MSC4103: Make threaded read receipts opt-in in /sync

proposals/4103-threaded-read-receipt-sync-filter.md

@@ -0,0 +1,75 @@
+# MSC4103: Make threaded read receipts opt-in in /sync
+
+## Problem
+
+When we added threaded RRs in [MSC3771](https://github.com/matrix-org/matrix-spec-proposals/pull/3771) we broke
+existing (i.e. non-thread-aware) clients from reliably being able to calculate whether a room is unread, given threaded
+read receipts are interpreted as unthreaded ones.
+
+For instance, let's assume Alice is logged in on two clients: a desktop client which is thread aware, and a mobile
+client which is not.
+
+ * Alice has a bunch of unread msgs on her main timeline and a bunch of unread threads.
+ * The threads are newer than the main timeline msgs.
+ * She reads a the most recently active thread on desktop, which sends a threaded RR.
+ * Mobile interprets this as a normal RR, and so marks the room as read (despite she hasn't read main timeline msgs)
+ * Desktop however correctly shows the room as still unread.
+
+The key thing is that **just because Alice happened to read a thread on a thread-aware client doesn't mean semantically
+that she's also read the main timeline**.
+
+Non-thread-aware clients simply do not care what messages users may have read on threads; they just care where the user
+has read up to in the main timeline (which for them, is a linearized view of the whole room, and for threaded clients
+is the main timeline).
+
+Another way of thinking of it which may or may not be helpful:
+
+ * If Alice reads a thread on desktop, then her mobile doesn't care.
+ * If Alice reads a main timeline msg on desktop, then her mobile will want to show the flattened timeline as read up to
+   that msg.
+ * If Alice reads a given msg on mobile, then she sends an unthreaded RR, which can mark all threads prior to that point
+   as read (for threaded clients). Non-thread-aware clients will be synced to the same RR point, as you'd expect.
+
+## Proposal
+
+Non-threade-aware clients should only act on unthreaded RRs.
+
+Prior to this MSC, there is no way for non-thread-aware clients to determine which RRs are unthreaded, other than seeing
+if the `thread_id` field is missing on the RR or present and equal to `main`.
+
+As a result, to get sensible room read state semantics for non-thread-aware clients, they would all need to become aware of
+the new `thread_id` field, breaking existing clients, even if they don't care about threads.  This feels contradictory and
+suboptimal.
+
+We could solve this by making threaded RRs a different event type (e.g. `m.read_thread`), so non-thread-aware clients
+would automatically ignore them.  However, this would require threaded clients to send double the read receipts whenever
+the user reads the main thread - both an `m.read` and an `m.read_thread`, which feels inefficient.  (It would however
+mean that [MSC4102](https://github.com/matrix-org/matrix-spec-proposals/pull/4102) would not be necessary, as the RR types
+would be clearly distinct).
+
+Instead, we propose that they continue to follow MSC3771 (i.e. putting a `thread_id` on the `m.read` RR).  However, we
+give threaded clients the ability to explicitly opt-in to threaded read receipts in the form of a new sync filter param.
+If clients explicitly include `threaded_read_receipts: true` on their sync filter, then the server will send them `m.read`
+events as received by the server.  If this filter is missing or false, then the server must only send the client `m.read` events
+whose `thread_id` is missing or `main`.
+
+This means that non-thread-aware clients are not spammed or confused with irrelevant threaded read receipts, and can
+calculate read state purely based on main timeline activity.
+
+## Alternatives
+
+We could split the event types into `m.read` and `m.read_thread` but instead have the server synthesise `m.read` events
+out of `m.read_thread` events with `thread_id: 'main'` when the `threaded_read_receipts` sync filter is false or
+absent. This would end up with the same end result, and would be cleaner in terms of type safety (and avoid the need
+for MSC4102), but would mean having to change every thread-aware client rather than just fixing it in servers, and back
+it out of the spec. So for pragmatism, this MSC proposes leaving MSC3771 as is, and instead let clients opt-in via sync
+filters to avoid breaking existing clients.
+
+## Security considerations
+
+If we ever encrypt RRs, then the server won't be able to filter them out.  But that's not planned currently, and we
+can always expose threadedness as metadata to aid this if it were to happen.
+
+## Unstable prefix
+
+`threaded_read_receipts` should be namespaced as `org.matrix.msc4013.threaded_read_receipts` while unstable.