test-patterns

Name the concrete failure each test must catch, then pick the cheapest evidence that reliably exposes it. Escalate to broader layers only when a named failure cannot be detected at a cheaper boundary, and record the cost of every layer chosen.

Read the actual test code and CI configuration, classifying tests by the resources they touch rather than directory names. Then identify missing, duplicated, or misplaced evidence and recommend one highest-leverage rebalancing move instead of a rewrite.

Use contract tests when producer and consumer deploy independently and interface drift is a real risk. If both sides ship together in one deployable, a shared type or narrow integration test is cheaper and equally truthful.

Only when the expected behavior comes from a source independent of the implementation, such as a specification, invariant, or human-reviewed example. Tests whose expected outputs were inferred from the code under test are coverage theatre and should be rejected or strengthened.

Split the system: test the deterministic scaffold (tool dispatch, parsing, retries, guardrails) with ordinary assertions, and evaluate semantic quality against curated cases and metrics. Treat an LLM judge as a measuring instrument requiring calibration against human-labelled examples.

When inspection shows the evidence is proportionate to the named risks, the review records the existing portfolio, the gates deliberately not opened, and their reopening signals. A clean bill of health is a complete outcome, not a failure to find work.