rakshithakho/bugfixerenv

🐛 BugFixerEnv

An OpenEnv-compatible reinforcement-learning environment where an agent reads a buggy Python function and submits a corrected version. Reward = fraction of test cases passed (0.0 – 1.0).

Table of Contents

• —Environment Description
• —Observation Space
• —Action Space
• —Tasks
• —Reward Function
• —API Reference
• —Setup & Running
• —Docker
• —Baseline Inference
• —Reproducible Scores

Environment Description

BugFixerEnv simulates real-world code debugging. At each episode:

1. 1.The environment presents a buggy Python function along with a plain-English description of what it should do.
2. 2.The agent submits fixed_code containing the corrected function.
3. 3.The environment executes the submission against hidden test cases and returns a reward equal to the fraction passed.
4. 4.The episode ends when the agent achieves a perfect score (1.0) or exhausts max_steps (5) attempts.

This mirrors the daily workflow of software engineers who must identify and fix bugs in unfamiliar code — a high-value, real-world task.

Observation Space

{
  "task_id":     "easy | medium | hard",
  "description": "What the function should do (natural language)",
  "buggy_code":  "The broken Python function",
  "steps_taken": 0,
  "max_steps":   5,
  "last_reward": 0.0,
  "done":        false
}

Action Space

{
  "fixed_code": "<corrected Python source code as a string>"
}

The submitted string must contain a valid Python function with the same name as the original.

Tasks

Difficulty progression

• —Easy — Single character change. All 4 test cases fail with buggy code.
• —Medium — Subtle algorithmic error. 2 of 5 tests pass with bug (partial reward = 0.40).
• —Hard — Two independent bugs. Fixing one yields partial reward (0.29–0.71), both needed for 1.0.

Reward Function

reward = tests_passed / total_tests   # range: 0.0 – 1.0

Partial progress is always rewarded. A submission passing 3 of 5 tests receives 0.60, not 0. This creates a meaningful gradient for learning agents.

API Reference

Setup & Running

git clone https://github.com/rakshithakh/bugfixerenv
cd bugfixerenv
pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

Browse docs at http://localhost:8000/docs.

Docker

docker build -t bugfixerenv .
docker run -p 7860:7860 bugfixerenv

Baseline Inference

export API_BASE_URL=https://rakshithakho-bugfixerenv.hf.space
export MODEL_NAME=meta-llama/Llama-3.3-70B-Instruct
export HF_TOKEN=hf_xxx

python inference.py

Baseline Scores

Reproducible Scores

All test cases are fully deterministic. temperature=0.0 ensures reproducible LLM outputs. Same input always produces same score.

Project Structure

bugfixerenv/
├── main.py          # FastAPI app
├── env.py           # Core environment logic
├── grader.py        # Scoring / test execution
├── tasks.py         # Task definitions
├── inference.py     # Baseline agent
├── openenv.yaml     # OpenEnv spec
├── pyproject.toml   # Project metadata
├── uv.lock          # Locked dependencies
├── requirements.txt
├── Dockerfile
├── README.md
└── server/
    └── app.py       # Server entry point