SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents
Overview
SWE-MiniSandbox is a container-free platform for software engineering automation, AI agents, and code-focused systems. It aims to provide:
- A container-free, reproducible mini sandbox environment for experiments;
- Integration with multiple popular SWE agents / RL / code-repair frameworks
(e.g., SWE-Agent, SkyRL, SWE-ReX); - A unified environment preparation pipeline for SWE-bench, SWE-smith, and custom datasets;
- Extensible CLI tools and Python APIs that you can plug into your own research or systems.
Typical use cases include:
- Train your own SWE-agent with SWE-MiniSandbox;
- Evaluating new models with MiniSadnbox on SWE-bench;
- Building custom pipelines over your own repositories and issue/patch data.
Repository Layout
The core layout of this repository is:
mkdocs.yml # MkDocs configuration
readme.md # Project overview and documentation entry
docs/ # Documentation source (MkDocs)
index.md # Documentation homepage
guide/ # Guides and API reference
img/ # Images used in docs
config/ # Configuration files (env, experiments, models, etc.)
data/ # Data processing scripts and intermediate results
dataset/ # Datasets (or download scripts)
mini-swe-agent/ # Mini SWE Agent core components
pic/ # Additional image resources
R2E-Gym/ # RL / environment components (e.g., R2E-Gym)
sandboxdev/ # Minimal sandbox development & experiments
sh/ # Shell scripts (setup, batch runs, etc.)
SkyRL/ # SkyRL related code
SWE-agent/ # Modified version of SWE-Agent
SWE-bench/ # SWE-bench utilities and scripts
SWE-ReX/ # Modified version of SWE-ReX
SWE-smith/ # Modified version of SWE-smith
zip/ # Compressed resources