When compiling under standalone compilation, generated runtime static declarations (static ɵfac and static ɵprov) are emitted into preprocessed TypeScript files and visible to the compiler during semantic typechecking.
When a subclass extends a base class where the subclass is not structurally subtype-compatible with the superclass (such as differing generic type constraints, contravariant method parameters, or EventEmitter<this>), TypeScript's class static side heritage check (TS2417) fails because ɵɵFactoryDeclaration<T> and ɵɵInjectableDeclaration<T> structurally referenced the instance type T.
This change updates ɵɵFactoryDeclaration to return any instead of T, and sets factory return and value types in ɵɵInjectableDeclaration to any. This decouples static side inheritance from T, resolving TS2417 errors across subclassed components and injectables while preserving .d.ts metadata indexing and assignability to ɵɵdefineInjectable. This brings ɵfac and ɵprov into alignment with other Ivy declarations (ɵcmp, ɵdir, ɵpipe, ɵinj), which already treat their generic parameters as phantom metadata.
(cherry picked from commit 66d505e287)
Foreign components are only supported in AOT mode. Using them in
JIT mode previously resulted in silent failures or confusing runtime
errors (such as unknown element errors or crashed template ingestion).
This commit adds explicit validation in JIT compilation:
- Throws an error during component compilation if `foreignImports` is
specified on `@Component`.
- Throws an error during standalone import verification if a foreign
component is mistakenly passed to `@Component.imports`.
(cherry picked from commit 915a03ae85)
For external templates (using `templateUrl`), primary diagnostics are
reported against the synthetic `ts.SourceFile` representing the HTML
template document. However, secondary related messages (such as those in
`foreign_component.ts` and `oob.ts`) were explicitly passing the
component's TypeScript file as `sourceFile`.
Because the character offsets (`start` and `end`) originate from the
HTML template AST, associating them with the TypeScript source file
caused IDEs and CLI diagnostics to map HTML offsets onto the `.ts` file,
resulting in corrupt or out-of-bounds source locations.
This commit resolves the issue by:
1. Making `sourceFile` optional in `makeTemplateDiagnostic` and related
checker interfaces (`TemplateTypeChecker`, `TemplateContext`).
2. Defaulting `relatedMessage.sourceFile` to the template's source file
(`sf` for external/indirect templates, or the component `.ts` file
for direct inline templates) when not explicitly provided.
3. Removing explicit `sourceFile: this.sourceMapping.node.getSourceFile()`
mappings from `foreign_component.ts` and DOM element checks in
`oob.ts`, allowing them to automatically resolve to the template file.
4. Adding unit test coverage for external templates encountering foreign
component conflicts with related messages.
(cherry picked from commit a46292af26)
This started from a real production crash trace:
TypeError: Cannot read properties of null (reading 'nextSibling')
Traced through a minified bundle back to siblingAfter() in
packages/core/src/hydration/node_lookup_utils.ts. It walks forward a
fixed number of DOM siblings during hydration, based on how many the
server rendered. Its only guard, validateSiblingNodeExists(), is
gated behind ngDevMode and stripped from production builds. If the
client DOM has fewer real siblings than the server-serialized data
expects — client/server render diverged, or something outside
Angular (a browser extension, an injected script) removed a node —
the loop ran fully unguarded in production: currentNode went null
partway through, and the next iteration's currentNode.nextSibling
threw a raw, uncoded TypeError instead of a coded, debuggable one.
Reproduced first, not just theorized: added a test that server-
renders a @for loop with 3 items, removes 2 of them before hydration
runs, and confirmed it actually throws that exact TypeError against
the original code before touching anything.
Fix: add a null check at the top of the loop, matching the pattern
already used for the existing NG0500/NG0502 hydration checks in
element.ts — the dev-mode check runs first as before (a no-op in
production), and the new check is a pure safety net that only fires
once that dev-mode check has already been compiled away. In dev mode
this is provably dead code, so nothing about dev behavior changes.
Coded as NG0501 (HYDRATION_MISSING_SIBLINGS, already existed).
