rjv-test-typeslisted
Install: claude install-skill rjvim/ai-skills
# Test types that keep agents honest
A test the agent wrote with the same blind spots as its code proves
little. Pick the type whose inputs or oracle the agent did not choose.
## Pick the type
```text
Does a user or caller go through a flow (pages, API calls, a job)?
├── Yes → end-to-end test, always
└── No
Is there a rule that must hold for every input?
├── Yes → property test
└── No
Is an old system doing the same job?
├── Yes → differential test
└── No → unit test on the public function, real inputs
```
More than one branch can apply. A money feature usually gets an
end-to-end flow and a property test for its invariant.
## End-to-end
Pick this first when you can only afford one type. Either the whole
feature works or it does not, and a failure shows what broke.
- Drive the running app through the repo's `verify` skill
(`rjv-verify-skill`), a browser, a simulator or an API harness.
- Assert on what the user sees or the caller receives, never on
internal calls.
- Use the repo's feature map as the rubric: each flow with its path,
selectors and expected result.
## Property
Write a statement that must never fail and let a library generate
thousands of inputs against it: fast-check (JS/TS), Hypothesis
(Python), proptest (Rust), PHP's Eris or a seeded loop.
| Property | Example |
|---|---|
| round trip | decode(encode(x)) equals x; parse(format(x)) equals x |
| invariant | a balance never goes negative; units times NAV equals am