AgentBox/README.md

🧠 CodeGuard Environment

📌 Overview

CodeGuard is a deterministic RL-style environment for evaluating code-fixing agents.
It simulates tasks like lint fixing, vulnerability patching, and structural refactoring with strict reward and grading rules.

---

🧩 Environment Schema

State

{
    "score": "float",
    "history": [
        {
            "step": "int",
            "action": "str",
            "reward": "float"
        }
    ]
}

Action

• Type: str
• Constraints:
  • Non-empty
  • Max length: 1000 chars

Reward

• Range: [-2.5, 1.0]
• Components:
  • Task score (grader)
  • -0.02 step penalty
  • -2.0 destructive penalty

🧪 Tasks & Difficulty

🟢 Easy — Lint Fix

• Goal: Fix syntax errors
• Grader: ast.parse()
• Output: 0.0 or 1.0

🟡 Medium — Vulnerability Patch

• Goal: Remove unsafe calls (eval, exec, etc.)
• Checks:
  • Unsafe calls removed
  • Function structure preserved
• Output: 0.0 – 1.0

🔴 Hard — Refactor + Type Hints

• Goal: Improve structure + add typing
• Checks:
  • Structural preservation
  • Type annotations present
• Output: 0.0 – 1.0

⚙️ Setup & Usage

Install Dependencies

pip install -r requirements.txt

Run API Server

uvicorn src.env:app --host 0.0.0.0 --port 7860

Run Inference Loop

export HF_TOKEN=your_token
python inference.py

📊 Baseline Scores (Mock)

Task	Score
Easy	0.88
Medium	0.71
Hard	0.54

Model: gpt-4.1-mini

🧪 Testing

Run validation tests:

pytest tests/

📌 Notes

• All graders are deterministic & stateless
• Execution time per grading <50ms
• No randomness in reward or evaluation