diff --git a/.github/workflows/integration_selenium.yaml b/.github/workflows/integration_selenium.yaml
index c6ca7e437397..88212edd124e 100644
--- a/.github/workflows/integration_selenium.yaml
+++ b/.github/workflows/integration_selenium.yaml
@@ -19,6 +19,10 @@ env:
   YARN_INSTALL_OPTS: --frozen-lockfile
   GALAXY_CONFIG_SQLALCHEMY_WARN_20: '1'
   GALAXY_DEPENDENCIES_INSTALL_WEASYPRINT: '1'
+  # TEMP: shake down SSE/notification system across full UI surface — revert before merge
+  GALAXY_CONFIG_OVERRIDE_ENABLE_NOTIFICATION_SYSTEM: '1'
+  GALAXY_CONFIG_OVERRIDE_ENABLE_SSE_HISTORY_UPDATES: '1'
+  GALAXY_CONFIG_OVERRIDE_ENABLE_SSE_ENTRY_POINT_UPDATES: '1'
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
diff --git a/.github/workflows/playwright.yaml b/.github/workflows/playwright.yaml
index 163c83b38fd3..44fddab54988 100644
--- a/.github/workflows/playwright.yaml
+++ b/.github/workflows/playwright.yaml
@@ -20,6 +20,10 @@ env:
   GALAXY_TEST_SELENIUM_HEADLESS: 1
   YARN_INSTALL_OPTS: --frozen-lockfile
   GALAXY_CONFIG_SQLALCHEMY_WARN_20: '1'
+  # TEMP: shake down SSE/notification system across full UI surface — revert before merge
+  GALAXY_CONFIG_OVERRIDE_ENABLE_NOTIFICATION_SYSTEM: '1'
+  GALAXY_CONFIG_OVERRIDE_ENABLE_SSE_HISTORY_UPDATES: '1'
+  GALAXY_CONFIG_OVERRIDE_ENABLE_SSE_ENTRY_POINT_UPDATES: '1'
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
diff --git a/.github/workflows/selenium.yaml b/.github/workflows/selenium.yaml
index 60003fdf51ea..7f5c8490872d 100644
--- a/.github/workflows/selenium.yaml
+++ b/.github/workflows/selenium.yaml
@@ -19,6 +19,10 @@ env:
   GALAXY_TEST_SKIP_FLAKEY_TESTS_ON_ERROR: 1
   YARN_INSTALL_OPTS: --frozen-lockfile
   GALAXY_CONFIG_SQLALCHEMY_WARN_20: '1'
+  # TEMP: shake down SSE/notification system across full UI surface — revert before merge
+  GALAXY_CONFIG_OVERRIDE_ENABLE_NOTIFICATION_SYSTEM: '1'
+  GALAXY_CONFIG_OVERRIDE_ENABLE_SSE_HISTORY_UPDATES: '1'
+  GALAXY_CONFIG_OVERRIDE_ENABLE_SSE_ENTRY_POINT_UPDATES: '1'
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
diff --git a/client/src/api/client/pendingRequestsMiddleware.ts b/client/src/api/client/pendingRequestsMiddleware.ts
index 04eb4b2c11aa..5c1dec00f8fb 100644
--- a/client/src/api/client/pendingRequestsMiddleware.ts
+++ b/client/src/api/client/pendingRequestsMiddleware.ts
@@ -18,6 +18,8 @@ export const pendingRequestsMiddleware: Middleware = {
             return new Request(request, { headers });
         }
         const shared = getPendingAbortSignal();
+        // Combine with any signal the caller may have set so we don't silently
+        // drop their cancellation semantics.
         const signal = typeof AbortSignal.any === "function" ? AbortSignal.any([request.signal, shared]) : shared;
         return new Request(request, { signal });
     },
diff --git a/client/src/api/schema/schema.ts b/client/src/api/schema/schema.ts
index b171e21e0f37..09a570680d07 100644
--- a/client/src/api/schema/schema.ts
+++ b/client/src/api/schema/schema.ts
@@ -1292,6 +1292,33 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/api/events/stream": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        /**
+         * Server-Sent Events stream for real-time updates.
+         * @description Opens a Server-Sent Events (SSE) connection that pushes real-time
+         *     updates for notifications, history changes, and other events.
+         *
+         *     On reconnect, the browser sends the ``Last-Event-ID`` header automatically.
+         *     If the notification system is enabled, any notifications created since that
+         *     timestamp are delivered as a catch-up ``notification_status`` event.
+         *
+         *     Anonymous users receive only broadcast events.
+         */
+        get: operations["stream_events_api_events_stream_get"];
+        put?: never;
+        post?: never;
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/api/exports": {
         parameters: {
             query?: never;
@@ -33209,6 +33236,46 @@ export interface operations {
             };
         };
     };
+    stream_events_api_events_stream_get: {
+        parameters: {
+            query?: never;
+            header?: {
+                "Last-Event-ID"?: string | null;
+                /** @description The user ID that will be used to effectively make this API call. Only admins and designated users can make API calls on behalf of other users. */
+                "run-as"?: string | null;
+            };
+            path?: never;
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content?: never;
+            };
+            /** @description Request Error */
+            "4XX": {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["MessageExceptionModel"];
+                };
+            };
+            /** @description Server Error */
+            "5XX": {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["MessageExceptionModel"];
+                };
+            };
+        };
+    };
     index_api_exports_get: {
         parameters: {
             query?: {
diff --git a/client/src/composables/useAuthNavigation.ts b/client/src/composables/useAuthNavigation.ts
index 244d5a0012ef..69d7f88a00da 100644
--- a/client/src/composables/useAuthNavigation.ts
+++ b/client/src/composables/useAuthNavigation.ts
@@ -16,8 +16,9 @@ import { useNotificationsStore } from "@/stores/notificationsStore";
  * will use a fresh signal and is not affected.
  */
 export function discardActiveConnectionsBeforeAuthNavigation() {
-    // Stop polling watchers first so they can't kick off new fetches, then
-    // abort any requests still in flight via the shared AbortController.
+    // Order: close SSE streams first (synchronous TCP close), then stop the
+    // polling watchers so they can't kick off new fetches, then abort any
+    // requests still in flight via the shared AbortController.
     useHistoryStore().stopWatchingHistory();
     useEntryPointStore().stopWatchingEntryPoints();
     useNotificationsStore().stopWatchingNotifications();
diff --git a/client/src/composables/useNotificationSSE.ts b/client/src/composables/useNotificationSSE.ts
new file mode 100644
index 000000000000..8bdae964f2e7
--- /dev/null
+++ b/client/src/composables/useNotificationSSE.ts
@@ -0,0 +1,191 @@
+import { onScopeDispose, ref } from "vue";
+
+import { withPrefix } from "@/utils/redirect";
+
+/**
+ * All SSE event types the server may emit.
+ */
+export const SSE_EVENT_TYPES = [
+    "notification_update",
+    "broadcast_update",
+    "notification_status",
+    "history_update",
+    "entry_point_update",
+] as const;
+
+export type SSEEventType = (typeof SSE_EVENT_TYPES)[number];
+
