The call's shortcuts were listened for on the window and allowed
whenever focus was inside the standalone app's `#root` — which, for a
component embedded in a host, is the host's own root, or nothing. So m,
v and space fired while the user typed in the host's composer, and two
Element Calls on one page both answered every key.
Listen on the element Element Call treats as its root instead (the body
standalone, so nothing changes there), and judge whether a key press is
spoken for by what has focus — a dialog or a text field — rather than by
where it sits in the DOM, since the modals are now portalled to that
same root.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The component built its `MediaDevices` once, in a scope nothing ever
ended, so every mount left device observers running for the rest of the
page's life. Building it in an effect ties the scope to the component's
lifetime — and to the options it was built with, which were previously
frozen at first render.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Everything downstream of the component's params — the mute state, the
call view model and with it the media connection — is keyed on the
identity of the params object, which was memoised on the identity of
the `config` prop. A host writing `config={{ ... }}` inline, which is
the natural way to write it, therefore tore the whole call down on
every render. The harness happened to pass a constant, so nothing
noticed.
`useStableValue` hands out the same object for as long as a deep
comparison says nothing changed, so an inline config costs nothing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`component/package.json` describes the built component as a package
(`@element-hq/element-call-component`: entry, stylesheet subpath, types,
peer dependencies) so that a host can depend on
`github:element-hq/element-call#<ref>&path:/component`. Its `prepare`
script builds on install, since nothing is published yet.
For that the component build now lands in `component/dist` instead of the
repository's `dist`, and `pnpm build:component` also emits the type
declarations (`component/tsconfig.build.json`, `build:component:types`),
which previously had to be produced by hand.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The widget tests cannot reach what makes a component different, because the
iframe used to guarantee it: that Element Call stays inside the space it was
given, and that two of them can exist in one page. So the development harness
gets driven by Playwright.
Three tests. Two components in one page holding a real call between two devices
of one account. The settings dialog and the reaction picker staying inside the
container the host gave, asserted by bounding box. And the host bridge
reporting in both directions, including a host-initiated mute coming back as a
report of the new state.
The containment test is the one worth having: both of the escapes found by
running the harness by hand — the settings dialog centred on the window, and
the reaction picker at 82vh landing below the container — would have failed it,
and neither was visible to typechecking, linting or the unit tests.
Users and rooms are created through the Synapse admin and client-server APIs
rather than by driving an interface, and the harness now takes its credentials
from its own query string so that a test can say which account and room to use.
A host reading its own URL is proper; it was Element Call doing so that was the
mistake.
Playwright gains a second web server for the harness on port 3001, a Vite dev
server whether or not the app itself is served from Docker, since the harness is
a development page with nothing to build.
Opening settings in an embedded call put the dialog in the middle of the host's
window, spilling outside the container, and with two calls on a page the second
drew over the first's dialog.
The container established a stacking context but not a containing block, which
are two different things and only the first had been done. `position: fixed`
resolves against the viewport unless an ancestor makes itself the containing
block, so the modal scrim and dialog in Overlay.module.css — `fixed`, `inset:
0`, centred, because in the standalone app they are meant to cover the page —
were positioned and sized against the window. Layout and paint containment
makes the container the containing block for those descendants and clips what
we paint to our own box. The second-call-on-top symptom goes with it, since the
dialog now stays inside the first call's box and there is nothing to overlap.
Two sibling components still cannot draw over one another by construction,
neither being able to leave its own stacking context, but that only shows if a
host overlaps them.
Containment clips a box measured in viewport units but cannot resize it, so
anything positioned that way is simply put somewhere outside the container and
disappears. That applied to the reactions overlay, a `100vw` by `100vh` box, and
to the reaction picker, which sits at `82vh` so as to appear near the footer it
belongs to. Both are positioned against whatever Element Call treats as its
root — `[data-overlay-container]` in the app, which is the size of the page, and
the host's container when embedded — so percentages mean the same thing there
and the right thing here. The earpiece overlay is `inset: 0` with no viewport
units, so containment is enough for it.
Three viewport-relative sizes remain, all of which need more than a change of
unit: the lobby's video preview is `50vh` on a flex item with an aspect ratio,
`--content-inset-*` ramps up to a desktop inset from the window width, and the
picker's `max-width` cap is the window's.
