RPG-ZeroRepo

RPG-ZeroRepo turns Repository Planning Graphs (RPG) into a control layer for long-horizon AI coding agents. An RPG connects requirements, features, architecture, files, and dependencies in one structured graph, so repository-level context persists across a long task instead of living in chat history and file search.

CoderMind is the agent-facing part of the project. It gives a coding agent a persistent RPG workspace for planning, multi-file generation, repository understanding, and graph-aware editing, reached through three interfaces: a cmind CLI that initializes a workspace in a new or existing repository, slash commands such as /cmind.feature_spec, /cmind.code_gen, /cmind.encode, and /cmind.rpg_edit, and MCP graph tools (search_rpg, explore_rpg, get_node_detail, list_rpg_tree). A post-commit hook can keep the graph in step with code changes. Claude Code and GitHub Copilot are listed as verified agents.

The repository also carries the standalone research code behind two papers, runnable without an agent CLI. ZeroRepo implements the forward pipeline from requirements to an RPG to a repository in three phases: feature planning that decomposes a requirement into a feature tree, architecture design that maps features onto modules, files, classes, interfaces, and data flows, and graph-guided generation that writes interdependent files in dependency order. RPG-Encoder implements the reverse pipeline, lifting an existing codebase back into an RPG through semantic lifting, structure reorganization, and artifact grounding, then maintaining it incrementally from commit-level diffs rather than re-encoding after each change.

RepoCraft is the accompanying benchmark for repository-level code generation. It holds 1,052 tasks across six real-world Python projects (scikit-learn, pandas, sympy, statsmodels, requests, and django) and scores coverage of reference feature categories, accuracy as a unit-test pass rate and a semantic voting rate, and code statistics such as file count, lines of code, and token count.