v3.4.5 — Production-Ready Test Density Scoring

What this release fixes

v3.4.4 introduced test density scoring but had a ship-blocking bug: the grep for counting test cases scanned node_modules, dist, and build artifacts. On a real React project (forge-ui), this inflated 4,187 real matches to 14,130 phantom matches — a 237% over-count.

Bug fixes

• P0: node_modules inflation — Test case grep now uses find with the same BUILD_ARTIFACT_EXCLUDES as source file counting. No more phantom matches from dependency packages.
• Double-counting — v3.4.4 used two grep passes that counted *.test.* files inside tests/ twice. Replaced with single find | xargs grep pipeline.
• Missing patterns — Now detects:
  • it.each(), test.skip(), test.only(), it.concurrent() (data-driven/skipped tests)
  • __tests__/ directory convention (Jest without .test. in filename)
  • *.cy.* files (Cypress E2E tests)
  • #[tokio::test], #[rstest], #[actix_web::test] (async Rust)

Validated against 3 real projects

Project	Stack	Tests	Source files	Density	Score
forge-ui	React 19	4,187	851	4.9/file	15/20 (solid)
forge-orchestrator	Rust	362	43	8.4/file	20/20 (thorough)
game.clicker	React + Vite	37	3	12.3/file	20/20 (thorough)

Known limitations (documented, not bugs)

• Distribution blindness — 100 tests in 1 file + 0 tests in 99 files averages to 1.0 density. Real line coverage (Tier 1) catches this; density (Tier 2) does not.
• Parameterized tests — test.each() with 10 rows counts as 1 test case. Go table-driven tests with t.Run() subtests are also undercounted.
• BDD/Cucumber — .feature files with Scenario: are not detected.

These limitations are inherent to any fast proxy. The scoring clearly recommends running --coverage for Tier 1 accuracy.

Upgrade

/forge:update

Then restart Claude Code.

---

Full changelog: v3.4.4...v3.4.5