README.md

metadata

title: Code Debugging Challenge
emoji: 🐛
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false
license: apache-2.0
tags:
  - openenv
  - reinforcement-learning
  - code-debugging
  - agentic-ai

🐛 Code Debugging Challenge - OpenEnv Environment

A production-ready OpenEnv environment where AI agents learn to debug Python code.

🎯 Overview

This environment challenges AI agents to identify and fix bugs in Python code snippets using the official OpenEnv framework from Meta-PyTorch and Hugging Face.

Key Features:

• ✅ Built with official OpenEnv library
• ✅ WebSocket-based client-server architecture
• ✅ Docker containerized for isolation
• ✅ Compatible with TRL, Torchforge, and other RL frameworks
• ✅ Production-ready with proper session management

🏗️ Environment Details

• Action Space: 4 discrete actions (analyze, fix, test, submit)
• Observation Space: Structured observations with code, errors, and feedback
• Reward Structure:
  • +1.0 for successful fix
  • -0.2 to -0.5 for failed attempts
  • +0.1 for analysis actions
  • -1.0 for premature submission
• Episode Length: Max 5 attempts per bug

🐞 Bug Types Included

1. Argument Count Errors - Wrong number of function arguments
2. Logic Errors - Incorrect loop variables and conditions
3. Exception Handling - Missing error handling for edge cases
4. Index Errors - Array/string index out of bounds
5. Infinite Recursion - Recursive calls without base case reduction
6. Type Errors - String/integer concatenation issues
7. Key Errors - Missing dictionary keys

🚀 Quick Start

Using Docker (Recommended)

from code_debug_env.client import DebugEnv

# Automatically starts Docker container and connects
env = DebugEnv.from_hub("openenv/code-debug-env")

# Reset to get first challenge
result = env.reset()
print(result.observation.buggy_code)
print(f"Expected output: {result.observation.expected_output}")

# Take action
from code_debug_env.models import DebugAction
action = DebugAction(action_type="test")
result = env.step(action)
print(f"Reward: {result.reward}")

# Cleanup
env.close()

🔧 Integration with RL Frameworks

With TRL (Transformer Reinforcement Learning)

from trl import OnlineDPOConfig, OnlineDPOTrainer
from code_debug_env.client import DebugEnv

config = OnlineDPOConfig(...)
trainer = OnlineDPOTrainer(
    config=config,
    env=DebugEnv.from_hub("openenv/code-debug-env"),
    # ... other args
)
trainer.train()

🏆 OpenEnv Challenge Submission

This environment is submitted to the OpenEnv Challenge: SOTA Environments to Drive General Intelligence (UC Berkeley AgentBeats Competition).

📜 License

Apache 2.0