microsoft/RPG-ZeroRepo

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

What it solves

Coding agents often struggle with long-horizon tasks because they rely on transient chat histories and simple file searches, leading to "requirements drift," lost architectural decisions, and inconsistent multi-file edits. RPG-ZeroRepo introduces a persistent Repository Planning Graph (RPG) that acts as a durable control layer, ensuring that requirements, architecture, and implementation remain aligned across the entire lifecycle of a codebase.

How it works

The system operates through three primary components:

  1. CoderMind: An agent-facing layer that integrates with tools like Claude Code and GitHub Copilot. It provides a persistent RPG workspace via CLI setup, slash commands (e.g., /cmind.code_gen), and MCP (Model Context Protocol) tools for searching and traversing the graph.
  2. ZeroRepo: A forward pipeline that transforms natural-language requirements into a structured feature tree, then into an architectural RPG, and finally into a complete repository generated in dependency-aware order.
  3. RPG-Encoder: A reverse pipeline that analyzes existing codebases to extract an RPG, capturing semantic intent and structural dependencies. It includes an evolution mechanism to incrementally update the graph based on commit diffs.

Who it’s for

It is designed for developers using AI coding agents (specifically Claude Code and GitHub Copilot) who need to build new repositories from scratch, understand complex existing codebases, or perform large-scale, graph-aware updates to existing projects.

Highlights

  • Persistent Planning: Replaces transient chat artifacts with a structured, searchable graph of the repository's design.
  • Bidirectional Pipeline: Supports both generating code from requirements (ZeroRepo) and extracting plans from existing code (RPG-Encoder).
  • Agent Integration: Provides specialized slash commands and MCP tools for seamless use within modern AI IDEs.
  • Incremental Updates: Uses commit-level diff parsing to keep the planning graph in sync with code changes without requiring full re-encoding.
  • RepoCraft Benchmark: Includes a specialized evaluation suite with over 1,000 tasks across major Python projects to measure repository-level generation quality.

Related

  • Project
    OthmanAdi/planning-with-files

    A persistent, file-based planning system for AI coding agents that stores task plans and progress on disk to prevent memory loss during context resets.

  • 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
    trailhq/Graft

    Graft is an open‑source TypeScript/Node.js tool that builds a local, markdown‑based knowledge graph of a codebase (summaries, key snippets, typed links) using tree‑sitter for structural parsing and optional LLM summarisation. The graph is cached in a `graft/` folder, refreshed only on file changes, and automatically wired into coding agents like Claude Code, Cursor, Codex, and Gemini. Benchmarks claim up to 46 % fewer tool calls, 42 % token savings, 60 % latency reduction, and a 12‑point correctness boost on SWE‑bench. Integration is via a single `graft init` command that writes agent‑specific instruction files; the CLI also offers search, visualization, and telemetry controls. The project is MIT‑licensed and targets faster, cheaper AI‑assisted development.

  • Project
    vitali87/code-graph-rag

    Code‑Graph‑RAG parses multi‑language codebases with Tree‑sitter, stores the AST‑derived structure in a Memgraph knowledge graph, and lets you query, edit and optimise the code via natural‑language prompts powered by an LLM. It supports dozens of languages, runtime trace overlay, and can run as an MCP server for tool integration.