from fastapi import FastAPI, HTTPException, Request from fastapi.responses import HTMLResponse from environment import CustomerSupportEnv from models import Action, Observation app = FastAPI(title="Customer Support OpenEnv API") env_instance = CustomerSupportEnv() @app.post("/reset") def reset_environment(): try: obs = env_instance.reset() return {"observation": obs} except Exception as e: raise HTTPException(status_code=500, detail=str(e)) @app.post("/step") def step_environment(action: Action): if env_instance.obs is None: raise HTTPException(status_code=400, detail="You must call /reset before calling /step.") try: obs, reward, done, info = env_instance.step(action) return {"observation": obs, "reward": reward, "done": done, "info": info} except Exception as e: raise HTTPException(status_code=500, detail=str(e)) @app.get("/state") def get_current_state(): if env_instance.obs is None: raise HTTPException(status_code=400, detail="Environment has not been initialized. Call /reset.") return {"observation": env_instance.state()} # ========================================== # OPENENV VALIDATION ENDPOINTS (HACKATHON) # ========================================== @app.get("/health") def health_check(): """Simple health check for the validator.""" return {"status": "healthy"} @app.get("/metadata") def get_metadata(): """Returns environment metadata.""" return { "name": "customer-support-openenv", "description": "A real-world Customer Support OpenEnv simulation", "version": "1.0.0", "tags": ["customer-support", "nlp", "openenv"] } @app.get("/schema") def get_schema(): """Returns the JSON schemas for the environment's Pydantic models.""" return { "action": Action.model_json_schema(), "observation": Observation.model_json_schema(), "state": Observation.model_json_schema() } @app.post("/mcp") async def mcp_endpoint(request: Request): """Minimal JSON-RPC 2.0 response to satisfy validator MCP checks.""" try: body = await request.json() req_id = body.get("id") except Exception: req_id = None return { "jsonrpc": "2.0", "id": req_id, "result": {} } @app.get("/", response_class=HTMLResponse) def home(): return """ Customer Support Command Center — OpenEnv Hackathon
OpenEnv Hackathon · Live Environment

CUSTOMER
SUPPORT
COMMAND CENTER

A clean, modern mission dashboard to train AI on real customer tickets with clear actions, live reward tracking, and policy-safe outcomes.

Tasks
6 Scenarios
Actions
5 Controls
Step Limit
10/Episode
Target Score
0.90+
CLASSIFY ISSUE· SEARCH KB· CHECK LOGS· RESOLVE TICKET· ESCALATE TO HUMAN· SENTIMENT SCORE· REWARD SIGNAL· CLASSIFY ISSUE· SEARCH KB· CHECK LOGS· RESOLVE TICKET· ESCALATE TO HUMAN· SENTIMENT SCORE· REWARD SIGNAL·
Mission Library
6 TICKET MISSIONS
01
Easy
Billing Receipt Request
Classify as Billing, optionally verify policy via search_kb, then resolve clearly.
02
Medium
Dashboard Error 404
Identify Technical issue, pull KB guidance, and provide the correct fix path.
03
Hard
Digital Item Refund Demand
For non-refundable digital downloads, policy-safe handling usually means escalation.
04
Hard
VIP Outage
Tests if the agent can correctly classify high-urgency technical issues without being distracted by account status keywords.
05
Medium
Vague Complaint
The user says "It is broken." Ask one clarifying question, then classify/search/resolve efficiently.
06
Hard
Subscription Plan Refund Difference
Reason over refund policy edge case and complete with policy-consistent action.
Control Panel
5 CORE ACTIONS.
POLICY-FIRST EXECUTION.
🏷️
classify_issue
Tag the ticket type first: Billing, Technical, or Refund_Request. Correct classification earns +0.03; incorrect earns +0.01.
📚
search_kb
Query the knowledge base with keywords. Returns the exact policy the agent needs to follow. Relevant query earns +0.05; off-target earns +0.02.
💬
ask_clarifying_question
Talk to the customer. Warning: ask too many times and their patience runs out fast.
✅
resolve_ticket
Close the ticket with a final solution. Resolved tickets get the terminal success score.
🚨
escalate_to_human
Hand off to a real person when policy requires it. For some refund tasks, this is the successful terminal action.
⚡
Reward Signal
Graduated rewards: ask_clarifying_question +0.01 · classify_issue +0.01/+0.03 · search_kb +0.02/+0.05 · correct terminal action +0.90 before step 10.
Interactive Demo
PLAY THE
ENVIRONMENT.
Ticket Status
NO TICKET IDLE
Tier—
CategoryNot classified
Steps0 / 10
Terminal Score—
Use a terminal action before step 10 to reach 0.90.
Progress0%
Sentiment
😊 Happy
😐 Neutral
😤 Upset
😡 Angry
Knowledge Base
Use search_kb to pull policy
Submit Action
🎧
No Ticket Yet
Hit "New Episode" to load a customer scenario and start playing.
Reward Log +0.00
Get Running
SETUP IN
60 SECONDS.
Install & Run
# Install dependencies
pip install fastapi uvicorn openai pydantic

# Start the server
uvicorn api:app --reload
Run Inference
set HF_TOKEN=your_token
set API_BASE_URL=https://router.huggingface.co/v1
set MODEL_NAME=Qwen/Qwen2.5-72B-Instruct

python inference.py
"""