--- name: testing description: "Use when writing, reviewing, or maintaining tests, verifying code, or preparing code/test changes for handoff." --- # Testing ## Philosophy - Tests before | during | after code → verify, explore, learn, validate. Test-first is optional. - Code settled → curate tests for original intent, bugs, regressions, safe refactoring. Keep protection, not construction history. - Prefer few high-leverage integrations covering substantial consumer behavior over many isolated unit checks. Integrations + useful unit tests generally form the bulk; keep unique edge cases, no fixed ratio. - Coverage + mutation → auxiliary diagnostics, not score targets. - Properties + reviewed snapshots → compact oracles for suitable code, replacing repetitive explicit expectations. ## Route - Development: keep local feedback fast; use existing project automation and required gates. - Before completion, handoff, or integration of code/test changes: read [suite review](references/suite-review.md). Also use it for existing-suite reviews; read-only means recommendations, not edits. - Prompts/model behavior: also read [prompt evaluation](references/prompt-evaluation.md). Technique adopted + relevant → read its reference before choosing or reviewing that strategy. Otherwise load only on explicit request. Establish adoption from tasks, config, tests, or reports; dependency presence alone ≠ adoption. Do not bulk-load references or automatically introduce tools, gates, or broad campaigns. - Code/test coverage; statement, line, branch reports → [coverage](references/coverage.md). - Mutation analysis; mutants, history/incremental runs → [mutation](references/mutation-testing.md). - Property/generative/QuickCheck-style checks → [properties](references/property-based-testing.md). - Approval tests, golden files/masters, reference output → [snapshots](references/snapshot-testing.md). - Module/component, API/service, contract, system/E2E, built-artifact consumers → [integration](references/integration-testing.md). Consumer harnesses count even without the integration label; inspect actual boundaries, not names. ## Boundaries - Derive expectations from requirements, invariants, or independent examples, not copied implementation logic or configured mock returns. - Local probes may inspect internals; they answer local questions, not broader correctness. Keep throwaway probe artifacts in OS temp. - Match completion claims to observed consumer behavior; local or synthetic success does not prove live configuration or portability. - Validate prerequisites before interpreting results; invalid setup makes a green result inconclusive. - Preserve supported behavior, known regressions, and relevant failure cases when replacing tests. - Diagnose failures before removal; never weaken assertions, skip failures, or refresh snapshots merely to get green. - Stay within authorized scope; unrelated weaknesses do not authorize edits, new gates, or full CI. Preserve required project gates. Investigate failed required checks; establish baseline evidence before calling failures pre-existing. Fix within scope or obtain explicit approval to broaden scope or defer; do not report completion with unresolved required failures.