The clipping cuts both ways: a menu near the edge of a small container is
trimmed rather than overflowing into the host. That is the trade an embedded
component makes.
The component seeded its parameters with `computeUrlParams()`, which reads
`window.location`. Element Call is not the page any more, so what it found
there was the host's URL: no `widgetId`, therefore not a widget, therefore no
intent, therefore the standalone app's preset.
The visible symptom was the Element Call logo in the footer of a call embedded
in someone else's application, since `showLogo` follows `header ===
HeaderStyle.Standard`. The rest of that preset mattered more. A hosted call
defaulted to `perParticipantE2EE: false` and `confineToRoom: false` — so
unencrypted, and willing to take the user out of the room a host had put them
in — and a host could not correct either without naming every parameter
itself, nor even name `header`, since the enums were not exported.
So the intent presets move out of `computeUrlParams` into
`configurationForIntent`, and the component builds its parameters from an
intent plus the properties a host has no URL to supply. `intent` becomes a
prop, defaulting to joining an existing group call: the lobby first, confined
to the room, encrypted per participant, and no Element Call branding in
someone else's interface. A host that knows which button the user pressed
should say which.
One deliberate difference from the widget presets: the background defaults to
solid rather than the gradient, which is drawn by a `position: fixed`
pseudo-element and would escape the container to cover the host.
`UserIntent.Unknown` still means the standalone preset, so the app and widget
are unchanged — including a widget that names no intent, which has always been
given those defaults.
`pnpm dev:component` serves a page that stands in for a host application: it
signs in twice against the development backend and shows two calls side by
side, in resizable boxes, with furniture of its own around them. Two devices of
one account, so a real call happens between the two components and anything
Element Call keeps once per process rather than once per call shows itself.
The host bridge is driven by hand and reports both directions in a log along
the bottom, which is the first exercise the theme, hang-up and device-mute
requests have had outside widget mode. Each pane can be unmounted and remounted
to see what Element Call leaves behind, and there is a `position: fixed` dialog
belonging to the host to see whether it covers the calls. The page uses none of
Element Call's design tokens, so anything that looks styled outside a pane came
from Element Call reaching out of its container. It reaches Element Call only
through the component's public interface, which is how the exports missing from
that interface came to light.
Three things about the component build the harness turned up on the way, all
too small to be worth their own commits:
- It copied `public/` into `dist/`, including the developer's own gitignored
config.json, into output we would publish. `publicDir: false`, as the
embedded build already does. The sdk build has the same leak; untouched.
- `pnpm lint:externals` now exists, which the build config already claimed it
did. It reads the external list out of that config and fails if the source
imports React, the Matrix SDK or LiveKit by a path the list does not name.
Since the bundler silently ignores the pattern form of that option, an
unnamed subpath is bundled with no warning at all — which is how a host would
end up with a second React.
- `lint:oxlint` ran over `src playwright`, so nothing in `component/` had ever
been linted.
Serving a page also meant the shared plugin list could no longer inject the
app's HTML entry point unconditionally, so that is now optional — and off for
the library build too, which never had an HTML page to inject it into.
The component built and typechecked in the previous commit, but only because
nothing had rendered it. Everything Element Call needs that `src/main.tsx`
side-loads was missing from it.
Its stylesheet: only main.tsx imported index.css, so the library build emitted
CSS-module styles with every `--cpd-*` and `--font-size-*` unresolved. Split
into base.css, which both the app and the component import, and the rules that
are about owning a page, which only the app does. The split is a straight move
— comment-stripped and sorted, the old file and the two new ones differ by
exactly one line — and that line is the deliberate part: `.no-scroll-body`
becomes `body.no-scroll-body`. Element Call adds that class to whatever it
treats as its root, and since the root can now be a container, `position:
fixed` would have taken that container out of the host's layout. Pinning the
page is what it always meant.
Its translations: `initializeElementCall` called `i18n.init` with neither
resources nor a backend, so every key would have rendered as itself. English is
bundled in. The app fetches locale files from URLs its own build emits, which a
host serving the library from somewhere else could not resolve, so how a host
picks a language is left open.