+interface SSEDebugGlobals {
+    __galaxy_sse_connected?: boolean;
+    __galaxy_sse_last_event_ts?: number;
+}
+
+function sseGlobals(): SSEDebugGlobals {
+    return window as unknown as SSEDebugGlobals;
+}
+
+// ---------------------------------------------------------------------------
+// Module-level shared EventSource.
+//
+// Every call to ``useSSE`` registers its handler against this one socket so
+// the tab opens a single ``/api/events/stream`` connection no matter how many
+// stores listen. HTTP/1.1 caps simultaneous connections per origin at six;
+// before this consolidation we burned three slots on SSE alone (history,
+// notifications, entry points), which is what starved the scratchbook iframe
+// flow — see the fix in ``client/src/entry/analysis/App.vue``.
+// ---------------------------------------------------------------------------
+
+type Handler = (event: MessageEvent) => void;
+
+let sharedSource: EventSource | null = null;
+const sharedConnected = ref(false);
+const subscribers: Map<SSEEventType, Set<Handler>> = new Map();
+// Track the per-type dispatchers we registered so ``closeSource`` removes the
+// exact same listeners (``addEventListener`` matches by reference).
+const dispatchers: Map<SSEEventType, Handler> = new Map();
+
+function openSourceIfNeeded() {
+    if (sharedSource) {
+        return;
+    }
+    sharedSource = new EventSource(withPrefix("/api/events/stream"));
+
+    for (const eventType of SSE_EVENT_TYPES) {
+        const dispatcher: Handler = (event) => {
+            // Selenium tests watch ``__galaxy_sse_last_event_ts`` to prove that
+            // an observable state change came from an SSE push and not the
+            // polling fallback (where the global would never advance).
+            sseGlobals().__galaxy_sse_last_event_ts = Date.now();
+            const subs = subscribers.get(eventType);
+            if (!subs) {
+                return;
+            }
+            for (const handler of subs) {
+                handler(event);
+            }
+        };
+        dispatchers.set(eventType, dispatcher);
+        sharedSource.addEventListener(eventType, dispatcher);
+    }
+
+    sharedSource.onopen = () => {
+        sharedConnected.value = true;
+        // Global readiness flag so Selenium tests can distinguish a working
+        // SSE pipeline from the polling fallback.
+        sseGlobals().__galaxy_sse_connected = true;
+    };
+
+    sharedSource.onerror = () => {
+        // EventSource auto-reconnects natively; SSE-vs-polling is a
+        // config-level decision (see historyStore / notificationsStore), so
+        // we must not give up on transient errors here — doing so would leave
+        // the client with no updates at all.
+        sharedConnected.value = false;
+        sseGlobals().__galaxy_sse_connected = false;
+    };
+
+    // Browser EventSource teardown during a full-page navigation
+    // (``window.location.href = …``) is not guaranteed to happen before the
+    // browser issues requests for the new page — we've seen Chrome keep the
+    // stream alive long enough that a login/register POST reload races the
+    // close, and the new page then loads with a stale auth view. Force a
+    // synchronous ``close()`` during ``pagehide`` (fires for both reloads and
+    // tab-close, unlike ``beforeunload``) to close that window.
+    if (typeof window !== "undefined") {
+        window.addEventListener("pagehide", closeSource);
+    }
+}
+
+function closeSource() {
+    if (!sharedSource) {
+        return;
+    }
+    for (const [eventType, dispatcher] of dispatchers) {
+        sharedSource.removeEventListener(eventType, dispatcher);
+    }
+    dispatchers.clear();
+    sharedSource.close();
+    sharedSource = null;
+    sharedConnected.value = false;
+    sseGlobals().__galaxy_sse_connected = false;
+    if (typeof window !== "undefined") {
+        window.removeEventListener("pagehide", closeSource);
+    }
+}
+
+function addSubscriber(onEvent: Handler, eventTypes: readonly SSEEventType[]) {
+    for (const eventType of eventTypes) {
+        let subs = subscribers.get(eventType);
+        if (!subs) {
+            subs = new Set();
+            subscribers.set(eventType, subs);
+        }
+        subs.add(onEvent);
+    }
+}
+
+function removeSubscriber(onEvent: Handler, eventTypes: readonly SSEEventType[]): boolean {
+    let anyRemaining = false;
+    for (const eventType of eventTypes) {
+        const subs = subscribers.get(eventType);
+        if (subs) {
+            subs.delete(onEvent);
+            if (subs.size === 0) {
+                subscribers.delete(eventType);
+            }
+        }
+    }
+    for (const subs of subscribers.values()) {
+        if (subs.size > 0) {
+            anyRemaining = true;
+            break;
+        }
+    }
+    return anyRemaining;
+}
+
+/**
+ * Composable for subscribing to events on the shared SSE stream.
+ *
+ * The browser's EventSource handles reconnection automatically and sends the
+ * ``Last-Event-ID`` header so the server can catch up on missed events. Only
+ * one EventSource is opened per tab regardless of how many callers invoke
+ * this composable; the composable multiplexes dispatch per event type.
+ *
+ * @param onEvent - callback invoked for every matching SSE event
+ * @param eventTypes - subset of event types to listen to (defaults to all)
+ */
+export function useSSE(onEvent: Handler, eventTypes: readonly SSEEventType[] = SSE_EVENT_TYPES) {
+    let connected_: boolean = false;
+
+    function connect() {
+        if (connected_) {
+            return;
+        }
+        connected_ = true;
+        addSubscriber(onEvent, eventTypes);
+        openSourceIfNeeded();
+    }
+
+    function disconnect() {
+        if (!connected_) {
+            return;
+        }
+        connected_ = false;
+        const anyRemaining = removeSubscriber(onEvent, eventTypes);
+        if (!anyRemaining) {
+            closeSource();
+        }
+    }
+
+    onScopeDispose(() => {
+        disconnect();
+    });
+
+    return { connect, disconnect, connected: sharedConnected };
+}
+
+/**
+ * @deprecated Use `useSSE` instead. This alias exists for backward compatibility.
+ */
+export const useNotificationSSE = useSSE;
diff --git a/client/src/entry/analysis/App.vue b/client/src/entry/analysis/App.vue
index cbc80f181cde..64bd2a784a02 100644
--- a/client/src/entry/analysis/App.vue
+++ b/client/src/entry/analysis/App.vue
@@ -48,7 +48,7 @@
 </template>
 <script>
 import { storeToRefs } from "pinia";
-import { ref, watch } from "vue";
+import { computed, ref, watch } from "vue";
 import { useRoute } from "vue-router/composables";
 