While investigating, found and fixed two more call sites with the
exact same shape of bug — a dev-only check that leaves production
completely unguarded:
- navigateToNode() (same file): walks an explicit "firstChild /
nextSibling" navigation path recorded for nodes that can't rely on
simple sibling-walking (e.g. content right after an <ng-content>
slot). Its only guard was also ngDevMode-only. Added the same
production safety net, both mid-loop (the raw-crash case) and
post-loop (where it could previously return null silently despite
the function's return type promising a non-null RNode). New code:
NG0509 (HYDRATION_MISSING_NODE_ON_PATH).
- getParentRElement() (render3/node_manipulation.ts): typed its
tNode parameter as always non-null, but a real production trace
showed it can be null at runtime, crashing on tNode.parent with no
useful information. Widened the type to TNode | null and added a
guard that throws a coded error instead. New code: NG0510
(PARENT_NODE_NOT_FOUND).
Unlike the other two, this one isn't gated behind ngDevMode at all
in the original code — it reproduces unconditionally. Decoded the
actual production minified stack trace against this source
(matched every frame character-for-character) to find the real
cause instead of guessing: an @if/@switch branch's content is its
own embedded template with its own TView, built lazily the first
time that branch renders. If an error interrupts that first pass —
here, a hydration mismatch on the branch's second child, after the
first child's TNode was already created — TView.firstCreatePass
still gets flipped to false in render.ts's `catch` block before the
error propagates, permanently marking that TView as corrupted.
Unlike a component's TView (rebuilt from scratch next time via
getOrCreateComponentTView()'s incompleteFirstPass check), nothing
rebuilds an embedded view's TView. The next time that exact branch
is selected again, its instructions read straight from the
corrupted tView.data instead of creating a fresh TNode, and any
node past the interruption point is still null.
The real test added for this (full_app_hydration_spec.ts) exercises
that whole chain for real: a hydration mismatch on an @if branch's
second child, then re-entering the same branch. Confirmed it throws
the exact byte-for-byte production error message ("Cannot read
properties of null (reading 'parent')") without the fix, and the
coded NG0510 with it. This replaces an earlier synthetic unit test
that just cast a variable to null to bypass the type system rather
than reproducing the actual failure.
Note: a more thorough fix would rebuild the corrupted embedded TView
itself (mirroring getOrCreateComponentTView()), which does make the
crash disappear — but doing that surfaces a second, separate bug:
the aborted first attempt's partially-created LView/DOM is never
torn down, so the retried branch's content gets appended alongside
leftover orphaned nodes instead of replacing them, corrupting
content silently instead of crashing loudly. That's a materially
bigger fix (proper LView cleanup after a mid-creation exception) and
is intentionally out of scope here; this commit keeps the narrower,
safe fix (a clear coded error) rather than trading a loud crash for
silent data corruption.
Along the way, deduplicated describeDomNode(), which existed as two
near-identical private copies (one in this file's own NG0500 check,
one in platform-browser's dom_renderer.ts insertBefore check) and
moved it into core's hydration/error_handling.ts, shared via the
private ɵdescribeDomNode export.
That function was then removed from every production code path
entirely, per review feedback: computing a human-readable node
description is debug-oriented work that shouldn't ship unconditionally
in production bundles, regardless of how cheap it is — production
should get the bare coded error only, the same as everywhere else in
this file. All three fixes above follow that: ngDevMode && 'message'
in production, no computed description. Confirmed the removal
actually shrinks output, not just assumed it: regenerated all 8
affected bundling symbol goldens, and each one shows exactly one
symbol removed — "describeDomNode" — and nothing else.
(cherry picked from commit 83f7695b2e)
Avoid trimming urlStr with String.prototype.trim() in resolveUrl to ensure URL parsing and resolution align with the WHATWG URL standard.
(cherry picked from commit 3e924cc8db)
Non-finite values all used NumberSymbol.Infinity, so formatNumber(NaN)
rendered as infinity. Locale data already defines NumberSymbol.NaN.
(cherry picked from commit 7596548e9b)
This fix ensures that metadata is properly retained when processing
strict standalone component errors for improved error diagnostics.
