The developer proposes storing project context directly within the repository rather than in chat history. This approach uses the GRACE methodology, enhanced by the doubled-graph utility, to ensure that project intent, module boundaries, and validation criteria remain linked to the codebase.
The system utilizes two types of graphs: a declared graph for project intent and a computed graph built from the code's Abstract Syntax Tree. A tool called detect_changes compares these graphs to identify discrepancies, such as unlinked code or missing module contracts, which can be integrated into CI/CD pipelines.
The methodology supports two operational modes. The migration mode allows an agent to analyze existing code and establish documentation, while the post-migration mode treats artifacts as the primary source of truth, blocking changes that deviate from defined contracts.
The framework relies on ten principles, including explicit intent, semantic code structure using module contracts and anchors, and observable belief states where the AI logs its reasoning directly within the code. This structure aims to provide consistent context for AI agents across different sessions.