 import { getGalaxyInstance } from "@/app";
@@ -105,9 +105,29 @@ export default {
         const uploadModal = ref(null);
         setGlobalUploadModal(uploadModal);
 
-        const embedded = useRouteQueryBool("embed");
+        // Treat any iframe context as embedded: scratchbook pops dataset
+        // displays into ``WinBox`` iframes that hit the same routes without
+        // an ``embed`` query param, and each one would otherwise open its own
+        // SSE + polling traffic, quickly saturating the HTTP/1.1 per-origin
+        // connection pool (e.g. ``test_scratchbook_window_persistence`` hangs
+        // indefinitely after two windows are open).
+        const inIframe = (() => {
+            if (typeof window === "undefined") {
+                return false;
+            }
+            try {
+                return window.top !== window.self;
+            } catch {
+                // Cross-origin access throws — that's definitely an iframe.
+                return true;
+            }
+        })();
+        const embeddedQuery = useRouteQueryBool("embed");
+        const embedded = computed(() => embeddedQuery.value || inIframe);
         const historyStore = useHistoryStore();
-        historyStore.startWatchingHistory();
+        if (!embedded.value) {
+            historyStore.startWatchingHistory();
+        }
 
         watch(
             () => embedded.value,
diff --git a/client/src/stores/_testing/sseStoreSupport.ts b/client/src/stores/_testing/sseStoreSupport.ts
new file mode 100644
index 000000000000..b3e5d97fcb89
--- /dev/null
+++ b/client/src/stores/_testing/sseStoreSupport.ts
@@ -0,0 +1,74 @@
+/**
+ * Shared test helpers for the SSE-driven stores (historyStore, notificationsStore).
+ *
+ * Both stores consume the same `useSSE` composable and need:
+ *  - a mock that captures the onEvent callback so tests can synthesize SSE messages;
+ *  - visibility-state patching without leaking across tests (JSDOM's `document`
+ *    is shared by every test in the same worker, so an unrestored
+ *    `Object.defineProperty` causes silent bleed).
+ *
+ * Because ``vi.mock`` is hoisted above module-level variables, tests must
+ * construct the SSE-mock state via ``vi.hoisted`` and then hand it to
+ * ``sseMockFactory`` from inside the ``vi.mock`` factory. See the ``.test.ts``
+ * files in this directory for the pattern.
+ */
+
+import { vi } from "vitest";
+import { type Ref, ref } from "vue";
+
+export interface SSEMockState {
+    onEvent: ((event: MessageEvent) => void) | null;
+    connect: ReturnType<typeof vi.fn>;
+    disconnect: ReturnType<typeof vi.fn>;
+    connected?: Ref<boolean>;
+}
+
+/** Build the factory used with ``vi.mock("@/composables/useNotificationSSE", ...)``. */
+export function sseMockFactory(state: SSEMockState) {
+    // Lazily initialize ``connected`` so existing callers that don't pass it
+    // still get a working ref.
+    if (!state.connected) {
+        state.connected = ref(false);
+    }
+    return {
+        useSSE: vi.fn((onEvent: (event: MessageEvent) => void) => {
+            state.onEvent = onEvent;
+            return { connect: state.connect, disconnect: state.disconnect, connected: state.connected };
+        }),
+    };
+}
+
+/** Synthesize an SSE message through the captured handler. */
+export function emitSse(state: SSEMockState, type: string, payload: unknown): void {
+    if (!state.onEvent) {
+        throw new Error("useSSE was not called by the store under test — cannot emit an SSE event");
+    }
+    state.onEvent(new MessageEvent(type, { data: JSON.stringify(payload) }));
+}
+
+/**
+ * Save the current ``document.visibilityState`` descriptor and return a restorer.
+ * Call the restorer in ``afterEach`` to prevent patching from leaking into later tests.
+ */
+export function useVisibilityPatch(): {
+    set: (state: "visible" | "hidden") => void;
+    restore: () => void;
+} {
+    const original = Object.getOwnPropertyDescriptor(document, "visibilityState");
+    return {
+        set(state: "visible" | "hidden") {
+            Object.defineProperty(document, "visibilityState", {
+                configurable: true,
+                get: () => state,
+            });
+            document.dispatchEvent(new Event("visibilitychange"));
+        },
+        restore() {
+            if (original) {
+                Object.defineProperty(document, "visibilityState", original);
+            } else {
+                delete (document as unknown as Record<string, unknown>).visibilityState;
+            }
+        },
+    };
+}
diff --git a/client/src/stores/entryPointStore.test.js b/client/src/stores/entryPointStore.test.js
index d0c6b69b4573..fdae2799d306 100644
--- a/client/src/stores/entryPointStore.test.js
+++ b/client/src/stores/entryPointStore.test.js
@@ -1,12 +1,25 @@
 import flushPromises from "flush-promises";
 import { createPinia, setActivePinia } from "pinia";
-import { beforeEach, describe, expect, it } from "vitest";
+import { beforeEach, describe, expect, it, vi } from "vitest";
 
 import { HttpResponse, useServerMock } from "@/api/client/__mocks__";
 
 import testInteractiveToolsResponse from "../components/InteractiveTools/testData/testInteractiveToolsResponse";
+import { sseMockFactory } from "./_testing/sseStoreSupport";
 import { useEntryPointStore } from "./entryPointStore";
 
+// ``vi.mock`` is hoisted above module-level declarations, so the capture-state
+// has to be built via ``vi.hoisted`` to be visible to the factory. Prevents
+// these tests from opening a real EventSource against ``/api/events/stream``
+// when ``useEntryPointStore()`` is invoked.
+const sseState = vi.hoisted(() => ({
+    onEvent: null,
+    connect: vi.fn(),
+    disconnect: vi.fn(),
+    connected: null,
+}));
+vi.mock("@/composables/useNotificationSSE", () => sseMockFactory(sseState));
+
 const { server, http } = useServerMock();
 
 describe("stores/EntryPointStore", () => {
diff --git a/client/src/stores/entryPointStore.ts b/client/src/stores/entryPointStore.ts
index 805124a713d5..2d0ee6c4ab7d 100644
--- a/client/src/stores/entryPointStore.ts
+++ b/client/src/stores/entryPointStore.ts
@@ -1,10 +1,12 @@
 import axios from "axios";
 import isEqual from "lodash.isequal";
 import { defineStore } from "pinia";
-import { computed, ref } from "vue";
+import { computed, ref, watch } from "vue";
 
 import { useResourceWatcher } from "@/composables/resourceWatcher";
+import { useSSE } from "@/composables/useNotificationSSE";
 import { getAppRoot } from "@/onload/loadConfig";
+import { useConfigStore } from "@/stores/configurationStore";
 import { rethrowSimple } from "@/utils/simple-error";
 
 const ACTIVE_POLLING_INTERVAL = 10000;
@@ -23,23 +25,8 @@ interface EntryPoint {
 }
 
 export const useEntryPointStore = defineStore("entryPointStore", () => {
-    const { startWatchingResource: startWatchingEntryPoints, stopWatchingResource: stopWatchingEntryPoints } =
-        useResourceWatcher(fetchEntryPoints, {
-            shortPollingInterval: ACTIVE_POLLING_INTERVAL,
-            enableBackgroundPolling: false, // No need to poll in the background
-        });
-
     const entryPoints = ref<EntryPoint[]>([]);
 
-    const entryPointsForJob = computed(() => {
-        return (jobId: string) => entryPoints.value.filter((entryPoint) => entryPoint["job_id"] === jobId);
-    });
-
-    const entryPointsForHda = computed(() => {
-        return (hdaId: string) =>
-            entryPoints.value.filter((entryPoint) => entryPoint["output_datasets_ids"].includes(hdaId));
-    });
-
     async function fetchEntryPoints() {
         const url = `${getAppRoot()}api/entry_points`;
         const params = { running: true };
@@ -51,6 +38,80 @@ export const useEntryPointStore = defineStore("entryPointStore", () => {
         }
     }
 
+    // SSE-driven path: on each entry_point_update signal, refetch the canonical
+    // list from REST. The event carries no data — it's a pure wake-up.
+    function handleEntryPointSSEEvent(_event: MessageEvent) {
+        fetchEntryPoints().catch((err) => console.error("Error refreshing entry points from SSE push:", err));
+    }
+    const {
+        connect: sseConnect,
+        disconnect: sseDisconnect,
+        connected: sseConnected,
+    } = useSSE(handleEntryPointSSEEvent, ["entry_point_update"]);
+    let stopPolling: (() => void) | null = null;
+    let stopConnectedWatcher: (() => void) | null = null;
+
+    let watchingInitialized = false;
+
+    // Callers opt in via ``startWatchingEntryPoints()`` (App.vue gates this on
+    // ``interactivetools_enable``). We then pick SSE or polling based on the
+    // server flag — mutually exclusive, mirroring historyStore / notificationsStore.
+    // ``useConfigStore`` is resolved lazily here so tests that only exercise
+    // the data methods don't need a ``/api/configuration`` handler registered.
+    function startWatchingEntryPoints() {
+        if (watchingInitialized) {
+            return;
+        }
+        watchingInitialized = true;
+        const configStore = useConfigStore();
+
+        const decide = () => {
+            if (configStore.config?.enable_sse_updates) {
+                // Baseline fetch + SSE. Reconnect-refetch closes the "user
+                // navigated away and missed events" window.
+                fetchEntryPoints().catch((err) => console.warn("Initial entry-point load failed", err));
+                sseConnect();
+                stopConnectedWatcher = watch(sseConnected, (isConnected, wasConnected) => {
+                    if (isConnected && !wasConnected) {
+                        fetchEntryPoints().catch((err) =>
+                            console.error("Error refreshing entry points on SSE reconnect:", err),
+                        );
+                    }
+                });
+            } else {
+                const { startWatchingResource, stopWatchingResource } = useResourceWatcher(fetchEntryPoints, {
+                    shortPollingInterval: ACTIVE_POLLING_INTERVAL,
+                    enableBackgroundPolling: false,
+                });
+                stopPolling = stopWatchingResource;
+                startWatchingResource();
+            }
+        };
+
+        if (configStore.isLoaded) {
+            decide();
+        } else {
+            const stop = watch(
+                () => configStore.isLoaded,
+                (loaded) => {
+                    if (loaded) {
+                        stop();
+                        decide();
+                    }
+                },
+            );
+        }
+    }
+
+    const entryPointsForJob = computed(() => {
+        return (jobId: string) => entryPoints.value.filter((entryPoint) => entryPoint["job_id"] === jobId);
+    });
+
+    const entryPointsForHda = computed(() => {
+        return (hdaId: string) =>
+            entryPoints.value.filter((entryPoint) => entryPoint["output_datasets_ids"].includes(hdaId));
+    });
+
     function updateEntryPoints(data: EntryPoint[]) {
         let hasChanged = entryPoints.value.length !== data.length ? true : false;
         if (entryPoints.value.length === 0) {
@@ -85,6 +146,22 @@ export const useEntryPointStore = defineStore("entryPointStore", () => {
         }
     }
 
