Test Quality Agent Skill

Test Quality

Config Resolution

Skill support project custom. Order:

• Look .lattice/config.yaml in repo root
• If find, check paths.test_quality for custom doc path
• If custom path exist, read doc & check YAML frontmatter for mode:
• mode: override (or no mode): Custom doc take full control. Use instead default. Must be complete -- is only reference.
• mode: overlay: Read ./references/defaults.md first, then apply custom doc on top. Custom sections replace match sections (by heading). New sections add after.
• If no config/path/file, read ./references/defaults.md
• Language adaptation: If paths.language_idioms exist in config, read "Testing Patterns" section and adapt §5 (Test Naming), §4 (Test Isolation), §6 (Test Data Builders) to language test framework idioms. Language idioms take precedence over pseudocode defaults.

Self-Validation Checklist

STOP after gen each test. Check ALL before continue. If fail, fix. If ambiguous (see Ambiguity Signals), flag -- show options & reasoning.

• AAA STRUCTURE: arrange, act, assert separate w/ blank lines? Any logic (if/loop/try) in arrange or assert?
• SINGLE BEHAVIOR: Test verify one behavior per loaded doc (default: one behavior per test, name need "and" = split)?
• ASSERTION QUALITY: Assert observable behavior, not implementation? Specific enough catch regression?
• ISOLATION: Test depend other test output/effects? All mutable state per-test?
• TEST NAME: Name follow team convention per loaded doc (default: describe behavior, not method)? Failure message clear?
• TEST DATA: Complex arrange uses builders/factories? Magic values → named constants? (Inline literals fine for trivial tests.)
• MOCK BOUNDARIES: Mock per loaded doc (default: only at arch boundaries — I/O, external — not internal collab)?
• TEST CODE AS FIRST-CLASS: Structured like production code? Shared constants at top, helpers extracted, no dead code, clear file organization?

Project-specific checks: if loaded doc contains a validation checklist section, apply those after base checklist.

Active Anti-Pattern Scan

After checklist, scan these. If find, fix before present.

• [ ] Test-per-Method: One test per method regardless behaviors → One test per scenario, named for behavior
• [ ] Assertion Roulette: Multiple unrelated asserts; unclear which broke → Split to one behavior per test
• [ ] Shared Mutable State: Pass alone, fail together → Isolate state; per-test setup; no static mutable
• [ ] Testing Implementation Details: Break on refactor w/ same behavior; mock call counts → Assert observable behavior, not method calls
• [ ] Mystery Guest: Depend external file/db/env var not visible → Inline data or use builders; all preconditions visible
• [ ] Slow Tests by Default: Unit suite take minutes; hit db/network/fs → Mock/fake I/O; use in-memory
• [ ] Conditional Test Logic: Test have if/loop/try -- test need own tests → Remove logic; use parameterized; let asserts fail natural
• [ ] Copy-Paste Tests: Near-identical w/ small changes → Extract shared setup to builders; use parameterized

Class-Level Review

Fire when: (1) completing all tests for a class — new or existing. (2) adding or editing any test in an existing class.

STOP before present. Per-test checks verify individual quality. This review verifies the test suite covers the class contract.

Full review (new class or significant additions)

• Behavior inventory — list every public method/behavior in class under test. If class not available, ask user to enumerate.
• Coverage matrix — map each test → behavior it covers. Behaviors with zero tests → blocking. Do not present until user addresses gap or explicitly accepts it.
• Error path check — scan class under test for: explicit throws, conditional error branches, edge guards. For each found: does a test exercise this path? If not → flag by name. Zero-coverage error paths are blocking unless user explicitly accepts.
• Behavioral duplication — compare "then" clauses across all tests. Same observable outcome regardless of structural differences → flag as likely duplication. Name both tests.
• Balance signal — any behavior with tests but none covering a failure or edge case → surface as question, not hard failure: "deleteUser has 1 test (happy path only) — does it have error cases?"

Edit-scoped review (adding or changing one test in existing class)

Run steps 3–5 only, scoped to the changed test:

• Does this test duplicate the observable outcome of any existing test?
• Does it cover a behavior or error path not previously covered?

Ambiguity Signals

Multiple valid outcomes. Present options, not choose silent.

• Unit vs Integration: Service coordinate components -- test isolate (mock) or real collab? Depend coupling & what verify.
• Mock Depth: Mock direct depend or let call through? Over-mock test implementation; under-mock create slow/flaky.
• Test Granularity: One test multi asserts vs multi tests one assert? When asserts verify facets same behavior, group ok.

Test Code as First-Class Code

Treat test files like production classes:

• Shared constants and boundary values at top of file (named, not magic)
• Shared builders/factories extracted to helpers -- not copy-pasted per test
• Setup methods or fixtures for repeated arrange patterns
• Logical grouping: related behaviors together (by feature, by scenario type)
• Dead tests removed, not commented out
• Refactoring applies: extract method when arrange is long, rename when intent unclear, move shared setup when duplicated

Refactoring opportunities to surface proactively:

• Multiple tests repeat same arrange → extract to builder or shared fixture
• Same assertion pattern across tests → extract custom assertion helper
• Test file grows beyond ~300 lines → split by behavior group
• Constants scattered inline → collect at top with descriptive names
• Deeply nested test structures → flatten with clear naming

See ./references/defaults.md for AAA structure examples, assertion patterns, isolation techniques, naming conventions, test data builder patterns, and pyramid distribution guidance.