Your Coding Agent Needs a Map, Not a Chat Log

What it does

RPG-ZeroRepo builds a persistent Repository Planning Graph that acts as a control layer for long-horizon AI coding agents. Instead of relying on chat history and file search, agents such as Claude Code and GitHub Copilot work against a structured graph tracking requirements, architecture, and cross-file dependencies. The repository bundles a practical agent integration called CoderMind with standalone research pipelines for generating code from requirements and encoding existing repositories back into graphs.

The interesting bit

The system closes a bidirectional loop: ZeroRepo turns natural-language requirements into an RPG and generates interdependent files in dependency order, while RPG-Encoder maps existing codebases back into the same graph format and updates them incrementally via commit diffs. This gives agents a durable, navigable map of the repository rather than a mental model rebuilt from scratch every prompt.

Key highlights

• CoderMind exposes the RPG through slash commands and MCP tools (search_rpg, explore_rpg, get_node_detail) and can keep the graph synchronized with code changes via a post-commit hook.
• ZeroRepo runs a three-phase forward pipeline—feature planning, architecture design, and graph-guided generation—to build repositories from requirements.
• RPG-Encoder reverse-engineers a repository into an RPG through semantic lifting and diff-based incremental evolution, avoiding full re-encoding after every change.
• RepoCraft benchmarks repository-level generation with 1,052 tasks across six real-world Python projects, measuring coverage, pass rate, and semantic accuracy.
• The underlying research was accepted at ICLR 2026 and ICML 2026.

Verdict

Teams using Claude Code or Copilot for multi-file, repository-scale work should evaluate this as a structured alternative to chat-driven context. Solo developers knocking out single-file scripts will find the graph overhead unnecessary.