests/common/mod.rs). CI's coverage step sets it, with nix installed, so a silent skip is impossible — see .github/workflows/ci.yml.
#Coverage and the Nix checks
Two octocov reports, deliberately separate so their histories never mix:
.octocov.yml— the SPA suite, readingdist/coverage/lcov.infoagainst a fixedacceptable: 96%floor..octocov.rust.yml— the crate, readingrust-coverage/lcov.infoas acurrent >= prevratchet.
nix flake check builds four checks from package.nix: test (cargo test), clippy, coverage, and app-test (an offline bun test against the vendored node_modules). The sandbox has no nix binary and the fixture flakes are outside the crate fileset, so the real-nix suites skip there by design — CI's out-of-sandbox cargo llvm-cov test step is where they must run.
Coverage needs --workspace; testing does not. cargo test and cargo clippy pick up both workspace members from default-members, so all 32 tests run without any flag. cargo llvm-cov is different: it scopes its report to the selected package, and because the workspace root is itself a package, the default selection is the root crate alone. Omitting --workspace therefore produces a report covering 6 files instead of 16 and silently drops every line of the extractor — the tests still run, they just go unattributed. It cost an 82.8% → 73.3% ratchet failure on the PR that introduced the split, and both invocations (the CI step and the coverage check in package.nix) now pass it explicitly.