The test split — scripts, then evaluation, then smoke

Three layers, each checking what the one before cannot (#262 AC15). No test calls a real model: real models answer only in evaluation and in the smoke tests, never in npm test.

Layer	What it checks	How	Models
Unit and application tests (*.test.ts, npm test, CI)	your code: the paths a run takes, what a step receives, its errors, limits and their .onError paths, a race, a component on its own	testWith(App) with scripts (mockLlm(X).thenReturn(…), when(…)), mockOf, TestConcurrency, the unit runners (testTool, testRouter, …) — Testing	scripted: free, offline, the same every run
Evaluation and golden sets (agboss eval score, agboss replay, agboss golden, agboss compare)	the models' behaviour: whether an agent obeys its prompt, picks the right oneOf variant, stops at the right depth; whether a prompt or a profile is better	stored Terns scored by a judge, a golden set replayed under two profiles — Terns and evaluation, Suites and Comparisons	real, from the workflow's eval account
Smoke tests (make smoke, never in CI)	what only a provider tells: real prices, rate limits, timeouts, a model's actual capabilities	a handful of real calls in tests/smoke/, with a key in .env	real, budget-capped

• A script is not a model: it proves that your code does the right thing with an answer, not that a model gives it. Whether it gives it is the evaluation's question.
• What a model would do with a prompt is never asserted in a unit test; what a step sent is (mockLlm(X).requests, toHaveBeenAskedWith).
• Observer events are tested with a test observer (@Application({ observers }) in a test application); the Langfuse UI is not tested.
• Recorded runs (VCR) sit between the first two layers: answers recorded once from real models, replayed offline — an integration test, not an evaluation.

Standalone examples with symmetric Spring-style tests/ mirroring src/: 01-hello-agboss, 03-support-refund-desk, and all 7 projects in agbossjs/agboss-examples — see Example.