FlowFixer-EDA: A Tool-Feedback-Augmented Large Language Model Agent for Automated Diagnosis and Repair of Open-Source RTL-to-GDS Design Flow Failures

Abstract

Debugging open-source RTL-to-GDS design flows remains challenging because failures may arise from heterogeneous sources, including RTL code, constraints, configuration files, floorplanning parameters, placement density, routing settings, and timing reports. Directly applying large language models (LLMs) to raw EDA logs is often unreliable, as long tool outputs contain redundant warnings, downstream symptoms, and stage-dependent messages. This paper proposes FlowFixer-EDA, a tool-feedback-augmented LLM agent for automated diagnosis and repair of RTL-to-GDS design flow failures. The framework integrates stage-aware log parsing, EDA knowledge retrieval, structured patch generation, patch validation, and iterative tool-feedback verification. The LLM is used to infer root causes and generate repair actions, while actual EDA tool reruns determine whether a repair is accepted. To support evaluation, we construct EDA-FlowBench, a benchmark covering multiple open-source designs, technology platforms, flow stages, and error categories. Experimental results on EDA-FlowBench show that FlowFixer-EDA achieves a flow completion repair success rate (FC-RSR) of 81.5% (163/200 tasks) and a strict quality repair success rate (SQ-RSR) of 78.5% (157/200 tasks), substantially outperforming the strongest equal-budget baseline, Iterative RAG-Only LLM, which achieves an FC-RSR of 61.5% and an SQ-RSR of 47.5%. FlowFixer-EDA also improves root-cause diagnosis accuracy and patch validity while substantially reducing harmful repairs. These results demonstrate that tool-grounded LLM agents can provide practical and reliable debugging support for open-source EDA workflows. These findings demonstrate feasibility in controlled open-source RTL-to-GDS workflows; their transferability to commercial EDA tools, proprietary PDKs, large-scale SoC designs, signoff flows, and industrial debugging environments remains to be established.