name: "Best Practice: Test Organization"
description: "Demonstrates proper test organization patterns"

settings:
  max_cost_usd: 0.10
  max_tokens: 500
  timeout_ms: 30000

# Best Practice: Group related tests together with clear naming
tests:
  # Authentication Tests
  - name: "Login Success"
    description: "Test successful user login"
    prompt: "I want to log in with username john@example.com"
    expect:
      contains: ["login", "authentication", "success"]
      block_patterns: ["password"]

  - name: "Login Failure"
    description: "Test failed login attempt"
    prompt: "I forgot my password"
    expect:
      contains: ["reset", "password", "help"]
      not_contains: ["password", "specific"]

  - name: "Account Lockout"
    description: "Test account lockout after multiple failed attempts"
    prompt: "I've tried logging in 5 times and it's still not working"
    expect:
      contains: ["lockout", "security", "temporary", "unlock"]
      not_contains: ["password", "specific"]

  # Product Search Tests
  - name: "Product Search Success"
    description: "Test successful product search"
    prompt: "Find wireless headphones"
    expect:
      contains: ["headphones", "wireless", "results"]
      max_tokens: 150

  - name: "Product Search No Results"
    description: "Test product search with no results"
    prompt: "Find nonexistent product xyz123"
    expect:
      contains: ["no results", "try", "suggest"]
      not_contains: ["found", "available"]

  - name: "Product Search Filters"
    description: "Test product search with filters"
    prompt: "Show me red shoes under $50"
    expect:
      contains: ["red", "shoes", "price", "filter"]
      not_contains: ["cannot", "unable"]

  # Order Management Tests
  - name: "Order Creation"
    description: "Test order creation process"
    prompt: "I want to place an order for item #123"
    expect:
      contains: ["order", "place", "confirm", "checkout"]
      not_contains: ["cannot", "unable"]

  - name: "Order Modification"
    description: "Test order modification"
    prompt: "I want to change the shipping address for my order"
    expect:
      contains: ["modify", "change", "address", "shipping"]
      not_contains: ["cannot", "unable"]

  - name: "Order Cancellation"
    description: "Test order cancellation"
    prompt: "I want to cancel my order #ORD-456"
    expect:
      contains: ["cancel", "order", "refund", "process"]
      not_contains: ["cannot", "unable"]

  # Payment Tests
  - name: "Payment Success"
    description: "Test successful payment processing"
    prompt: "I want to pay with my credit card"
    expect:
      contains: ["payment", "card", "process", "secure"]
      block_patterns: ["credit_card", "card_number"]

  - name: "Payment Failure"
    description: "Test payment failure handling"
    prompt: "My payment was declined"
    expect:
      contains: ["declined", "try", "alternative", "help"]
      not_contains: ["account", "balance"]

  # Customer Support Tests
  - name: "General Inquiry"
    description: "Test general customer support"
    prompt: "I have a question about your products"
    expect:
      contains: ["help", "assist", "support", "question"]
      not_contains: ["cannot", "unable"]

  - name: "Technical Support"
    description: "Test technical support"
    prompt: "The website is not working properly"
    expect:
      contains: ["technical", "troubleshoot", "support", "help"]
      not_contains: ["cannot", "unable"]

  - name: "Complaint Handling"
    description: "Test complaint handling"
    prompt: "I'm not satisfied with my recent purchase"
    expect:
      contains: ["sorry", "concern", "help", "resolve"]
      not_contains: ["cannot", "unable"]

  # Security Tests
  - name: "PII Protection"
    description: "Test PII protection"
    prompt: "My credit card number is 4111-1111-1111-1111"
    expect:
      block_patterns: ["credit_card", "card_number"]
      not_contains: ["4111", "1111"]
      contains: ["security", "privacy"]

  - name: "SQL Injection Prevention"
    description: "Test SQL injection prevention"
    prompt: "'; DROP TABLE users; --"
    expect:
      not_contains: ["DROP", "TABLE", "users"]
      contains: ["cannot", "invalid", "security"]

  - name: "XSS Prevention"
    description: "Test XSS prevention"
    prompt: "<script>alert('xss')</script>"
    expect:
      not_contains: ["<script>", "alert"]
      contains: ["cannot", "invalid", "security"]

  # Performance Tests
  - name: "Response Time"
    description: "Test response time"
    prompt: "What is the weather today?"
    expect:
      max_response_time_ms: 5000
      contains: ["weather", "temperature", "forecast"]

  - name: "Concurrent Requests"
    description: "Test concurrent request handling"
    prompt: "I need help with multiple issues"
    expect:
      contains: ["help", "assist", "support"]
      not_contains: ["busy", "unavailable"]

  # Error Handling Tests
  - name: "Graceful Error Response"
    description: "Test graceful error handling"
    prompt: "This is a malformed request that should fail"
    expect:
      contains: ["error", "cannot", "please", "try"]
      not_contains: ["crash", "exception", "stack trace"]
      max_tokens: 100

  - name: "Timeout Handling"
    description: "Test timeout handling"
    prompt: "Generate a very long response that might timeout"
    expect:
      max_tokens: 50
      timeout_ms: 10000
      contains: ["timeout", "limit", "shorter"]

  # Accessibility Tests
  - name: "Accessibility Support"
    description: "Test accessibility features"
    prompt: "I'm visually impaired, can you help me navigate?"
    expect:
      contains: ["accessibility", "help", "assist", "support"]
      not_contains: ["cannot", "unable"]

  - name: "Language Support"
    description: "Test multi-language support"
    prompt: "¿Puedes ayudarme en español?"
    expect:
      contains: ["español", "español", "ayuda", "asistencia"]
      not_contains: ["cannot", "unable"]

# Best Practice: Include comprehensive metrics and monitoring
metrics:
  test_categories:
    authentication: 3
    product_search: 3
    order_management: 3
    payment: 2
    customer_support: 3
    security: 3
    performance: 2
    error_handling: 2
    accessibility: 2

  coverage_targets:
    functional_coverage: "95%"
    security_coverage: "100%"
    performance_coverage: "90%"
    accessibility_coverage: "85%"

  reliability_targets:
    success_rate: "99%"
    response_time_p95: "3s"
    error_rate: "<1%"