How RepoMaster Enables AI Agents to Master GitHub Repositories for Complex Tasks

RepoMaster Core Framework

RepoMaster enables AI agents to treat a selected GitHub repository as a reusable tool. The workflow consists of three tightly coupled stages.

Stage 1 – Hierarchical Repository Analysis

Hierarchical Code Tree (HCT) : extracts package‑module‑class‑function hierarchy.

Function Call Graph (FCG) : records caller‑callee relationships.

Module Dependency Graph (MDG) : captures import‑level dependencies.

Repositories are scored on dependency density, code complexity and update frequency to prioritize core modules.

Stage 2 – Autonomous Exploration & Execution

Context‑aware code view: on‑demand inspection of any file, class or function.

Dependency tracing using FCG and MDG to follow call chains.

Keyword‑based code search within the repository.

Interactive feedback loop: the agent iteratively writes code, runs it, inspects logs and adjusts actions based on success or failure.

Stage 3 – Multi‑Level Information Filtering

Code reduction : extracts only the AST sub‑trees relevant to the current task.

Document reduction : splits large documentation into chunks and retrieves the most pertinent fragments.

Log reduction : keeps only the head and tail of execution logs that contain error messages.

Experimental Evaluation

Two benchmarks that require reuse of existing code were used.

MLE‑R : derived from OpenAI’s MLE‑Bench, focuses on machine‑learning tasks solvable inside real GitHub projects.

GitTaskBench : newly built suite covering tasks such as old‑photo restoration and speech denoising; introduces the Task Pass Rate metric to measure end‑to‑end delivery quality.

RepoMaster achieved a highest task success rate of 62.96 % (up from 40.74 % of the strongest baseline) and reduced token consumption to ≈57 % of that of SWE‑Agent.

Case Study – 3D Pose Estimation

In a 3‑D pose‑estimation task, baseline agents either failed due to blind trial‑and‑error or deviated from the core algorithm because they lacked a global view of the repository. RepoMaster’s hierarchical maps quickly identified the critical components, enabling efficient task completion.

Resources

Paper:

RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving

ArXiv PDF: https://arxiv.org/pdf/2505.21577 GitHub repository:

https://github.com/QuantaAlpha/RepoMaster