(cherry picked from commit 74b294cd51)
Store transfer state and serialization callbacks in null-prototype dictionaries, and only read values that belong to the store. This keeps special state keys from changing lookup behavior or exposing inherited cache entries.
Fixes#70265
(cherry picked from commit 168a324cce)
Infer URL parameter values as strings and type preloading work as Observable<void> to match the values produced by these internal flows.
(cherry picked from commit f275d289a0)
When a derived class inherits from a base class, TypeScript enforces that static
properties on the derived class are assignable to those on the base class.
Because `ɵɵFactoryDeclaration<T, ...>` had a return type of `T`, static `ɵfac`
members across inheritance hierarchies could result in type incompatibilities
(for example, when dealing with generics or differing class shapes where the
derived factory return type is not compatible with the base class factory).
Updating the return type of `ɵɵFactoryDeclaration` from `T` to `any` avoids
these static type conflicts across inheriting classes.
(cherry picked from commit 0d7ae16350)
The semantics of optional chaining changes when there are extra parenthesis. We need to make sure that we do not introduce some unnecessary ones.
fixes#70143
(cherry picked from commit e9ba39d671)
`stableTypeOrdering` was recently enabled in the internal builds. While there aren't any breakages in Angular, these changes enable it so we can catch potential issues earlier.
(cherry picked from commit c18261f3fd)
The regular expressions in `_scopeAnimationRule` expect an `animation` or `animation-name` property to be preceded by whitespace or a semicolon, and its value to end at a semicolon. Minified CSS breaks both assumptions. Inside an at-rule the property follows a `{`, and the last declaration of a block has no trailing semicolon, so the closing `}` lands inside the captured value. The keyframe name is then left unscoped while the `@keyframes` rule itself is renamed, so the animation does nothing in a production build.
Accept `{` as a leading boundary and stop the value at `}`. The prefix is written back unchanged, and a declaration value cannot contain an unescaped `}`.
Fixes#70316
(cherry picked from commit 58b0cb4735)
The readonly rule previously used 'when' in configOrLogic to determine if a configuration object was provided. Under property renaming/minification, the string literal property lookup fails and causes the rule to fall back to being permanently readonly.
This change switches to the dot-access form (configOrLogic?.when), matching hidden() and ensuring property renaming works correctly.
(cherry picked from commit 1e35de536d)
Currently the standalone migration only copies symbols referenced from the NgModule metadata into the main file. Top-level `registerLocaleData` calls, and the default imports they depend on, are dropped silently when the module file is pruned, which breaks locale resolution at runtime with a "Missing locale data" error.
Copies top-level `registerLocaleData` calls from the file of the bootstrapped module into the main file, and adds support for carrying over default imports which were previously skipped silently by the reference resolution.
Fixes#50886
(cherry picked from commit bdc09e8183)
The quick fix and completion auto-import always generated the module specifier with double quotes, ignoring the user's quote preference and the style used in the file.
The generated import now respects the `quotePreference` from the TypeScript user preferences. When the preference is `auto` (or absent), the style is inferred from the first existing import in the file, mirroring TypeScript's own behavior. Files with no imports
keep the previous double-quote default.
Fixes#67108
(cherry picked from commit 0f0d52e9c1)
Use the Subscription type returned by Observable.subscribe instead of storing async validator subscriptions as any.
(cherry picked from commit a4a428ece2)
When Angular type-checks host bindings and template class bindings
(`[class.foo]="expr"`), the Type Check Block (TCB) emits the binding
as a standalone expression statement (e.g. `(expr);`). This verifies that
the expression itself is syntactically and semantically valid (e.g. properties
exist on the component instance), but does not constrain the
expression to `boolean` because Angular evaluates class bindings using standard
JavaScript truthiness.
In classic `ngtsc`, the runtime Ivy instructions (`ɵɵdefineComponent`) were
generated only during the JS emit phase, after TypeScript type-checking had
completed. Thus, `tsc` never validated the arguments passed to `ɵɵclassProp`.
Under standalone commpilation, Ivy definitions are generated
directly into the TypeScript AST and type-checked by `tsc`. This causes `tsc`
to check the emitted `ɵɵclassProp('foo', expr)` call against the instruction's
declared signature. Because `ɵɵclassProp` was strictly typed as
`boolean | undefined | null`, any valid truthy non-boolean expression
(e.g., `1`, `items.length`, or non-empty strings) produces a `TS2345` compiler error.
