Idiomatic Ts Lit

1.2 as any Is a Fire Alarm

as any in production code means the type system has a gap. Four remaining instances in our src/ are all accounted for — if you need another, explain why in a comment. But be prepared for pushback from the user.

// ❌ Silences the compiler, hides the bug
const result = someValue as any;

// ✅ Narrow first, assert only when necessary
if (isValidResult(result)) {
  const typed: SpecificType = result;
}

Acceptable escapes:

• Test mocks — as unknown as AppServices with a partial mock is fine; that's the point of a mock.
• Library boundaries — When a third-party type is wrong, cast with a // TODO(types): comment linking the issue.
• Trust boundaries — Trusted types (CSP) require casts by design.

1.3 as unknown as T — The Double Cast

25 occurrences in our source. Some are legitimate (decorators, trusted types, signal-backed collection wrappers). Some are hiding a missing generic. Before writing one, ask:

Can I add a type parameter, use a type guard, or restructure the API to avoid this?

// ❌ Symptom of a missing generic
const transformer = raw as unknown as DataPartTransformer;

// ✅ Fix the source — add the generic to the factory
const transformer = createTransformer<DataPartTransformer>(raw);

1.4 Discriminated Unions Over Type Assertions

When you have a value that could be one of several shapes, use discriminated unions with exhaustive checks.

// ❌ Casting your way through variants
if ((result as ErrorResult).$error) { ... }

// ✅ Exhaustive discrimination
type Outcome<T> = { value: T } | { $error: string };

function handle(outcome: Outcome<Data>) {
  if ("$error" in outcome) {
    return err(outcome.$error);
  }
  return ok(outcome.value);    // TypeScript narrows automatically
}

2. Import Hygiene — Dependencies at a Glance

2.1 Static import type Over Inline import()

We codified this as "Standard 5.30" in our linting patterns. Inline imports clutter signatures and hide dependencies.

// ❌ Cluttered signature — reader can't scan the dependencies
export function mapState(
  state: import("@breadboard-ai/types").NodeLifecycleState
): import("@breadboard-ai/types").NodeRunStatus { ... }

// ✅ Types at the top — dependencies are visible at a glance
import type { NodeLifecycleState, NodeRunStatus } from "@breadboard-ai/types";

export function mapState(state: NodeLifecycleState): NodeRunStatus { ... }

2.2 Dynamic import() — Only for Lazy Loading

Dynamic imports are for code splitting and lazy loading. If the module is always used, it should be a static import.

// ❌ Dynamic import for something you always need
const { html } = await import("lit");

// ✅ Dynamic import for optional heavyweight deps
const pdfLib = await import("pdfjs-dist/build/pdf.mjs");

Current legitimate uses: bootstrap entry points, polyfills, PDF viewer, debug tools. If your dynamic import doesn't fall into "user may never trigger this code path," make it static.

2.3 Explicit Exports at the Top

Canonical pattern in this repo: imports first, then explicit exports, then implementations. This makes the public API scannable without scrolling.

import { foo } from "./foo.js";
export { bar, baz };

function bar() {
  /* ... */
}
function baz() {
  /* ... */
}

3. Lit Components — Thin Rendering Shells

3.1 Components Are Not Controllers

If your component class has a Map, a Set, or any business logic beyond DOM intrinsic state (scroll position, focus, animation frames), it's doing too much.

// ❌ Business logic trapped in the DOM
@customElement("my-list")
class MyList extends LitElement {
  private items = new Map<string, Item>();   // ← state belongs in a Controller
  private async loadItems() { ... }          // ← logic belongs in an Action
}

// ✅ Thin shell — state and logic live in SCA
@customElement("my-list")
class MyList extends SignalWatcher(LitElement) {
  @consume({ context: scaContext })
  accessor sca!: SCA;

  render() {
    const items = this.sca.controller.editor.items;  // signal read
    return html`${items.map(i => html`<my-item .data=${i}></my-item>`)}`;
  }
}

3.2 Definite Assignment for Context

Context is guaranteed by the application shell. Optional chaining on this.sca is noise — it hides real bootstrap errors behind silent undefined.

// ❌ "Safety" that silences real bugs
@consume({ context: scaContext })
protected accessor sca: SCA | undefined = undefined;
// ... this.sca?.controller?.editor?.graph  ← defensive noise

// ✅ Definite assignment — fail loudly if bootstrap is broken
@consume({ context: scaContext })
protected accessor sca!: SCA;
// ... this.sca.controller.editor.graph    ← clean and honest

3.3 SignalWatcher Is Non-Negotiable

If a component reads from SCA controllers, it must extend SignalWatcher(LitElement). Without it, signal reads don't trigger re-renders.

The one exception: pure "service locator" consumers that only call imperative methods (not read signals) in event handlers. Document this with an eslint disable comment.

3.4 No @consume Reads in Constructors or Initializers

Context values aren't injected until connectedCallback. Accessing them during class setup phase gives you undefined.

