Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autonomous Issue Resolver: Towards Zero-Touch Code Maintenance

Aliaksei Kaliutau

TL;DR

This work reframes repository-scale automated program repair by replacing traditional file-centric representations with a Data-First Transformation Graph (DTG) that traces data lineage through transformations. The Autonomous Issue Resolver (AIR) combines a multi-agent, RL-informed framework with DTG-driven navigation to autonomously diagnose and repair code within CI/CD platforms, addressing the semantic trap of control-centric tools. Empirical results on SWE-bench Verified show AIR achieving 87.1% repair resolution, with ablations highlighting the critical roles of data-centric graphs and learned navigation. The paper also discusses practical integration into GitHub/GitLab ecosystems and outlines a roadmap toward polyglot DTGs, dynamic environment refinement, and broader security/refactoring applications. Overall, this approach lays a scalable, data-driven foundation for near-zero-touch software maintenance in increasingly complex codebases.

Abstract

Recent advances in Large Language Models have revolutionized function-level code generation; however, repository-scale Automated Program Repair (APR) remains a significant challenge. Current approaches typically employ a control-centric paradigm, forcing agents to navigate complex directory structures and irrelevant control logic. In this paper, we propose a paradigm shift from the standard Code Property Graphs (CPGs) to the concept of Data Transformation Graph (DTG) that inverts the topology by modeling data states as nodes and functions as edges, enabling agents to trace logic defects through data lineage rather than control flow. We introduce a multi-agent framework that reconciles data integrity navigation with control flow logic. Our theoretical analysis and case studies demonstrate that this approach resolves the "Semantic Trap" inherent in standard RAG systems in modern coding agents. We provide a comprehensive implementation in the form of Autonomous Issue Resolver (AIR), a self-improvement system for zero-touch code maintenance that utilizes neuro-symbolic reasoning and uses the DTG structure for scalable logic repair. Our approach has demonstrated good results on several SWE benchmarks, reaching a resolution rate of 87.1% on SWE-Verified benchmark. Our approach directly addresses the core limitations of current AI code-assistant tools and tackles the critical need for a more robust foundation for our increasingly software-dependent world.

Autonomous Issue Resolver: Towards Zero-Touch Code Maintenance

TL;DR

This work reframes repository-scale automated program repair by replacing traditional file-centric representations with a Data-First Transformation Graph (DTG) that traces data lineage through transformations. The Autonomous Issue Resolver (AIR) combines a multi-agent, RL-informed framework with DTG-driven navigation to autonomously diagnose and repair code within CI/CD platforms, addressing the semantic trap of control-centric tools. Empirical results on SWE-bench Verified show AIR achieving 87.1% repair resolution, with ablations highlighting the critical roles of data-centric graphs and learned navigation. The paper also discusses practical integration into GitHub/GitLab ecosystems and outlines a roadmap toward polyglot DTGs, dynamic environment refinement, and broader security/refactoring applications. Overall, this approach lays a scalable, data-driven foundation for near-zero-touch software maintenance in increasingly complex codebases.

Abstract

Recent advances in Large Language Models have revolutionized function-level code generation; however, repository-scale Automated Program Repair (APR) remains a significant challenge. Current approaches typically employ a control-centric paradigm, forcing agents to navigate complex directory structures and irrelevant control logic. In this paper, we propose a paradigm shift from the standard Code Property Graphs (CPGs) to the concept of Data Transformation Graph (DTG) that inverts the topology by modeling data states as nodes and functions as edges, enabling agents to trace logic defects through data lineage rather than control flow. We introduce a multi-agent framework that reconciles data integrity navigation with control flow logic. Our theoretical analysis and case studies demonstrate that this approach resolves the "Semantic Trap" inherent in standard RAG systems in modern coding agents. We provide a comprehensive implementation in the form of Autonomous Issue Resolver (AIR), a self-improvement system for zero-touch code maintenance that utilizes neuro-symbolic reasoning and uses the DTG structure for scalable logic repair. Our approach has demonstrated good results on several SWE benchmarks, reaching a resolution rate of 87.1% on SWE-Verified benchmark. Our approach directly addresses the core limitations of current AI code-assistant tools and tackles the critical need for a more robust foundation for our increasingly software-dependent world.

Paper Structure

This paper contains 39 sections, 1 equation, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Problem Overview and Background
  3. Limitations of existing tools
  4. Related Work
  5. Methodology
  6. A data-first graph representation
  7. Why Reinforcement Learning is crucial
  8. Why holistic reasoning is the key
  9. Algorithm details
  10. Code property graphs and representation bloat
  11. Data-centric optimization and flow-based programming
  12. Agentic navigation on graphs
  13. The Data-First Transformation Graph
  14. Formal definition
  15. Edge set: transformations
  16. ...and 24 more sections

Figures (5)

  • Figure 4.1: The partial visualization of the multi-dimensional Data Transformation Graph for a snippet of code. Nodes contains full description of processed data. Edges are effectively transformation functions (sanitize, handle)
  • Figure 6.2: The AIR Multi-Agent Architecture. The system employs a decoupled "Plan-Navigate-Execute" loop to manage context window limitations. The Context Agent (pink) operates exclusively within the graph domain to localize faults without reading file content. Once the fault subgraph is identified, the Maintenance Agent (yellow) formulates a high-level repair plan, which is translated into concrete syntax edits by the Editor Agent (gray). This separation prevents the "Semantic Trap" by ensuring the planner is conditioned only on causally relevant data lineages rather than raw repository noise.
  • Figure 7.3: Our product leverages the existing GitHub platform and workflows (Issues, Branches, PRs). There is no need for developers to learn new tools or interfaces. The Bot acts like another member of the team, communicating and contributing via established mechanisms. Such approach does not break existing automation. Note, that human developers can still review the Bot's code changes directly within the PR interface, leave comments, or request modifications if needed. This step maintains human oversight where desired.
  • Figure 10.4: Testing results on GitHub Issues. Numbers indicate how many issues were resolved. We randomly selected 10 issues from each repository.
  • Figure A.1: The reconstructed sample of Data Transformation Graph given the raw data to build Nodes and Edges from Table \ref{['tab:constructs']}.