README.md

metadata

title: DiskPanic OpenEnv
emoji: 💥
colorFrom: red
colorTo: yellow
sdk: docker
app_port: 8000
pinned: false
license: apache-2.0

DiskPanic — SRE Incident Response OpenEnv

A real-world RL environment where an LLM agent plays an on-call Site Reliability Engineer responding to a production incident: the root filesystem is full and app.service has crashed. The agent must free space, restart the service, and preserve business-critical audit logs — the wrong rm -rf tanks the reward.

Built for the OpenEnv Round 1 Hackathon by Yash Pravin Pawar's team.

Why this env

Every SRE has lived this exact 3am nightmare. The env tests three skills:

1. Diagnosis — finding the bloated file with du / ls / find
2. Surgical deletion — removing the right thing without touching protected dirs
3. Recovery — restarting services and (on hard) dropping a logrotate config to stop a runaway writer

The reward signal is dense: the agent sees its score climb as disk usage drops, gets a bonus for restoring the service, and is penalized if the SHA-256 of /var/log/audit/ changes.

Tasks

ID	Scenario	Graded on
easy	One 8.7 GiB rotated nginx log is filling the disk.	Disk usage < 80% + audit dir untouched
medium	Disk full + app.service has failed.	disk(0.4) + service(0.4) + audit(0.2)
hard	Same + a runaway writer grows /var/log/app/runaway.log by 100 MiB every tick.	disk(0.3) + service(0.3) + audit(0.2) + logrotate config(0.2)

All graders return a scalar in [0.0, 1.0].

Action space

DiskPanicAction(command: str) — a single bash-lite command per step. Supported:

df                         ls <path>        du <path>
cat <path>                 find <path>      sha256sum <path>
rm [-rf] <path>            systemctl is-active|status|start|restart <svc>
echo "content" > <path>    (for writing files like logrotate configs)

Observation space

DiskPanicObservation:

• stdout: str — output of the last command
• df_output: str — current simulated df -h /
• service_status: str — active / inactive / failed
• task_id: str — current task (easy | medium | hard)
• step: int
• last_error: Optional[str]

Safety & sandbox

The env does not touch the real filesystem. Everything is a Python dict representing a virtual filesystem. Commands are parsed via shlex and dispatched to whitelisted operations — no subprocess, no shell expansion, no escape surface. This keeps the env deterministic, safe, and fast (runs easily on 2 vCPU / 8 GB RAM).

Running locally

# 1. Install
pip install -r requirements.txt

# 2. Build the Docker image
docker build -t disk-panic:latest .

# 3. Set env vars
export HF_TOKEN=<your-key>                   # Groq key or HF token
export API_BASE_URL=https://api.groq.com/openai/v1
export MODEL_NAME=llama-3.3-70b-versatile
export IMAGE_NAME=disk-panic:latest

# 4. Run inference (all 3 tasks)
python inference.py

Deployment

The env is deployed as a Hugging Face Space (Docker SDK). The FastAPI server is wired by openenv.core.create_fastapi_app and exposes the standard OpenEnv endpoints: /reset, /step, /state, /schema, /health, /ws, /metadata, /web.

Layout

8-DiskPanic/
├── inference.py             # required at root per hackathon spec
├── Dockerfile
├── openenv.yaml
├── requirements.txt
├── README.md
└── disk_panic/
    ├── __init__.py          # exports DiskPanicEnv, DiskPanicAction, DiskPanicObservation
    ├── models.py            # Pydantic Action + Observation
    ├── client.py            # EnvClient subclass
    └── server/
        ├── app.py           # FastAPI app via create_fastapi_app
        ├── environment.py   # DiskPanicEnvironment
        ├── scenarios.py     # the 3 task builders
        ├── graders.py       # deterministic reward functions
        └── vfs.py           # in-memory virtual FS + command parser