+    // Closes the SSE stream and stops the polling watcher; paired with login
+    // /register flows so background traffic doesn't outlive the navigation
+    // and clobber the freshly authenticated session cookie.
+    function stopWatchingEntryPoints() {
+        sseDisconnect();
+        if (stopPolling) {
+            stopPolling();
+            stopPolling = null;
+        }
+        if (stopConnectedWatcher) {
+            stopConnectedWatcher();
+            stopConnectedWatcher = null;
+        }
+        watchingInitialized = false;
+    }
+
     return {
         entryPoints,
         entryPointsForJob,
diff --git a/client/src/stores/historyStore.test.ts b/client/src/stores/historyStore.test.ts
new file mode 100644
index 000000000000..bf04fe8fc0c8
--- /dev/null
+++ b/client/src/stores/historyStore.test.ts
@@ -0,0 +1,194 @@
+import flushPromises from "flush-promises";
+import { createPinia, setActivePinia } from "pinia";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+import { useServerMock } from "@/api/client/__mocks__";
+
+import { emitSse, sseMockFactory, useVisibilityPatch } from "./_testing/sseStoreSupport";
+import { useHistoryStore } from "./historyStore";
+
+// ``vi.mock`` is hoisted above module-level ``const`` declarations, so the
+// capture-state has to be built via ``vi.hoisted`` to be visible to the factory.
+const sseState = vi.hoisted(() => {
+    return {
+        onEvent: null as ((event: MessageEvent) => void) | null,
+        connect: vi.fn(),
+        disconnect: vi.fn(),
+    };
+});
+
+vi.mock("@/composables/useNotificationSSE", () => sseMockFactory(sseState));
+
+// `watchHistory(app)` is the polling handler invoked on the short/long
+// interval. We mock it so each invocation is observable without pulling in
+// the history-items store, dataset store, and Galaxy app instance.
+const mockWatchHistory = vi.fn().mockResolvedValue(undefined);
+const mockRefreshHistoryFromPush = vi.fn().mockResolvedValue(undefined);
+vi.mock("@/watch/watchHistory", () => ({
+    ACTIVE_POLLING_INTERVAL: 3000,
+    INACTIVE_POLLING_INTERVAL: 60_000,
+    watchHistory: (app: unknown) => mockWatchHistory(app),
+    refreshHistoryFromPush: (app: unknown) => mockRefreshHistoryFromPush(app),
+}));
+
+vi.mock("@/app", () => ({
+    getGalaxyInstance: () => ({ name: "fake-galaxy" }),
+}));
+
+const { server, http } = useServerMock();
+
+function registerDefaultHandlers({ enableSse }: { enableSse: boolean }) {
+    server.use(
+        http.get("/api/configuration", ({ response }) => {
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            return response(200).json({ enable_sse_updates: enableSse } as any);
+        }),
+    );
+}
+
+async function primeStore(startFn: () => void): Promise<void> {
+    startFn();
+    // Config load is async; let the watch fire and the initial fetch complete.
+    await flushPromises();
+    await vi.runOnlyPendingTimersAsync();
+    await flushPromises();
+}
+
+describe("historyStore — config-driven SSE vs polling", () => {
+    let visibility: ReturnType<typeof useVisibilityPatch>;
+
+    beforeEach(() => {
+        setActivePinia(createPinia());
+        sseState.connect.mockClear();
+        sseState.disconnect.mockClear();
+        sseState.onEvent = null;
+        mockWatchHistory.mockClear();
+        mockRefreshHistoryFromPush.mockClear();
+        vi.useFakeTimers();
+        visibility = useVisibilityPatch();
+    });
+
+    afterEach(() => {
+        visibility.restore();
+        vi.useRealTimers();
+    });
+
+    describe("when enable_sse_updates is true (SSE scenario)", () => {
+        beforeEach(() => {
+            registerDefaultHandlers({ enableSse: true });
+        });
+
+        it("primes the store with one initial load, connects SSE, and does not keep polling", async () => {
+            const store = useHistoryStore();
+            await primeStore(() => store.startWatchingHistory());
+
+            expect(sseState.connect).toHaveBeenCalledTimes(1);
+            // One-shot initial fetch so the history panel isn't empty before
+            // the first SSE event arrives.
+            expect(mockWatchHistory).toHaveBeenCalledTimes(1);
+
+            // Advance past the short polling interval (3s) several times and
+            // confirm the polling handler is not invoked a second time in SSE mode.
+            vi.advanceTimersByTime(30_000);
+            await flushPromises();
+            expect(mockWatchHistory).toHaveBeenCalledTimes(1);
+        });
+
+        it("does not start polling when the tab regains visibility", async () => {
+            const store = useHistoryStore();
+            await primeStore(() => store.startWatchingHistory());
+            expect(mockWatchHistory).toHaveBeenCalledTimes(1);
+
+            // Simulate a tab hide/show cycle. `useResourceWatcher` registers
+            // a `visibilitychange` listener whose handler calls
+            // `startWatchingResourceIfNeeded` — in SSE mode that would
+            // silently resume polling. Because we never instantiated the
+            // watcher, no listener should exist and no poll should fire.
+            visibility.set("hidden");
+            visibility.set("visible");
+
+            await flushPromises();
+            vi.advanceTimersByTime(30_000);
+            await flushPromises();
+            expect(mockWatchHistory).toHaveBeenCalledTimes(1);
+        });
+
+        it("triggers refreshHistoryFromPush when an SSE event names the current history", async () => {
+            // This test asserts the store's *decision* to refresh, not the refresh
+            // itself — ``refreshHistoryFromPush`` is mocked so we can observe the
+            // dispatch. The real refresh is covered end-to-end in the Selenium
+            // SSE integration tests (see test/integration_selenium/test_history_sse.py).
+            const store = useHistoryStore();
+            await primeStore(() => store.startWatchingHistory());
+            // Drive the store to a known current-history id so the handler has
+            // something to match against. ``currentHistoryId`` is a computed
+            // that only returns the stored id when the history is present in
+            // ``storedHistories``, so the history has to be registered too.
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            store.setHistory({ id: "hist-1" } as any);
+            store.setCurrentHistoryId("hist-1");
+
+            mockRefreshHistoryFromPush.mockClear();
+            emitSse(sseState, "history_update", { history_ids: ["hist-1", "hist-2"] });
+            await flushPromises();
+
+            expect(mockRefreshHistoryFromPush).toHaveBeenCalledTimes(1);
+        });
+
+        it("ignores SSE history events that do not include the current history", async () => {
+            const store = useHistoryStore();
+            await primeStore(() => store.startWatchingHistory());
+            store.setCurrentHistoryId("hist-1");
+
+            mockRefreshHistoryFromPush.mockClear();
+            emitSse(sseState, "history_update", { history_ids: ["hist-2"] });
+            await flushPromises();
+
+            expect(mockRefreshHistoryFromPush).not.toHaveBeenCalled();
+        });
+    });
+
+    describe("when enable_sse_updates is false (polling scenario)", () => {
+        beforeEach(() => {
+            registerDefaultHandlers({ enableSse: false });
+        });
+
+        it("does not connect SSE and polls on the configured interval", async () => {
+            const store = useHistoryStore();
+            await primeStore(() => store.startWatchingHistory());
+
+            expect(sseState.connect).not.toHaveBeenCalled();
+
+            // The resource watcher invokes the handler immediately on start
+            // and then re-schedules after each completion. Advance past the
+            // short interval and confirm repeated invocations.
+            const initialCalls = mockWatchHistory.mock.calls.length;
+            expect(initialCalls).toBeGreaterThanOrEqual(1);
+
+            await vi.advanceTimersByTimeAsync(3000);
+            await flushPromises();
+            expect(mockWatchHistory.mock.calls.length).toBeGreaterThan(initialCalls);
+        });
+
+        it("calling startWatchingHistory again is idempotent (no second SSE, polling tick count +1 only)", async () => {
+            const store = useHistoryStore();
+            await primeStore(() => store.startWatchingHistory());
+
+            const pollsAfterFirst = mockWatchHistory.mock.calls.length;
+
+            store.startWatchingHistory();
+            await flushPromises();
+
+            expect(sseState.connect).not.toHaveBeenCalled();
+            // Calling again must not schedule a second independent polling loop.
+            // Advance past one interval and confirm only one handler tick fires,
+            // not two.
+            await vi.advanceTimersByTimeAsync(3000);
+            await flushPromises();
+            // Exactly one additional poll after the 3000ms advance — anything
+            // else means a second independent polling loop was scheduled.
+            const deltaAfterSecond = mockWatchHistory.mock.calls.length - pollsAfterFirst;
+            expect(deltaAfterSecond).toBe(1);
+        });
+    });
+});
diff --git a/client/src/stores/historyStore.ts b/client/src/stores/historyStore.ts
index 019c57de8ae4..f0f6b73fc4b6 100644
--- a/client/src/stores/historyStore.ts
+++ b/client/src/stores/historyStore.ts
@@ -1,5 +1,5 @@
 import { defineStore } from "pinia";
-import { computed, del, ref, set } from "vue";
+import { computed, del, ref, set, watch } from "vue";
 
