[jitl]

> event.timeStamp < lastKeyDownTimeStamp + ANDROID_COMPOSITION_LATENCY

I tried so hard to avoid adding timeouts/timestamp logic for Android because it feels like a nondeterministic hack. In the end I gave up and did something similar.

I also found beforeInput doesn’t offer much solace on Android or with CJK language, because messing around with preventDefault or the DOM during composition disturbs the user.

One weird thing I never figured out is that if the user puts the selection in an empty block, GBoard wants to jump to the nearest word in a different paragraph. We do some brutal hacks to get around that :-/

I read your comment on ProseMirror forum about preventing interference from Granmarly, etc. Do you do this by reverting MutationRecords? Where is the code/docs for that?

[trueadm]

Yeah, I really wanted to avoid them too. There doesn't seem a good way. The issue you describe with selection jumping is a well known issue with GBoard. Here's the logic for the mutation handling:

https://github.com/facebook/lexical/blob/main/packages/lexic...

[jitl]

https://github.com/facebook/lexical/blob/5802651c24d88b2f7d2...

> // Check for any random auto-added <br> elements, and remove them.

> // These get added by the browser when we undo the above mutations

> // and this can lead to a broken UI.

This made me laugh out loud. Browsers!

Until here our approach looked quite similar. Leave the observer connected during reverts and avoiding double-peocessing using .takeRecords never occurred to me, I always disconnect it. Maybe some of these brs are causing chaos…

[trueadm]

Now thinking about it, I don't know why we didn't just do that. There must have been a reason, but I really can't recall what it was. If you're ever interested in contributing, this would be an epic contribution! :D