Component-Driven Development assumed a render model that signal-based Angular has quietly left behind. The tooling has not caught up, and I am not sure simply patching it will be enough.

The moment that made me write this

I was building ng-prism, an Angular-native component showcase tool I maintain, and I needed to update a component's inputs from a controls panel. The naive version looked roughly like this:

const ref = vcr.createComponent(ButtonComponent);
const instance = ref.instance;

instance.variant = 'primary';
instance.label  = 'Save';
ref.changeDetectorRef.detectChanges();

It works on a pre-signal component. On a signal-based component it quietly breaks the component. variant and label are no longer plain properties. They are InputSignal objects, callable as variant(). Angular stores the InputSignal as a plain class field; there is no defineProperty setter or proxy to intercept the write. So the assignment just overwrites the field reference with a string, and Angular never finds out. The next render then blows up the first time the template evaluates variant(), because variant is no longer a function. I verified this against @angular/core 21.2.0: createInputSignal() returns a plain function with a [SIGNAL] symbol attached, no setter trap in sight.

The fix is one line:

ref.setInput('variant', 'primary');
ref.setInput('label',   'Save');

That is the version that lives in ng-prism today, in the renderer effect that drives every showcase:

// packages/ng-prism/src/app/renderer/prism-renderer.component.ts
// Simplified; real version also handles content projection and unknown-input warnings.
effect(() => {
  const inputs = this.rendererService.inputValues();
  const ref    = this.componentRef;
  if (!ref) return;

  performance.mark('prism:rerender:start');
  for (const [key, value] of Object.entries(inputs)) {
    ref.setInput(key, value);
  }
  // setInput() already marks dirty and schedules CD. The explicit
  // detectChanges() here forces it to run synchronously so the
  // performance.mark below wraps the actual render, not just the
  // dirty-marking. Also keeps timings predictable under zoneless.
  ref.changeDetectorRef.detectChanges();
  performance.mark('prism:rerender:end');
});

This is a tiny detail, but it is the symptom of a much bigger thing. The way Component-Driven Development (CDD) thinks about a component (props in, render out, args table on the side) was modelled on a world where component inputs were plain properties. That world is gone in Angular. And once you stop assuming it, a lot of the tooling we have built over the last decade starts to look like it is solving the wrong problem.

The implicit model behind classic CDD

Look at what almost every CDD tool, Storybook included, has agreed on for years:

1. A component is a pure-ish function of its inputs.
2. Inputs are configured as a flat dictionary, the “args”.
3. A “story” is a specific value of that dictionary.
4. Changing args triggers a discrete re-render cycle.
5. Addons (a11y, viewport, knobs, controls) hook into that cycle.

In the pre-Hooks era this was the right abstraction. React class components had setState. Angular had @Input() decorators and ngOnChanges. Vue had options API. All of them had a clear, discrete moment where a property got assigned, the framework noticed, and the component re-rendered top-down. Args tables map onto that perfectly. Whatever knob you turn becomes a property assignment, which becomes an ngOnChanges call, which becomes a render.

This is also why Storybook’s args model felt so natural for so long. Args are properties. Properties are state. State drives render. Story = scenario = property snapshot.

It was a beautiful, simple model. It is also, for Angular today, the wrong abstraction.

What Signals actually changed

I do not want to retread the “signals are great” territory. What matters for CDD is more specific.

Inputs are no longer properties. An InputSignal is a callable getter. From outside the component, the only correct way to push a value into it is ComponentRef.setInput(name, value). There is no property to assign anymore. Tools that still build their abstractions on “assign this prop” are not even wrong yet, they just produce broken components.

Components are nodes in a reactive graph, not pure functions of inputs. Half of a real-world signal-based component is computed() derived state. Those derived signals are part of the component’s behaviour. They are not inputs, but they are also not opaque internal state. An args table cannot represent them. A story shapes inputs, not derived state.

Lifecycle work is increasingly split between explicit hooks and reactive subscriptions. ngOnInit and ngOnDestroy still exist and are still idiomatic for a lot of cases (subscriptions, teardown, one-shot setup). What has changed is that ngOnChanges is largely irrelevant for signal inputs, and a growing share of the work that used to live in lifecycle hooks now lives inside effect() or computed() that is read by the template. The places where work happens have multiplied, and a chunk of it fires on a much finer-grained schedule than mount → change → destroy. CDD tools that visualise lifecycle as those three states are visualising a real but shrinking slice of what the component actually does.

Re-renders are continuous, not cycle-based. Zoneless Angular still has ApplicationRef.tick() and a ChangeDetectionScheduler. What is gone is the implicit tick driven by zone.js intercepting every async operation. In a zoneless app, change detection runs because the scheduler decided to run it, which in practice is because a signal flagged a view dirty. From the outside that looks less like a single render cycle and more like a graph of signals notifying their dependents at fine granularity. An addon that says “re-run my check on each render cycle” has no clean event to listen to, because there is no single render cycle visible at the public surface of the component.

The pre-signal mental model is not catastrophically wrong. It is just lossy. And every layer of CDD tooling pays the price of that loss somewhere.

Where the tooling lags, concretely

Three examples I have hit while building ng-prism.

