DEV Community
Follow
I built a tiny CI tool to keep AI agent configs from drifting in my repo
Managing AI coding agents often leads to fragmented rules scattered across various documents. These rules dictate tool usage, agent interactions, and escalation procedures. Crucially, the CI pipeline does not enforce adherence to these living documents. To address this, a new tool has been developed that centralizes these rules. The rules are now defined in a YAML file within the repository. This format specifies allowed tools for different agents and lists blocked tools. A Python validator checks for violations, failing the CI pipeline if an agent attempts unauthorized actions. These checks include using ungranted tools or calling disallowed agents. Additionally, the system can enforce these rules at runtime before tools execute. A Python module allows integration into agent runners to block forbidden actions. The project is designed to be a lightweight contract-testing layer, not a full agent framework or sandbox. The creator is seeking feedback on the YAML structure and the effectiveness of the implemented checks. They are also interested in identifying potential gaps in what the system can catch. Contributions through issues and pull requests are welcomed.
