Bump to v2026.9.26

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

backend/app/main.py

@@ -58,8 +58,12 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     # ASGITransport, used by the plain REST test fixtures -- only
     # TestClient's websocket_connect-based tests actually run lifespan).
     # The background listener task genuinely needs a running event loop
-    # though, so that part stays here.
+    # though, so that part stays here. run_forever() (not listen() directly)
-    listener_task = asyncio.create_task(app.state.broadcaster.listen())
+    # -- #76: a bare listen() here died permanently and silently the moment
+    # its Redis connection ever dropped (e.g. Redis itself restarting),
+    # with nothing to notice or reconnect; run_forever() supervises it and
+    # restarts on failure instead.
+    listener_task = asyncio.create_task(app.state.broadcaster.run_forever())
 
     yield
 

backend/app/ws/broadcaster.py

@@ -1,13 +1,26 @@
+import asyncio
+import contextlib
 import json
+import logging
 import uuid
 
 from redis.asyncio import Redis
 
 from app.ws.connection_manager import ConnectionManager
 
+logger = logging.getLogger(__name__)
+
 ROOM_CHANNEL_PREFIX = "room:"
 USER_CHANNEL_PREFIX = "user:"
 
+# #76: a run shorter than this before dying doesn't count as "stable" for
+# resetting the backoff in run_forever() below -- long enough that a real,
+# working connection has clearly settled in, short enough not to leave a
+# just-recovered listener exposed to a full backoff climb if Redis blips
+# again shortly after reconnecting.
+STABLE_RUN_SECONDS = 60
+MAX_BACKOFF_SECONDS = 30
+
 
 class Broadcaster:
     """Cross-instance message fan-out (ARCHITECTURE.md phase 5).
@@ -36,6 +49,43 @@ class Broadcaster:
     async def publish_to_user(self, user_id: uuid.UUID, payload: dict) -> None:
         await self._redis.publish(f"{USER_CHANNEL_PREFIX}{user_id}", json.dumps(payload))
 