1. Controls and args still treat inputs as a flat property bag.

The Storybook args object and ng-prism’s own variant config both look like this:

{ variant: 'primary', label: 'Save', disabled: false }

That is fine as a starting state. It is wrong as a model of how the component actually consumes its inputs. A signal-based input has a default expression, can be required, can be tied through computed() into other state, and can be read multiple times per render. None of that is in the args object. The args object is a snapshot of a tree it cannot see.

2. A11y, perf, and visual-diff addons assume “render happened, now check”.

axe-core, layout measurements, screenshot capture: they all rely on a moment where the DOM has settled and you can inspect it. Angular does give you some primitives here. afterNextRender() fires once after the next render. afterRender() fires after every render. ApplicationRef.isStable exposes a stable-state observable. None of those quite match what an a11y or visual-diff addon actually wants, which is closer to “the reactive graph has been quiet for N milliseconds across multiple microtasks, and I am now safe to walk the DOM”. That concept does not exist as a framework primitive. So you debounce, you wait for animation frames, you hope. In ng-prism the built-in a11y audit runs after a 500ms debounce on signal changes (A11yAuditService.scheduleAudit, default debounceMs = 500), which works in practice but feels like a workaround. I am not sure a true “graph is quiescent” signal even makes sense in a system designed to update continuously.

3. The “scenario” unit is too coarse.

A Storybook story is one set of args. A ng-prism variant is one set of input values. Neither can express “this component embedded in a parent that emits a signal stream over time”. The interesting bug in a signal-based component is rarely the static case. It is the transition. It is the moment when an upstream computed() updates twice in the same microtask and your effect runs once instead of twice. You cannot represent that with { variant: 'primary' }.

What I tried in ng-prism, honestly

ng-prism started life with the same implicit model as Storybook. Components have inputs. Inputs have values. Variants are named tuples of input values. The decorator looks similar:

@Showcase({
  title: 'Button',
  variants: [
    { name: 'Primary', inputs: { variant: 'primary', label: 'Save' } },
    { name: 'Danger',  inputs: { variant: 'danger', disabled: true } },
  ],
})

What ng-prism does that I think is on the right track:

• Uses setInput() for every input push, never property assignment. So at least the signal contract is respected. This is the absolute floor and a surprising number of Angular-adjacent tools still get it wrong.
• Drives the rendering loop with effect() on a signal of input values, not with a render-cycle event. The renderer reacts to whatever upstream signal happens to change.
• Treats the scanner as a build-time concern. Signal inputs are recognised via the TypeScript Compiler API at build time, so the runtime never has to read decorator metadata. The decorator itself is literally a no-op, just a marker.
• Supports zoneless via an opt-in flag in the ng add schematic (--zoneless), which wires up provideZonelessChangeDetection() and drops zone.js from polyfills. The default still ships with zone.js, but nothing in the renderer depends on an implicit zone tick; change detection runs because something signal-shaped told it to.

What ng-prism does not solve yet, and where I think the whole category is still stuck:

• The variant model is still a flat input snapshot. I cannot describe a component as “embedded in a parent that pushes this signal stream over 2 seconds”.
• computed() derived state is invisible in the UI. You see the inputs you set. You do not see the graph that hangs off them. For some components, that graph is the interesting part.
• The a11y, perf, and box-model panels all hook into renderer output the way an old addon would. They debounce. They do not subscribe to the actual reactive graph the component is part of.
• The “code snippet” feature generates a template string from input values. It cannot show how the component would behave in a parent where one of those inputs is itself a signal.

I am being deliberate about this in the docs. I do not want ng-prism to claim a level of insight into signal-based components that it does not yet deliver.

What I think the next generation should look like

This is the speculative part, so take it as opinion.

I think the unit of CDD will stop being “the component with these props” and start being “the component embedded in a reactive context”. A scenario will look more like a tiny fixture: this component, mounted under this parent, with these signal sources feeding it, over time. Closer to a Playwright scenario than a Storybook story.

I think tooling will need to render the reactive graph, not just the visual output. For a serious component, the dependency graph between inputs, computed(), and effect() is the documentation. We render the box and the controls. We should also be able to render the graph.

I think the addon model needs to flip. Instead of hooking into a render lifecycle that no longer exists, addons should subscribe to specific signals. Visual diff subscribes to the renderedElement signal. A11y subscribes to the same. Perf subscribes to a “graph quiescent” signal that the framework would have to expose. None of this lives in Storybook’s current architecture, and bolting it on is not obviously the right answer.

And I think “args” as a UI metaphor has run its course. A signal-based component is not configured by setting properties. It is wired up. The control surface should reflect that.

Closing

I am not arguing that Storybook is bad, or that the people maintaining it have missed something obvious. They built the right tool for the framework world of 2018, and that tool became a standard. Standards lag by definition. That is fine.

What I am arguing is that the abstraction is starting to leak in ways that matter. Signal-based Angular is a different kind of component runtime, not a faster version of the old one, and the surface area that CDD tooling was designed against has changed underneath it.

I am building ng-prism partly to test that hypothesis in code, not just in prose. I am also fully prepared to find out that I am wrong, that an args table plus setInput() plus a debounce is good enough for 95% of cases, and that the rest is academic. Possible. But I do not think it is, and I would rather find out by building than by predicting.

If you have hit the same kind of friction, or you have a counter-example where the args model still maps cleanly onto a signal-heavy component, I would genuinely like to hear it. That is more useful than another “signals are great” thread.

ng-prism is open source on GitHub. Feedback and counter-examples especially welcome.