Local Test Review

Local Test Review

Answer "is this project tested the way it should be?" with evidence, not vibes: what exists, what passes, what's missing that matters, and a plan sized to the risk. This is the thinking layer over run-tests (the execution layer) — it never runs a suite by hand, and run-tests never designs strategy.

0. Delegate strategy design if testing-strategy is available

Check the skills available in this session for one named testing-strategy. If it exists, plan to delegate the strategy design to it — but invoke it only when you reach section 3, with the inventory (1) and baseline (2) in hand: strategy designed blind of the project's facts is the failure mode this ordering exists to prevent. If it's not available, use the inline fallback in section 3.

Either way, these house rules apply on top of whatever the strategy says:

• Determinism boundary: unit suites must run with no network, no model calls, no live services — that work belongs in separated integration/eval tiers, gated behind markers or dedicated commands.
• Escalation ladder: plans should support fast → scoped → full execution, not only an all-or-nothing suite.
• No invented thresholds: hold the project to a coverage bar only when it documents one — flag meaningful gaps, don't measure green suites against imaginary percentages.
• Tests are never disabled to get to green, and a repo that declares "no tests by design" keeps that status unless the user decides otherwise — when that status is declared, do not invent testing gaps or recommend a new suite unless the user explicitly opts in.

1. Inventory the current testing state

Establish the facts the same way run-tests does (its detection steps and references/detection.md apply):

• Read the project's own instructions — AGENTS.md / CLAUDE.md, .cursor/rules/*.mdc, CONTRIBUTING.md, CI workflows — for the intended testing story.
• Map what actually exists: which tiers (unit / integration / E2E / evals), which frameworks, where the tests live, roughly how much there is per tier, and which tiers need infrastructure (databases, Docker, env vars).
• Note the discrepancies both ways: documented-but-missing (aspirational docs) and present-but-undocumented.
• Short-circuit for intentional testlessness. If the project declares "no tests by design", the review ends here: report the policy and the documented alternative validation, and skip sections 2–4 entirely — no pyramid, no gap ranking, no add-tests plan — unless the user explicitly opts in to reconsidering the policy (some repos require an RFC first).

2. Baseline with run-tests

If a runnable suite exists, invoke the run-tests skill to establish what passes today — a red baseline changes every recommendation that follows (fix before extend). Skip the baseline only when the suite needs infrastructure the user hasn't agreed to start; report that it was skipped and why.

3. Design the target strategy

Via the delegated skill when available; otherwise:

• Pyramid shape: many fast unit tests, some integration, few E2E — high confidence at the top, fast feedback at the bottom.
• By component type: API endpoints → unit tests for business logic, integration for the HTTP layer, contract tests for consumers; data pipelines → input validation, transformation correctness, idempotency; frontend → component, interaction, visual regression, accessibility; infrastructure → smoke, load.
• Cover: business-critical paths, error handling, edge cases, security boundaries, data integrity. Skip: trivial getters/setters, framework code, one-off scripts.

Ground the strategy in this project: its component types, its risk profile, its existing conventions — not a generic ideal.

4. Gap analysis and plan

Compare inventory (1) + baseline (2) against strategy (3):

• Rank gaps by risk: untested business-critical path > untested error handling > missing tier > style issues. A handful of high-risk gaps beats an exhaustive inventory of trivial ones.
• For each gap worth closing: what to test, the tier it belongs in, and one or two concrete example cases — specific enough that an implementer doesn't have to re-derive the analysis.
• Call out over-testing too: slow, flaky, or redundant tests that cost more than they catch.

5. Report and hand off

Deliver: current state, baseline result (or why skipped), ranked gaps, the recommended plan, and any repo-quirk warnings for whoever implements. Writing the tests is separate work — hand the plan to the user, or into the gather-requirements → create-plan → execute-plan flow for anything substantial. Don't start implementing tests inside this skill. For the implementer, note the relevant how-to skills when installed: python-testing-patterns for pytest fixtures/mocking/parametrization, webapp-testing for ad-hoc Playwright verification of web UIs.