title	Service Incident Triage Environment
emoji	🚨
colorFrom	purple
colorTo	red
sdk	docker
pinned	false
app_port	7860
tags	openenv
base_path	/web

Service Incident Triage — OpenEnv Environment

An AI agent acts as an on-call engineer performing service incident triage. Given partial logs and metrics from a multi-service system, the agent must strategically inspect services, identify the root cause, and declare the correct severity — all while minimizing unnecessary diagnostic steps.

Environment type: POMDP (partially observable Markov decision process) Episode length: 3–6 steps depending on difficulty

---

Motivation

Real incident response requires quickly narrowing down which service is failing from noisy, incomplete signals. This environment captures that challenge:

• Partial observability — only one service's logs are visible at the start.
• Information-gathering actions cost steps.
• Misleading signals at higher difficulties.
• Shaped reward encourages efficient investigation.

---

Action Space

Every action is a JSON object with three fields:

Field	Type	Values	Description
action	string	inspect_logs, inspect_metrics, declare	What to do
target	string	service name (e.g. "db", "auth")	Which service to act on
severity	string	P0, P1, P2	Required only for declare

Action semantics

Action	Effect
inspect_logs(target)	Reveals the full log text for that service
inspect_metrics(target)	Reveals latency, error rate, CPU, memory for that service
declare(target, severity)	Ends the episode with a root-cause declaration

---

Observation Space

Field	Type	Description
visible_logs	dict[str, str]	Service → log text (only for inspected services)
visible_metrics	dict[str, dict[str, float]]	Service → {latency_ms, error_rate, cpu_pct, memory_pct}
services	list[str]	All services in the current incident
step_count	int	Steps taken so far
max_steps	int	Step budget for this task
message	str	Human-readable status after each action
incident_id	str	Current incident identifier
done	bool	Whether the episode has ended
reward	float	Reward for the latest step

---

Reward Function

Intermediate rewards

Signal	Reward
Useful inspection (reveals anomaly)	+0.1
Redundant / uninformative inspection	−0.05

Terminal rewards

Outcome	Reward
Correct service and severity	+1.0
Correct service, wrong severity	+0.6
Wrong service	0.0 − 0.3 penalty

Penalties

Condition	Penalty
Exceeding step budget	−0.2

---

Tasks

Task	Difficulty	Max Steps	Description
easy	🟢 Easy	3	One service clearly spiking; obvious anomaly in logs and metrics
medium	🟡 Medium	5	Multiple noisy signals; correlated failures across services
hard	🔴 Hard	6	Misleading logs; hidden root cause; requires careful multi-service investigation

Each task has 5 pre-generated incidents (15 total) with deterministic seeding for reproducible evaluation.

---

Setup & Usage

Prerequisites

pip install openenv-core[core]>=0.2.2

Start the server locally

cd service_incident
uvicorn server.app:app --host 0.0.0.0 --port 7860

Run with Docker

cd service_incident
docker build -f server/Dockerfile -t service-incident-env .
docker run -p 7860:7860 service-incident-env

Interact via HTTP

# Reset with a task
curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"task_id": "easy", "seed": 42}'

# Step — inspect logs
curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"action": "inspect_logs", "target": "db"}'

# Step — declare
curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"action": "declare", "target": "db", "severity": "P0"}'

# Get state
curl http://localhost:7860/state

Deploy to Hugging Face Spaces

openenv push service_incident --space-id your-username/service-incident-triage

Run the baseline inference

export API_BASE_URL="https://api.openai.com/v1"
export MODEL_NAME="gpt-4o-mini"
export HF_TOKEN="your-api-key"
export ENV_URL="http://localhost:7860"

python inference.py

---

Baseline Scores

Task	Average Score
easy	(run inference.py to fill)
medium	(run inference.py to fill)
hard	(run inference.py to fill)

---

Project Structure

service_incident/
├── __init__.py                 # Package exports
├── client.py                   # WebSocket client (EnvClient subclass)
├── data.py                     # Pre-generated incident dataset (15 incidents)
├── models.py                   # Action & Observation Pydantic models
├── openenv.yaml                # OpenEnv spec metadata
├── pyproject.toml              # Python project configuration
├── tasks.py                    # Task definitions (easy / medium / hard)
├── README.md                   # This file
└── server/
    ├── __init__.py
    ├── app.py                  # FastAPI application (create_app)
    ├── Dockerfile              # Container build
    ├── requirements.txt        # Server dependencies
    └── service_incident_environment.py  # Core POMDP environment logic