Test Audit
Keep only tests that earn their maintenance cost. Two modes, one value bar: the authoring gate for every new or changed test, and the audit that sweeps existing ones. Optimize for confidence, not deletion count. The tdd skill defines a good test and the regression rule.
Authoring gate
Before adding a test, answer all four; a missing answer means wait:
- What observable behavior, invariant or independent contract does it protect?
- What credible regression makes it fail?
- Why does existing coverage miss that failure? Each contract has one owner test at the strongest boundary; another layer needs its own risk. Extend a table-driven case or shared fixture rather than adding a near-duplicate.
- Does it need a production seam (export, flag, wrapper, hook) no production caller needs? Then test at the real boundary instead.
A test that matches a junk pattern fails the gate unless the retention bar names the contract it alone guards.
Junk patterns
- assertion-free coverage probes, self-comparisons, identity copiers;
- copied fixtures, inventories, manifests or export lists; exact source, import or string greps;
- private predicates or call shapes already tested at a real boundary; duplicate calls of one contract;
- tests that only keep test-only exports, globals or wrappers alive; production code whose only callers are tests;
- expected values produced by the code under test; mocks that implement the asserted behavior;
- fixtures that supply what the owner should produce, or assert a store the path never writes;
- capability tests that restate declared flags instead of exercising the delivery;
- negative controls that pass for an unrelated reason;
- names that promise more than the input exercises.
Retention bar
Keep a test that independently enforces a public API, protocol, config, migration, storage, security, platform, default, generated-code, package, release or architecture contract. Also keep observable call ordering, regressions with a credible failure mode, and a source check when it is the cheapest independent guard (it fails when the user-facing key, byte or path changes and survives a rename). Static or slow is not a reason to delete. A kept test that fails on the baseline may be a product bug: reproduce it and fix the owner.
Audit
- Read root and scoped
AGENTS.md, then the whole test, its production owner, callers, siblings, overlapping tests, CI routing and history. Stay read-only until the evidence is in. - For each candidate record: name and location; what failure it can detect; non-test callers of its seam; the stronger proof that remains; why it exists; what deletion it unlocks; risk and the focused command. A missing field means not ready.
- Edit one coherent batch: delete obsolete test-only seams and dead paths (no aliases), move kept regressions to their owners, consolidate repeated assertions. Prefer net-negative production lines; add no replacement that restates the implementation.
- Validate: run the smallest owner and sibling tests, the real script behind any removed grep, formatting and
git diff --check, then the checks the repo requires. Never edit files a running test reads. - Report: what went and why, false positives kept and why, the proof run, production versus test lines (
git diff --numstat), follow-ups. Land only when authorized, one batch at a time.
Layered verification
Tests are one layer; stack cheaper ones before human attention, scaled to risk: deterministic guardrails (types, lint, tests, secret scans); a fresh-context LLM review; deploy-time flags, canaries and rollback; runtime traces and SLOs; then focused human review of intent and architecture.