title	Admission Helpdesk OpenEnv
sdk	docker
pinned	false

Admission Helpdesk OpenEnv

A real-world OpenEnv benchmark environment that simulates admissions support operations. The agent must triage, route, and resolve student tickets while respecting policy and SLA constraints.

Why this environment

This is not a toy game. It models an actual workflow run by admission teams:

• classify urgency (low, medium, high)
• assign the right queue (admissions, finance, tech, hostel)
• draft policy-safe support replies
• escalate only when needed
• close tickets only after meaningful progress

Action Space

The environment accepts one textual command per step:

• set_priority(low|medium|high)
• assign_team(admissions|finance|tech|hostel)
• draft_reply(<text>)
• escalate(<reason>)
• close_ticket()
• next_ticket()

Observation Space

Each step returns typed observation fields:

• task_name
• objective
• step
• active_ticket_id
• ticket_text
• student_tier
• channel
• sla_minutes_left
• queue_remaining
• history
• echoed_message

Tasks and Graders

The environment includes 3 deterministic graded tasks:

1. easy_priority_routing (easy)
2. medium_resolution (medium)
3. hard_sla_queue (hard)

Each task returns normalized score in [0.0, 1.0] via info.normalized_score.

Explicit agent graders are provided in graders.py:

• grade_easy_priority_routing
• grade_medium_resolution
• grade_hard_sla_queue

Each task in openenv.yaml includes explicit grader references (grader, grader_fn, agent_grader).

To select a task at runtime, call:

• POST /reset with {"task_name": "easy_priority_routing"} (or medium/hard task id)

Reward Design

Step rewards provide partial progress:

• positive for correct priority/team routing
• positive for high-quality policy-safe replies
• positive for required escalations
• positive for valid close actions
• penalties for unsafe wording, incorrect routing, unnecessary escalation, and SLA breaches

Required Submission Files

• inference.py (root-level)
• openenv.yaml
• Dockerfile
• README.md

Local Run

Install dependencies:

pip install -r requirements.txt

Run API server (for /reset / /step / /state):

uvicorn app:app --host 0.0.0.0 --port 7860

Run inference for one task:

set MY_ENV_V4_TASK=easy_priority_routing
set HF_TOKEN=<your_token>
python inference.py

Run deterministic baseline over all tasks:

python scripts/run_baseline.py

Deployment

This repository includes a working Dockerfile suitable for Hugging Face Spaces (Docker runtime). The container exposes port 7860 and serves the environment API with FastAPI.

Hugging Face Space Setup (Submission Ready)

Use this exact flow to deploy quickly:

1. Create a new Space on Hugging Face.
2. Select Space SDK as Docker.
3. Name it (example: admission-helpdesk-openenv).
4. Push this project to that Space repository.
5. Wait for build to complete and open the Space URL.

Option A: Connect and push with git

1. Create the Space first from UI.

2. In your local project folder, add HF remote and push:

   git init

   git add .

   git commit -m "Initial OpenEnv submission"

   git remote add hf https://huggingface.co/spaces/<username>/<space-name>

   git push hf main

Automated Windows flow (recommended)

If you are short on time, use these scripts in order:

1. Create Space (Docker SDK):

python scripts/create_hf_space.py --username <hf_username> --space <space_name> --token <hf_token>

2. Configure required Space variables/secrets:

python scripts/configure_hf_space.py --username <hf_username> --space <space_name> --token <hf_token> --hf-secret <hf_token>

3. Commit and push code to Space:

scripts\deploy_hf.cmd <hf_username> <space_name> "OpenEnv submission"

After push, wait for build completion in Space logs.

Option B: Upload from Hugging Face UI

1. Open your Space.
2. Go to Files.
3. Upload all repository files.
4. Confirm build logs show success.

Space Variables and Secrets

In Space settings, add these:

Variables:

• API_BASE_URL (example: https://router.huggingface.co/v1)
• MODEL_NAME (example: Qwen/Qwen2.5-72B-Instruct)
• MY_ENV_V4_BENCHMARK (optional, default already set)

Secrets:

• HF_TOKEN (or API_KEY)

Optional:

• LOCAL_IMAGE_NAME only if your inference flow uses from_docker_image(...)

Quick Post-Deploy Checks

After Space is live, verify:

1. POST /reset returns HTTP 200.
2. POST /step works with a sample action payload.
3. GET /state returns valid JSON state.
4. Local validator script passes against your Space URL.

Sample checks:

• curl -X POST https://<space-url>/reset -H "Content-Type: application/json" -d "{}"
• curl https://<space-url>/state

Validation

Use prevalidation script:

bash scripts/validate-submission.sh <your_hf_space_url>

If you are on Windows without WSL/Git Bash, run the same script from a Linux shell in CI.