title	PyPatch
emoji	🐛
colorFrom	red
colorTo	yellow
sdk	docker
pinned	false
license	mit

PyPatch 🐛

An OpenEnv-compliant RL environment for AI-powered Python code debugging.

PyPatch challenges AI agents to identify and fix bugs in Python code across three difficulty levels. Reward is proportional to the fraction of hidden test cases the agent's fix passes — giving rich, continuous feedback for reinforcement learning.

Live links:

• Hugging Face Space: PyPatch Space
• GitHub Repository: Anurag-M1/PyPatch
• Judge UI: /ui

Why PyPatch

• Clean difficulty ladder from syntax repair to algorithmic debugging
• Dense partial-credit reward rather than brittle pass/fail scoring
• Deterministic task graders with explicit metadata
• FastAPI environment with clear OpenEnv-style interaction flow
• Baseline inference that emits structured logs and uses the evaluator proxy

---

Environment Overview

Property	Value
Tasks	3 (easy → medium → hard)
Max steps per task	5
Reward range	0.0 – 1.0
Termination	All tests pass OR max steps reached
API	OpenEnv-compliant REST (FastAPI)
Judge UI	/ui
Task metadata	Includes grader and non-boundary baseline score

---

Tasks

Task 1 — Fix Factorial Function (Easy)

The agent receives a factorial function with syntax errors (missing colon, wrong assignment operator in condition).
Reward: fraction of 4 test cases passing.

Task 2 — Fix Second Largest Element (Medium)

The agent receives a function with a logic error that returns the largest element instead of the second largest.
Reward: fraction of 5 test cases passing.

Task 3 — Fix Binary Search Algorithm (Hard)

The agent receives a binary search with 4 distinct algorithmic bugs: off-by-one boundary, wrong while condition, and two incorrect pointer updates.
Reward: fraction of 7 test cases passing.

---

Action & Observation Spaces

Observation

{
  "task_id": "task_hard_binary_search",
  "task_name": "Fix Binary Search Algorithm",
  "difficulty": "hard",
  "task_description": "...",
  "buggy_code": "def binary_search(arr, target):\n    ...",
  "error_hint": "Hint: There are 4 bugs — check ...",
  "step_count": 1
}

Action

{
  "fixed_code": "def binary_search(arr, target):\n    ..."
}

---

API Endpoints

Method	Path	Description
GET	/	Health check
GET	/ui	Judge-facing landing page
GET	/health	Health check
GET	/tasks	List all tasks
POST	/reset	Start episode (optional: {"task_id": "..."})
POST	/step	Submit fix {"fixed_code": "..."}
GET	/state	Current episode state

---

Local Setup

git clone https://github.com/Anurag-M1/PyPatch
cd PyPatch

python3 -m pip install -r requirements.txt

python3 -m uvicorn main:app --host 0.0.0.0 --port 7860

# Optional: in another terminal — run inference
export API_BASE_URL=https://your-litellm-proxy.example/v1
export API_KEY=your_proxy_key
export MODEL_NAME=gpt-4o-mini
export ENV_URL=http://localhost:7860
python3 inference.py

Open locally:

• http://127.0.0.1:7860/ui
• http://127.0.0.1:7860/docs

Docker

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

---

Reward Design

Partial credit is intentional. If an agent fixes the syntax error in Task 3 but misses two algorithmic bugs, it might pass 4/7 test cases and receive reward 0.57. This dense feedback signal makes PyPatch suitable for RL training, not just one-shot evaluation.

Baseline Behavior

inference.py is designed to satisfy evaluation constraints safely:

• emits [START], [STEP], and [END] blocks to stdout
• uses API_BASE_URL and API_KEY when provided by the evaluator
• produces per-task scores that are strictly between 0 and 1
• keeps behavior deterministic for reproducibility

---

Required Environment Variables (inference)

Variable	Description
API_BASE_URL	Evaluator-provided LiteLLM proxy URL
API_KEY	Evaluator-provided proxy API key
MODEL_NAME	Model identifier
ENV_URL	PyPatch server URL

---

Built for OpenEnv Round 1 — April 2026