// ❌ Context not available yet
@customElement("my-thing")
class MyThing extends SignalWatcher(LitElement) {
  @consume({ context: scaContext })
  accessor sca!: SCA;

  private tools = this.sca.controller.tools; // 💥 undefined at init time!
}

// ✅ Defer access — lazy evaluation in arrow functions is safe
@customElement("my-thing")
class MyThing extends SignalWatcher(LitElement) {
  @consume({ context: scaContext })
  accessor sca!: SCA;

  // Access in render(), event handlers, or explicit lifecycle methods
  render() {
    const tools = this.sca.controller.tools; // ✅ connected by now
    return html`...`;
  }
}

4. Signal Patterns — Reactivity with Precision

4.1 @field for All Controller State

No SignalMap, SignalSet, SignalArray, or @signal from @lit-labs/signals in SCA controllers. They bypass hydration and persistence.

// ❌ Reactivity island — invisible to persistence
private _myMap = new SignalMap<string, string>();

// ✅ Unified lifecycle
@field({ deep: true })
private accessor _myMap: Map<string, string> = new Map();

4.2 void for Intentional Signal Reads

When you read a signal purely for subscription (side effect), the linter flags it as an unused expression. Use void to communicate intent.

// ❌ Linter: "Expected an assignment or function call"
controller.editor.graph.version;

// ✅ Explicit intent — "I'm subscribing, not using the value"
void controller.editor.graph.version;

4.3 Version + 1 for Triggers

A trigger condition that returns 0 is falsy — the trigger never fires. Always offset by 1.

// ❌ The Sticky Trigger Hazard — version 0 is falsy, trigger never fires
return controller.editor.graph.version;

// ✅ Always truthy, always changes
return controller.editor.graph.version + 1;

5. Error Handling — Outcomes, Not Exceptions

5.1 Return Outcomes, Don't Throw

Actions return Outcome<T> — either ok(value) or err("message"). This makes error paths explicit and testable.

// ❌ Untyped exception — caller has no idea what to catch
throw new Error("No editor available");

// ✅ Typed outcome — caller knows the shape
if (!editor) return err("No editor available");

5.2 Fail-Early Guards

The first lines of an Action should validate preconditions. Don't nest deep before discovering the state is invalid.

const { controller, services } = bind;
const editor = controller.editor.graph.editor;

if (!editor) return err("No editor available");
if (!metadata) return err("Missing metadata");
// ... proceed with confidence

6. Testing — No Coverage Theater

6.1 mock.method — Never Direct Overwrites

// ❌ Direct overwrite — no automatic restore, leaks between tests
globalThis.fetch = () => Promise.resolve(new Response("{}"));

// ✅ Restorable mock
mock.method(globalThis, "fetch", () => Promise.resolve(new Response("{}")));
// afterEach: mock.restoreAll()

6.2 as unknown as AppServices for Partial Mocks

This is the one place double-casting is encouraged. Partial mocks should only implement what the test exercises.

Asset.bind({
  controller,
  services: { googleDriveBoardServer: { dataPartTransformer: () => mock } }
    as unknown as AppServices,
});

6.3 Verify Assertions Actually Ran

A test that passes because a guard skipped the logic provides zero coverage. Add a console.log or explicit counter during development to confirm the assertion block executed.

7. Naming — Make the Code Read Like Prose

7.1 Name Your Bugs

Named bugs are memorable bugs. "The Unseen Cast" is instantly recognizable. "Fixed event handler typing" is forgettable.

7.2 Actions Are Verbs, Controllers Are Nouns

board.save()          ← Action: what the user does
editor.graph          ← Controller: what the system holds
graphStore.fetch()    ← Service: infrastructure verb

8. Architectural Discipline

8.1 No Cross-Action Imports

Actions talk to each other through Triggers and Pending Signals, never through direct imports. This prevents circular dependencies and enforces the "Action independence" invariant.

8.2 Destructure bind at the Top

// ❌ Chained access — hard to trace dependencies
bind.controller.editor.graph.setEditor(e);

// ✅ Destructure first — dependencies are visible
const { controller, services } = bind;
controller.editor.graph.setEditor(e);

8.3 Shared Utilities Before New Code

Before writing a helper, check packages/utils and src/utils/. The Déjà Code system will flag you if you reimplement something that already exists — but catching it before you write it is better.

9. Lit Patterns — The Framework-Specific Craft

The fundamental contract: State lives in SCA. Rendering lives in components. That's not a suggestion — it's the architectural boundary that makes everything else work.

A component reads from Controllers (via signals) and dispatches through Actions. It never owns business state, never fetches data, never decides what happens next. Those are Controller, Service, and Action concerns respectively. The component's only job is to turn signals into pixels and user gestures into action calls.

 ┌──────────────────────────────────┐
 │         UI Component             │
 │  reads signals → renders HTML    │
 │  handles events → calls Actions  │
 │  owns ONLY: scroll, focus, anim  │
 └──────────┬───────────────────────┘
            │
 ┌──────────▼───────────────────────┐
 │      SCA (Controllers/Actions)   │
 │  owns ALL business state         │
 │  orchestrates ALL workflows      │
 │  coordinates ALL services        │
 └──────────────────────────────────┘

If you find yourself adding a Map, a fetch(), or an if/else business rule inside a component — stop. That logic has a home, and it's not here.

