Dojo SDK Core

Core models and types for the Dojo RL environment framework. Not intended for standalone use by end users.

Overview

dojo-sdk-core provides the core components shared across the Dojo ecosystem:

• Action Types - Standardized action models for computer use (click, type, scroll, etc.)
• Reward Functions - Reward functions for dojos that need custom validation
• Shared Types - Common data structures for tasks, scores, and configurations
• Settings - Loading of environment variables used by the Dojo SDK
• Task Loaders - Loading of tasks from local files or remote datasets

Installation

uv add dojo-sdk-core

Key Components

Action Types

Set of computer-use actions: KEY, CLICK, RIGHT_CLICK, DOUBLE_CLICK, MIDDLE_CLICK, DRAG, MOVE_TO, PRESS, HOTKEY, SCROLL, TYPE, DONE, WAIT

Reward Functions

Reward functions provide custom validation logic for tasks that cannot be completed by simple state comparison. Each reward function follows a standard signature:

def reward_function(initial_state: Dict[str, Any], final_state: Dict[str, Any]) -> Tuple[float, str]:
    """
    Args:
        initial_state: The state before any actions were executed
        final_state: Current state of the environment

    Returns:
        Tuple of (score, reason) where:
        - score: float between 0.0 and 1.0 where environment is considered completed if score is 1.0
        - reason: human readable string explaining the validation result
    """

Available reward functions:

• _validate_get_2048: Checks if a 2048 tile is present on the game board
• _validate_search_for_dzaka: Validates successful search for a specific user in LinkedIn
• _validate_drag_to_different_column: Validates issue movement to a different column in Linear
• _validate_drag_two_issues_same_user: Validates multiple issues moved within same user's board

Documentation

Visit docs.trydojo.ai for complete documentation.