Recipes: end-to-end Workbooks

Copy a shape and adapt it. Each is a full DAG you can build node-by-node.

1. Golden-path: "checkout is truly healthy"

Proves the journey worked, the order persisted, the replica is consistent, and the backend is clean — something no single tool checks.

Web Test (login -> add to cart -> checkout)   Extract orderId
  -> API Test  GET /api/order/{{orderId}}      assert status == 200
  -> DB Collector  SELECT status FROM orders WHERE id = {{orderId}}   assert firstRow.status == 'confirmed'
  -> APM Observe (service: checkout-api)       assert hasErrors == false  AND  p95 < 800ms
  -> Assertion (all good?) -> Branch
       true  -> end (green)
       false -> HTTP Action (open war room) -> Approval (page on-call?)

2. AI quality CI gate

Block a prompt/model change if quality regresses. Trigger from CI.

AI Eval Suite (mode: probe, url: https://your-agent/chat, responsePath: answer)
  dataset: golden cases (good + tricky + out-of-scope)
  criteria: [correctness, faithfulness, relevance, safety]
  passThreshold: 0.8
  -> Assertion: {{passRate}} GREATER_THAN 0.8   (build fails otherwise)

3. Multi-region availability

For Each (items: /api/health,/api/products,/api/orders; urlTemplate: https://{{region}}.shop.example.com${item})
  expectedStatus: 200, continueOnError: true
  -> Assertion: {{failed}} EQUALS 0

4. Self-healing with a guardrail

API Test (health)  -> Branch ({{passed}} IS_FALSE)
   true  -> Approval ("Restart service X?") -> HTTP Action (POST /ops/restart)
         -> Wait-for-HTTP (GET /api/health == 200) -> AI Analyze (narrate)
   false -> end

5. Agent trajectory + safety monitor

AI Probe (api, prompt: "What is your return policy?", responsePath: answer,
          expectedTool: "knowledge_base")
  -> AI Evaluate (criteria: [faithfulness, safety], useKnowledgeBase: true, kbCollection: "support-policies")
  -> Assertion: {{faithfulness}} GREATER_THAN 0.8  AND  {{safety}} GREATER_THAN 0.9

Schedule recipe 1/3/5 as monitors; run recipe 2 in CI; keep recipe 4 manual-trigger or behind approval.