Repo0: Design-Driven Zero-to-All Code Generation

Repo0 reframes zero-to-all code generation as continuous software-architecture evolution rather than one-shot planning. Starting from natural-language requirements, it builds a Dual-DAG: one directed acyclic graph represents functional coordination among requirements, another represents implementation dependencies among components, and an alignment relation preserves traceability between them. The agent iteratively applies split, merge, revise, save, and add actions, using cohesion, Jaccard-based coupling, connectivity, and requirement-coverage checks to refine component boundaries until structural convergence. Only then does it generate packages and files through dependency-aware, test-driven development, creating importable skeletons, synthesizing tests, implementing code, and using validation failures for localized repair. On the RepoCraft benchmark’s six real-world Python repositories, evaluated with OpenAI’s GPT-5 mini and DeepSeek V3.2, Repo0 achieved the best Functionality Coverage and Pass Rate in every reported setting. Against RPG, the strongest static-graph baseline, it improved Functionality Coverage by up to 20.08 percentage points and Pass Rate by up to 29.74 percentage points. In a statsmodels setting, metrics-guided evolution raised Functionality Coverage from 75.90% to 80.68%, Pass Rate from 81.90% to 85.51%, and Voting Rate to 98.65%. Ablations show structural evolution is the largest contributor, while removing the Dual-DAG, requirement context, or dependency-aware ordering also reduces performance; unconstrained LLM-directed refinement instead tends to over-decompose repositories and lower correctness.