TraceCoder Enables Explainable, Auditable AI Code Generation.

Current AI coding agents often produce code without revealing the underlying rationale or the iterative process of its creation, making auditing and understanding difficult. TraceCoder introduces a novel approach to make AI-generated code transparent and accountable. The system employs three core mechanisms: first, a relational schema that meticulously records the history of each code snippet, including benchmark references, repair rounds, failure descriptions, and the LLM's explanations. Second, a browser-based visualization tool renders this history directly onto the source code, using heatmaps and hover annotations to show changes. Third, a unique position-key indexing scheme assigns stable, lexicographically ordered identifiers to code snippets, enabling precise tracking of modifications without affecting surrounding lines. Evaluated on algorithmic programming tasks, TraceCoder demonstrated its ability to explain how specific benchmark failures influenced the final program's lines. This mechanism transforms the internal "narrative" of automated code generation into an auditable and replayable process, which is crucial for building trust and ensuring accountability in production environments.