We forgot to tell React that we need the audio renderer to react to changes in the set of MatrixRTC participants; instead we had it referencing rtcSession.memberships non-reactively.
Now, I'm not 100% confident that this is going to fix the "speaking from the void" issues observed in the wild, because I can't reproduce them and, in my testing, the InCallView component always seemed to be rendered redundantly when the MatrixRTC participants change, even though we hadn't explicitly stated that it needs to react. (This makes sense as we haven't memoized the component.) But it's worth a shot.
* Avoid reactivity bugs in how we track external state
Many of our hooks which attempt to bridge external state from an EventEmitter or EventTarget into React had subtle bugs which could cause them to fail to react to certain updates. The conditions necessary for triggering these bugs are explained by the tests that I've included.
In the majority of cases, I don't think we were triggering these bugs in practice. They could've become problems if we refactored our components in certain ways. The one concrete case I'm aware of in which we actually triggered such a bug was the race condition with the useRoomEncryptionSystem shared secret logic (addressed by a1110af6d5).
But, particularly with all the weird reactivity issues we're debugging this week, I think we need to eliminate the possibility that any of the bugs in these hooks are the cause of our current headaches.
* Reuse useTypedEventEmitterState in useLocalStorage
* Fix type error
This gives us the additional insurance of breaking the Safari media acquisition loop at the source by admitting that they can be spurious in practice. Safari, why!?
Now that the toger5/add-room-key-fallback-on-encryption-manager-not-supported branch has been merged, we can reset to develop. (And need to, actually, because that branch is deleted.)
It doesn't check whether it's actually used inside a GroupCallErrorBoundary, and it's generally useful for interacting with any error boundary, so I'm giving it a generic name to reflect this.
We were previously using the useGroupCallErrorBoundary hook to surface errors that happened during joining, but because that part is outside the GroupCallErrorBoundary it just ended up sending them to the app-level error boundary where they got displayed with a more generic message.
Whenever we use a Git dependency, we should reference a specific tag or commit rather than a branch name, because Yarn Classic is bad at recognizing when a Git dependency needs to be re-installed due to it resolving to a new version. (It seems to simply use the version from package.json as a cache key.)
This is relevant only for developers who run 'yarn install' on top of a previously up-to-date working copy, not for CI which re-installs the project on every run.
Following a75952cf77, this is one more upgrade to the widget communication that I'd like to make within this release cycle.
The motivating issue is https://github.com/element-hq/element-web/issues/29429. Fundamentally, without a 'join' action, the only info Element Web can use to determine whether it's joined the call is whether a MatrixRTC membership exists. But membership state events can inaccurately represent the client's actual state (whether because delayed events aren't supported, or because the delayed event hasn't timed out yet), so I suggest we send a 'join' action here just as we do in the Element Web Jitsi wrapper (https://github.com/element-hq/element-web/blob/e9a3625bd6e9a64f216e3caeabca66f48b649332/src/vector/jitsi/index.ts#L503) to let Element Web tap directly into the widget's local state. (This will need additional Element Web changes, but is certainly backwards compatible.)
The one place where we should log out of PostHog and reset our analytics ID is when the user is logging out. This matches the behavior in Element Web and makes sense, I think, because logging out is essentially a request for the app to forget who you are. This means we should also start analytics at the point of logging in / reauthenticating.
I noticed while making this change that there was an unused branch in setClient, so I cleaned it up rather than making myself update it.
We shouldn't be calling PosthogAnalytics.instance.logout() when leaving the call in widget mode, because all this does is reset your analytics ID. In the big picture this is probably inflating our user count metrics.
2bb5b020e6 refactored the end of the GroupCallView render function to not use any early returns, and clumsily failed to account for the fall-through case that makes returnToLobby work (as opposed to sitting on a blank screen).
We need compatible versions of RxJS in our dependencies and LiveKit's dependencies, but LiveKit has pinned it to a very specific version which is now holding us back from updating. See https://github.com/livekit/components-js/issues/1101 for a request for a proper solution.
I was touching this file and noticed a chance to clean up this little TODO comment by swapping in our useLatest hook and giving it better documentation.
Currently, if you set hideHeader to true, you get a big empty <header> on error pages which serves no purpose except to push the screen's content off center.
By keeping 'hangup' and 'close' as separate actions, we can allow Element Call widgets to stay on an error screen after the user has been disconnected without the widget completely disappearing from the host's UI. We don't have to request any additional capabilities to use a custom widget action like this one.
Robin