It's much less likely we'll get another conflict if we retry
with ?oldid=12345 than if we retry omitting the oldid again.
Change-Id: I10e40f714fd44d7f0ac16c9d0dc6239a95c0ec51
Also introduces paction=metadata in ApiVisualEditor which
doesn't hit Parsoid but gets all the other information we need.
Bug: T90374
Change-Id: I71edbc0f8dc15a6f8d3672c10ca3c641cc172be7
Split up beta and meta dialog show methods so that beta dialog
is displayed as soon as possible, regardless of the surface being
ready. Also make sure that we destroy the temporary window
manager on destroy.
Bug: T90454
Change-Id: Ib8f94518af431487ce940a74a8c268dbdbe403d2
With RESTbase we'll be able to retry the older one, but right now we have to do
both with the same request.
This commit retries loading once, then starts emitting errors.
Bug: T92993
Change-Id: Ifbf44b777279dafbe31ab438838d70acd44de726
Due to changes in the way VisualEditor is loaded, the trace.activate.{enter,exit}
events no longer enclose the entire VisualEditor initialization process. This
change ensures that trace.activate.enter is emitted as soon as the user clicks
on an edit link / tab and that the trace.activate.exit is not emitted until the
toolbar has been fully activated.
Change-Id: Ief798faa95a58898b9ae4dcbbbd30506c5dbd9a7
On short pages which don't extend below the fold, the progress
bar should center within the #content element.
Change-Id: I0b99e42f5bab8177d3b4ca6dd372d6403ae9b136
When we moved to mw.Api, edit conflict errors stopped going via onSave and
started going to onSaveError instead. So the case to check for them in onSave
would not have worked - instead, they would go through onSaveError and get
picked up by our unknown error code, which just picks out error.info and shows
it, instead of showing our edit conflict screen.
Also return in a couple of other error cases where we probably should've been
(but not necessarily must've been... that didn't necessarily show to the user
if another error emit already caused saveDeferred to be rejected)
Change-Id: Iae7a66a8aa96ee777e9fa780005feeb429129d5f
We're currently rewriting "mwtiming.performance.system.domLoad" to
"timing.ve.undefined.performance.system.domLoad". The "undefined" comes from
the missing targetName property on the event. I'm not sure why it is missing,
but having it default to 'mwTarget' will Do the Right Thing™.
Task: T93156
Change-Id: If70ff601b6c54ec8f95171cbc43c82c87a177508
Relies on I69cf0a88
Using our own new message because TitleBlacklist's own one includes text that
we can't parse on the client and relies on a parameter we don't get from the
API.
This relies on WikimediaMessages' version of the Edit schema being updated at
the same time.
Change-Id: I4c75369b8b97973b72899bfaecbd5a996a440c68
Set 've-activated' as soon as edit is clicked, with a 've-loading' state.
This necessitates moving the relevant styles to mw.ViewPageTarget.init.css.
Change-Id: Ic9757cdbf63a2f72eda0dd03ff5588d79028ba0e
Move requestPageData from mw.Target to TargetLoader, call it
in init init, and pass the promise it returns into load()
via activate().
Bug: T90372
Change-Id: I828b8474e5a76b3d0d7d08735b4d865c29d2f820
This introduces TargetLoader, which manages plugins and RL modules
in a slightly more generic fashion so that Targets themselves don't
have to. This allows us to load all RL modules in one load.php
request, rather than first loading ViewPageTarget which then
loads the other modules.
TargetLoader loads in the bottom queue, so it will be loaded
as part of the main load.php request, but in VPT.init.js we
still have to wait for it with using() because it might not
have arrived yet. This also degrades gracefully on cached pages
where TargetLoader isn't in the bottom queue: it'll be loaded
as a separate request instead, which is suboptimal but no
worse that what we were doing before.
Right now TargetLoader is small enough that it could also be in
the top queue, but in the future we want to add things like
the action=visualeditor API request to it, and mw.Api is
relatively big.
Note: this also makes a breaking change to the plugin API:
plugin callbacks no longer receive the target instance
as a parameter, as they're now executed before the target
has been constructed rather than after. In the long term,
if we want to give plugins access to the target instance,
we could give them the target promise somehow. For now,
I've killed this feature because nothing used it and
the change from a direct object reference to a promise
would have been a breaking change anyway.
Also fixed incorrect documentation index for ve.init.mw.ViewPageTarget.init.
Bug: T53569
Change-Id: Ibfa6abbeaf872ae2aadc6ed9d5beba7473ea441a
ve-ce-surface is a child of ve-ui-surface. The ve-ui-surface has
margins that would remain in the document flow causing a temporary
growth outwards to the bottom during deactivation (causing the
white area and blue border on Vector's grey background to move down)
after which it would restore again.
Follows-up 77f016a.
Bug: T91442
Change-Id: I5b999b580c968fcd24f07d9a895cfa17afc80c0d
It seems to have been intended to allow overriding of which icons
are used, but this isn't used anywhere.
Change-Id: I312f6c8e69d5a4d9c11f4af5f9487d0890a1f4e1
Move the surface focus() call to be done after VE has loaded, to
make sure that the position of the content editable field is at the
top of the page and not -- as happens in Firefox -- after the read
page contents.
Bug: T90420
Change-Id: If91cea42c083d67b1ee6396402c3f607dde70471
Depends on Ib9471bc0 in VisualEditor core. Without that patch
this is actually necessary, because we were removing some event
handlers in the meantime.
Change-Id: I145f1891efd1c91eeb6154e11e96e3fc160e2be3
While showing the content earlier as soon as possible is nice, the ve-ce-surface
remained visible until toolbar and surface were torn down.
This avoids the split second where the new pageview content *and*
CE surface were visible in the page (this it also messed with the scroll position,
and would temporarily cause a scrollbar to appear).
Bug: T91442
Change-Id: I0a7232cd0264bff28ad66aa328de29d339891aa1
The #tryWithPreparedCacheKey method was assuming that the request
could only fail because of invalid cache key. It can also fail because
of a bad token, and probably for a number of other reasons.
So only retry only once, and then fail. If it's a 'badtoken' error
during save, the caller will handle it and retry with a working token.
If it's something else, who knows, but we don't go into infinite loop
at least.
Failing in a way that will be handled is weird because 8e48f945 changed
the signatures of the promises, but not the functions that use them.
This must be fixed later.
Bug: T91158
Change-Id: I103cf888d339b44e3fd4fe2376edf5e37ce4157f
New changes:
a65ad7c Move special character inserter to toolbar dialog manager
… and add it to the toolbar as a terminal option.
Change-Id: I35834d866a13c5dea7f5a520c63b8a99451fcf6d
We used to attempt to not break Firefox's bfcache, but this
didn't really work very well, and it's not clear that avoiding
bfcache breakage is even a good idea. It's more sensible and
consistent to deliberately break bfcache while VE is active.
On top of that, the Firefox people believe that our trick
shouldn't even have worked to begin with:
https://bugzilla.mozilla.org/show_bug.cgi?id=1102664
Because the onbeforeunload handler is removed when VE
is deactivated, bfcache still works if you first click
Read, then navigate away.
Also clean up the management of the unload handler, using
addEventListener() and removing the return value fallback
code that is needed for beforeunload but not unload.
For beforeunload this is harder to clean up because
the addEventListener()-based API for returning a value
isn't consistent across browsers.
Change-Id: Ie4fe9ea3a59a54ba462733aa5e42bfc0ed5b15eb
Still misses some preinit aborts because we need to figure out
a way to attach the unload handler early enough to catch these.
Change-Id: I0ce721e24e69c31318064c6b443c1bfe01077546