At runtime, `ɵɵclassProp` delegates to `checkStylingProperty`, which evaluates
the value via truthiness (`!!value`). Widening the parameter type to `any`
aligns the instruction's type signature with Angular's binding semantics and
prevents type-checking failures during in-place compilation.
(cherry picked from commit 779f67777d)
This forwards the `AbortSignal` natively supplied by Chrome 153.0.8009.0's update to `registerTool` executions down into the Angular context. This helps tools gracefully handle execution cancellations initiated by agents or users, preventing unnecessary background work and leaking resources.
We combine the injector/component destruction abort signals with the native WebMCP execution cancellation signals so the underlying tools correctly cancel long-running requests without producing invalid states.
Justification:
https://groups.google.com/a/chromium.org/g/chrome-ai-dev-preview/c/9291sjhIRz0/m/RFuyRrs5AAAJ
(cherry picked from commit e041f9483e)
Introduces support for route-level resources via the `resources` property on route definitions, enabled with `withRouterResources()`.
Router resources provide a reactive, signal-based alternative to resolvers, allowing routes to declare resources tied to route parameters and route lifecycle:
- The `resources` function executes during navigation transitions within an injection context scoped to the route (`_localInjector`).
- It receives a `ResourceContext` containing signals for `params`, `queryParams`, `fragment`, and `data`, alongside the static `snapshot`.
- For newly created routes, `resources` runs once and attaches to `ActivatedRoute.resources`. For reused routes, parameter signals update reactively to trigger new data fetches while keeping resource references stable.
- Wrapped resources (`routerResource`) provide transactional stability: snapshots are frozen during active navigations to prevent UI jitter, unfreezing on `NavigationEnd`.
- On cancelled navigations or errors, rollback recovery retains the frozen snapshot until reverted signals settle, avoiding flashes of loading state. Manual reloads are rejected while frozen.
- Local injectors are automatically cleaned up if navigations are cancelled or rolled back.
Note that this commit only implements non-blocking resources (marked with `nonBlocking()`), deferring blocking resource resolution to future work to keep the initial changeset smaller and less complex.
PR Close#70211
Activated Route Injector feature was developed for resources. This moves the internal
naming (files and vars) to reflect this now that we have landed the core bits of plumbing
and are ready to implement router resources.
PR Close#70211
`findConnectedBlocks` scans siblings after an `@if` to collect connected
`@else`/`@else if` blocks. Whitespace-only text nodes encountered during
the scan were eagerly added to `processedNodes`, marking them as "do not
emit", before confirming whether a connected block actually followed.
By the time `findConnectedBlocks` runs, `WhitespaceVisitor` has already
converted `&ngsp;` (and ` `) into a plain space character, making
them indistinguishable from insignificant whitespace via `.trim().length`.
When the next sibling was a second, unrelated `@if` instead of `@else`,
the scan stopped but the text node was already silently dropped.
Fix by deferring the `processedNodes` insertion into a pending buffer and
only committing those nodes once a connected block is confirmed to follow.
Fixes#55791
(cherry picked from commit 53fc371142)
Defer blocks that ask to run during browser idle time with the same
options get grouped into one batch ("bucket") and processed together
the next time the browser is idle. While that batch runs, if one
callback (or the change-detection check right after it) asks to
schedule more idle work for that same batch, Angular used to register
a second idle-callback handle instead of reusing the one already in
flight.
That second handle became orphaned. By the time the batch finished
running, its queue was empty, so Angular threw away the bookkeeping
for it — including the only reference that could have cancelled the
handle. If the app was destroyed before the browser got around to
calling it, it fired anyway, against a scheduler that no longer
existed.
The root cause was a marker (idleId) that IdleScheduler uses to know
"I already have a browser callback pending for this batch, don't
request another one." That marker was being cleared to null right at
the start of processing a batch, before any of its callbacks had
actually run. So a callback that re-entrantly asked to schedule more
work mid-batch saw "nothing pending" and requested a redundant handle.
