← ClaudeAtlas

rr-sabotagelisted

Measure what a test suite actually covers by planting defects in a COPY of the repository and checking whether the tests go red. Use when asked whether tests are any good, what they really cover, whether a green suite can be trusted, to audit or stress-test checks, or to find tests that cannot fail. Plants one small edit per claim (weaken an obligation, hand-edit a generated artifact, make a record lie, flip a declaration), runs the suite, and reports CAUGHT, MISSED or VOID, with a baseline run first so a suite that was already red cannot produce a meaningless result. A sample, never coverage; never writes to the audited tree.
yaerhf/research-ratchet · ★ 1 · AI & Automation · score 72
Install: claude install-skill yaerhf/research-ratchet
# Find out what your tests actually cover — by breaking things on purpose A green test suite is evidence that nothing broke it. It is evidence of nothing else. This skill runs the other half: **plant a defect in the code a check claims to cover, run the checks, and see whether they go red.** A check that stays green with a defect in place has been *demonstrated* not to cover it — not argued, demonstrated, which is why the finding survives an argument with whoever wrote the test. **Everything happens in a copy. The repository you point this at is never written to.** ## How to run it `audit.py` is bundled with this skill. It needs the command that runs the checks, and one or more sabotages: ```bash python audit.py --command "pytest -q" \ --plant "src/cart.py::if total > limit::if total >= limit" \ --label "boundary loosened by one" ``` Or a plan file, which is the normal way once there is more than one: ```json [ { "label": "discount can now exceed the total", "file": "src/pricing.py", "find": "discount = min(discount, total)", "replace": "discount = discount", "claims": "test_discount_never_exceeds_total" } ] ``` ```bash python audit.py --command "npm test" --plan sabotages.json ``` The tool runs the **unmodified** copy first. **If the checks are already red it refuses**, because nothing can be concluded about a check that was failing before you touched anything. ## Choosing what to break — this is the whole skill The tool plants and run