Large-Scale Language Models (LLMs) are adept at generating individual functions or single code files, but they struggle to generate complete code repositories from scratch. This paper addresses the generation of a complete repository, a key challenge for building consistent software systems from high-level specifications and maximizing the potential of automated code generation. This requires two stages of planning: defining features and modules (the proposal phase) and defining implementation details (the implementation phase). We point out that existing natural language-based approaches suffer from ambiguity and lack of structure, leading to unclear specifications, component inconsistencies, and design vulnerabilities. To overcome these limitations, we propose the Repository Planning Graph (RPG), a structured representation that encodes functions, file structures, data flows, and functions as a unified graph. RPG enables consistent long-term planning for repository creation through explicit blueprints. Based on RPG, we develop the ZeroRepo framework, which performs graph-based code generation through proposal-level planning, implementation-level configuration, and test verification. We evaluated ZeroRepo on the RepoCraft benchmark, which consists of six real-world projects and 1,052 tasks. ZeroRepo generated 36,000 lines of code and 445,000 code tokens, an average of 3.9x more than the previous best-performing model (Claude Code) and 68x more than other baselines. It also achieved 81.5% coverage and 69.7% test accuracy, outperforming Claude Code by 27.3 and 35.8 points, respectively. Further analysis revealed that RPG models complex dependencies, enables more sophisticated planning that scales nearly linearly, and accelerates localization by improving the agent's understanding of the repository.