The fix: leave the marker set for the entire batch instead of clearing
it up front. Only clear it once every callback in the batch has run —
at that point, if there's leftover work, it's safe to request a fresh
handle for it.
Added a test for the basic re-entrant scenario, plus five more for
edge cases: a re-entrant add() into a different batch, and re-entrant
remove() of a sibling callback that hasn't run yet versus one that
already has.
(cherry picked from commit c2b14b7ab4)
Before: getSignalGraph only exposed debuggableFn on computed and template
nodes, preventing DevTools from linking to source code for other node types.
After: debuggableFn is also populated for linkedSignal, effect, and
afterRenderEffectPhase nodes.
(cherry picked from commit 0ddbc47e7f)
When a componentless parent route has children rendered into a named outlet
(e.g. `outlet: 'inner'`), the `canDeactivate` guard received `null` as the
component argument instead of the actual component instance.
The bug was in `deactivateRouteAndItsChildren`: for componentless routes,
each child was passed the same `context` inherited from the parent lookup,
which was `null` when the parent component only registered a named outlet.
The children's actual outlet contexts were never consulted.
The fix adds a `parentContexts: ChildrenOutletContexts | null` parameter so
that for componentless routes, each child is looked up by its own outlet name
in `parentContexts` (e.g. `parentContexts.getContext('inner')`). For
component routes, `context.children` is passed as the new `parentContexts`
on recursion, ensuring correct context resolution across component boundaries.
This re-addresses #34614. A previous fix (#36302) was reverted because it
passed `parentContexts` unchanged through component boundaries; this fix
updates `parentContexts` to `context.children` when descending into a
component route, preventing the wrong contexts from propagating into deeper
componentless levels.
Fixes#34614
(cherry picked from commit f1a4c85212)
We keep getting PRs that target single usages of `hasOwnProperty` and we have ~100 of them. These changes aim to address the issue centrally by swapping out all the instances and adding a lint rule against introducing new ones.
(cherry picked from commit 732e505018)
When a view has more than one effect scheduled to run, and one of them
destroys the view (e.g. by calling `componentRef.destroy()`), the
remaining effects in that same flush could still run afterward,
against a view that no longer exists. In some cases this crashed
outright with `TypeError: view[EFFECTS] is not iterable`.
Here's why: `runEffectsInView` walks a view's effects in a `for...of`
loop, wrapped in an outer `while` loop that re-checks for any effects
that became dirty as a side effect of ones that already ran. When an
effect destroys its view, `view[EFFECTS]` gets set to `null` as part
of tearing the view down.
First attempt checked for that inside the `for...of` loop, before each
effect runs. That covers a sibling effect later in the *same* pass,
but misses a second case: if the effect that destroys the view *also*
dirties another effect on that same view in the process (e.g. by
writing a signal the sibling depends on), the outer `while` loop sees
`HasChildViewsToRefresh` set and tries to restart — and immediately
crashes re-entering `for (const effect of view[EFFECTS])` on a
now-null value, before the in-loop check ever gets a chance to run.
Reproduced that exact crash with a test first: two effects on one
view, the second one writes a signal the first depends on and then
destroys the view in the same call — confirmed it throws before
touching the fix.
Fixed by checking right after `effect.run()` instead of before it,
covering both cases in one place: the remaining effects in the current
pass, and the loop trying to restart afterward. As soon as one effect
destroys the view, nothing else runs against it again.
This is intentionally narrow in scope. An earlier version of this fix
also tried to guarantee that `onCleanup()` callbacks still ran even
when registered after an effect destroyed its own view. That's been
dropped — destroying your own view and then continuing to register
more work for it isn't something the framework should have to paper
over. If you need to do both, register `onCleanup` first, then
destroy.
