Vurnix — the honest gate for AI-written code

AI coding agents don't just write bugs. Asked to "make the tests pass", they take shortcuts a human reviewer rarely thinks to check for: they delete the failing assert, skip the test, pad the suite with assert True, or import modules that don't exist anywhere. The suite goes green; the code is still broken.

Vurnix is a set of deterministic, zero-dependency checkers that make those shortcuts impossible to hide. No model in the loop, no heuristics you have to trust — every check is plain code you can read, with a hard exit code.

pip install vurnix
vurnix gate ./myproject        # compile + phantom-import + honest test count, one verdict

The design rule everything here follows: a check that cannot run is not a check that passed. Missing toolchain? That's a labelled SKIP, never a silent OK. Any failing check is a BLOCK and a non-zero exit — there is no "mostly green".

The checkers

vurnix integrity — the anti-weakening guard

Snapshot the test file before handing it to an agent; compare after:

vurnix integrity snapshot tests/test_api.py > before.json
# ... agent "fixes" the failing tests ...
vurnix integrity compare before.json tests/test_api.py
# WEAKENED: asserts dropped 17 -> 3
# integrity: test file WEAKENED — do NOT count this round as fixed.  (exit 1)

Delta-based: pre-existing skips are fine; only new weakening flags. Strengthening (adding tests) always passes. When a test function disappears, it is named:

WEAKENED: test function(s) deleted/missing: test_redirect_302 (6 -> 5)

vurnix coverage — the honest test-count floor

Prints one integer: the number of distinct, non-trivial test functions. A test counts only if it makes a real assertion (or expects a raise), and identical bodies collapse to one — so "write N easy copies" and "write N assert True" both count as what they are.

vurnix mutation — test strength, measured

vurnix mutation run ./app --max 24 --strict --threshold 0.34 -- python -m pytest -q
# mutation: 9/13 killed — score 69%

A suite that only checks status == 200 scores ~15% on a real CRUD app; a suite that checks behaviour scores ~70%. Survivors are listed by file and mutation so the fix is actionable. Stdlib AST for Python; conservative masked-regex operators for JS/Go/Java (mutants that break syntax simply fail the run — the safe direction).

vurnix phantom — imports that exist nowhere

Real example from our benchmarks: a 12B local model wrote from models import UrlRequest and from db import Database — no models.py, no db.py, no such packages. Syntax checks pass (it is valid syntax); the failure surfaces later as a cryptic install error. This prints the file and the invented module, before anything runs. Host site-packages are deliberately excluded from resolution, so "works on my machine" doesn't mask a missing dependency.

Why we built this

These checkers are extracted from the gate of a local-first autonomous coding pipeline we run against small (12B-class) local models — an environment where every failure mode of agentic coding shows up early and often. All of them earned their place by catching real incidents, including:

• a cloud model, brought in to rescue a failing build, rewriting a 17-assertion test file down to 3 assertions to go green — blocked by integrity;
• theatre suites (assert-the-status-code-and-nothing-else) passing every gate until the mutation score exposed them (15% vs 69% on the same app);
• build-tool resolution failures being mistaken for "tests passed" — which is why SKIP is a first-class, labelled result and never an OK.

The orchestration pipeline itself is not part of this package; these are its trust primitives, usable with any agent, harness, or CI.

Install

pip install vurnix        # Python >= 3.10, zero runtime dependencies
vurnix gate --help

js/go/java checks use your existing node/go/mvn+javac toolchains when present, and report labelled SKIPs when not.

An npm i -g @vurnix/cli stub is also published under the project's npm scope; the Python package is primary.

License

MIT — see LICENSE.