wtf is proofpatch?

ProofPatch is a verification tool built for a world where AI coding agents write bug fixes. Its core problem statement is simple: an agent saying a patch fixed a bug does not mean the bug is actually fixed, so ProofPatch sets up a controlled process to prove it before the patch is trusted. The workflow runs in five steps. First, a bug report is turned into something testable. Second, the failure has to actually be reproduced before any patching starts, so there is a real baseline to compare against. Third, the coding agent writes its patch inside a disposable Git clone rather than touching the real repository. Fourth, a fresh isolated workspace checks that the original failure is gone and that other existing checks still pass. Fifth, ProofPatch writes out a proof receipt recording the patch and the verification evidence, so someone can review it later. Other features include Docker based protected execution so patches run in a sandboxed environment, Git based isolation so the original codebase is never touched directly, support for different agent adapters, and general regression checking to catch a fix that breaks something else. Setting it up requires Python 3.12 or newer, Git, and Docker with Linux container support for the protected execution step. After installing it with pip, you initialize a starter configuration for your project, review and adjust the generated settings file, then kick off a run pointing at a described issue. ProofPatch gives back a run ID you can use to check status, inspect the stored receipt, and if everything checks out, apply the verified patch to a new branch in your actual repository. The project is licensed under the Apache License 2.0.