Parse errors allow a custom control to communicate that it is currently
unable to produce a valid value.
Parse errors are reported by implementing the optional `parseErrors`
property on the `FormUiControl`. The property should be a signal of the
current parse errors.
Also renames several `*Field` types to `*FieldTree`. This aligns with the new naming of the concept after `Field` was renamed
to `FieldTree`.
Add the host/container comment node to the `DeferBlockData`. This node can be used as a `@defer` block locator in the DOM tree in the absence of root nodes.
PR Close#66546
Extends the `focus` method of form fields and custom controls to accept and propagate `FocusOptions`.
This enables developers to control focus behavior more precisely, for example, preventing scrolling when focusing an element.
Currently migrating from `[ngClass]` to `[class]` isn't entirely supported, because `[ngClass]` supports `Set` values while `[class]` ignores them.
These changes add a bit of logic to bring them closer together and make the migration easier.
This PR adds the ability to manually register a binding with the
`FormField` directive. This is useful for a lower-level implementation
that takes the field tree as an `input()` rather than relying on the
automatic binding from `FormUiControl`.
This completes the rename started in #66136. `[field]` is too generic of
a selector for the forms system to own, and likely to cause naming
collisions with existing components. Therefore it is being renamed to
`[formField]`
Reworks how we store arrow functions in the following ways:
1. Rather than the `storeCallback` and `getCallback` instructions, we generate a single `arrowFunction` instruction.
2. The `arrowFunction` instruction uses a factory to create a new instance of the function when a function is read for the first time.
3. We now keep arrow functions in listeners in line so that they have access to `$event`.
Adds the `ɵɵstoreCallback` instruction that allows for a callback to be stored for later usage, as well as `ɵɵgetCallback` which can be used to retrieve it.
The `thisArg` in pure functions isn't used so we can drop it. We still need to keep it on the underlying implementation, because pipe instructions rely on it.
The animations code currently tracks which views have running leave animations by adding them to a `Set`. This can leak memory if we don't clean something up on time.
These changes switch to tracking the views by their ID which doesn't risk retaining the view.
Fixes#66255.
This will replace the `[field]` directive, since `[field]` is a very
generic name for signal forms to commandeer
refactor(forms): hook up `formField` directive in compiler
Hooks up the `formField` direcive to get the same treatment as the
`field` directive in the compiler.
apply updated formatting
The framework will now bind all field state properties to their
corresponding native properties (if any) on interop form controls (those
using `ControlValueAccessor`), excluding those handled explicitly by
`ControlValueAccessor` such as `disabled`.
Since the `Field` directive manages binding `FieldState` properties to
the underlying form control automatically, the type checker prohibits
explicit bindings to the same properties to avoid conflicts. This proved
problematic in cases where developers wanted to bind these properties to
the inputs of other directives on form controls. Now the framework will
bind the field state properties to all matching directive inputs on
native controls.
Fix#65617
* Recognize directives with non signal-based models as valid custom controls
* Relax type checker to allow non signal-based models
The `FormValueControl` and `FormCheckboxControl` interfaces still
require a `model()`-input, however, a custom control need not implement
either interface to be bound by the `Field` directive.
All of the following examples can be used to define a custom control:
```ts
// Preferred: model()
class MyFormControl implements FormValueControl<string> {
readonly value: model.required<string>();
}
// Supported: input() + output()
class MyFormControl {
readonly value: input.required<string>();
readonly valueChange: output<string>();
}
// Supported: @Input() + @Output()
class MyFormControl {
@Input({required: true}) value!: string;
@Output() valueChange: new EventEmitter<string>();
}
```
The latter two may still choose to implement `FormUiControl` for other
properties, but again it is not required.
