MSC3773: Notifications for threads (#3773)

* Add initial MSC for notfications for threads.

* No longer depends on MSC3664.

* Fix events in diagram.

* Major simplification.

* Add missing words.

* Add a note about reactions.

* Clarify homeserver behavior when processing push rules.

* Link to a spec version instead of latest.

Co-authored-by: Travis Ralston <[email protected]>

* Update 3773-notifications-for-threads.md

typoe

* Clarify wording.

* Update how homeservers are to partition threads.

Co-authored-by: Travis Ralston <[email protected]>
Co-authored-by: Matthew Hodgson <[email protected]>

Author: 3 people

Diff for: proposals/3773-notifications-for-threads.md

@@ -0,0 +1,162 @@
+# MSC3773: Notifications for threads
+
+Since the unread notification count does not consider threads, a client is unable
+to separate the unread message counts into threads (as defined by
+[MSC3440](https://github.com/matrix-org/matrix-doc/pull/3440))
+without iterating over every missing message. Without this, clients are unable to:
+
+* Let users know that a thread has new messages since they last read it.
+* Accurately display a count of unread messages in a room (or a thread).
+
+## Proposal
+
+### Modification to push rule processing
+
+When an event which is part of a thread matches a push rule which results in a
+`notify` action then the homeserver should partition the resulting notification
+count per-thread. (This is needed for the
+[proposed `/sync` changes](#unread-thread-notifications-in-the-sync-response)).
+
+It is recommended that at least 3 relations are traversed when attempting to
+find a thread, implementations should be careful to not infinitely recurse.[^1]
+
+Similar behavior should be applied for an event which results in `notify` action
+with a `highlight` tweak set.
+
+This MSC does not propose any changes to the payload sent to push gateways.
+
+### Unread thread notifications in the sync response
+
+Threaded clients can opt into receiving unread thread notifications by passing
+a new `unread_thread_notifications` parameter
+[as part of the `RoomEventFilter`](https://spec.matrix.org/v1.2/client-server-api/#filtering).
+(This is [similar to `lazy_load_members`](https://spec.matrix.org/v1.2/client-server-api/#lazy-loading-room-members),
+but only applies to the `/sync` endpoint.):
+
+* `unread_thread_notifications`:  If `true`, enables partitioning of unread notification
+  counts by thread. Defaults to false.
+
+If this flag is set to `true`, for each ["Joined Room" in the `/sync` response](https://spec.matrix.org/v1.3/client-server-api/#get_matrixclientv3sync)
+a new field is added:
+
+* `unread_thread_notifications`: Counts of unread thread notifications for this
+  room, an object which maps thread ID (the parent event ID) to
+  `Unread Notification Counts`.
+
+Additionally, the `unread_notifications` dictionary is modified to only include
+unread notifications from events which are not part of a thread.
+
+An example of a joined room from a sync response:
+
+```json5
+{
+  "account_data": {
+    // ...
+  },
+  "ephemeral": {
+    // ...
+  },
+  "state": {
+    // ...
+  },
+  "summary": {
+    // ...
+  },
+  "timeline": {
+    "events": [
+      {
+        "event_id": "$143273582443PhrSn:example.org",
+        // other fields ...
+      },
+      {
+        "event_id": "$SGNxGPGUopcPBUoTTL:example.org",
+        "m.relates_to": {
+          "event_id": "$143273582443PhrSn:example.org",
+          "rel_type": "m.thread"
+        }
+        // other fields ...
+      }
+    ]
+  },
+  "unread_notifications": {
+    "highlight_count": 2,
+    "notification_count": 18
+  },
+  "unread_thread_notifications": {
+    "$143273582443PhrSn:example.org": {
+      "highlight_count": 0,
+      "notification_count": 1
+    }
+  }
+}
+```
+
+## Potential issues
+
+### Scalability
+
+Rooms with many unread threads could cause some downsides:
+
+* The size of the `/sync` response would increase without bound.
+* The effort to generate and process the receipts for each room would increase
+  without bound.
+
+This is not dissimilar to rooms which are never read, however, as their unread
+counts are continually tracked and returned as part of the `/sync` response.
+
+### Clearing unread notifications
+
+This MSC does not attempt to modify how unread notifications (for a thread or
+otherwise) are cleared. It currently assumes the rules set forth by
+[read receipts](https://spec.matrix.org/v1.3/client-server-api/#receiving-notifications)
+still apply. This will cause some flakiness with unread notifications, as the current
+receipt infrastructure assumes that a room's timeline is linear, which is no
+longer true.
+
+[MSC3771](https://github.com/matrix-org/matrix-spec-proposals/pull/3771) is a
+potential solution for this.
+
+## Alternatives
+
+### Using push rules
+
+It might seem that a new push rule `action` (or `tweak`) should be used to control
+the behavior of whether an event generates a notification for a thread or the
+room itself. There are issues with either approach though:
+
+A new `action` (e.g. `notify_thread`) would mean that additional logic would
+need to be defined and added for events which aren't part of a thread but attempt
+to use this action. It also conflicts with [MSC3768](https://github.com/matrix-org/matrix-spec-proposals/pull/3768),
+which attempts to define another `action` which should also work fine for threads.
+
+A new `tweak` (e.g. `threaded`) was discarded as an option since there is no need to
+pass this through to the push server, which is at odds with the current `tweaks`
+mechanism.
+
+Regardless, the main issue with using push rules is that it becomes necessary to
+define rules which match threaded events. Whenever adding a new rule, matching rules
+would need to be added, but as a thread-specific version.
+
+## Security considerations
+
+N/A
+
+## Unstable prefix
+
+While this feature is in development the following unstable prefixes should be used:
+
+* `unread_thread_notifications` --> `org.matrix.msc3773.unread_thread_notifications`
+
+To detect server support, clients can either rely on the spec version (when stable)
+or the presence of a `org.matrix.msc3773` flag in `unstable_features` on `/versions`.
+
+## Dependencies
+
+N/A
+
+[^1]: Three relations is relatively arbitrary, but is meant to cover an edit or
+reaction to a thread (to an event with no relations, i.e. the root of a thread):
+`A<--[m.thread]--B<--[m.annotation]--C`.
+With an additional leftover for future improvements. This is considered reasonable
+since threads cannot fork, edits cannot modify relation information, and generally
+annotations to annotations are ignored by user interfaces.