+    async def run_forever(self) -> None:
+        """Supervises listen(), restarting it if it ever dies.
+
+        #76's real root cause, found after the per-message try/except below
+        turned out not to be enough: this is the *one* pubsub listener for
+        the whole process (see main.py's lifespan), created once at startup
+        with nothing watching over it afterward. A connection-level failure
+        -- Redis restarting, a network blip between the app and Redis --
+        raises from the `async for` iteration in listen() itself, outside
+        that per-message guard, and previously killed this task permanently
+        and silently: HTTP/REST and login kept working fine (neither
+        touches this), but live message delivery was gone for every room
+        and every user on this instance until the whole process was
+        restarted. Confirmed live in production: Redis got restarted by an
+        OS update, and `PUBSUB NUMPAT` against it afterward showed zero
+        pattern subscriptions across all 4 gunicorn workers -- every one of
+        them had silently lost its listener with nothing to bring it back.
+        """
+        backoff_seconds = 1.0
+        while True:
+            started_at = asyncio.get_event_loop().time()
+            try:
+                await self.listen()
+                logger.warning("Broadcaster.listen() exited without error -- restarting")
+            except asyncio.CancelledError:
+                raise
+            except Exception:
+                logger.error("Broadcaster.listen() failed -- restarting", exc_info=True)
+            ran_for = asyncio.get_event_loop().time() - started_at
+            # Sleep at the *current* backoff first, then grow it for next
+            # time -- computing the next value before sleeping would skip
+            # straight past the first (shortest) retry delay.
+            await asyncio.sleep(backoff_seconds)
+            backoff_seconds = (
+                1.0 if ran_for > STABLE_RUN_SECONDS else min(backoff_seconds * 2, MAX_BACKOFF_SECONDS)
+            )
+
     async def listen(self) -> None:
         pubsub = self._redis.pubsub()
         await pubsub.psubscribe(f"{ROOM_CHANNEL_PREFIX}*", f"{USER_CHANNEL_PREFIX}*")
@@ -43,14 +93,37 @@ class Broadcaster:
             async for message in pubsub.listen():
                 if message["type"] != "pmessage":
                     continue
-                channel = message["channel"]
+                # #76: this is the one pubsub listener for the whole process
-                payload = json.loads(message["data"])
+                # (see main.py's lifespan) -- ConnectionManager.broadcast/
-                if channel.startswith(ROOM_CHANNEL_PREFIX):
+                # send_to_user already guard against one dead socket taking
-                    room_id = uuid.UUID(channel.removeprefix(ROOM_CHANNEL_PREFIX))
+                # the rest of a single fan-out down, but this belt-and-
-                    await self._manager.broadcast(room_id, payload)
+                # suspenders catch is for anything else unexpected (a
-                elif channel.startswith(USER_CHANNEL_PREFIX):
+                # malformed payload, e.g.) doing the same. An uncaught
-                    user_id = uuid.UUID(channel.removeprefix(USER_CHANNEL_PREFIX))
+                # exception here previously meant *this whole async for*
-                    await self._manager.send_to_user(user_id, payload)
+                # loop died silently -- no more live delivery to any room on
+                # this instance, for anyone, until the process restarted.
+                try:
+                    channel = message["channel"]
+                    payload = json.loads(message["data"])
+                    if channel.startswith(ROOM_CHANNEL_PREFIX):
+                        room_id = uuid.UUID(channel.removeprefix(ROOM_CHANNEL_PREFIX))
+                        await self._manager.broadcast(room_id, payload)
+                    elif channel.startswith(USER_CHANNEL_PREFIX):
+                        user_id = uuid.UUID(channel.removeprefix(USER_CHANNEL_PREFIX))
+                        await self._manager.send_to_user(user_id, payload)
+                except Exception:
+                    logger.error("Error handling pubsub message on channel %s", message.get("channel"), exc_info=True)
         finally:
-            await pubsub.punsubscribe(f"{ROOM_CHANNEL_PREFIX}*", f"{USER_CHANNEL_PREFIX}*")
+            # #76: reached via a connection that's already dead just as
-            await pubsub.aclose()
+            # often as a clean shutdown (that's the whole failure mode
+            # run_forever() now recovers from) -- punsubscribe/aclose over
+            # a broken connection would themselves raise, masking whatever
+            # actually killed the loop above and, since run_forever() only
+            # starts backing off *after* listen() returns/raises, delaying
+            # the reconnect for no benefit (a fresh pubsub object is about
+            # to be created on retry regardless of whether this cleanup
+            # succeeded).
+            with contextlib.suppress(Exception):
+                await pubsub.punsubscribe(f"{ROOM_CHANNEL_PREFIX}*", f"{USER_CHANNEL_PREFIX}*")
+            with contextlib.suppress(Exception):
+                await pubsub.aclose()

backend/app/ws/chat.py

@@ -1,3 +1,4 @@
+import asyncio
 import uuid
 
 from fastapi import APIRouter, Depends, WebSocket, WebSocketDisconnect
@@ -31,6 +32,18 @@ router = APIRouter(tags=["ws"])
 
 WS_UNAUTHENTICATED = 4401
 
+# #76: comfortably under the ~60s idle-connection timeout common to reverse
+# proxies/load balancers sitting in front of this endpoint, and short enough
+# that the frontend's own zombie-detection watchdog (useChatSocket.ts) can
+# notice a dead connection and reconnect well within a session -- a socket
+# can go silently dead (no close frame at all, e.g. a laptop sleeping or a
+# NAT dropping an idle mapping) without this, and no live message reaches it
+# again until *something* else finally triggers a close, which used to be
+# "switch rooms and back" purely by accident (that remounts the message
+# list, which refetches history over plain REST -- unrelated to the socket
+# itself actually recovering).
+WS_PING_INTERVAL_SECONDS = 30
+
 
 class ClientEnvelope(BaseModel):
     type: str
@@ -124,7 +137,18 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
 