 import {
     type AnyHistory,
@@ -15,7 +15,9 @@ import type { ArchivedHistoryDetailed } from "@/api/histories.archived";
 import { getGalaxyInstance } from "@/app";
 import { HistoryFilters } from "@/components/History/HistoryFilters";
 import { useResourceWatcher } from "@/composables/resourceWatcher";
+import { useSSE } from "@/composables/useNotificationSSE";
 import { useUserLocalStorage } from "@/composables/userLocalStorage";
+import { useConfigStore } from "@/stores/configurationStore";
 import {
     createAndSelectNewHistory,
     getCurrentHistoryFromServer,
@@ -30,6 +32,7 @@ import { sortByObjectProp } from "@/utils/sorting";
 import {
     ACTIVE_POLLING_INTERVAL,
     INACTIVE_POLLING_INTERVAL,
+    refreshHistoryFromPush as refreshHistoryFromPushSuppliedApp,
     watchHistory as watchHistorySuppliedApp,
 } from "@/watch/watchHistory";
 
@@ -391,14 +394,94 @@ export const useHistoryStore = defineStore("historyStore", () => {
         return watchHistorySuppliedApp(app);
     }
 
-    const {
-        startWatchingResource: startWatchingHistory,
-        stopWatchingResource: stopWatchingHistory,
-        isWatchingResource: isWatchingHistory,
-    } = useResourceWatcher(watchHistory, {
-        shortPollingInterval: ACTIVE_POLLING_INTERVAL,
-        longPollingInterval: INACTIVE_POLLING_INTERVAL,
-    });
+    // SSE-driven history updates: when we receive a history_update event,
+    // immediately trigger a refresh of the current history
+    const SSE_HISTORY_EVENT_TYPES = ["history_update"] as const;
+    const { connect: sseHistoryConnect, disconnect: sseHistoryDisconnect } = useSSE(
+        handleHistorySSEEvent,
+        SSE_HISTORY_EVENT_TYPES,
+    );
+    let stopHistoryPolling: (() => void) | null = null;
+    let stopIsWatchingWatcher: (() => void) | null = null;
+
+    function handleHistorySSEEvent(event: MessageEvent) {
+        try {
+            const data = JSON.parse(event.data);
+            const changedHistoryIds: string[] = data.history_ids ?? [];
+            if (currentHistoryId.value && changedHistoryIds.includes(currentHistoryId.value)) {
+                // SSE is itself the signal that the history changed — force the
+                // refresh so the update_time short-circuit in watchHistoryOnce
+                // can't suppress the contents fetch.
+                const app = getGalaxyInstance();
+                refreshHistoryFromPushSuppliedApp(app).catch((err) =>
+                    console.error("Error refreshing history from SSE push:", err),
+                );
+            }
+        } catch (e) {
+            console.error("Error handling history SSE event:", e);
+        }
+    }
+
+    // Choose between SSE and polling based on the server config flag
+    // `enable_sse_updates`. SSE success at the socket level is not a
+    // reliable proxy: the `/api/events/stream` endpoint accepts connections
+    // even when the HistoryAuditMonitor is disabled, so relying on the
+    // EventSource `connected` state would silently stop polling without any
+    // events ever arriving.
+    //
+    // `useResourceWatcher` is instantiated lazily because it registers a
+    // `visibilitychange` listener that calls `startWatchingResourceIfNeeded`
+    // every time the tab regains focus — in SSE mode that would re-start
+    // polling we explicitly don't want.
+    const isWatchingHistory = ref(false);
+    let watchingInitialized = false;
+    function startWatchingHistoryWithSSE() {
+        if (watchingInitialized) {
+            return;
+        }
+        watchingInitialized = true;
+
+        const configStore = useConfigStore();
+        const decide = () => {
+            if (configStore.config?.enable_sse_updates) {
+                // SSE delivers incremental updates only; the store still needs
+                // a baseline fetch so the history panel isn't empty until the
+                // first change arrives.
+                watchHistory().catch((err) => console.warn("Initial history load failed", err));
+                sseHistoryConnect();
+            } else {
+                // The resource watcher fires its handler once immediately and
+                // then re-schedules on the polling interval, which covers the
+                // initial load as well as ongoing updates.
+                const { startWatchingResource, stopWatchingResource, isWatchingResource } = useResourceWatcher(
+                    watchHistory,
+                    {
+                        shortPollingInterval: ACTIVE_POLLING_INTERVAL,
+                        longPollingInterval: INACTIVE_POLLING_INTERVAL,
+                    },
+                );
+                stopHistoryPolling = stopWatchingResource;
+                stopIsWatchingWatcher = watch(isWatchingResource, (v) => (isWatchingHistory.value = v), {
+                    immediate: true,
+                });
+                startWatchingResource();
+            }
+        };
+
+        if (configStore.isLoaded) {
+            decide();
+        } else {
+            const stop = watch(
+                () => configStore.isLoaded,
+                (loaded) => {
+                    if (loaded) {
+                        stop();
+                        decide();
+                    }
+                },
+            );
+        }
+    }
 
     async function loadHistoryById(historyId: string) {
         if (!isLoadingHistory.has(historyId)) {
@@ -497,6 +580,23 @@ export const useHistoryStore = defineStore("historyStore", () => {
         return contentStats;
     }
 
