How it uses Jev
The test selector sends a git diff as Jev state and asks one Score question per discovered test. Each q-prefixed key identifies a test; the four ordered levels range from unrelated to likely failing. The selector keeps tests above its chosen risk cutoff.
What Jev decides
Can this change alter the named test outcome? The numeric suffix is generated per test.
A git diff; each question additionally names one test file, test name, line range and optional prior-failure file hints.
App workflow
- Discover tests and diff
Read changed code and enumerate test cases using the runner’s names and locations.
- Ask Jev in batches
Send diff state with one four-level Score question per test, splitting oversized requests as needed.
- Select tests
Compare scores with the gate cutoff and retain malformed or uncertain answers conservatively.
Why it is interesting
It applies a shared diff context to many bounded test-impact judgments, reducing test runs while preserving a fail-safe path when a model answer is malformed.