More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Test Blindspots: What Green Tests Miss

    1

    Find what a green test suite does not actually prove: map assertions against behavior, hunt the classic blindspot patterns, probe the riskiest gaps, report defects honestly.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Test Blindspots: What Green Tests Miss

    Test Blindspots: What Green Tests Miss

    Example session with this skill installed

    Our payments service has 94% line coverage and a fully green suite, yet we had two production incidents last month — a double charge and a stuck refund. What is the suite not proving?

    • Read your context and instructions
    • Compiled the test blindspots

    HUNT REPORT — payments service, green suite, two incident classes.

    WHAT THE TESTS ACTUALLY CLAIM: Reading the suite, the assertions cover: charge succeeds on a valid card (mocked gateway), refund issues a gateway call (mocked), retries fire on gateway 5xx (mocked clock). Every claim stops at a mock. Nothing asserts behavior at the real gateway boundary, across a process restart, or under a repeated request.

    GAP 1 — Shared assumption, consequence 5 × doubt 5: The idempotency key is built from the order ID in both the code and the tests, and every test reuses one fixed order ID. Probe: submit the same checkout twice with a regenerated order ID, as the mobile client's retry does. Result: two gateway charges, keys differ. CONFIRMED DEFECT — this is the double charge; the suite passed precisely because the fixture never varied the input that mattered.

    GAP 2 — Untested transition: Refund state is asserted per state, never through a restart mid-refund. Probe: kill the worker between the gateway call and the state write, then restart. Result: the refund re-issues, the gateway rejects the duplicate, and the refund sits in "processing" forever. CONFIRMED DEFECT — no reconciliation step exists; this is the stuck refund.

    GAP 3 — Specification question: No test or document states whether a partial refund after a partial capture is allowed. The code permits it; support says it should not. Classified as a SPEC QUESTION for the owner — not a defect until someone decides, and exactly where the next incident report comes from.

    NOT EXAMINED: concurrency under real load, rounding beyond USD. Note the shape of the result: 94% coverage, and none of the three gaps is a line a coverage report could see.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    A passing suite feels like proof. It is evidence for exactly what its assertions check, under exactly the setup they use — and the incidents live in the gap: the boundary replaced by a mock, the transition never exercised, the wrong assumption the test shares with the code it tests. This skill hunts that gap with focused exploration instead of a coverage chase. It starts by pinning down what the tests actually claim, in plain language, separately from what the code does. Then it reads implementation against tests looking for the characteristic blindspot patterns: the mock that is the boundary, happy-path-only error handling, untested state transitions, missing consumer expectations, shared untested assumptions, concurrency the sequential suite cannot see, and data-shaped edges fixtures never contain. Candidate gaps are scored on consequence and doubt so the hunt stays proportional, and each consequential gap gets the smallest probe that settles it — against the real component, not a bigger mock. Results are classified with discipline: confirmed defects with minimal reproducers, untested risks labeled as risks, and specification questions routed to the human instead of being invented away. The report ends with what was not examined, because the reader deserves to know what confidence they are buying. A complete worked example — a green retry-with-backoff suite hiding two defects — shows the whole method in one page.

    • It is a focused hunt, not an exhaustive audit: findings are bounded by the scope examined, and the report states what was not examined.
    • Probes need a runnable environment; where none exists, gaps are reported as untested risks rather than settled either way.
    • It starts from a green suite. If the suite is already red, that failure is the finding — debug it first, then hunt.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 24 days ago

    Permissions

    No special permissions declared or detected

    Listed24 days ago
    Updated21 days ago

    What's inside

    Frequently Asked Questions

    A review judges the diff; coverage counts lines that ran. This starts from a passing suite and asks what the passing actually establishes, comparing assertions against real behavior at boundaries, transitions, and shared assumptions — the places both reviews and coverage numbers miss.

    No — deliberately. Findings are reported with a minimal reproducer, and a regression test is added when test edits are in scope. Quietly patching mid-hunt would destroy the evidence and the report's credibility; fixes are a separate step the owner approves.

    That is a valid, useful result. "The five riskiest unasserted behaviors were probed and they hold" is real information with named boundaries — it tells you exactly how much confidence the green suite has earned, and where it has not.

    New in Testing & Debugging