+    // Closes SSE and stops polling so the watcher can't emit a trailing
+    // anonymous-cookie request that would overwrite the authenticated
+    // ``galaxysession`` cookie set by the login/register response.
+    function stopWatchingHistory() {
+        sseHistoryDisconnect();
+        if (stopHistoryPolling) {
+            stopHistoryPolling();
+            stopHistoryPolling = null;
+        }
+        if (stopIsWatchingWatcher) {
+            stopIsWatchingWatcher();
+            stopIsWatchingWatcher = null;
+        }
+        isWatchingHistory.value = false;
+        watchingInitialized = false;
+    }
+
     return {
         histories,
         changingCurrentHistory,
@@ -525,7 +625,7 @@ export const useHistoryStore = defineStore("historyStore", () => {
         restoreHistory,
         restoreHistories,
         handleTotalCountChange,
-        startWatchingHistory,
+        startWatchingHistory: startWatchingHistoryWithSSE,
         stopWatchingHistory,
         isWatchingHistory,
         loadCurrentHistory,
diff --git a/client/src/stores/notificationsStore.test.ts b/client/src/stores/notificationsStore.test.ts
new file mode 100644
index 000000000000..a88fceac595f
--- /dev/null
+++ b/client/src/stores/notificationsStore.test.ts
@@ -0,0 +1,185 @@
+import flushPromises from "flush-promises";
+import { createPinia, setActivePinia } from "pinia";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+import { useServerMock } from "@/api/client/__mocks__";
+import type { UserNotification } from "@/api/notifications";
+
+import { emitSse, sseMockFactory, useVisibilityPatch } from "./_testing/sseStoreSupport";
+import { useNotificationsStore } from "./notificationsStore";
+
+// ``vi.mock`` is hoisted above module-level ``const`` declarations, so the
+// capture-state has to be built via ``vi.hoisted`` to be visible to the factory.
+const sseState = vi.hoisted(() => {
+    return {
+        onEvent: null as ((event: MessageEvent) => void) | null,
+        connect: vi.fn(),
+        disconnect: vi.fn(),
+    };
+});
+
+vi.mock("@/composables/useNotificationSSE", () => sseMockFactory(sseState));
+
+// Realistic fixture: a single unread notification, as returned by
+// GET /api/notifications. Shape mirrors UserNotification.
+function makeNotificationFixture(overrides: Partial<UserNotification> = {}): UserNotification {
+    return {
+        id: "notif-1",
+        source: "galaxy_test",
+        category: "message",
+        variant: "info",
+        create_time: "2026-01-01T00:00:00",
+        update_time: "2026-01-01T00:00:00",
+        publication_time: "2026-01-01T00:00:00",
+        expiration_time: null,
+        seen_time: null,
+        deleted: false,
+        content: { category: "message", subject: "hello", message: "welcome" },
+        ...overrides,
+    } as UserNotification;
+}
+
+const SCENARIO_NOTIFICATION = makeNotificationFixture();
+const SCENARIO_STATUS_SINCE = {
+    total_unread_count: 1,
+    notifications: [SCENARIO_NOTIFICATION],
+    broadcasts: [],
+};
+
+const { server, http } = useServerMock();
+
+const statusSpy = vi.fn();
+
+function registerDefaultHandlers({ enableSseUpdates }: { enableSseUpdates: boolean }) {
+    server.use(
+        http.get("/api/configuration", ({ response }) => {
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            return response(200).json({
+                enable_notification_system: true,
+                enable_sse_updates: enableSseUpdates,
+            } as any);
+        }),
+        http.get("/api/notifications", ({ response }) => {
+            return response(200).json([SCENARIO_NOTIFICATION]);
+        }),
+        http.get("/api/notifications/broadcast", ({ response }) => {
+            return response(200).json([]);
+        }),
+        http.get("/api/notifications/status", ({ response }) => {
+            statusSpy();
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            return response(200).json(SCENARIO_STATUS_SINCE as any);
+        }),
+    );
+}
+
+/** Config load + initial fetch + store-decision watch needs a couple of ticks. */
+async function primeStore(startFn: () => Promise<void> | void): Promise<void> {
+    // Let the config-store fetch resolve before the store's `watch` runs.
+    await vi.runOnlyPendingTimersAsync();
+    await startFn();
+    // Two flush cycles: one for the config watch, one for the resulting fetch.
+    await flushPromises();
+    await vi.runOnlyPendingTimersAsync();
+    await flushPromises();
+}
+
+describe("notificationsStore — config-driven SSE vs polling", () => {
+    let visibility: ReturnType<typeof useVisibilityPatch>;
+
+    beforeEach(() => {
+        setActivePinia(createPinia());
+        sseState.connect.mockClear();
+        sseState.disconnect.mockClear();
+        sseState.onEvent = null;
+        statusSpy.mockClear();
+        vi.useFakeTimers();
+        visibility = useVisibilityPatch();
+    });
+
+    afterEach(() => {
+        visibility.restore();
+        vi.useRealTimers();
+    });
+
+    describe("when enable_sse_updates is true (SSE scenario)", () => {
+        beforeEach(() => {
+            registerDefaultHandlers({ enableSseUpdates: true });
+        });
+
+        it("connects SSE and does not poll the status endpoint", async () => {
+            const store = useNotificationsStore();
+            await primeStore(() => store.startWatchingNotifications());
+
+            expect(sseState.connect).toHaveBeenCalledTimes(1);
+
+            // Advance well past the polling interval (30s) and confirm
+            // the status endpoint is never polled while SSE is the active channel.
+            vi.advanceTimersByTime(120_000);
+            await flushPromises();
+            expect(statusSpy).not.toHaveBeenCalled();
+        });
+
+        it("does not start polling when the tab regains visibility", async () => {
+            const store = useNotificationsStore();
+            await primeStore(() => store.startWatchingNotifications());
+
+            visibility.set("hidden");
+            visibility.set("visible");
+
+            await flushPromises();
+            vi.advanceTimersByTime(120_000);
+            await flushPromises();
+            expect(statusSpy).not.toHaveBeenCalled();
+        });
+
+        it("ingests notification_update events into the store state", async () => {
+            const store = useNotificationsStore();
+            await primeStore(() => store.startWatchingNotifications());
+
+            const pushed = makeNotificationFixture({
+                id: "notif-2",
+                content: { category: "message", subject: "pushed via sse", message: "hi" },
+            });
+            emitSse(sseState, "notification_update", pushed);
+            await flushPromises();
+
+            expect(store.notifications.map((n) => n.id)).toContain("notif-2");
+            expect(store.totalUnreadCount).toBeGreaterThan(0);
+        });
+
+        it("ingests notification_status catch-up events on reconnect", async () => {
+            const store = useNotificationsStore();
+            await primeStore(() => store.startWatchingNotifications());
+
+            emitSse(sseState, "notification_status", {
+                total_unread_count: 42,
+                notifications: [makeNotificationFixture({ id: "notif-catchup" })],
+                broadcasts: [],
+            });
+            await flushPromises();
+
+            expect(store.totalUnreadCount).toBe(42);
+            expect(store.notifications.map((n) => n.id)).toContain("notif-catchup");
+        });
+    });
+
+    describe("when enable_sse_updates is false (polling scenario)", () => {
+        beforeEach(() => {
+            registerDefaultHandlers({ enableSseUpdates: false });
+        });
+
+        it("does not connect SSE and polls the status endpoint on the configured interval", async () => {
+            const store = useNotificationsStore();
+            await primeStore(() => store.startWatchingNotifications());
+
+            expect(sseState.connect).not.toHaveBeenCalled();
+
+            // Advance past the short polling interval (30s) and confirm
+            // the status endpoint is hit by the resource watcher.
+            vi.advanceTimersByTime(30_000);
+            await flushPromises();
+            expect(statusSpy).toHaveBeenCalled();
+        });
+    });
+});
diff --git a/client/src/stores/notificationsStore.ts b/client/src/stores/notificationsStore.ts
index 2a154885d999..3f5acf414f26 100644
--- a/client/src/stores/notificationsStore.ts
+++ b/client/src/stores/notificationsStore.ts
@@ -1,9 +1,11 @@
 import { defineStore } from "pinia";
-import { computed, ref } from "vue";
+import { computed, ref, watch } from "vue";
 
 import { GalaxyApi } from "@/api";
 import type { NotificationChanges, UserNotification, UserNotificationsBatchUpdateRequest } from "@/api/notifications";
 import { useResourceWatcher } from "@/composables/resourceWatcher";
+import { useSSE } from "@/composables/useNotificationSSE";
+import { useConfigStore } from "@/stores/configurationStore";
 import { rethrowSimple } from "@/utils/simple-error";
 import { mergeObjectListsById } from "@/utils/utils";
 
