Coverage Ratchet
A build/test gate proves the tests that exist pass. It says nothing about whether tests were written — a phase with no test task ships green on build success alone. The coverage ratchet closes that hole deterministically: line coverage is measured on every verify, compared against a committed baseline, and may never drop. It only rises.
Why a Ratchet, Not a Target
An absolute threshold ("80%") fails one of two ways on a real codebase: it fails day one (so it gets disabled), or it gets set below current reality (so it gates nothing). A ratchet starts from wherever the project actually is and only tightens. Nobody argues with "don't get worse."
When This Skill Activates
Use this skill when:
- A project has a test target but coverage has never been measured
- Agent-written features keep arriving with build-passing, test-free code
- Setting up the deterministic gauntlet (with
testing/fitness-functions/andswift/code-size/)
Setup
- Copy
templates/coverage-gate.shinto the project (e.g.Scripts/coverage-gate.sh), make it executable, and fill in the config block (scheme, destination, app target name). - Bootstrap the baseline:
Scripts/coverage-gate.sh --init # runs tests with coverage, writes .coverage-baseline- Commit both the script and
.coverage-baseline.
.coverage-baseline holds a single decimal fraction (e.g. 0.6231) — the app target's line coverage. Trailing whitespace ignored; nothing else in the file.
The Gate
Scripts/coverage-gate.sh # exit 0 = pass, exit 1 = coverage dropped- Runs the test suite with coverage enabled, extracts the app target's line coverage from the
xccovJSON report. - Fails if
current < baseline − 0.0025(the epsilon absorbs measurement jitter from unchanged code). - Suggests a ratchet-up when
current > baseline + 0.01: update.coverage-baselineto the new number in the same commit as the tests that earned it. That's the ratchet clicking forward. - Fails loudly on an unexpected report shape or a missing target — a broken measurement must never read as a pass.
Ratchet-up etiquette: raising the baseline is routine and encouraged; lowering it requires a written reason in the commit message (e.g. deleting a well-covered module) — treat it like widening a fitness-function allowlist.
CRAP Report (Advisory)
templates/crap-report.sh combines per-function cyclomatic complexity (SwiftLint JSON output) with per-function coverage (the same xccov report) into the CRAP score — Change Risk Anti-Patterns:
CRAP(f) = complexity(f)² × (1 − coverage(f))³ + complexity(f)A complex and untested function scores explosively; simple-and-tested stays near its complexity. The report prints the top 20 — the exact functions where a test is worth the most. Advisory only, never a gate: the name-join between the two tools is heuristic.
Common Pitfalls
| Pitfall | Problem | Solution |
|---|---|---|
| Absolute threshold on existing code | Fails day one or gates nothing | Ratchet from measured reality |
| Gating total project coverage | Test-target/generated code pollutes the number | Gate the app target's own lineCoverage |
| Silent pass on tool failure | Broken measurement reads as green | Script fails loudly on unexpected JSON |
| Chasing the number with assertion-free tests | Coverage without verification | Pair with mutation spot-checks (testing/mutation-testing/) |
| Never ratcheting up | Baseline fossilizes at day-one level | Raise it in the same commit as new tests |
References
testing/mutation-testing/— do the covering tests actually assert anything (advisory)swift/code-size/— the complexity half; small functions are coverable functionstesting/fitness-functions/— architecture invariants as testsgenerators/ci-cd-setup/— CI wiring for the gate