README.md

metadata

title: Git Conflict Resolver
emoji: 🔀
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
tags:
  - openenv

🔀 Git Conflict Resolver — OpenEnv Environment

An RL environment where AI agents learn to resolve Git merge conflicts.
Built for the OpenEnv Hackathon (Meta × Hugging Face × Scaler).

---

🧠 Environment Description & Motivation

Merge conflicts are a daily reality for every software team. Resolving them correctly requires understanding both the syntactic structure of code and the semantic intent of diverging branches — a genuinely hard task for AI agents.

This environment presents agents with real Python files containing <<<<<<< HEAD, =======, and >>>>>>> conflict markers. The agent must produce a clean, fully resolved file — no markers, valid syntax, and correct logic.

Unlike toy environments, this simulates:

• Easy conflicts: Accept an obvious incoming change (e.g. updated timeout value)
• Medium conflicts: Apply different resolution strategies per conflict block
• Hard conflicts: Combine additive changes from both branches (not just pick one)

---

📐 Action & Observation Space

Observation Space (structured JSON)

Field	Type	Description
task_name	string	Current task identifier
task_description	string	Natural language instructions for the agent
filename	string	Name of the file being resolved
file_language	string	Language of the file (python, text)
conflicted_content	string	Full file content with conflict markers
branch_ours	string	Name of the HEAD (current) branch
branch_theirs	string	Name of the incoming branch
num_conflicts	integer	Number of <<<<<<< blocks in the file
last_attempt	string | null	Agent's previous resolution (for retry)
last_error	string | null	Grading feedback from last step
step	integer	Current step number
max_steps	integer	Maximum allowed steps (10)
done	boolean	Whether the episode is finished

Action Space

{
  "resolved_content": "<full resolved file as a string>"
}

The agent outputs the complete file content with all conflict markers removed.

---

📋 Task Descriptions

Task 1: single_conflict — Easy

• File: config.py
• Conflicts: 1 block
• Description: A timeout value was changed from 30s to 60s on a feature branch. The agent must accept the incoming change.
• Expected difficulty: Any capable LLM should solve this in 1–2 steps.

Task 2: multi_conflict — Medium

• File: user_service.py
• Conflicts: 3 blocks
• Description: Authentication was refactored. Each block requires a different resolution: accept new import, keep original constant, accept new function implementation.
• Expected difficulty: Requires reading context across blocks.

Task 3: logic_conflict — Hard

• File: data_pipeline.py
• Conflicts: 2 blocks
• Description: Both branches added valid, additive features. The agent must combine them — not simply pick one side. Requires understanding code semantics.
• Expected difficulty: Frontier models (GPT-4, Qwen-72B) score ~0.5–0.7 without specific tuning.

---

🏆 Reward Function

The reward is shaped — agents get feedback at every step, not just at the end.

Signal	Value	Trigger
Improvement bonus	+0.75 × score_delta	Score improves over previous step
Marker-free bonus	+0.10	First time no markers remain
Perfect match bonus	+0.25	Score reaches 1.0
Stagnation penalty	-0.10	Identical submission as previous step
Step cost	-0.01 × (step/max_steps)	Every step

Grading Breakdown (per step)

Component	Score	Criterion
no_markers	0.25	No <<<<<<<, =======, >>>>>>> in output
valid_syntax	0.25	File parses as valid Python (AST check)
similarity	0.25	Fuzzy match ratio vs. expected resolution
exact_match	0.25	Character-exact match with expected output

---

🚀 Setup & Usage

Local Setup

# 1. Install dependencies
pip install -r server/requirements.txt

# 2. Start the server
uvicorn server.main:app --host 0.0.0.0 --port 7860 --app-dir server

# 3. Test the endpoints
curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"task": "single_conflict"}'

curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"resolved_content": "# your resolved content here"}'

curl http://localhost:7860/state

Docker

# Build
docker build -t git-conflict-resolver .

# Run
docker run -p 7860:7860 git-conflict-resolver

# Verify
curl http://localhost:7860/health

Run Baseline Inference

export HF_TOKEN=your_token_here
export API_BASE_URL=https://router.huggingface.co/v1
export MODEL_NAME=Qwen/Qwen2.5-72B-Instruct
export ENV_URL=http://localhost:7860

python inference.py

---

📊 Baseline Scores

Scores obtained using Qwen/Qwen2.5-72B-Instruct via HF Inference API.

Task	Score	Steps	Success
single_conflict	1.00	1	✅
multi_conflict	0.75	3	❌
logic_conflict	0.50	5	❌
Average	0.75	—	—

---

📁 Project Structure

openenv_hackathon/
├── server/
│   ├── main.py          # FastAPI server — /reset /step /state /health
│   ├── env.py           # Core environment logic (reset/step/state/close)
│   ├── models.py        # Pydantic Observation, Action, Reward models
│   ├── tasks.py         # Task definitions (3 tasks with conflict content)
│   ├── graders.py       # Deterministic graders (marker, AST, similarity, exact)
│   ├── reward.py        # Shaped reward function
│   └── requirements.txt
├── inference.py         # Baseline inference script (root — required)
├── openenv.yaml         # OpenEnv metadata (required for openenv validate)
├── Dockerfile           # Container build (port 7860 for HF Spaces)
└── README.md

---

🔧 API Reference

POST /reset

Start a new episode.

{ "task": "single_conflict" }

Returns: ConflictObservation

POST /step

Submit a conflict resolution.

{ "resolved_content": "# full resolved file..." }

Returns: { observation, reward, done, info }

GET /state

Returns current episode state (step, total_reward, history).

GET /health

Returns { "status": "ok" } — used for HF Space validation.

GET /tasks

Returns { "tasks": ["single_conflict", "multi_conflict", "logic_conflict"] }

---

👥 Team

Agent Smith — OpenEnv Hackathon, April 2026

• Ganesh Doosa (Team Lead)
• Gajula Akanksha
• Yashwanth Kumar