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

The Core Challenge

Building an entire software project from scratch using an LLM agent is harder than it seems. Existing systems assume the repository structure already exists—you feed them a predefined folder layout and ask them to fill in the code. But real zero-to-all generation means the agent must figure out what files and folders to create in the first place while keeping everything modular and coherent. The authors' key insight: explicitly evolve the repository architecture before writing code, guided by modularity metrics, rather than hoping good code naturally emerges from a bag of requirements.

How Repo0 Works

The system maintains an architectural blueprint called a Dual-DAG—think of it as two interconnected graphs. The first graph organizes requirements (high-level features from the user); the second organizes components (modules and files). These two graphs stay aligned throughout the process.

The workflow has two phases. Phase 1: Structural Evolution. Starting from natural-language requirements, Repo0 iteratively adjusts component boundaries—deciding which modules should exist, how they depend on each other, and what each should contain. This isn't random; it's guided by modularity metrics that reward clean separation and low coupling. The system keeps refining structure until it reaches convergence—essentially, when further tweaks don't meaningfully improve modularity. Phase 2: Code Generation. Once the architecture stabilizes, the agent generates actual code for each component, informed by the finalized structure. It uses test-driven development as a guide: write tests first, then code to satisfy them.

The Dual-DAG is the lever here. By explicitly tracking both what you're building (requirements) and how you're organizing it (components), the system avoids the trap of building code in isolation. Every structural decision is justified by modularity; every component knows its place in the requirement landscape.

Why It Matters and the Results

Most LLM code agents generate functions or scripts, not full projects. The few that tackle entire repositories assume someone has already decided "you need a models/ folder, a utils/ folder, etc." Real-world development doesn't work that way. Repo0 closes that gap.

The authors tested on six real-world repositories using GPT-4o mini and DeepSeek V3.2. Against RPG—the strongest prior baseline for repository-level planning—Repo0 achieved significant gains:

• Functionality Coverage (how many required features are present) improved by up to 20.08 percentage points.
• Pass Rate (tests that actually pass) improved by up to 29.74 percentage points.

Ablations confirm both the Dual-DAG representation and the modularity-guided structural evolution are essential; removing either drops performance substantially. The structural-evolution analysis shows the framework genuinely converges on reasonable architectures rather than flailing randomly.

The practical implication: LLM agents can now scaffold realistic multi-file projects end-to-end, with coherent module organization, not just individual code snippets.