Fix#65478
Change the order of precedence that `[field]` binds to from
1. Custom control (`FormValueControl`, `FormCheckboxControl`)
2. Interop control (`ControlValueAccessor`)
3. Native control (`<input>`, `<select>`, `<textarea>`)
to
1. Interop control (`ControlValueAccessor`)
2. Custom control (`FormValueControl`, `FormCheckboxControl`)
3. Native control (`<input>`, `<select>`, `<textarea>`)
This ensures that Reactive Forms controls authored to use
`ControlValueAccessor` work correctly with Signal Forms, even if they
happen to conform to the `FormValueControl` interface.
Currently circular references in user code manifest themselves with an error like `Cannot read properties of undefined (reading 'ɵcmp')`. This is a bit cryptic so these changes add an assertion mentioning circular references.
Relates to #65917.
Wrap operationsCounter method calls (recordCreate, recordDestroy, reset)
with ngDevMode guards to ensure they are tree-shaken in production builds.
This aligns with the existing pattern where operationsCounter is only
initialized in development mode, and eliminates unnecessary method call
overhead in production.
The optional chaining (?.) is retained as TypeScript doesn't narrow types
based on ngDevMode checks, but the entire expression will be removed during
production builds.
Replace the 'any' type with ClassDebugInfo for the debugInfo parameter in
stringifyTypeFromDebugInfo function. This removes the TODO comment that was
tracking the need for proper typing without creating circular dependencies.
The change improves type safety while avoiding circular imports since
stringify_utils.ts doesn't export anything that definition.ts imports.
Move the instructions used to dynamically bind a `Field` directive to a
form control onto the `Field` itself. This way the instructions are only
retained if the app uses the `Field` directive.
PR Close#65599
When registering providers, the DI system assumes that `viewProviders` are registered before plain `providers`. This was reinforced by components always being first in the array of directive matches, only one component being allowed per node and the fact that only components can have `viewProviders`.
This breaks down if there are host directives with `providers` on the component, because they'll execute earlier, throwing off the order of operations.
These changes fix the issue by separating out the resolvers for `viewProviders` and plain `providers` and explicitly running the component's `viewProviders` resolver before any others. This also has the benefit of not attempting to resolve `viewProviders` for directives which are guaranteed not to have them.
Fixes#65724.
This commit implements a security fix to prevent XSS vulnerabilities where SVG animation elements (`<animate>`, `<set>`, etc.) could be used to modify the `href` or `xlink:href` attributes of other elements to `javascript:` URLs.
Adds a mark for signal forms so we can track adoption. Also moves the call for `@let` into `declareLet` since we don't need it to fire as often as in `storeLet`.
Adds a DI configuration option for signal forms that allows the
developer to specify CSS classes that should be automatically added
by the `Field` directive based on the field's status.
We track all effects that are created for debugging purposes in the `resolverToEffects` map. This ends up leaking memory for effects registered on long-living resolvers (e.g. on the root injector), because they stay in the array, even if the effect itself has been destroyed.
These changes add a callback to clean up the references.
Fixes#65265.
Support binding `[field]` to directives that implement
`FormValueControl` or `FormCheckboxControl`.
The `[field]` binds to whichever directive (or component) matches first in the
event there are multiple implementations. We are considering whether to make
this an error state, which could be reported during type checking.
Closes#63910, Closes#64992
Although the prior commit has made more profiler events guaranteed symmetric
through the use of finally-blocks, there continue to be some situations
that could potentially result in asymmetric events, e.g. application
bootstrap doesn't guarantee symmetric events. This commit makes the profiler
lenient to these situations by unrolling the stack past the asymmetric event
data, eventually reaching the expected start event.
Profiler events are expected to be symmetric, yet in the case of errors this symmetry may break
if events aren't always kept in sync with their corresponding start event. This commit moves
various end events to be run from a finally-block, allowing them to notify the profiler even
when an error has occurred.
Fixes#62947
The `getControlDirective` is called multiple times, both at init and during each update run. Under the hood it performs a linear search for the `Field` directive.
We can speed this up by finding its index once and reusing it since the array of directive matches is static.
