| name | add-cpp-test |
| description | Use when adding a unit or integration test, writing a new Catch2 TEST_CASE, scaffolding a test fixture, starting TDD work, OR figuring out how to run / filter / debug Catch2 tests flexibly. Trigger on "add a test", "write a test for", "I need test coverage", "TDD this", "run just this test", "how do I filter tests by tag", "Catch2 generators", "Catch2 sections", "Catch2 matchers", "reproduce a random-order failure". Enforces Catch2 v3.14.0 patterns, the one-test-one-implementation TDD discipline, the layer-direction rule for test sources, and the tests/<layer>/<file>_test.cpp layout convention. The Catch2 idiom cheat sheet and the bench-binary invocation reference live as separate files under `references/`, loaded on demand when the workflow steps point to them. |
| allowed-tools | Read, Glob, Grep, Edit, Write, Bash |
Adding a test to Lancet2
This skill formalizes the canonical procedure for adding a Catch2 test. The discipline matters because Claude's default is implementation-first; under context pressure it may rewrite a test to match buggy code rather than fix the code. The one-test-one-implementation rhythm prevents that failure mode.
Step 1 — Locate the right test file
Tests live in tests/. Mirror the structure of src/lancet/: a test for src/lancet/caller/genotyper.cpp lives at tests/caller/genotyper_test.cpp (the convention is the _test.cpp suffix; verify by globbing tests/). If the unit under test does not yet have a test file, create one matching the existing naming convention. If it does, append to the existing file.
Step 2 — Read existing tests in the same area
Before writing, read at least two existing tests in the relevant subdirectory in full. Catch2 idioms used across the project will reveal themselves: the choice between TEST_CASE with SECTION versus TEMPLATE_TEST_CASE, the convention for fixture setup, the use of REQUIRE versus CHECK, and the project's tolerance conventions for floating-point comparison.
Step 3 — Write ONE failing test
Write a single TEST_CASE for the specific behavior you intend to exercise. Do not write multiple tests speculatively; under context pressure you will rewrite them to match the implementation rather than reflect the actual specification. The test should fail for the expected reason — a REQUIRE that compares against the expected output, with a clear failure message.
For correctness-critical math (genotype likelihoods, Phred conversions, scoring), include numerical-equivalence checks against a known-correct reference value computed offline. Use Approx (or the project's tolerance helper) with explicit absolute and relative tolerances.
Step 4 — Build and run, confirm failure
Run pixi run test. The output should show your new test failing for the expected reason. If it fails for a different reason (compilation error, wrong fixture, unrelated bug), fix that first; you cannot trust an implementation against a test that does not fail correctly.
Step 5 — Implement the minimum change
In src/lancet/<layer>/, make the smallest change that turns the test green. Respect the layer-direction rule (the at-write hook will catch violations, but knowing the rule prevents wasted effort). Do not refactor surrounding code; resist the temptation to add helpers, extract methods, or improve names. Keep this commit narrow.
Step 6 — Build and run, confirm green
Run pixi run test again. The new test must pass. All previously-passing tests must continue to pass. If any other test fails, the change has unintended scope; revert and reconsider.
Step 7 — Format and lint
The PostToolUse hook should have already run pixi run fmt-fix. Run pixi run lint-check manually before commit to catch any clang-tidy violations the at-write naming hook missed.
Step 8 — Commit
Use the project's chglog-filtered commit prefix. Test-only commits use chore: (the catch-all that renders as "Refactoring" in CHANGELOG.md, since the project's .chglog/config.yml does not filter on test). A test that lands together with the code it exercises uses fix: or feat: matching the source change. Keep the commit message imperative and specific: chore: add Phred-overflow test for posterior_base_qual is better than add tests.
Step 9 — Stop here
Resist the temptation to add more tests in the same change. Each additional test is a separate commit and ideally a separate session. The vertical-slice discipline is what makes the workflow reliable; widening scope is how it breaks down.
When NOT to use this skill
Do not use this skill for benchmarks; benchmarks live in benchmarks/ and the procedure is different (use the profile-and-optimize skill, which incorporates the canonical benchmark workflow). Do not use it for end-to-end pipeline tests that exercise the binary against real BAMs; that is the /e2e-pipeline-test slash command's job.
Layer-direction reminder
Test sources in tests/<layer>/ may include from any layer of src/lancet/. Tests are exempt from the layer-direction rule because they exist to exercise the layered code, not to be part of it.
