Metatron

Metatron is designed to represent the actual implementation decisions made within a codebase, focusing on aspects like preferred patterns, rejected approaches, edge cases, and internal conventions. It does this by structuring these decisions as priors and enabling coding agents to access them, ensuring they can write code similarly to a senior engineer who is already familiar with the codebase. This minimizes the need for rediscovery of conventions each time a developer addresses new parts of the code.

The system operates entirely on-premises, allowing it to work securely with private codebases. During the extraction process, only structural signals such as imports and commit subjects are sent to the model, ensuring sensitive information remains private. Feedback from agents is stored locally in a SQLite database, and all decisions begin as candidates for curation, ensuring that nothing is automatically promoted without human oversight.

Metatron allows for a unique decision curation process where human intervention is necessary to validate and promote decisions. Decisions are maintained in a structured format, supporting fields for pattern, scope, rationale, confidence, and source references. Users can engage with the system via commands that support mirroring decisions in a portable Open Knowledge Format (OKF) bundle, making it easier to share conventions across teams.

Installation can be automated via a shell script or be carried out manually, and Metatron can also be run using Docker for easy management. The tool is designed for developers who need to maintain high standards of code quality and adherence to team conventions while working on complex software projects.