AGENTS.md

Unit test conventions for this package. They apply to the Angular shell in src/, package-level infrastructure specs under test/, and the libraries under projects/ that have no file of their own: zeppelin-notebook-core and zeppelin-sdk, which are framework-neutral, and zeppelin-visualization, which is mostly so apart from one @Component base class.

Two subtrees override this file: e2e/AGENTS.md for the Playwright suite, and projects/zeppelin-react/AGENTS.md for the React remote, which has a different CI status and one exception of its own.

The repository root AGENTS.md asks every change to include unit tests. This file is how.

Layout

  • A product-code spec lives next to its source: foo.ts / foo.spec.ts. Specs for package-level test and reporting infrastructure live under test/.
  • The shell runner is Vitest on jsdom. There is no Karma. test/test-setup.ts loads zone.js and reflection metadata, initializes Angular TestBed, and resets the test environment after each spec.
  • npm run test:shell covers src/, package-level specs under test/, projects/zeppelin-sdk and projects/zeppelin-visualization. The two libraries have no runner of their own; they ride on the shell config because their code needs nothing extra. projects/zeppelin-react is separate. It has its own Vitest config and its own file here.
  • projects/zeppelin-notebook-core has a dedicated Vitest config (vitest.notebook-core.config.mts) using Node without Angular test setup. It covers the package's contract specs and the dependency-boundary checks under test/notebook-core. These specs are not part of the shell or React suites. The core compiler excludes DOM libraries and application path aliases.

Running

CommandPurpose
npm run test:shellRun the unit tests for src/, test/ and the two libraries
npm run test:shell -- --coverageSame, with a coverage report
npm run test:shell -- foo.spec.tsRun one file
npm run test:notebook-coreRun the dedicated notebook-core Node suite
npm run typecheck:notebook-coreCheck core source and specs, rebuild the package, and check the React type-only contract against built declarations and the same core source

test:shell, test:notebook-core, and typecheck:notebook-core are bound to the Maven test phase (pom.xml), so a spec added here starts running in CI the day it merges. It does not run where you would expect. frontend.yml builds this module with -DskipTests, which frontend-maven-plugin honours by skipping test-phase executions, so the run that counts is mvnw verify -Pweb-e2e inside the run-playwright-e2e-tests job. A failing spec surfaces there, under an e2e job name. Giving the unit tests a step of their own is ZEPPELIN-6566.

Where a test belongs

The frontend has two test layers, not three. There is no integration tier.

LayerRunnerAnswers
UnitVitest + jsdom (shell/React) or Node (notebook-core)Is the judgement we wrote correct?
E2EPlaywrightDoes the page actually work in a browser?

Prefer a unit test when the question can be answered without a browser. Reach for e2e when the answer depends on wiring: routing, mounting a federated remote, authentication, or anything a user would have to click.

The two layers do not substitute for each other, and neither replaces the cross-framework parity checks the React migration needs.

The notebook-core host/remote contract is currently a type-only scaffold; it does not implement a notebook runtime.

A third kind is planned but does not exist yet: contract specs that replay captured WebSocket traffic against the notebook runtime, arriving with ZEPPELIN-6627. Those will run on Vitest as well and live under test/contract/, with the captured traffic beside them. Conventions for them are added once they exist.

What to test

The judgement we wrote: branches, boundaries, error paths.

If a function has an if, it is worth a spec. If it only forwards to a framework API, it usually is not.

Boundaries are where the bugs are. HumanizeBytesPipe switches units at 1000 but divides by 1024, so transform(1000) renders 0.98 KB. A spec pins that down, a reviewer's intuition does not.

What not to test

  • Framework internals. Change detection, lifecycle ordering, dependency injection itself. Test the code we wrote, not Angular.
  • Template render snapshots. Markup changes when a surface is restyled or migrated; a snapshot only records that it changed.
  • Flows already covered by e2e. Wiring between components, navigation, and anything that needs a real browser belongs in e2e/.
  • Functions with no logic to check. element.ts feature-detects a DOM API and forwards to it. A spec would assert that a mock was called.