Headline rules from docs_dev/style/test_style.md
The full project-wide test-style guide lives at docs_dev/style/test_style.md. Read it once end-to-end before adding a non-trivial test file. The headline rules below summarise the dimensions Claude is most likely to get wrong by default; each paragraph ends with a pointer to the full guidance.
File layout. One test file per source unit at tests/<layer>/<unit>_test.cpp. The file's body lives inside namespace lancet::<layer>::tests { ... }, mirroring the source-layer namespace. File-private helpers go in an anonymous namespace { ... } block inside the named namespace; cross-file helpers go in a header under tests/<layer>/. See docs_dev/style/test_style.md § 1 for full guidance.
TEST_CASE naming. Present-tense full sentence describing the verified behavior — "RevComp returns the canonical complement for each base", not "test foo". The name surfaces in CI logs and failure reports. See docs_dev/style/test_style.md § 2 for full guidance.
Tags. Required and only: [lancet][<layer>][<unit>]. The <unit> is the PascalCase name of the primary public symbol the TEST_CASE exercises (per-symbol rule). No semantic tags like [property], [edge], [golden]. See docs_dev/style/test_style.md § 3 for full guidance.
Test types and the quality bar. Bar 1 (default): one happy-path TEST_CASE, one edge-case TEST_CASE per input type, one error-path TEST_CASE per documented failure mode. Bar 3 (numerical kernels where it pays): a property test with a pinned seed, plus cross-validation against an in-tree reference, a std-lib equivalent, or a precomputed scientific-Python TSV. See docs_dev/style/test_style.md § 4 for full guidance.
SECTION vs separate TEST_CASE. Use SECTION for variants that share setup; use separate TEST_CASEs for conceptually-independent scenarios. A SECTION-heavy TEST_CASE that mixes happy + edge + error paths is worse than three separate TEST_CASEs. See docs_dev/style/test_style.md § 5 for full guidance.
Edge-case taxonomy. A table indexed by input type (string, float, integer, span, group, file, state machine) listing the standard edge-case classes. Pick at least one per input type the unit accepts. See docs_dev/style/test_style.md § 6 for full guidance.
Random-input rules. Pin the seed with a const literal. No std::random_device() in tests/. Don't use random generation for happy-path fixtures. 100–1000 iterations per property as the default range. See docs_dev/style/test_style.md § 7 for full guidance.
Assertion macros. REQUIRE for invariants subsequent code depends on; CHECK for non-fatal multi-property verification; REQUIRE_THROWS_AS for error-path type checks; INFO/CAPTURE for failure-log context. See docs_dev/style/test_style.md § 8 for full guidance.
Determinism discipline. No std::random_device. No wall-clock dependencies — use the codebase's clock-injection seam where available. No own-filesystem dependencies outside tests/data/ (read-only) and the temp directory (writable) — use IO-sink seams (e.g. std::ostream& ctor overloads) where available. Tests are independent of execution order. See docs_dev/style/test_style.md § 9 for full guidance.
Skipping. SKIP("reason: ...") with an explicit reason; never return; silently. Skip-on-missing-fixture is fine when the fixture is intentionally out-of-band. See docs_dev/style/test_style.md § 10 for full guidance.
Test fixtures and generator scripts. Generators live in tests/scripts/<name>.py; generated data lives in tests/data/<layer>/<name>.tsv (or analogous). The script's docstring documents the regeneration command (pixi run -e test-fixtures python tests/scripts/<name>.py). The fixture is committed alongside the script — tests do not depend on the script being runnable at test time. See docs_dev/style/test_style.md § 11 for full guidance.
References
These are loaded on demand when this skill points to them.
references/catch2_idioms.md — Catch2 v3.14.0 cheat sheet: assertion families, TEST_CASE/SECTION, tags, type-parametrised tests, generators, fixtures, matchers, logging, runtime skipping, plus the note about why benchmarks are out of scope. Read this when writing or extending a test (Steps 2 and 3).references/catch2_invocation.md — Running the test binary directly: filtering by name/tag, running a specific section or generator value, discovery (list tests/tags/reporters), reproducing a failed random-order run, debugging modes (break/abort/no-throw), reporters, sharding, and a recipe table for common scenarios. Read this when iterating on a failing test (Steps 4 and 6).