And the types a host needs: `HostBridge` alone is not enough to implement
`HostBridge` — `HostRequest`, `DeviceMuteState`, `DeviceMuteRequest` and
`JoinCallData` all appear in its signatures, and `ConfigOptions` in
`initializeElementCall`'s.
Adds component/index.tsx as a fourth build target: <ElementCall client
roomId /> and an initializeElementCall to await once beforehand. It gives
Element Call everything it would otherwise take from the page it is on —
the parameters, the host bridge, media devices, translations, a container
to confine itself to — and hands it the host's client rather than finding
one of its own.
React, the Matrix SDK and LiveKit stay external, since the host has them
and a second copy of any would not merely be wasteful: React would hold
two sets of hooks and the client would run two sync loops. Every subpath
has to be listed by name, because the pattern and callback forms of
rollupOptions.external are silently ignored here — a lesson worth the
comment that records it.
Element Call's own navigation runs in a MemoryRouter, so being embedded
cannot disturb the host's URL. ClientContext and GroupCallView both
navigate, so some router has to be present.
The bundle is not yet a reasonable size: library mode base64-inlines
assets referenced through import.meta.url, so MediaPipe's vision runtime
lands in it whole. Left for its own change, since the fix — loading the
background blur transformer lazily — is worth doing for the standalone app
too.
ElementCallView took muteStates as a prop, which would have meant an
embedder building one before it could show a call. It could not simply make
its own: RoomPage created one on mount whichever branch it went on to
render, and two would both report the user's mute state to the host,
talking over each other.
Move the construction into a useMuteStates hook, and give the lobby shown
while waiting to be let into a room its own component. Each lobby now holds
mute state only while it is on screen, so there is never a second one, and
the component can own the call's.
KnockLobbyView also takes the room summary and label handling that RoomPage
was assembling on its behalf, leaving the page with arriving at a call
rather than being in one.
Six selectors named `body` directly — the gradient backdrop and the
platform font overrides in index.css, and the iOS adjustments in
AppBar.module.css and Modal.module.css — so they only applied when Element
Call owned the page. A host mounting it into a container would have got an
interface decorated correctly and styled incorrectly, with nothing to show
that anything was wrong.
Mark the root element with data-element-call-root and match on that
instead. Scoping this way keeps the selectors more specific than they were,
rather than less: widening them to a bare [data-platform=…] would have
dropped specificity from (0,1,1) to (0,1,0) and changed which rules win.
The platform attribute moves with them, from the initializer's write onto
document.body to a layout effect on the root, alongside the theme — so it
still lands before anything is painted.
No visual change while Element Call owns the page: the root is the body,
which now carries the attribute, so every rewritten selector matches the
element it always did.
Being in a call and deciding to show one are different jobs, but RoomPage
did both: routing, authentication, resolving room aliases and knocking, and
then the call itself. Only the first set belongs to whatever is hosting
Element Call.
Add ElementCallView as the seam between them. It takes a client and a call
to join, and owns whether the user has joined — which is the call's own
business rather than its host's. RoomPage keeps everything about arriving
at a call and renders this for the call itself.
Nothing else moves yet. muteStates is still passed in, because the
standalone shell shares one with the lobby it shows while waiting to be let
into a room, and two instances would both report the user's mute state to
the host.
RootElementProvider was exported but never used: nothing supplies a root
element, so every consumer falls back to the document body. Knip reports
it, which is what is failing CI.
M1 adds a provider back along with the component that mounts Element Call
into a container. Until then there is nothing to provide.
RootElementContext's documentation promised that Element Call confines its
decoration to the given element, but several selectors still name body
directly and the initializer writes data-platform onto it. A non-body root
is decorated correctly and styled incorrectly, with no error to show for
it, so the documentation should say so until the stylesheets are scoped.
close() sent io.element.close and then stopped the transport, so a
rejected send skipped the stop. Both call sites this replaced stopped it
unconditionally — ErrorView in a finally, GroupCallView outside its
try/catch — because a host that never acknowledges the request would
otherwise leave the messaging live and the close button doing nothing.
Restore that with a finally.
The SDK stopped sending join notifications: it threaded callIntent into
createCallViewModel$ but not its pair sendNotificationType, which
enterRTCSession used to read for itself, so an explicit
?sendNotificationType=ring — or an intent that implies one — no longer
reached joinRTCSession.
