TraceCAD Improves Agentic CAD Generation with Persistent Repair State

Large Language Model (LLM)-based agents for Computer-Aided Design (CAD) often struggle with correction loops, losing critical information about design requirements, errors, and previous repair attempts. This paper introduces TraceCAD, a novel recovery layer designed to address this issue. TraceCAD maintains a persistent state that meticulously links requested features, individual modeling steps, evidence of failures, and potential repair outcomes. This comprehensive tracing allows the system to accurately diagnose problematic operations and conduct targeted, localized edits within the affected dependency regions. The system validates proposed repairs through execution and preservation checks, and crucially, it stores both successful and failed repair attempts in a reusable skill memory. Experimental results on DeepCAD-derived benchmarks show that TraceCAD achieves competitive geometric quality. The research highlights that persistent state and localized search are critical for improving recovery scores and reducing geometric regressions, while initializing the skill store further optimizes retries, token costs, and latency.