@@ -13,11 +15,6 @@ const ACTIVE_POLLING_INTERVAL = 30000; // 30 seconds
 const INACTIVE_POLLING_INTERVAL = ACTIVE_POLLING_INTERVAL * 20; // 10 minutes
 
 export const useNotificationsStore = defineStore("notificationsStore", () => {
-    const { startWatchingResource: startWatchingNotifications, stopWatchingResource: stopWatchingNotifications } =
-        useResourceWatcher(getNotificationStatus, {
-            shortPollingInterval: ACTIVE_POLLING_INTERVAL,
-            longPollingInterval: INACTIVE_POLLING_INTERVAL,
-        });
     const broadcastsStore = useBroadcastsStore();
 
     const totalUnreadCount = ref<number>(0);
@@ -28,6 +25,45 @@ export const useNotificationsStore = defineStore("notificationsStore", () => {
 
     const unreadNotifications = computed(() => notifications.value.filter((n) => !n.seen_time));
 
+    // --- SSE setup (listen only for notification event types) ---
+    const NOTIFICATION_EVENT_TYPES = ["notification_update", "broadcast_update", "notification_status"] as const;
+    const { connect: sseConnect, disconnect: sseDisconnect } = useSSE(handleSSEEvent, NOTIFICATION_EVENT_TYPES);
+    let stopPolling: (() => void) | null = null;
+
+    function handleSSEEvent(event: MessageEvent) {
+        try {
+            const data = JSON.parse(event.data);
+            switch (event.type) {
+                case "notification_update":
+                    notifications.value = mergeObjectListsById(
+                        notifications.value,
+                        [data as UserNotification],
+                        "create_time",
+                        "desc",
+                    );
+                    updateUnreadCount();
+                    break;
+                case "broadcast_update":
+                    broadcastsStore.updateBroadcasts([data]);
+                    break;
+                case "notification_status":
+                    // Full catch-up on reconnect (same shape as GET /api/notifications/status)
+                    totalUnreadCount.value = data.total_unread_count;
+                    notifications.value = mergeObjectListsById(
+                        notifications.value,
+                        data.notifications as UserNotification[],
+                        "create_time",
+                        "desc",
+                    );
+                    broadcastsStore.updateBroadcasts(data.broadcasts);
+                    break;
+            }
+            lastNotificationUpdate.value = new Date();
+        } catch (e) {
+            console.error("Error handling SSE event:", e);
+        }
+    }
+
     async function loadNotifications() {
         const { data, error } = await GalaxyApi().GET("/api/notifications");
 
@@ -76,6 +112,75 @@ export const useNotificationsStore = defineStore("notificationsStore", () => {
         }
     }
 
+    // Choose between SSE and polling based on the server config flag
+    // `enable_sse_updates`. The `/api/events/stream` endpoint accepts
+    // connections regardless of the flag, so we cannot rely on EventSource
+    // connectivity to decide — config is the source of truth. Callers gate
+    // on `enable_notification_system` before reaching here, so we only have
+    // to pick the transport.
+    //
+    // `useResourceWatcher` is instantiated lazily because it registers a
+    // `visibilitychange` listener that calls `startWatchingResourceIfNeeded`
+    // every time the tab regains focus — in SSE mode that would re-start
+    // polling we explicitly don't want.
+    let watchingInitialized = false;
+    function ensureWatchingWithConfig() {
+        if (watchingInitialized) {
+            return;
+        }
+        watchingInitialized = true;
+
+        const configStore = useConfigStore();
+        const decide = () => {
+            if (configStore.config?.enable_sse_updates) {
+                sseConnect();
+            } else {
+                const { startWatchingResource: startPolling, stopWatchingResource } = useResourceWatcher(
+                    getNotificationStatus,
+                    {
+                        shortPollingInterval: ACTIVE_POLLING_INTERVAL,
+                        longPollingInterval: INACTIVE_POLLING_INTERVAL,
+                    },
+                );
+                stopPolling = stopWatchingResource;
+                startPolling();
+            }
+        };
+
+        if (configStore.isLoaded) {
+            decide();
+        } else {
+            const stop = watch(
+                () => configStore.isLoaded,
+                (loaded) => {
+                    if (loaded) {
+                        stop();
+                        decide();
+                    }
+                },
+            );
+        }
+    }
+
+    async function startWatchingNotifications() {
+        // Always do an initial load first
+        if (!lastNotificationUpdate.value) {
+            try {
+                loadingNotifications.value = true;
+                await broadcastsStore.loadBroadcasts();
+                await loadNotifications();
+                updateUnreadCount();
+                lastNotificationUpdate.value = new Date();
+            } catch (e) {
+                console.error(e);
+            } finally {
+                loadingNotifications.value = false;
+            }
+        }
+
+        ensureWatchingWithConfig();
+    }
+
     async function updateBatchNotification(request: UserNotificationsBatchUpdateRequest) {
         const { error } = await GalaxyApi().PUT("/api/notifications", {
             body: request,
@@ -88,7 +193,11 @@ export const useNotificationsStore = defineStore("notificationsStore", () => {
         if (request.changes.deleted) {
             notifications.value = notifications.value.filter((n) => !request.notification_ids.includes(n.id));
         }
-        startWatchingNotifications();
+        // When polling is the active transport, trigger a poll to refresh
+        // state after a local mutation. SSE delivers the update on its own.
+        if (!useConfigStore().config?.enable_sse_updates) {
+            startWatchingNotifications();
+        }
     }
 
     async function updateNotification(notification: UserNotification, changes: NotificationChanges) {
@@ -99,6 +208,19 @@ export const useNotificationsStore = defineStore("notificationsStore", () => {
         totalUnreadCount.value = notifications.value.filter((n) => !n.seen_time).length;
     }
 
+    // Closes the SSE stream and stops the polling watcher so nothing running
+    // in the background can outlive a full-page navigation (login/register).
+    // A late-arriving response from an anonymous-cookie request would otherwise
+    // overwrite the just-issued authenticated ``galaxysession`` cookie.
+    function stopWatchingNotifications() {
+        sseDisconnect();
+        if (stopPolling) {
+            stopPolling();
+            stopPolling = null;
+        }
+        watchingInitialized = false;
+    }
+
     return {
         notifications,
         totalUnreadCount,
diff --git a/client/src/watch/watchHistory.js b/client/src/watch/watchHistory.js
index 2860325e4610..523d3003163a 100644
--- a/client/src/watch/watchHistory.js
+++ b/client/src/watch/watchHistory.js
@@ -39,13 +39,29 @@ export async function watchHistory(app) {
 }
 
 export async function watchHistoryOnce(app) {
+    return _fetchHistoryAndChangedItems(app, { force: false });
+}
+
+/**
+ * Forces a fresh history + changed-items fetch, ignoring the `lastUpdateTime`
+ * short-circuit. Use this when an out-of-band signal (SSE `history_update`)
+ * already told us the history changed — we shouldn't re-gate on update_time.
+ */
+export async function refreshHistoryFromPush(app) {
+    return _fetchHistoryAndChangedItems(app, { force: true });
+}
+
+async function _fetchHistoryAndChangedItems(app, { force }) {
     const historyStore = useHistoryStore();
     const historyItemsStore = useHistoryItemsStore();
     const datasetStore = useDatasetStore();
     const collectionElementsStore = useCollectionElementsStore();
 
-    // get current history
     const checkForUpdate = new Date();
+    // Always pass the `since` cursor so the server can short-circuit cheaply
+    // when nothing has changed; SSE-driven `force` only bypasses the
+    // client-side update_time equality gate, not the server-side cursor —
+    // we don't want to refetch items the client already has.
     const history = await historyStore.loadCurrentHistory(lastUpdateTime);
     const { lastCheckedTime } = storeToRefs(historyItemsStore);
     lastCheckedTime.value = checkForUpdate;
@@ -53,39 +69,35 @@ export async function watchHistoryOnce(app) {
         return;
     }
 
