Customer Face Resolution Agent Airline Disruption — Design skill for Claude Code

AeroDesk — Airline Customer Resolution Agent

AIONOS Assignment 3 — Customer-Facing Resolution Agent (Streamlit prototype).

Objective

Build a customer-facing agent for airline disruptions that:

• Understands customer intent from free text
• Uses only the supplied customer, booking and policy data
• Applies policy-based decisions (never invents rules or compensation)
• Handles frustrated or confused customers
• Executes and records allowed actions
• Escalates requests outside the agent's authority

Architecture ()

The agent is built as two deliberately separate layers, so policy can never drift even if the language model changes what it says:

Customer message
      │
      ▼
┌─────────────────────┐     reads      ┌───────────────┐
│  Intent parser        │◄──────────────│   data.py       │
│  (policy_engine.py)    │               │  customers,     │
│  keyword/rule based    │               │  flights, fare  │
└─────────┬───────────┘               │  quotes, rules  │
          │ decision (facts, actions,  └───────────────┘
          │  escalations) — plain dict
          ▼
┌─────────────────────┐
│  Reply phrasing        │
│  • Template (default)  │
│  • Claude API (optional)│  ← phrasing ONLY, never decides policy
└─────────┬───────────┘
          │
          ▼
   Chat reply + ledger entry
   (action/escalation + the policy it relied on)

1. data.py — the supplied data pack (customers, bookings, fare-difference quotes, allowed/prohibited action lists), unmodified from the assignment.
2. policy_engine.py — the only place that decides anything. analyze() reads the customer's message, detects intent (cancellation, delay, refund, rebooking, hotel, fare difference, legal threat, sentiment), and returns a plain dict of facts / allowed actions / escalations, each one traceable to a specific rule in the data pack. This function never calls an LLM.
3. app.py — the Streamlit UI. It calls analyze() to get the decision,