SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration

🔗 Github Link: https://github.com/SKYLENAGE-AI/SWE-CI

🔗 Paper Link: Coming soon ...

🏆 What is SWE-CI? A Coding benchmark for continuous integration

SWE-CI is a repo-level coding benchmark designed to evaluate how AI agents handle long-term code evolution through a Continuous Integration (CI) loop.

SWE-CI introduces a unique dual-agent workflow to simulate the operation of a professional software team. The architect agent analyzes CI-feedback (test failures, backtracking, errors) and transforms it into high-level requirements documents (natural language). The programmer agent, guided by the architect's requirements, modifies existing source code to meet the ever-changing needs of the codebase. By repeatedly executing the cycle of run_tests -> define_requirements -> modify_code, SWE-CI effectively simulates the real-world software development cycle and measures the agent's ability to maintain the software repository over the long term.

🏆 Why SWE-CI? Beyond Static Issue Solving

Most current benchmarks (e.g., SWE-bench) follow an "Issue-to-PR" paradigm: given a bug report at a single point in time, the model must fix it. However, real-world software engineering is rarely a one-shot task.

SWE-CI introduces four key shifts:

• From Snapshots to Evolution: Instead of fixing a single bug, SWE-CI focuses on the trajectory between two commits (from ‘current’ commit to ‘target’ commits). It models how codebases actually grow, refactor, and evolve over time.

• From Static Descriptions to CI-Driven Requirements: Instead of relying on human-written issue descriptions, SWE-CI uses the "Test Gap" as the primary driver. By running the Target Tests on the Current Source, we generate raw CI failure reports. In the SWE-CI framework, an Architect Agent then translates these technical logs into natural language requirements, which a Programmer Agent uses to implement the changes.

• Iterative Refinement (The CI Loop): SWE-CI is built for a closed-loop evaluation. The model doesn't just "guess" the fix; it must interact with the environment, analyze test failures, and refine the code iteratively—exactly like a developer working against a CI pipeline.

• From One-off Fixes to Sustainable Maintenance: SWE-CI emphasizes the long-term maintainability of the software. It evaluates whether agents can accurately comprehend complex systems and propose the most essential requirements. It challenges agents to write readable and extensible code that facilitates future evolution cycles, rather than just "hacking" a solution that passes the current tests but hinders future growth.

Dataset description

SWE-CI-full contains a total of 137 tasks. Each task consists of three parts: metadata, source code, and environment images.

• Metadata (recorded in /metadata/*.csv files): Contains 9 fields in total:

  • task_id: The unique ID used to identify the task.
  • repo_name: The name of the open-source repository (hereinafter referred to as "the repository") referenced to build the task.
  • url: The GitHub link of the repository.
  • licence: The license used by the repository.
  • current_sha: A relatively early commit selected from the repository's commit history, serving as the starting point for the CI-Loop. current_sha represents the SHA256 value of this commit.
  • target_sha: A relatively late commit selected from the repository's commit history, serving as the GroundTruth for the CI-Loop. target_sha represents the SHA256 value of this commit.
  • test_gap: The number of test cases that pass in the repository branch represented by target_sha but fail in the branch represented by current_sha.
  • image_sha: The hash value of the environment image file, used to verify download integrity.
  • code_sha: The hash value of the source code archive, used to verify download integrity.

• Source Code (/data/**/code.zip files): Packaged and downloaded based on the url field.

• Environment Images (/data/**/image.tar.gz files): We designed a complete pipeline that allows us to automatically generate the environment dependencies required to run the repository based on the source code. Subsequently, we packaged these environments into Docker image files to ensure reproducibility.

For more detailed information: including the data collection, cleaning, and filtering processes, the construction of environment images, and further explanations of the metadata, please refer to our paper.

Licensing

This dataset is released under a multi-licensing model to account for the different sources and types of content:

1. Metadata (/metadata/*.csv files)

The metadata, including the curation, task formulations, and the specific selection of commits, is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

• What this means: You are free to share and adapt the material, but you must give appropriate credit (citation) to this project.

2. Source Code (/data/**/code.zip files)

The source code contained in code.zip consists of sampled files from various open-source repositories.

• Ownership: These files remain the property of their original authors.
• License: They are governed by the original open-source licenses (e.g., MIT, Apache 2.0, BSD) of the respective upstream projects.
• Disclaimer: We redistribute these code snippets for research and benchmarking purposes under the redistribution permissions granted by their original licenses. Content Removal
• Content Removal: We respect open-source intellectual property. If you are a copyright holder and wish to have your repository removed from this dataset, please open an Issue in the "Community" tab of this Hugging Face repository. We will process valid requests promptly.

3. Environment Images (/data/**/image.tar.gz files)

The pre-built Docker images are provided as Aggregate Works for reproducibility.

• Curation Work: The logic, environment configurations (e.g., specific library versions, dependency resolution), and setup scripts integrated within these images are licensed under the Apache License 2.0.
• Third-Party Components: The underlying Operating System (e.g., Debian/Ubuntu), system libraries, and pre-installed packages within the images are subject to their own respective upstream licenses.