HGARena: Budgeted Test-Driven Multi-Agent Repository Issue Resolution with Heterogeneous Graph-Augmented Retrieval.
Abstract
Real GitHub issue resolution requires a system to read issue reports, search large repositories, edit one or more files, and check the patch with executable tests. We present HGARena, a role-structured multi-agent framework that combines heterogeneous graph-augmented retrieval with budgeted iterative patch refinement. HGARena builds a typed RepoGraph over files, symbols, tests, and edit hunks. The framework retrieves graph-grounded context and runs a submitter-reviewer-executor loop where Docker test execution is the only final acceptance check, and reviewers only guide search. Using GPT-4o under a budget of 20 LLM calls, HGARena achieves a 25.3% resolve rate on SWE-bench Lite. Controlled variants show that iteration is essential, RepoGraph helps control token cost in iterative settings, and reviewer guidance improves the RR-token trade-off. Crossbackbone runs overall show similar resolve rates but different token costs across three commercial backbones. Index Terms-Retrieval-Augmented Generation, Graph retrieval, Program repair, Repository-level issues I. I NTRODUCTION Repository-level issue resolution is fundamentally different from function-level code generation. Given a repository snapshot and a natural-language issue description, a system must locate the relevant code, generate a patch, and rely on executable tests to determine whether the patch is correct [1]. This setting is closer to real software maintenance than selfcontained coding benchmarks. It requires fault localization across many files, coordinated edits under repository dependencies, and reasoning from sparse binary feedback produced by test execution. A central obstacle is that the evidence needed to fix an issue is inherently heterogeneous. Useful signals may come from issue text, source files, functions or classes, call sites, failing tests, stack traces, and previous edit attempts. If a repository is treated as a flat text corpus, these signals are mixed together and their relations are weakened. For example, a failing test may point to a covered symbol, a file may define several relevant functions, and a symbol may call another symbol that is closer to the actual bug. We therefore model