Specs that cannot fail

A spec with no assertion, or one whose assertion sits inside an if, passes by skipping the check it exists to make. These are caught by lint, not review:

RuleCatches
vitest/expect-expecta test with no assertion
vitest/no-conditional-expectan assertion only some runs reach
vitest/no-identical-titlea duplicate name silently shadowing another
vitest/valid-expectexpect(x) with no matcher
vitest/no-focused-testsit.only left behind, hiding the rest
vitest/no-disabled-testsit.skip left behind (warning)

The e2e suite gets the same protection from eslint-plugin-playwright.

monaco-editor and path aliases in specs

vitest.shell.config.mts mirrors the paths block in tsconfig.base.json, which Vite does not read; add an alias there when you add a path. Eleven files under src/ import monaco-editor, two behind the @zeppelin/services barrel, so a spec that reaches the editor or notebook area loads it: a few seconds on first import and a marked.umd.js.map sourcemap warning, both monaco's, not ours. Mock it with vi.mock('monaco-editor', ...) when the spec only needs to assert the editor was called. The runner resolves monaco to editor.api, which skips the language registrations editor.main performs, so do not assert that a built-in language is registered.

Determinism

No clock, no randomness, no network. A spec that reads Date.now() or fetches will eventually fail for reasons unrelated to the code under test.

Naming

The failure message must say what broke. should work does not.

it('renders a dash for null and undefined', ...)     // good
it('handles input', ...)                             // not a test name

Angular classes with and without TestBed

Construct directly testable directives, pipes and services as plain classes when Angular framework wiring is not the subject of the spec. Pass mocks to the constructor instead of starting TestBed unnecessarily.

const loader = { loadModule } as Pick<ReactRemoteLoaderService, 'loadModule'>;
const directive = new ReactMountDirective(host, ngZone, loader as ReactRemoteLoaderService);

See src/app/share/react-mount/react-mount.directive.spec.ts for a worked example, and src/app/share/pipes/humanize-bytes.pipe.spec.ts for a pipe.

Use TestBed when the behavior depends on template bindings, dependency injection, change detection, or Angular lifecycle wiring. src/app/share/react-mount/react-mount.directive.testbed.spec.ts shows that path with a decorated host component. Keep the direct-construction spec beside it for behavior that does not need Angular wiring.

Migration (Angular to React)

Write the spec while Angular is still the source of truth.

A spec written after a surface moves to React pins the new implementation's behaviour, not the behaviour we were trying to preserve. That turns a regression into the expected result. The migration (ZEPPELIN-6627) gates on “does this still behave the same?”, which cannot be judged when the previous behaviour was never written down.

src/app/pages/workspace/notebook/paragraph/paragraph-patch.spec.ts is the pattern: a regression was fixed and the behaviour was pinned in the same change.

Coverage

--coverage produces a v8 report under coverage/. It is measured, not gated. There are no thresholds.

Read the percentage carefully: it is not whole-tree coverage. The denominator is only the files the specs actually load, because include is left unset (see vitest.shell.config.mts). Coverage parsing is separate from the test transform, so the whole tree must be deliberately remeasured before that setting is widened. Most of the tree is absent from the report rather than counted as zero, so the figure reads far better than the real state, and it can fall as specs are added, since each new spec pulls more files into the denominator. Expect a large drop when include is eventually turned on.

That is deliberate. src/ currently holds a few specs against more than 200 source files, so any threshold set today is either meaningless or permanently red. The intended progression is: measure only, then ratchet so the number cannot fall, then require coverage on changed files. A whole-tree percentage is the wrong target during a migration, because much of the tree is going to be rewritten anyway.

This is a different measurement from e2e/reporter.coverage.ts, which counts annotated component pages rather than lines. The two numbers are not comparable and are not merged.

Adding a Test (Agents Start Here)

  1. Pick a target with logic to check: a branch, a boundary, an error path.
  2. Create foo.spec.ts next to foo.ts.
  3. Import from vitest (describe, expect, it), not from Jasmine or Jest. Check the target has callers before you invest in it. get-keyword-positions.spec.ts is a worked example of a function that turned out to have none.
  4. Construct the class directly unless the behavior depends on Angular wiring; use TestBed when it does.
  5. Run npm run test:shell for shell/SDK/visualization changes. For notebook-core changes, run npm run test:notebook-core and npm run typecheck:notebook-core. Confirm the relevant checks pass before opening a PR.