0.1.1

Three fixes, all found by running the tool against a corpus of merged pull requests written by coding agents rather than against its own examples.

A real failure could be filed as "never ran." The two lists behind the provisional verdict were merged rather than ordered, so an assertion failure whose message happened to contain is not a function or undefined: — which is what a great many honest failures say — was read as a test that never started. Such a test left the report as provisional instead of as an alibi, silently. Evidence that a test body was entered now settles the question.

A python project with no config file was refused. A tests/ directory full of test_*.py and no pytest.ini or pyproject.toml produced "no test runner found" — a refusal over a missing config file rather than a missing runner.

Installing the skill copied the whole repository. 98 files to deliver one page of instructions. The skill now lives in skills/alibi/, and an install is one file.

Also adds experiments/corpus/ — the harness that found all three.

0.1.0

First public release.

Reverts the source to the base revision, keeps the tests the change added, runs each of them on its own and reads nothing but the exit code. A test that passes against the old source is not evidence for the change.

Install

npx skills add BOTIROFF-D/alibi     # for a coding agent
npm install -g @botiroff/alibi      # for you and CI

In this release

• The alibi run, with three verdicts — alibi, provisional, none — so that a test written against a module that did not exist yet is never counted as proof.
• Diff checks for deleted tests, newly skipped tests and weakened assertions.
• A static reading for assertions that cannot fail. It can never produce a lie; the severity ladder is fixed by provenance.
• An optional mutation pass over the lines the change touched.
• The claims ledger, .alibi/claims.md.
• Runners: vitest, jest, mocha, node --test, pytest, go test, cargo test, rspec, phpunit, and a shell template for anything else.
• A museum of six ways to buy a green suite, each built into a real repository by the test suite.

No runtime dependencies, no network calls, no model in the loop.