Groundtruth – checks your AI coding agent's claims against the Git diff (github.com)

Groundtruth, a new plugin for Claude Code, challenges the reliability of AI coding agents by verifying their claims against actual Git diffs, without relying on additional models or complex workflows. It operates deterministically, assessing whether the agent’s reported accomplishments—such as successful code additions and tests—truly align with the produced changes. Groundtruth identifies discrepancies, such as missing files, unmet security rules, and incomplete tests, effectively ensuring that agents do not misrepresent their outputs. This instant verification allows developers to catch errors as they occur, before the coding session concludes. The significance of Groundtruth lies in its ability to tackle the common issue of AI agents' overconfidence in their outputs—often falsely claiming deliverables that aren't actually present or functional. By employing a first-principles approach, it audits claims against specific project rules and code changes, enhancing accountability without introducing false positives typical in traditional review tools. Furthermore, the plugin operates entirely offline, improving security and reliance by ensuring its verdicts are built solely on verifiable evidence, thereby giving developers a reliable tool to maintain code quality while leveraging AI assistance.