9.1 Templates: nothing Over Empty Strings

Lit's nothing sentinel removes nodes from the DOM entirely. An empty string leaves an empty text node — a subtle difference that matters for CSS selectors, childNodes.length, and layout.

// ❌ Empty text node lingers in the DOM
render() {
  return html`${this.showIcon ? html`<span>icon</span>` : ""}`;
}

// ✅ Node cleanly removed
import { nothing } from "lit";

render() {
  return html`${this.showIcon ? html`<span>icon</span>` : nothing}`;
}

9.2 repeat() for Keyed Lists, .map() for Static

repeat() from lit/directives/repeat.js keeps DOM identity stable across re-renders when items shift position. Use it for lists the user can reorder, add to, or remove from. .map() is fine for static read-only lists.

// ✅ Keyed — DOM nodes move with their data
import { repeat } from "lit/directives/repeat.js";

${repeat(
  messages,
  (msg) => msg.id,       // stable key
  (msg) => html`<div>${msg.text}</div>`
)}

// ✅ Static — no reordering, .map() is simpler
${items.map(item => html`<span>${item.label}</span>`)}

9.3 classMap() and styleMap() Over String Interpolation

String interpolation for classes and styles is fragile and hard to read. Lit provides directives that handle the bookkeeping.

// ❌ Manual string building
class=${`g-icon ${rotate ? "rotate" : ""} ${filled ? "filled" : ""}`}

// ✅ Declarative and clean
import { classMap } from "lit/directives/class-map.js";

class=${classMap({
  "g-icon": true,
  rotate: this.isRotating,
  filled: this.isFilled,
})}

9.4 Lifecycle Hooks — Know Your Options

Rules of thumb:

• willUpdate > updated for derived state — it runs before paint.
• Always pair connectedCallback setup with disconnectedCallback teardown.
• Never read layout in willUpdate (DOM isn't updated yet).

connectedCallback() {
  super.connectedCallback();
  this.#resizeObserver = new ResizeObserver(this.#onResize);
  this.#resizeObserver.observe(this);
}

disconnectedCallback() {
  super.disconnectedCallback();
  this.#resizeObserver?.disconnect();
}

9.5 Custom Events — The eventInit + static eventName Pattern

Every event class uses a shared eventInit object ({ bubbles: true, cancelable: true, composed: true }) and a static eventName for listener registration. This is non-negotiable for Shadow DOM traversal.

const eventInit = { bubbles: true, cancelable: true, composed: true };

export class BoardSaveEvent extends Event {
  static eventName = "bbboardsave";

  constructor(public readonly url: string) {
    super(BoardSaveEvent.eventName, { ...eventInit });
  }
}

Why composed: true? Without it, events dispatched inside Shadow DOM stop at the shadow boundary. Every component in this codebase uses Shadow DOM, so every event that needs to reach a parent must be composed.

Why static eventName? It provides a single source of truth for the string. Listeners use BoardSaveEvent.eventName instead of a magic string.

9.6 The StateEvent Type-Safe Event Bus

For cross-cutting events that flow through the SCA system, use the typed StateEvent<T> class. The discriminated union StateEventDetailMap ensures that every event type has a known payload shape.

// Dispatching — the detail type is enforced by the generic
this.dispatchEvent(new StateEvent(new Board.Create({ title: "My Board" })));

// Receiving — evt.detail is fully typed
async (evt?: StateEvent<"board.create">): Promise<void> => {
  const { title } = evt!.detail; // type-safe access
};

9.7 @property() vs @state() — Public API vs Internal

@property() is the component's public API — attributes that consumers set. @state() (or private fields) is internal rendering state.

@customElement("my-widget")
class MyWidget extends LitElement {
  // Public — consumers set this via HTML or JS
  @property({ type: Boolean, reflect: true })
  accessor active = false;

  // Internal — only this component manages it
  @state()
  accessor #menuOpen = false;
}

Use reflect: true sparingly — only when the attribute needs to be visible in the DOM for CSS selectors (:host([active])) or accessibility.

9.8 CSS: static styles Is the Only Way

Lit's static styles property enables style sharing and adoptedStyleSheets for performance. Never use inline <style> tags in templates.

// ❌ Inline styles — re-parsed every render
render() {
  return html`<style>.foo { color: red }</style><div class="foo">...</div>`;
}

// ✅ Static — parsed once, shared across instances
static styles = css`
  :host {
    display: block;
  }

  .foo {
    color: var(--my-color, red);
  }
`;

CSS custom properties are the theming API. Expose --bb-* variables from :host and consume them internally. This is how child components inherit design tokens without prop-drilling styles.

The Short Version

Say what you mean. Type parameters over casts. Static imports over dynamic. Definite assignment over optional chaining. Outcomes over exceptions. Thin components over fat ones. Named patterns over anonymous fixes.

If the compiler can check it, let it. If a human can name it, name it. If a tool can enforce it, write the rule.