A production-grade, OpenEnv-compliant reinforcement learning environment and interactive platform for building, evaluating, and visualizing AI agents specialized in SRE (Site Reliability Engineering) and Auto-Healing.

This project simulates real-world DevOps system failures, providing a rich training ground for AI agents to diagnose and resolve incidents including memory leaks, database locks, and CPU scaling bottlenecks.

Ensure you have uv installed.

uv run uvicorn server.app:app

Port 8000 (standard for OpenEnv).

uv run streamlit run streamlit_app.py

Accessible at http://localhost:8501.

# Requires GROQ_API_KEY or OpenAI-compatible endpoint
uv run python3 inference.py

The platform is built on the openenv-core SDK and utilizes a Dual-State Model:

1. Observable State: CPU (%), Memory (%), DB Latency, and Service Status (API/DB/Cache).
2. Hidden Root Causes: Memory leaks, database locks, and CPU scaling bottlenecks.

• Failure Propagation: Metrics evolve dynamically based on hidden failures. A memory leak (memory > 85%) increases the objective chance of a service crash.
• Multi-Step Chains: Resolving 'Expert' tasks requires specific action sequences (e.g., investigating logs → clearing cache → restarting the API).
• Log Generator: A 3-depth log system that only reveals the true 'Root Cause' message after the agent executes check_logs.

• restart_service:[api|database|cache]
• scale_up:cpu
• optimize_database
• clear_cache
• check_logs
• do_nothing (only valid when system is healthy)

Episodes are graded out of 1.0 via a deterministic 5-component scoring model:

• Outcome (35%): Reaching a 'healthy' status.
• Logic Match (25%): Using the correct corrective action for the specific hidden cause.
• Efficiency (15%): Minimizing unnecessary steps.
• Health Bonus (15%): Maximizing resource optimization.
• Diagnostic (10%): Rewarding log investigation before fixing.

The project includes a production-grade inference engine (inference.py) that supports:

• Groq (LLaMA-3): Blazing-fast inference for real-time SRE responses.
• Reasoning Protocol: The agent identifies the root_cause, cites evidence_logs, and proposes a plan before taking an action.
• Safety Fallbacks: Hardcoded SRE heuristics protect the system if the LLM fails or hallucinations are detected.

.
├── env/                   # Core Logic
│   ├── environment.py     # Rewards, State Transitions, Failures
│   ├── tasks.py           # Dynamic Scenario Generation
│   └── models.py          # Pydantic Schemas
├── server/
│   └── app.py             # FastAPI Endpoints (/reset, /step, /auto-run)
├── streamlit_app.py       # Interactive Web UI
├── inference.py           # AI Agent & Reasoning Engine
├── openenv.yaml           # Manifest for Hugging Face Deployment
└── requirements.txt       # Dependencies

This environment is optimized for deployment as a Docker Space:

# Register the environment and push to HF
openenv push

The deployed space includes:

• Interactive UI at /web
• FastAPI Documentation at /docs
• OpenEnv Specification at /spec

MIT License. Created for the OpenEnv DevOps/SRE Auto-Healing Benchmark.