Show HN: Overreach – audit your AI agent's diff against the prompt you gave it (github.com)

A new standalone tool named Overreach has been released, designed to catch "scope creep" in AI-generated code. When provided with a prompt and the resulting code diff, Overreach analyzes whether the changes made by the AI agent exceed the originally defined scope. This is particularly significant for developers utilizing AI coding assistants, as it prevents unapproved modifications such as adding endpoints, environment variables, or other unintended features, effectively ensuring that AI-generated outputs adhere strictly to user intent. Overreach operates in three stages: it first extracts the authorized scope from the prompt using a language model, parses the diff for additions, and finally compares these to produce findings. The tool can be used in a standalone manner or integrated into CI/CD pipelines, marking any commits that contain significant scope creep. It accommodates multiple LLM providers and can run in a more conservative "paranoid mode" without an API key. By providing this functionality, Overreach enhances developer oversight, mitigates risks before code commits, and supports a more disciplined approach to AI-assisted software development.