openhands/runtime/README.md

OpenHands Runtime

Introduction

The OpenHands Runtime folder contains the core components responsible for executing actions and managing the runtime environment for the OpenHands project. This README provides an overview of the main components and their interactions. You can learn more about how the runtime works in the Docker Runtime documentation.

Main Components

1. base.py

The base.py file defines the Runtime class, which serves as the primary interface for agent interactions with the external environment. It handles various operations including:

• Bash sandbox execution
• Browser interactions
• Filesystem operations
• Environment variable management
• Plugin management

Key features of the Runtime class:

• Initialization with configuration and event stream
• Asynchronous initialization (ainit) for setting up environment variables
• Action execution methods for different types of actions (run, read, write, browse, etc.)
• Abstract methods for file operations (to be implemented by subclasses)

2. impl/action_execution/action_execution_client.py

The action_execution_client.py file contains the ActionExecutionClient class, which implements the Runtime interface. It is an abstract implementation, meaning it still needs to be extended by a concrete implementation to be used.

This client interacts with an action_execution_server (defined below) via HTTP calls to actually perform runtime actions.

3. action_execution_server.py

The action_executor_server.py file contains the ActionExecutor class, which is responsible for executing actions received via the /execute_action HTTP endpoint. It returns observations in the HTTP response.

Key features of the ActionExecutor class:

• Initialization of user environment and bash shell
• Plugin management and initialization
• Execution of various action types (bash commands, IPython cells, file operations, browsing)
• Integration with BrowserEnv for web interactions

4. Other Implementations

The ./impl/ directory contains a few different Runtime implementations, all of which extend the ActionExecutionClient class. These implementations handle the lifecycle of a Docker container or other environment running the ActionExecutor server.

There are currently four implementations:

• Docker (runs locally in a Docker container)
• Remote (runs via a custom HTTP API for creating, pausing, resuming, and stopping runtimes in a remote environment)
• Modal (uses the Modal API)
• Runloop (uses the Runloop API)

Workflow Description

1. Initialization:

   • The Runtime is initialized with configuration and event stream.
   • Environment variables are set up using ainit method.
   • Plugins are loaded and initialized.

2. Action Handling:

   • The Runtime receives actions through the event stream.
   • Actions are validated and routed to appropriate execution methods.

3. Action Execution:

   • Different types of actions are executed:
     • Bash commands using run method
     • IPython cells using run_ipython method
     • File operations (read/write) using read and write methods
     • Web browsing using browse and browse_interactive methods

4. Observation Generation:

   • After action execution, corresponding observations are generated.
   • Observations are added to the event stream.

5. Plugin Integration:

   • Plugins like Jupyter and AgentSkills are initialized and integrated into the runtime.

6. Sandbox Environment:

   • The ActionExecutor sets up a sandboxed environment inside a Docker container.
   • User environment and bash shell are initialized.
   • Actions received from the OpenHands backend are executed in this sandboxed environment.

7. Browser Interactions:

   • Web browsing actions are handled using the BrowserEnv class.

Important Notes

• The runtime uses asynchronous programming (asyncio) for efficient execution.
• Environment variables can be added dynamically to both IPython and Bash environments.
• File operations and command executions are abstracted, allowing for different implementations in subclasses.
• The system uses a plugin architecture for extensibility.
• All interactions with the external environment are managed through the Runtime, ensuring a controlled and secure execution environment.

Runtime Types

Docker Runtime

The Docker Runtime is designed for local execution using Docker containers:

• Creates and manages a Docker container for each session
• Executes actions within the container
• Supports direct file system access and local resource management
• Ideal for development, testing, and scenarios requiring full control over the execution environment

Key features:

• Real-time logging and debugging capabilities
• Direct access to the local file system
• Faster execution due to local resources

This is the default runtime used within OpenHands.

Remote Runtime

The Remote Runtime is designed for execution in a remote environment:

• Connects to a remote server running the ActionExecutor
• Executes actions by sending requests to the remote client
• Supports distributed execution and cloud-based deployments
• Ideal for production environments, scalability, and scenarios where local resource constraints are a concern

Key features:

• Scalability and resource flexibility
• Reduced local resource usage
• Support for cloud-based deployments
• Potential for improved security through isolation

At the time of this writing, this is mostly used in parallel evaluation, such as this example for SWE-Bench.

Related Components

• The runtime interacts closely with the event system defined in the openhands.events module.
• It relies on configuration classes from openhands.core.config.
• Logging is handled through openhands.core.logger.

This section provides an overview of the OpenHands Runtime folder. For more detailed information on specific components or usage, please refer to the individual files and their docstrings.