In what I hoped would be a five-minute refactor to help clean up an annoying
not-really-used codepath, but turned into a bit of a hackathon on the tests,
create Olm sessions lazily in Olm rooms, just as we do in megolm rooms, which
allows us to avoid having to get the member list before configuring e2e in a
room.
When we got a redaction event, we were adding the entire (circular) MatrixEvent
object for the redaction to the redacted event, which would then cause
exceptions down the line (particularly when dealing with gappy timelines).
We should only be adding the raw event.
Fixes (hopefully) https://github.com/vector-im/vector-web/issues/1389.
Use another-json instead of awful manual json building. Sign the device keys at
the point of upload, instead of having to keep the signed string in
memory. Only upload device keys once (they are correctly merged with the
one-time keys by synapse).
Starts work on a class which is intended to just wrap the Matrix apis with very
simple functions.
There is a lot more work to be done here. For now, I have just taken methods
which don't refer to anything in MatrixClient except _http. This excludes a
bunch of things which refer to $userId, as well as the login stuff because of
the deviceId stuff I've just added :/.
For now, it's an internal class. I don't really see any reason it can't be
exposed to applications, though.
When we send an encrypted message with Olm, we need to know who to send it
to. Currently that is based on a user list passed to setRoomEncryption. We want
to be able to change the list as people join and leave the room; I also want to
not have to keep the member list in local storage.
So, use the member list at the point of enabling encryption to set up sessions,
and at the point of sending a message to encrypt the message.
Further work here includes:
* updating the react-sdk not to bother setting the member list
* monitoring for new users/devices in the room, and setting up new sessions
with them
Separate the helper functions from the tests, and order them by functionality
rather than by test order.
I've been struggling to find the tests among the helpers, and struggling to
find the helpers among the other helpers. Hopefully this will help.
Make `MatrixEvent.event` contain the *encrypted* event, and keep the plaintext
payload in a separate `_clearEvent` property.
This achieves several aims:
* means that we have a record of the actual object which was received over
the wire, which can be shown by clients for 'view source'.
* makes sent and received encrypted MatrixEvents more consistent (previously
sent ones had `encryptedType` and `encryptedContent` properties, whereas
received ones threw away the ciphertext).
* Avoids having to make two MatrixEvents in the receive path, and copying
fields from one to the other.
*Hopefully* most clients have followed the advice given in the jsdoc of not
relying on MatrixEvent.event to get at events. If they haven't, I guess they'll
break in the face of encrypted events.
(The pushprocessor didn't follow this advice, so needed some tweaks.)
If an exception was thrown by the encryption process, the event would be
queued, but the exception would not be handled. This meant that the event got
stuck as a grey 'sending' event in the UI.
Fixing this correctly is slightly more complex than just handling the exception
correctly. A naive approach would mean that the event would be shown as a red
'unsent' message, and clicking 'resend' would then send the message *in the
clear*. Hence, move the encryption to _sendEvent, where it will be called again
when the event is resent.
Add a 'verified' property to the response from MatrixClient.listDeviceKeys, and
add MatrixClient.setDeviceVerified to set it. Also changes the format of data
stored for user devices in the session store slightly (in a
backwards-compatible way).
Use different mock http backends for the two different clients, so that we can
better control what each of them is doing (in particular, this is a
prerequisite for having them both running /sync loops)
The pyramid of doom was getting unmanageable, not to mention the difficulty of
diagnosing why tests were failing, so replace the callbacks with promises.
Fix a bug in the error handling in downloadKeys: If the http request failed,
then the exception would get silently swallowed and the promise would never
resolve.
Also: tests!
... and add some sanity checks
Two things here:
1. Clean up the Room API for adding new events to the timeline. Where before
we had addEvents and addEventsToTimeline, whose purposes were unclear, we now
have addLiveEvents which must be used for adding events to the end of the live
timeline, and addEventsToTimeline which should be used for pagination (either
back-pagination of the live timeline, or pagination of an old timeline).
2. Add some sanity checks for the live timeline. Today we have seen problems
where somehow the live timeline had gained a forward pagination token, or the
live timeline had got joined to another timeline, leading to much confusion -
and I would like to notice these sooner.
In TimelineWindow.paginate, keep a count of the number of API requests we have
made, and bail out if it gets too high, to ensure that we don't get stuck in a
loop of paginating right back to the start of the room.
Hopefully this will improve our recovery time after a laptop is suspended. The
idea is to treat the timeouts on the http apis as being in realtime, rather
than in elapsed time while the machine is awake.
To do this, we add a layer on top of window.setTimeout. We run a callback every
second, which then checks the wallclock time and runs any pending callbacks.
David Baker