(cherry picked from commit c658d73210)
Previously, when you passed an object typed like
Record<'a' | 'b', number> into the `keyvalue` pipe, TypeScript would
"forget" that the keys could only ever be 'a' or 'b', and just tell
you the key was a plain `string` instead. So code like this used to
fail to compile, even though it's correct:
```ts
const input: Record<'a' | 'b', number> = {a: 1, b: 2};
const result = pipe.transform(input);
const key: 'a' | 'b' = result[0].key; // error: string is not 'a' | 'b'
```
This happened because the pipe has multiple overloaded versions of
transform(), and TypeScript checks them top to bottom, using the
first one that matches. The "number keys" overload was listed first,
and it happened to also match string-keyed objects by accident, so
it "won" before the correct "string keys" overload ever got a
chance to run.
The fix just reorders those two overloads so the string-keys one is
checked first. Nothing about runtime behavior changes — objects with
actual numeric keys (e.g. Record<1 | 2, string>) still correctly
report their keys as plain `string`, matching what Object.keys()
really returns at runtime.
(cherry picked from commit 46d2cb7ff0)
Prevent arbitrary MetaDefinition properties from writing on* handlers directly to meta elements. Browser events can execute these handlers, including on meta elements rendered in the document body.
(cherry picked from commit 6f9a6bea50)
The "Value transformation" section taught readers to hand-roll
transformation with `linkedSignal()` and a manual parse method, even
though `@angular/forms/signals` ships `transformedValue()` for exactly
this case. Readers ended up with a weaker version of a feature the
framework already provides — notably, no parse error reporting.
Rewrite the section around `transformedValue()` and document the parts
the manual pattern could not cover: returning `{error}` from `parse` to
surface parse errors on the field's `errors()`, and `reset()` clearing
them. This also makes good on the cross-reference from the validation
guide, which pointed here for parse error details the section never
covered.
Fixes#70206
Co-authored-by: Matthieu Riegler <kyro38@gmail.com>
(cherry picked from commit 182c371d82)
Replace unsafe `any` type annotations with `unknown` across the localize
utility layer to improve type safety and catch potential type errors at
compile time rather than at runtime.
Changes in `messages.ts`:
- `ParsedMessage.substitutions`: `Record<string, any>` → `Record<string, unknown>`
- `parseMessage` parameter `expressions`: `readonly any[]` → `readonly unknown[]`
- Local `substitutions` variable: `{[key: string]: any}` → `Record<string, unknown>`
Changes in `translations.ts`:
- `isMissingTranslationError` parameter: `any` → `unknown`, with proper
narrowing (`typeof e === 'object' && e !== null`) before property access
- `MissingTranslationError.type` visibility: `private` → `readonly` to allow
access through the narrowed `unknown` type in the type guard
- `translate` parameter and return type: `readonly any[]` → `readonly unknown[]`
- `makeTemplateObject` cast: `cooked as any` → `cooked as unknown as TemplateStringsArray`
Fix in `mock_message.ts` (test helper):
- `substitutions: []` → `substitutions: {}` — the array literal was only
assignable because the field was typed as `any`; the correct empty value
for a `Record<string, unknown>` is an object literal
(cherry picked from commit b10021c79b)
NgModel injects ControlContainer with @Host(), which stops the injector at
the component host element boundary. When NgForm lives in a parent component
and ngModel lives in a child component, the injection returns null silently
and the control acts standalone — never registering with the form.
To surface this invisible failure, emit a dev-mode warning (NG01354) when
ngModel's @Host() injection finds nothing but the element Injector can still
reach a ControlContainer further up the hierarchy. The warning identifies the
cross-boundary issue and points developers to the viewProviders fix or the
standalone option.
Adds the NG01354 reference page explaining why the warning fires and providing
two remediation paths: bridging ControlContainer via viewProviders, or opting
out with [ngModelOptions]="{standalone: true}".
Fixes#47580
(cherry picked from commit 38d093232c)
This test makes sure that when Angular is served from Electron or just from the file system that the router works correctly.
(cherry picked from commit 70b252cac3)
StateValue and AnimationTransitionNamespace.trigger detect the {value,
params} object form of a trigger binding by calling hasOwnProperty on the
bound value. When that value is an object from untrusted data (for example
a parsed JSON payload) carrying an own hasOwnProperty key, the shadowed
property is called as a method and throws, breaking the animation flush.
Use Object.hasOwn for the check so a shadowing key no longer matters.
(cherry picked from commit c73a001fbf)
Angular Robot