PatchProof

A passing test is useful. A test that fails before the fix is evidence.

PatchProof checks whether the same supplied regression test fails for an expected reason on a base Git commit and passes after a patch. It creates disposable snapshots and saves JSON, HTML, original JUnit reports, and execution evidence.

The distribution name is patchproof-ai. The Python import and command are both patchproof. Python 3.10+, Git, and Linux or macOS are required for verification. No model account is needed.

Install

python -m pip install 'patchproof-ai[runner]'
patchproof --help

The runner extra installs pytest. The base package has no runtime dependencies. Target-project dependencies must already exist in your reviewed environment or container image; PatchProof does not install them automatically.

Verify reviewed code

patchproof run \
  --repo /path/to/reviewed-project --base HEAD \
  --patch /path/to/fix.diff --test /path/to/test_regression.py \
  --expect 'BUG_ADD' --repeat 3 \
  --output /tmp/patchproof-run-001 --trusted-local

Use a fresh output directory outside the source repository. The regression test must contain a meaningful assertion whose failure message includes BUG_ADD. An unrelated failure does not count as reproduction. Existing test suites and per-case expectations can also be compared.

Trusted-local execution is not a sandbox. It runs reviewed Python code as your current user. For an explicitly prepared local Docker image, select --docker-image IMAGE instead. Docker restrictions reduce exposure but are not a hostile-code VM boundary. Read the execution guide before running unfamiliar code.

Python SDK

from pathlib import Path
from patchproof import VerificationRequest, verify_patch
from patchproof.runners import TrustedLocalRunner

# Only for repository and test code you own and have reviewed.
result = verify_patch(
    VerificationRequest(
        repo=Path('/path/to/reviewed-project'),
        base='HEAD',
        patch=Path('/path/to/fix.diff'),
        test=Path('/path/to/test_regression.py'),
        expected_failure='BUG_ADD',
        output=Path('/tmp/patchproof-sdk-001'),
    ),
    runner=TrustedLocalRunner(trust_confirmed=True),
)
print(result['verdict'], result['reason'])

Interpret the evidence

Input and runner errors also exit 2. Setup/collection errors, missing reports, skips, timeouts, and unexpected base failures never count as successful reproduction. Repeated runs preserve every attempt.

Scope and optional proposals

PatchProof supplies regression evidence, not a proof of complete correctness. Snapshot preparation leaves the source checkout intact, but trusted-local test code can access host resources. Windows process cleanup, binary/rename/copy patches, symlink/submodule snapshots, and arbitrary project fixtures in the injected regression test are not supported.

Optional Claude test proposals require explicit remote opt-in. Selected files are sent to the configured provider; schema/syntax-checked outputs remain untrusted .proposed files. They are never executed automatically. Verification still requires human review and an explicit execution adapter.

Documentation and source

• Quick start and authored demo
• CLI, Docker, and SDK recipes
• Architecture
• Verification status
• Security boundaries
• Report a bug

Original project code is MIT licensed. Third-party dependencies retain their own licenses.