The mechanism is worth fixing rather than the instance. The defaults on
CallViewModelOptions describe a standalone Element Call, so a widget
caller that misses a field gets standalone behaviour rather than an error,
and the SDK is only ever a widget. Give both callers one shared mapping so
they cannot drift, and cover the whole chain from URL to options in tests.
autoLeaveWhenOthersLeft and waitForCallPickup stay out of it: the view
model never read those from the parameters, so enabling them for the SDK
would be a change in its behaviour rather than a fix.
Element Call reached the widget API through a mutable module-level
binding, which every consumer imported directly. Nothing outside the app
shell needs it any more, so hand it back from initializeWidget and thread
it through: the initializer returns it, main passes it to App, and App
uses it to build the host bridge and to await the client the host is
lending us.
ClientContext's loadClient is now only about restoring or creating a
session of Element Call's own, since a widget's client arrives as a prop
like any other host's would.
Also fixes an early return added in the previous commit, which skipped
starting the analytics settings listener when a client was supplied. That
was harmless until now, but would have stopped analytics following the
user's choices in widget mode.
sdk/main.ts asked the host to close by hand; it now uses the bridge, which
also stops the transport as the app does.
ClientProvider found its own client: from the widget API, or by restoring
or creating a session. A host that embeds Element Call already has one,
and owns the user's session, so accept it as a prop and skip all of that.
A supplied client seeds the state synchronously, since there is no session
of ours to restore and so nothing to wait for.
Also guard the broadcast that shuts down other instances of the app. It
protects Element Call's own session and crypto stores, which is why it was
already skipped in widget mode — a host's client has the same property, so
without this an embedded Element Call would close down the user's other
tabs on mount.
Adds the first tests for ClientContext, covering both
PosthogAnalytics read its own configuration out of the environment on
first use: the URL parameters, config.json, and the widget global. An
embedded Element Call has none of those to offer, and analytics that
configure themselves cannot be switched off by a host that does its own
reporting.
Take an AnalyticsConfig through PosthogAnalytics.configure() instead,
called from the initializer once the config has loaded. Unconfigured
analytics stay off.
Note the two halves of that config are decided differently, and have to
be: where the credentials come from depends on the package, but who owns
the user's analytics identity depends on how Element Call is running,
since the full package can be used as a widget too.
Drop the widget check around cryptoVersion, which never did anything —
widget mode never initialises crypto, so getCrypto() is already undefined
there.
Move the tests covering which package reads which credential source onto
analyticsConfigFromEnvironment, where that decision now lives.
The remaining reads of the widget global were all asking one of two
different questions, so they get two different answers.
The app shell — the auth hooks, automatic guest registration, the group
call loader's diagnostic and the initial mute state — wants to know
whether Element Call was launched as a widget. That is a property of the
URL it was launched with, so expose the isWidget that computeUrlParams
already computed internally, documented as being for shell use only.
The call interface — whether to offer the profile settings tab — wants to
know something about its host, so it asks the bridge. A host that can
dismiss Element Call owns the user's account, so their profile is not ours
to edit; this reuses the close capability as a proxy, with a TODO
alongside the others.
ClientContext also takes supportsReactions from the bridge rather than
checking four widget capabilities itself, which removes widgetApi from
InitResult — a field that was always null outside widget mode.
Note this changes behaviour for a malformed widget URL: one carrying a
widget ID and parent URL but missing the room, user, device or base URL
would previously have fallen back to registering a guest user, and will
now not.
MuteStates, CallViewModel and LocalMember reached the host through the
widget global. None of them are React components, so they take the bridge
as an explicit parameter: a constructor argument for MuteStates, a field
on CallViewModelOptions, and one on createLocalMembership$'s props.
src/state no longer refers to the widget API.
The conditionals around it mostly disappear: nullHostBridge's observables
are NEVER, so there is nothing to guard, and a request carries its own
reply rather than needing the transport and the original event.
CallViewModelWidget.test.ts drove hangup by emitting on the mocked
widget's action emitter, so it now injects a bridge instead, and checks
that the request is acknowledged.
