What is nowreck?

NoWreck is a command line tool that checks whether an AI coding assistant's explanation of a code change actually matches what happened in the code. When an AI tool edits your files and describes what it did, NoWreck compares that description against a structural analysis of the real changes, rather than asking another AI to judge the work. The tool works in three steps. First it scans all Python files in a before and after snapshot of a project and parses them to build a list of every function, class, and method. Second, it compares the two scans to detect what actually changed, such as functions or classes being added or removed, new files appearing, or new function calls showing up. This comparison becomes the single record of truth that the rest of the tool relies on. Third, it checks each claim the AI made against that record. If a claim matches a detected change, it is marked confirmed. If a claim contradicts what was detected, for example the AI says something was added when it was actually removed, it is marked contradicted. If nothing matches either way, the claim is marked unverifiable rather than guessed at. NoWreck currently supports only Python projects, since it relies on parsing Python's abstract syntax tree. It can be installed from a cloned copy of the repository using pipx and requires Python 3.10 or later. It works with any OpenAI-compatible model endpoint, including Groq, DeepSeek, Ollama, LM Studio, OpenRouter, and OpenAI itself, so a user can plug in their own AI provider and API key. There are two main ways to use it. In prompt mode, you describe a change in plain language and NoWreck calls the configured model, gets back a set of claims, and verifies them automatically. In pre and post mode, you supply two folder snapshots directly, which is useful for checking a change that already happened. Output can be shown as colored terminal text or as structured JSON for use in automated pipelines. The project is upfront that static analysis has real limits. It cannot see through dynamic behavior such as exec, eval, dynamic imports, or getattr calls, since these are not visible to a plain structural scan.