ReasoningBank Documentation

The mini-swe-agent package provides several environment implementations for executing code within different isolation levels. Depending on your security and infrastructure requirements, you can choose from the following:

• local.py: Executes code directly on the host machine using subprocess.run. This is the least isolated method.
• docker.py: Executes code within a Docker or Podman container.
• singularity.py: Executes code within a Singularity or Apptainer container.

In addition to the core environments, there are specialized implementations available in the extra/ directory:

• extra/swerex_docker.py: Executes environments using Docker via swerex.
• extra/bubblewrap.py: Executes environments using bubblewrap.

To test the agent in a local environment without container isolation, use the following scripts:

• hello_world.py: An extremely simple example demonstrating how to use the default.py agent.
• mini.py: Provides an interactive experience using either the interactive.py or interactive_textual.py agent/simple UI.

# Example usage (conceptual)
python hello_world.py
python mini.py

To solve GitHub issues within a controlled, isolated environment, use the github_issue.py script. This script utilizes a Docker environment to provide the necessary sandbox for the agent to operate safely.

python github_issue.py

SWE-Bench implementation is built upon mini-swe-agent. To set it up:

1. Navigate to the ./third_party directory.
2. Install it from source using editable mode:

pip install -e .

This will install dependencies specified in pyproject.toml.

cd third_party
pip install -e .

WebArena requires browsergym and a Docker environment.

1. Install BrowserGym: Follow the official documentation.
2. Docker Configuration: Download and configure the Docker environment following the webarena-setup tutorial. Execute scripts in numerical order and ensure website addresses are configured in the scripts as instructed.

ReasoningBank supports multiple model families via environment variables.

GPT Models

To use gpt-3.5-turbo, gpt-4, or gpt-4o, set your OpenAI API key:

export OPENAI_API_KEY="your-openai-api-key"

Gemini & Claude (via Vertex AI)

To use Gemini models (gemini-2.5-flash, gemini-2.5-pro) or Claude (claude-3-7-sonnet@20250219) on Vertex AI, follow these steps:

1. Authenticate via Google Cloud CLI: gcloud auth application-default login.
2. Set the required project and location variables:

export GOOGLE_CLOUD_PROJECT="your-project-id"
export GOOGLE_CLOUD_LOCATION="your-region"
export GOOGLE_GENAI_USE_VERTEXAI="True"

# Example for Vertex AI configuration
gcloud auth application-default login
export GOOGLE_CLOUD_PROJECT="your-project-id"
export GOOGLE_CLOUD_LOCATION="your-region"
export GOOGLE_GENAI_USE_VERTEXAI="True"

The repository includes a patched webarena/ harness in third_party/webarena/ that includes fixes for shopping annotations, wishlist evaluation, and click stability. To ensure the system uses this patched version instead of a pip-installed browsergym.webarena, add the third_party directory to your PYTHONPATH.

export PYTHONPATH="$(pwd)/third_party:$PYTHONPATH"

To prepare WebArena tasks, download the raw test files from this source and place them in the WebArena/config_files directory. Alternatively, use the patched version located at third_party/webarena/test.raw.json.

Then, run the following script to generate the necessary config files:

python WebArena/config_files/generate_config_files.py

# Assuming you are in the root directory
python WebArena/config_files/generate_config_files.py