     try:
         while True:
-            raw = await websocket.receive_json()
+            try:
+                raw = await asyncio.wait_for(
+                    websocket.receive_json(), timeout=WS_PING_INTERVAL_SECONDS
+                )
+            except asyncio.TimeoutError:
+                # Nothing received in a while -- ping rather than assume the
+                # connection's still good. A genuinely dead socket fails this
+                # send (raising, caught below same as any other disconnect);
+                # a merely quiet one just gets proded and the loop continues
+                # waiting.
+                await websocket.send_json({"type": "ping"})
+                continue
             try:
                 try:
                     envelope = ClientEnvelope.model_validate(raw)
@@ -146,6 +170,14 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
                     joined_rooms.add(envelope.room_id)
                     await websocket.send_json({"type": "joined", "room_id": str(envelope.room_id)})
 
+                elif envelope.type == "pong":
+                    # #76: the client's reply to our "ping" above -- nothing
+                    # to do with it. Its only purpose is resetting the
+                    # frontend's own liveness watchdog; the timeout on the
+                    # receive_json() above already gets reset by receiving
+                    # any frame at all, this one included.
+                    pass
+
                 elif envelope.type == "leave":
                     if envelope.room_id is None:
                         await websocket.send_json({"type": "error", "detail": "room_id required"})

backend/app/ws/connection_manager.py

@@ -1,8 +1,11 @@
+import logging
 import uuid
 from collections import defaultdict
 
 from fastapi import WebSocket
 
+logger = logging.getLogger(__name__)
+
 
 class ConnectionManager:
     """Local, single-process WebSocket socket registry.
@@ -43,8 +46,22 @@ class ConnectionManager:
 
     async def broadcast(self, room_id: uuid.UUID, payload: dict) -> None:
         for websocket in list(self._rooms.get(room_id, ())):
-            await websocket.send_json(payload)
+            # #76: one dead/racing socket (e.g. its own disconnect cleanup
+            # hasn't finished removing it from _rooms yet) must not take the
+            # rest of this broadcast down with it -- this runs inside
+            # Broadcaster.listen()'s single, never-restarted per-process
+            # pubsub loop, so an unhandled exception here previously meant
+            # one bad socket silently killed *live delivery for every room
+            # on this instance*, not just the one socket, until the process
+            # was restarted.
+            try:
+                await websocket.send_json(payload)
+            except Exception:
+                logger.warning("Dropping dead socket during broadcast to room %s", room_id, exc_info=True)
 
     async def send_to_user(self, user_id: uuid.UUID, payload: dict) -> None:
         for websocket in list(self._users.get(user_id, ())):
-            await websocket.send_json(payload)
+            try:
+                await websocket.send_json(payload)
+            except Exception:
+                logger.warning("Dropping dead socket sending to user %s", user_id, exc_info=True)

backend/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ds-chat"
-version = "2026.9.4"
+version = "2026.9.26"
 description = "DS Chat backend service"
 license = { text = "AGPL-3.0-or-later" }
 requires-python = ">=3.11"

