microsoft/RPG-ZeroRepo

[ICLR 2026] RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation

What it solves

AI coding agents often lose repository-level context during long tasks: requirements drift, architecture decisions get forgotten, and edits miss hidden dependencies. This project gives agents a persistent graph-based planning workspace, so they can keep track of the whole repository throughout a session.

How it works

The core idea is a Repository Planning Graph (RPG) that serves as a structured, searchable map of the repository—linking requirements, features, architecture, files, and dependencies. The project provides three main pieces:

  • CoderMind is the agent-facing layer. It works with Claude Code and GitHub Copilot, giving them a persistent RPG workspace via a CLI, slash commands, and MCP graph tools. Agents can plan, generate code, understand an existing codebase, and make graph-aware edits.
  • ZeroRepo is the forward pipeline: starting from a natural-language requirement, it decomposes it into features and architecture, builds the RPG, then generates the repository in dependency order.
  • RPG-Encoder is the reverse pipeline: it takes an existing codebase and encodes it back into an RPG, so agents can navigate and maintain repositories.
  • RepoCraft is a benchmark for evaluating repository-level code generation, measuring coverage, accuracy, and code statistics.

Who it’s for

  • Developers using AI coding assistants (Claude Code or GitHub Copilot) who need to handle large, multi-file projects without losing the big picture.
  • Researchers reproducing the papers behind RPG, ZeroRepo, or RPG-Encoder, or working on repository-level code generation.
  • Teams that want to evaluate or improve long-horizon coding agents.

Highlights

  • Persistent RPG workspace replaces chat history and file search as the agent’s memory for repository context.
  • Supports both directions: generating a repository from requirements (ZeroRepo) and mapping an existing repository to a graph (RPG-Encoder).
  • Integrates with Claude Code and GitHub Copilot via slash commands and MCP tools.
  • Includes RepoCraft, a benchmark with 1,052 tasks across 6 real-world Python projects (scikit-learn, pandas, sympy, statsmodels, requests, django).
  • Backed by two peer-reviewed papers (ICLR 2026 and ICML 2026) and an MIT license.

Related

  • Project
    Autoloops/greplica

    A persistent, searchable engineering memory system for AI coding agents that reduces redundant context-gathering and token usage by storing durable repository knowledge.

  • Project
    conductor-oss/conductor

    An open-source durable execution platform for microservices and AI agents that turns complex workflow graphs into persisted, inspectable, and recoverable executions.

  • Project
    AgentMemoryRepo/agentmemoryrepo

    A Git-based memory standard that allows AI agents to maintain persistent, linked knowledge across sessions, users, and agent swarms.

  • Project
    nurettincoban/ai-prd-workflow

    A structured, document-driven workflow for AI coding agents that enforces RFC-based development to ensure project consistency, traceability, and long-term planning.

  • Project
    OthmanAdi/planning-with-files

    A persistent file-based planning system for AI coding agents that stores task state on disk to prevent memory loss during context resets, crashes, or long-running sessions.