title	SupportDesk OpenEnv
emoji	🤖
colorFrom	blue
colorTo	green
sdk	docker
app_port	7860

SupportDesk OpenEnv

SupportDesk OpenEnv is a real-world customer support operations environment built for OpenEnv-style agent training and evaluation. Instead of a toy game, the agent must inspect support tickets, consult policy knowledge, assign urgency, route ownership, write a customer reply, and resolve the case correctly.

Why this environment

This environment models work that real support teams do every day:

• Billing dispute handling
• Security incident escalation
• Logistics recovery for VIP customers

The environment is deterministic and reproducible, but still requires multi-step reasoning across ticket facts, contract notes, recent events, and policy snippets.

Tasks

• billing_duplicate_charge (easy): identify a duplicate annual charge, route to billing, approve the refund, and communicate the settlement window.
• security_compromised_account (medium): triage a likely account takeover, escalate to Security Ops, and communicate containment steps.
• logistics_missing_order_vip (hard): resolve a missing event-critical shipment, dispatch a replacement, and apply a shipping credit.

Action Space

The agent interacts through typed actions:

• open_ticket
• search_policy
• set_priority
• route_ticket
• send_reply
• resolve_ticket

You can inspect the machine-readable schema at GET /action_schema.

Observation Space

Each observation contains:

• Task instruction
• Available ticket IDs
• The active ticket once opened
• Latest search results from the policy base
• Recent action log
• Partial score so far
• Remaining step budget

Reward Design

The reward function emits partial progress instead of a sparse pass/fail signal:

• 0.05 for opening the correct ticket
• 0.10 for consulting required policy knowledge
• 0.15 for correct priority
• 0.20 for correct team routing
• 0.20 for reply coverage of required policy details
• 0.30 for the correct final resolution

Total score is capped at 1.0, and /grader returns a deterministic component breakdown.

Endpoints

• POST /reset
• POST /step
• GET /state/{episode_id}
• GET /tasks
• POST /grader
• POST /baseline
• GET /health
• GET /action_schema

Local setup

python -m venv .venv
. .venv/bin/activate
pip install -e .
uvicorn main:app --reload --host 0.0.0.0 --port 7860

On Windows PowerShell:

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e .
uvicorn main:app --reload --host 0.0.0.0 --port 7860

Baseline

Run the deterministic baseline agent:

python baseline.py

It executes the built-in heuristic policy against all 3 tasks and prints reproducible scores.

Hugging Face Spaces deployment

This repository includes a Dockerfile that launches the environment with Uvicorn on port 7860, which is suitable for Spaces container deployment.