backend/tests/test_broadcaster_supervision.py

@@ -0,0 +1,61 @@
+import asyncio
+
+import pytest
+
+from app.ws.broadcaster import Broadcaster
+
+
+async def test_run_forever_restarts_after_listen_raises(monkeypatch):
+    # #76: listen() dying (a dropped Redis connection, most commonly -- e.g.
+    # Redis itself restarting) must not permanently kill live delivery for
+    # the whole process. run_forever() has to notice and restart it, not
+    # just let the background task disappear silently while everything else
+    # (login, REST endpoints) keeps working fine.
+    broadcaster = Broadcaster(redis=None, manager=None)
+    calls = 0
+
+    async def fake_listen():
+        nonlocal calls
+        calls += 1
+        if calls < 3:
+            raise ConnectionError("redis went away")
+        # Stop the loop once the retry behavior below is proven, rather
+        # than looping forever in a test.
+        raise asyncio.CancelledError()
+
+    monkeypatch.setattr(broadcaster, "listen", fake_listen)
+    sleep_calls = []
+
+    async def fake_sleep(seconds):
+        sleep_calls.append(seconds)
+
+    monkeypatch.setattr(asyncio, "sleep", fake_sleep)
+
+    with pytest.raises(asyncio.CancelledError):
+        await broadcaster.run_forever()
+
+    assert calls == 3
+    # Slept once after each of the first two failures (exponential:
+    # 1s, then 2s) -- not after the third, which is the CancelledError
+    # that stops the loop before reaching the sleep.
+    assert sleep_calls == [1.0, 2.0]
+
+
+async def test_run_forever_does_not_swallow_cancellation(monkeypatch):
+    # A genuine shutdown (the app's own lifespan cancels this task) must
+    # propagate immediately, not get treated as just another failure to
+    # retry after a backoff.
+    broadcaster = Broadcaster(redis=None, manager=None)
+
+    async def fake_listen():
+        raise asyncio.CancelledError()
+
+    monkeypatch.setattr(broadcaster, "listen", fake_listen)
+
+    async def fail_if_called(seconds):
+        raise AssertionError("should not sleep/retry on cancellation")
+
+    monkeypatch.setattr(asyncio, "sleep", fail_if_called)
+
+    with pytest.raises(asyncio.CancelledError):
+        await broadcaster.run_forever()

backend/tests/test_ws_chat.py

@@ -109,6 +109,24 @@ def test_history_tiebreaks_identical_timestamps_by_id(ws_client):
     assert [m["content"] for m in history] == [m["content"] for m in expected_order]
 
 
+def test_ws_sends_heartbeat_ping(ws_client, monkeypatch):
+    # #76: with nothing else sent on the connection, a "ping" must arrive
+    # proactively once the idle timeout elapses -- shrunk here so the test
+    # doesn't actually wait the real 30s. Replying "pong" (the frontend's
+    # own behavior) must be accepted silently and leave the connection
+    # fully usable afterward, not treated as an unknown message type.
+    monkeypatch.setattr("app.ws.chat.WS_PING_INTERVAL_SECONDS", 0.05)
+    username = _unique("alice")
+    _register(ws_client, username=username)
+    room = ws_client.post("/api/rooms", json={"name": _unique("general")}).json()
+
+    with ws_client.websocket_connect("/ws/chat") as ws:
+        assert ws.receive_json() == {"type": "ping"}
+        ws.send_json({"type": "pong"})
+        ws.send_json({"type": "join", "room_id": room["id"]})
+        assert ws.receive_json() == {"type": "joined", "room_id": room["id"]}
+
+
 def test_ws_message_without_join_errors(ws_client):
     _register(ws_client, username=_unique("alice"))
     room = ws_client.post("/api/rooms", json={"name": _unique("general")}).json()

frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "frontend",
   "private": true,
-  "version": "2026.9.4",
+  "version": "2026.9.26",
   "license": "AGPL-3.0-or-later",
   "type": "module",
   "scripts": {

frontend/src/ws/useChatSocket.ts

@@ -10,6 +10,13 @@ interface UseChatSocketOptions {
 const RECONNECT_BASE_DELAY_MS = 1000
 const RECONNECT_MAX_DELAY_MS = 30000
 
+// #76: the backend pings every 30s (WS_PING_INTERVAL_SECONDS in chat.py) --
+// this needs to comfortably tolerate one missed ping plus ordinary network
+// jitter before declaring the connection dead, not fire on the very first
+// late beat.
+const ZOMBIE_TIMEOUT_MS = 65000
+const ZOMBIE_CHECK_INTERVAL_MS = 10000
+
 const desktopMode = isDesktopNotificationsSupported()
 
