Files
atomic-design-poc/src/docs/machines.mdx
T
ehoandClaude Sonnet 5 44eb2d2186
CI / frontend (push) Successful in 1m46s
CI / storybook-a11y (push) Successful in 4m23s
CI / backend (push) Successful in 1m14s
CI / codeql (csharp) (push) Has been cancelled
CI / codeql (javascript-typescript) (push) Has been cancelled
CI / api-client-drift (push) Has been cancelled
CI / e2e (push) Has been cancelled
chore(deps): update npm packages within declared ranges; reformat for prettier 3.9.4
npm update brought every package to the latest version its existing package.json
range allows (Angular tooling 22.0.2/22.0.4 -> 22.0.5, prettier 3.8.4 -> 3.9.4,
typescript-eslint 8.62.0 -> 8.62.1); package.json itself needed no range changes.

Auditing actual deprecation warnings (not just outdated versions) found nothing
further to fix: @angular/platform-browser-dynamic and @angular-devkit/build-angular
are deprecated by Angular but still required peer dependencies of the latest
published @storybook/angular (10.4.6 — peer range still `>=18.0.0 < 22.0.0`,
already why .npmrc sets legacy-peer-deps); jest-process-manager/expect-playwright
are transitive-only through @storybook/test-runner's latest stable (0.24.4). No
newer version of either Storybook package exists yet that drops them. The
remaining npm audit advisory (@babel/core, low severity) is the same
already-documented, deliberately-left issue in README.md (fixing it downgrades
Angular). Left package.json's overrides untouched.

The prettier bump alone changed formatting opinions on files this session didn't
otherwise touch (a stale markdown italics marker, a few object-literal wrap
points) — reformatted everything so `format:check` (part of CI) doesn't regress.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-07-05 10:29:36 +02:00

105 lines
5.1 KiB
Plaintext

import { Meta } from '@storybook/addon-docs/blocks';
<Meta title="Foundations/State Machines (TEA)" />
# State machines (The Elm Architecture, in Angular)
Every form or wizard with validation or submission in this app is wired the **same
way**: one Model, one Msg union, one pure `reduce`, one command per side effect. Pick any
one — `herregistratie.machine.ts` is the fullest worked example — and the shape
transfers everywhere else.
## Model / Msg / reduce
```ts
// Model — everything the UI needs to render, as ONE tagged union
export type WizardState = { tag: 'step1'; draft: Draft } | { tag: 'step2'; valid: Valid } | …;
// Msg — every way the Model is allowed to change
export type WizardMsg = { tag: 'FieldChanged'; field: string; value: string } | { tag: 'NextStep' } | …;
// reduce — PURE: (current, message) -> next. No I/O, no Date.now(), no randomness.
export function reduce(s: WizardState, m: WizardMsg): WizardState { … }
```
Because the whole state is one value, a bug reproduces from a message log; because
`reduce` is pure, every transition is a one-line assertion in a spec — no `TestBed`, no
mocked HTTP, just `expect(reduce(state, msg)).toEqual(next)`.
## Commands: side effects stay OUT of the reducer
`reduce` only ever answers "what is the new state" — it never calls `fetch`. A
**command** (an `application/submit-*.ts` file, or a store method) does the I/O, then
dispatches a message describing the outcome:
```ts
// command = "go do it, then say what happened" — reduce never sees the HTTP call itself
async function submit(store: Store<WizardState, WizardMsg>) {
const r = await adapter.submit(toDto(store.model()));
store.dispatch(
r.ok
? { tag: 'SubmitConfirmed', referentie: r.value }
: { tag: 'SubmitFailed', error: r.error },
);
}
```
This is also how a machine receives **server-owned config** without becoming aware of
HTTP: `intake.machine.ts`'s scholing threshold has an offline fallback
(`SCHOLING_THRESHOLD_DEFAULT`) baked into the model, and a plain `SetPolicy` message
that overwrites it once the real value arrives — the machine doesn't know or care that
the value came from a `resource()` fetch.
## `createStore`: the one wiring idiom
```ts
private store = createStore<WizardState, WizardMsg>(initial, reduce);
readonly model = this.store.model; // Signal<WizardState> — template reads this
dispatch = this.store.dispatch; // template calls this, on click/input/etc — never mutates
```
A page or component **never** hand-rolls `signal(initialModel)` plus its own local
`dispatch` function that calls `reduce` inline — that's the same idea reinvented with a
worse name, and it's the thing a newcomer copies if two idioms are visible side by side.
Wire every machine through `createStore`, full stop.
`dispatch` uses `model.update(…)`, not `model.set(reduce(model(), msg))` — the latter
reads `model()` _inside_ the call, which means an `effect()` that both reads `model` and
calls `dispatch` would subscribe to its own write and livelock. `.update()`'s callback
receives the current value directly, untracked.
## Naming
- A top-level machine's types are **context-prefixed**: `ChangeRequestState`,
`ChangeRequestMsg`, `WizardState`, `WizardMsg` — never bare `State`/`Msg`. A bare name
reads fine in the one file that defines it and then collides (or forces an import
alias) the moment two machines are open side by side.
- A top-level machine exports `initial` (the starting Model) and `reduce` — unprefixed,
since the file/module already disambiguates them at the import site
(`import { initial, reduce } from './herregistratie.machine'`).
- A **composable sub-machine** — one embedded _inside_ a parent Model, like
`upload.machine.ts`'s upload-widget state living inside the registratie wizard's own
Model — keeps **prefixed value exports** instead: `initialUpload`, `reduceUpload`.
The parent machine already imports several machines' `initial`/`reduce`; prefixing the
sub-machine's exports avoids a wall of `as` import aliases at the composition site.
## Derive, don't store
If a value can be computed from the Model, it is **not** a field on the Model. The
wizard's visible steps are `visibleSteps(answers)`, a pure function of the current
answers — not a `visibleSteps: Step[]` field someone has to remember to keep in sync
every time an answer changes. The reflex: before adding a field, ask "could this just be
a function of what I already have?"
## Where RemoteData fits in
A machine owns the **domain** lifecycle of what it holds once it exists (draft →
submitted → approved, in the brief's case). It should generally _not_ also own the
**fetch** lifecycle (loading/failed) for the initial GET that produces it — that's a
generic concern `RemoteData` already models once, consistently, across the app (see
[Foundations/RemoteData & Async](?path=/docs/foundations-remotedata-async--docs)). Where
a machine's own state happens to have `loading`/`failed` tags that purely mirror that
fetch, project them onto a `RemoteData` at the store layer for `<app-async>` to render
(`BriefStore.remoteData` is the worked example) rather than teaching every consumer to
hand-roll a `@switch` over the machine's own tags.