title	Code Debug Environment
emoji	🐛
colorFrom	blue
colorTo	purple
sdk	docker
app_port	7860
tags	openenv code-debugging reinforcement-learning

Code Debug Environment

A Python code debugging environment built on OpenEnv. An AI agent receives broken Python code, fixes it, and the environment scores the fix by running it against test cases.

Built for the Meta x Scaler OpenEnv Hackathon.

---

How It Works

1. The environment presents broken Python code to the agent
2. The agent submits fixed code
3. The environment runs the fixed code against hidden test cases
4. The agent receives a score (0.0 to 1.0) and feedback on which tests failed
5. The agent can retry up to 5 times per task

There are 8 tasks across 3 difficulty levels. The agent must fix syntax errors, logic bugs, data structure misuse, and interdependent bugs across multiple functions.

---

Project Structure

code_debug_env/
  models.py            - Pydantic models (Action, Observation, State)
  client.py            - OpenEnv client for connecting to the server
  inference.py         - LLM-based agent that solves all 8 tasks
  openenv.yaml         - OpenEnv manifest
  pyproject.toml       - Python project config
  Dockerfile           - Container config for deployment
  server/
    app.py             - FastAPI app entry point
    environment.py     - Core environment logic, tasks, and grading

---

Tasks

Easy (3 tasks)

easy_001 - Fix syntax errors (missing colons, parentheses) in a calculate_average function. 5 tests.

easy_002 - Fix a missing return statement and a wrong comparison operator in find_max. 5 tests.

easy_003 - Fix off-by-one errors in repeat_string and first_n_chars. 5 tests.

Medium (3 tasks)

medium_001 - Fix logic errors in is_palindrome (wrong comparison) and count_vowels (wrong increment). Code runs but produces wrong results. 5 tests.

medium_002 - Fix a mutable state bug in merge_dicts (modifies original dict) and missing deduplication in unique_sorted. 5 tests.

medium_003 - Fix a wrong list operation in flatten_list (append instead of extend) and an off-by-one slice in chunk_list. 5 tests.

Hard (2 tasks)

hard_001 - Fix 3 interdependent bugs across compress_stream, decompress_stream, and stream_stats. The bugs compensate for each other, so the broken code passes all tests as-is. Fixing only 1 or 2 bugs breaks everything. All 3 must be fixed together. 6 tests.

hard_002 - Fix 3 bugs in a data pipeline: a boundary condition error in filter_and_sort (> vs >=), wrong sort order (ascending vs descending), and wrong key in summarize (sorts by age instead of score). 6 tests.

---

Scoring

score = tests_passed / total_tests

• Partial credit is given. Passing 3 out of 5 tests = 0.6
• If the submitted code has a syntax error or crashes, score = 0.0
• Code that runs longer than 3 seconds is killed (catches infinite loops)
• An episode ends when score reaches 1.0 or after 5 steps

---

Action Space

What the agent sends to the environment:

Field	Type	Description
fixed_code	str	The corrected Python code
task_id	str	Which task is being solved

Observation Space

What the environment sends back:

Field	Type	Description
broken_code	str	The original broken code
description	str	What the task is about
score	float	0.0 to 1.0
tests_passed	int	How many tests passed
total_tests	int	Total test cases
feedback	str	Which tests failed and why
done	bool	Whether the episode is over
difficulty	str	easy, medium, or hard

---

Setup

Prerequisites

• Python 3.11+
• uv (recommended) or pip

Install dependencies

uv sync

Or with pip:

pip install openenv-core fastapi uvicorn openai

---

Running Locally

Step 1: Start the server

uv run server

Leave this running in a terminal. The server starts on port 7860.

Step 2: Run the agent

In a separate terminal:

export HF_TOKEN=your_huggingface_token
export MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
export ENV_URL=http://localhost:7860

uv run python inference.py

You should see output like:

[START] task=easy_001 env=code-debug-env model=meta-llama/Llama-3.1-8B-Instruct
[STEP] step=1 action=fix_code reward=0.99 done=true error=null
[END] success=true steps=1 score=0.99 rewards=0.99
...
=== BASELINE SCORES ===
easy_001: 0.99
easy_002: 0.99
easy_003: 0.99
medium_001: 0.99
medium_002: 0.99
medium_003: 0.99
hard_001: 0.99
hard_002: 0.83
Average: 0.97

---

Docker

Build

docker build -t code-debug-env .

Run

docker run -p 7860:7860 code-debug-env

The server will be available at http://localhost:7860.

---

Deploy to Hugging Face Spaces

openenv push --repo-id your-username/code-debug-env

The Dockerfile is configured to expose port 7860, which is required by Hugging Face Spaces.

---

Baseline Scores

Tested with meta-llama/Llama-3.1-8B-Instruct:

Task	Difficulty	Score	Steps
easy_001	Easy	0.99	1
easy_002	Easy	0.99	1
easy_003	Easy	0.99	1
medium_001	Medium	0.99	1
medium_002	Medium	0.99	1
medium_003	Medium	0.99	1
hard_001	Hard	0.99	2
hard_002	Hard	0.83	5
Average	-	0.97	-

---

Environment Variables

Variable	Required	Default	Description
HF_TOKEN	Yes	-	Hugging Face API token
MODEL_NAME	Yes	-	Model to use for inference
API_BASE_URL	No	https://router.huggingface.co/v1	LLM API endpoint
ENV_URL	No	http://localhost:7860	Environment server URL