 // Room join/leave (live message delivery) depends only on visibility --
@@ -51,6 +58,16 @@ export function useChatSocket({ onUnauthenticated }: UseChatSocketOptions) {
   // live WebSocket delivery the user can't actually see will do the job.
   const desiredRoomsRef = useRef(new Set<string>())
   const isVisibleRef = useRef(isVisible())
+  // #76: last time *anything* arrived on the socket (a real message, a
+  // "ping", doesn't matter which) -- the only signal that a connection
+  // sitting at readyState OPEN is actually still alive rather than a
+  // zombie. Without this, a connection that dies without ever sending a
+  // close frame (laptop sleep, a NAT silently dropping an idle mapping)
+  // looks identical to a healthy-but-quiet one: onclose never fires, so the
+  // already-correct reconnect+rejoin logic below never runs, and new
+  // messages just stop arriving until something unrelated (e.g. switching
+  // rooms, which refetches history over plain REST) papers over it.
+  const lastActivityRef = useRef(Date.now())
 
   const sendRoomFrame = useCallback((type: 'join' | 'leave', roomId: string) => {
     const ws = socketRef.current
@@ -99,6 +116,7 @@ export function useChatSocket({ onUnauthenticated }: UseChatSocketOptions) {
         if (hasConnectedBefore) checkForUpdate()
         hasConnectedBefore = true
         reconnectDelay = RECONNECT_BASE_DELAY_MS
+        lastActivityRef.current = Date.now()
         setConnected(true)
         // Re-join whatever rooms were joined before a reconnect -- the
         // server has no memory of a dropped connection's prior state. Only
@@ -122,7 +140,17 @@ export function useChatSocket({ onUnauthenticated }: UseChatSocketOptions) {
 
       ws.onmessage = (event) => {
         if (socketRef.current !== ws) return
-        const envelope = JSON.parse(event.data) as ServerEnvelope
+        lastActivityRef.current = Date.now()
+        const parsed = JSON.parse(event.data)
+        // #76: the server's heartbeat -- purely transport-level, not a
+        // domain event, so it's answered and swallowed right here instead
+        // of being forwarded to subscribers (ServerEnvelope's own type
+        // doesn't include it for exactly that reason).
+        if (parsed.type === 'ping') {
+          ws.send(JSON.stringify({ type: 'pong' }))
+          return
+        }
+        const envelope = parsed as ServerEnvelope
         for (const handler of subscribersRef.current) handler(envelope)
       }
 
@@ -143,10 +171,9 @@ export function useChatSocket({ onUnauthenticated }: UseChatSocketOptions) {
         if (stopped) return
 
         // Unexpected close -- a deploy restarting the app server, a brief
-        // network blip, or (absent any app-level ping/pong) an idle
+        // network blip, or an idle connection getting recycled by a reverse
-        // connection getting recycled by a reverse proxy in front of it.
+        // proxy in front of it. Retry with exponential backoff instead of
-        // Retry with exponential backoff instead of leaving the chat
+        // leaving the chat silently dead until the user manually reloads.
-        // silently dead until the user manually reloads.
         reconnectTimer = setTimeout(connect, reconnectDelay)
         reconnectDelay = Math.min(reconnectDelay * 2, RECONNECT_MAX_DELAY_MS)
       }
@@ -154,9 +181,21 @@ export function useChatSocket({ onUnauthenticated }: UseChatSocketOptions) {
 
     connect()
 
+    // #76: catches the case onclose can't -- a connection that dies without
+    // ever sending a close frame at all (readyState stays OPEN forever) is
+    // otherwise invisible to this hook. Forcing a close here just routes it
+    // through the exact same onclose/reconnect/rejoin path as a normal
+    // disconnect, rather than needing separate recovery logic of its own.
+    const zombieCheckTimer = setInterval(() => {
+      if (Date.now() - lastActivityRef.current > ZOMBIE_TIMEOUT_MS) {
+        socketRef.current?.close()
+      }
+    }, ZOMBIE_CHECK_INTERVAL_MS)
+
     return () => {
       stopped = true
       if (reconnectTimer) clearTimeout(reconnectTimer)
+      clearInterval(zombieCheckTimer)
       socketRef.current?.close()
       socketRef.current = null
     }