harrry24/sysadmin-openenv

SysAdmin OpenEnv Simulator

This project is a submission for the Meta PyTorch OpenEnv Hackathon x Scaler Round 1. It implements a complete, real-world OpenEnv environment simulating a Junior System Administrator's workspace.

Environment Description

The SysAdmin Simulator evaluates an AI agent's ability to diagnose and repair simulated system, application, and process-level problems using a bash terminal interface. The agent is placed within an isolated Unix-like workspace where three distinct tasks of escalating difficulty are simulated.

This is not a toy game: investigating logs, fixing configuration files, and terminating rogue processes are daily realities for engineering and operations teams.

Tasks

1. 1.Easy (`easy_log_analysis`): The agent must grep or analyze /var/log/auth.log to find the exact IP address responsible for repeated failed SSH login attempts, and submit the IP.
2. 2.Medium (`med_config_fix`): The agent must fix a misconfigured Nginx/web server port in /workspace/app/config/server.conf from 9090 to the proper 8080, then restart the service.
3. 3.Hard (`hard_cpu_hog`): A rogue script (evil_miner.sh) is running and consuming fake CPU cycles. The agent must locate the process ID (PID) of this script and issue a kill command.

Action & Observation Spaces

The environment follows an interactive bash-shell paradigm.

Action Space

• —The agent outputs a single bash command string.
• —Schema: Action(command: str)
• —Example: {"command": "cat /var/log/auth.log"} or {"command": "submit 192.168.1.105"}

Observation Space

• —The environment executes the command within a strict 5-second timeout in an isolated sandbox. It returns standard bash output.
• —Schema: Observation(stdout: str, stderr: str, exit_code: int, task_id: str, reward: float, done: bool, info: dict)
• —Reward Strategy:
• —1.0: Task completed successfully.
• —0.0: Progressing / standard step.
• —-0.05: The executed command failed (exit code != 0, giving the agent a slight penalty for hallucinating commands or syntax errors).

Setup & Execution Instructions

Prerequisites

• —Docker OR Python 3.10+
• —An OpenAI-compatible API key (e.g., standard OpenAI or a proxy like Hugging Face Router as provided in the hackathon).

Running Locally with Docker (Recommended)

You can test the entire pipeline locally, just as it would run on a Hugging Face Space.

1. 1.Build the Environment Container:

   docker build -t openenv-sysadmin .

1. 1.Run the Environment:

   docker run -p 8000:8000 openenv-sysadmin

1. 1.Run Inference (in a separate terminal):

   export API_BASE_URL="https://api.openai.com/v1"  # Or your specific router
   export MODEL_NAME="gpt-4o-mini" # Or an open model
   export HF_TOKEN="your_api_key_here"
   
   python3 inference.py

Running Locally without Docker

1. 1.Install dependencies: pip install -r requirements.txt
2. 2.Start the FastAPI server: uvicorn app:app --host 0.0.0.0 --port 8000
3. 3.Set environment variables and run python inference.py

Baseline Scores

The environment was evaluated using the internal inference.py running the gpt-4o-mini baseline agent:

• —Easy Task (`easy_log_analysis`): [Score TBD by user]
• —Medium Task (`med_config_fix`): [Score TBD by user]
• —Hard Task (`hard_cpu_hog`): [Score TBD by user]

Judging Validation

This environment strictly adheres to the OpenEnv JSON model standard over HTTP endpoints (/reset, /step, /state). It meets the runtime limit requirement (the FastAPI server is highly lightweight, consuming <50MB RAM).