ErrorView needed a widget only to decide whether to offer a close button
or a link home. That prop was threaded down through ErrorPage, RichError
and GroupCallErrorBoundary from seven call sites, to answer the single
question of whether the host can dismiss Element Call — which the host
bridge now answers directly through the presence of close().
Read the bridge from context in ErrorView and remove the prop, along with
the plumbing that carried it.
No functional change: the rendered output is unchanged, as the existing
snapshot confirms.
Move GroupCallView onto the host bridge
GroupCallView asked the widget API to keep it on screen, to close it, and
to tell it when a preloaded call should join. Route all three through the
host bridge and drop the widget prop.
useRoomEncryptionSystem is used from GroupCallView, inside the part of
Element Call that will become the embeddable component, but it reached for
getUrlParams() via getKeyForRoom(). Extract the lookup into a helper taking
the room ID and password explicitly: the hook supplies them from the
parameters context, while getKeyForRoom keeps reading the URL for its one
remaining caller in the app shell.
No functional change.
The view models reached for getUrlParams() — and so window.location —
from deep inside the call path: CallViewModel, MediaDevices, Publisher,
LocalMember and the footer view model. An embedded Element Call has no URL
of its own, so these values have to arrive as arguments instead.
Add the relevant options to CallViewModelOptions, to the MediaDevices and
Publisher constructors, to createLocalMembership$ and enterRTCSession, and
to createCallFooterViewModel. The remaining React consumers read the
context added in the previous commit. AppViewModel now takes its audio
output options too, moving that URL read out to main.tsx, where the app
shell can act as the adapter.
The new CallViewModelOptions fields are optional, defaulting to what the
URL parameters resolve to outside widget mode; the MediaDevices and
Publisher arguments are required, so that every construction site has to
be explicit.
useTheme.test.ts mocked the UrlParams module with a factory, so it needed
updating to mock the hook rather than getUrlParams.
No functional change.
Element Call's parameters come from its URL, which works while it owns
the page but leaves an embedder with nowhere to put them. Add a context
so they can be provided directly, falling back to parsing
window.location when no provider is present.
This also decouples the thirteen consumers from react-router: useUrlParams
called useLocation, so each of them required a router ancestor, which the
embedded build will not have. The standalone and widget builds keep their
URL-derived behaviour via useUrlParamsFromLocation, provided in App.
No functional change.
Element Call writes its theme classes, background and layout attributes
straight onto document.body, and portals its modals there too. That is
only correct while it owns the page; embedded in a host application it
has to confine itself to the container it was mounted into.
Add a RootElementContext, defaulting to document.body so that the
standalone and widget builds are unaffected, and point the theme classes,
data-background, no-scroll-body and the fullscreen target at it.
Give the Modal and Toast portals an explicit container as well. Radix and
vaul both default to document.body, so without this every modal, drawer
and toast would render outside the container and lose the theme and
platform attributes set on it.
No functional change: the root element is document.body until an embedder
provides otherwise.
Config.init() derives the location of config.json from window.location,
which only makes sense while Element Call owns the page. When it is
embedded in a host application the host owns the configuration, so add
Config.initWith() to accept it directly.
Share the defaulting and validation between both paths via resolveConfig(),
so that an injected config behaves identically to a fetched one, and mark
initialization as complete so that the init() calls already on the startup
path resolve immediately instead of fetching over the top of it.
initDefault() becomes initWith({}), which also stops it handing out a
shallow copy of DEFAULT_CONFIG whose nested objects were shared with the
module-level default.
Element Call configured the global i18next singleton. When Element Call
runs embedded in a host application rather than as its own page, that
singleton belongs to the host, so configuring it would clobber the
host's translations.
Create Element Call's own instance in utils/i18n.ts, configure it in the
initializer, and pass it to components via <I18nextProvider>. Drop
.use(initReactI18next) from the initializer: it registers the instance
as react-i18next's global default, which is the global we are trying to
avoid. Tests and stories keep using it, so that they do not need to wrap
every render in a provider.
Two modules imported `t` directly from "i18next" and so were bound to
the global instance: utils/errors.ts now calls i18n.t() on the instance
(reached at call time, since i18next only assigns `t` during init), and
QrCode uses useTranslation() like every other component.
No functional change.