-    // continue if the history update time has changed
-    if (!lastUpdateTime || lastUpdateTime < history.update_time) {
-        const historyId = history.id;
-        lastUpdateTime = history.update_time;
-        historyItemsStore.setLastUpdateTime();
-        // execute request to obtain recently changed items
-        const params = {
-            v: "dev",
-            limit: limit,
-            q: "update_time-ge",
-            qv: lastRequestDate.toISOString(),
-        };
-        // request detailed info only for the expanded datasets
-        const detailedIds = getCurrentlyExpandedHistoryContentIds();
-        if (detailedIds.length) {
-            params["details"] = detailedIds.join(",");
-        }
-        const url = `/api/histories/${historyId}/contents`;
-        lastRequestDate = new Date();
-        const payload = await urlData({ url, params });
-        // show warning that not all changes have been obtained
-        if (payload && payload.length == limit) {
-            console.debug(`Reached limit of monitored changes (limit=${limit}).`);
-        }
-        // pass changed items to attached stores
-        historyStore.setHistory(history);
-        datasetStore.saveDatasets(payload);
-        historyItemsStore.saveHistoryItems(historyId, payload);
-        collectionElementsStore.saveCollections(payload);
-        // trigger changes in legacy handler
-        if (app) {
-            app.user.loadFromApi(app.user.id || "current");
-        }
+    if (!force && lastUpdateTime && lastUpdateTime >= history.update_time) {
+        return;
+    }
+
+    const historyId = history.id;
+    lastUpdateTime = history.update_time;
+    historyItemsStore.setLastUpdateTime();
+    const params = {
+        v: "dev",
+        limit: limit,
+        q: "update_time-ge",
+        qv: lastRequestDate.toISOString(),
+    };
+    const detailedIds = getCurrentlyExpandedHistoryContentIds();
+    if (detailedIds.length) {
+        params["details"] = detailedIds.join(",");
+    }
+    const url = `/api/histories/${historyId}/contents`;
+    lastRequestDate = new Date();
+    const payload = await urlData({ url, params });
+    if (payload && payload.length == limit) {
+        console.debug(`Reached limit of monitored changes (limit=${limit}).`);
+    }
+    historyStore.setHistory(history);
+    datasetStore.saveDatasets(payload);
+    historyItemsStore.saveHistoryItems(historyId, payload);
+    collectionElementsStore.saveCollections(payload);
+    if (app) {
+        app.user.loadFromApi(app.user.id || "current");
     }
 }
 
diff --git a/client/vite-plugin-galaxy-dev-server.js b/client/vite-plugin-galaxy-dev-server.js
index b46e13defa64..e2ca53831713 100644
--- a/client/vite-plugin-galaxy-dev-server.js
+++ b/client/vite-plugin-galaxy-dev-server.js
@@ -99,16 +99,33 @@ export function galaxyDevServerPlugin() {
                 const originalWrite = res.write.bind(res);
                 const originalEnd = res.end.bind(res);
 
-                // Buffer to collect response body
+                // Buffer to collect response body. We only buffer when the
+                // upstream response is HTML; everything else (JSON, binary,
+                // and crucially `text/event-stream`) must pass straight
+                // through, because streaming responses never call
+                // `res.end()` and would otherwise stall indefinitely.
                 const chunks = [];
-                let isHtml = false;
+                // Tri-state: null = undecided (first write hasn't landed yet),
+                //             true = stream it through untransformed,
+                //             false = buffer for HTML rewrite on end().
+                let passthrough = null;
+
+                function decidePassthrough() {
+                    if (passthrough !== null) {
+                        return;
+                    }
+                    const contentType = res.getHeader("content-type");
+                    passthrough = !contentType || !contentType.toString().includes("text/html");
+                }
 
-                // Override write to collect chunks
                 res.write = function (chunk, encoding, callback) {
+                    decidePassthrough();
+                    if (passthrough) {
+                        return originalWrite(chunk, encoding, callback);
+                    }
                     if (chunk) {
                         chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk, encoding));
                     }
-                    // Don't write yet - we'll write in end()
                     if (typeof encoding === "function") {
                         encoding(); // encoding is actually the callback
                     } else if (typeof callback === "function") {
@@ -119,41 +136,39 @@ export function galaxyDevServerPlugin() {
 
                 // Override end to transform and send response
                 res.end = function (chunk, encoding, callback) {
+                    decidePassthrough();
+                    if (passthrough) {
+                        return originalEnd(chunk, encoding, callback);
+                    }
+
                     if (chunk) {
                         chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk, encoding));
                     }
 
-                    // Check content type
-                    const contentType = res.getHeader("content-type");
-                    isHtml = contentType && contentType.toString().includes("text/html");
-
                     // Combine all chunks
                     let body = Buffer.concat(chunks);
 
-                    // Transform HTML responses that contain Galaxy bundles
-                    if (isHtml) {
-                        // Decompress gzip responses (common from remote Galaxy servers)
-                        const contentEncoding = res.getHeader("content-encoding");
-                        if (contentEncoding === "gzip") {
-                            try {
-                                body = gunzipSync(body);
-                            } catch (e) {
-                                // If decompression fails, continue with original body
-                                console.warn("[galaxy-dev-server] Failed to decompress gzip response:", e.message);
-                            }
+                    // Decompress gzip responses (common from remote Galaxy servers)
+                    const contentEncoding = res.getHeader("content-encoding");
+                    if (contentEncoding === "gzip") {
+                        try {
+                            body = gunzipSync(body);
+                        } catch (e) {
+                            // If decompression fails, continue with original body
+                            console.warn("[galaxy-dev-server] Failed to decompress gzip response:", e.message);
                         }
+                    }
 
-                        let htmlString = body.toString("utf-8");
-                        if (htmlString.includes("bundled.js") || htmlString.includes("/static/dist/")) {
-                            htmlString = transformGalaxyHtml(htmlString);
-                            body = Buffer.from(htmlString, "utf-8");
+                    let htmlString = body.toString("utf-8");
+                    if (htmlString.includes("bundled.js") || htmlString.includes("/static/dist/")) {
+                        htmlString = transformGalaxyHtml(htmlString);
+                        body = Buffer.from(htmlString, "utf-8");
 
-                            // Update content-length header
-                            res.setHeader("content-length", body.length);
+                        // Update content-length header
+                        res.setHeader("content-length", body.length);
 
-                            // Remove content-encoding since we've decompressed it
-                            res.removeHeader("content-encoding");
-                        }
+                        // Remove content-encoding since we've decompressed it
+                        res.removeHeader("content-encoding");
                     }
 
                     // Send the response
diff --git a/doc/source/admin/galaxy_options.rst b/doc/source/admin/galaxy_options.rst
index 47e3aaacf343..bff4e8e3e55a 100644
--- a/doc/source/admin/galaxy_options.rst
+++ b/doc/source/admin/galaxy_options.rst
@@ -281,6 +281,23 @@
 :Type: int
 
 
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+``kombu_sqla_transport_cleanup_interval``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+:Description:
+    Time (in seconds) between attempts to delete fully-consumed rows
+    from the Kombu SQLAlchemy transport tables (``kombu_message``).
+    Only relevant when ``amqp_internal_connection`` uses a
+    ``sqlalchemy+*`` scheme (the default with an on-disk
+    control.sqlite); the SQLAlchemy transport has no built-in TTL, so
+    without this task the tables grow unbounded. The task no-ops on
+    non-SQLAlchemy brokers (RabbitMQ/Redis honor per-message
+    expiration natively). Set to 0 to disable the cleanup task.
+:Default: ``900``
+:Type: int
+
+
 ~~~~~~~~~~~~~
 ``file_path``
 ~~~~~~~~~~~~~
@@ -3408,6 +3425,18 @@
 :Type: bool
 
 
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+``queue_metrics_interval``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+:Description:
+    How often (in seconds) the Celery beat task emits queue-depth,
+    SSE-connection, and WorkerProcess gauges. Only active when
+    statsd_host is set. Set to 0 to disable.
+:Default: ``15``
+:Type: int
+
+
 ~~~~~~~~~~~~~~~~~~~~~~
 ``library_import_dir``
 ~~~~~~~~~~~~~~~~~~~~~~
@@ -5804,6 +5833,41 @@
 :Type: str
 
 
+~~~~~~~~~~~~~~~~~~~~~~
+``enable_sse_updates``
+~~~~~~~~~~~~~~~~~~~~~~
+
+:Description:
+    Enables real-time updates via Server-Sent Events (SSE), replacing
+    the history (3 s), entry-point (10 s) and notification (30 s)
+    polling loops with push events delivered over a single
+    ``/api/events/stream`` connection per browser tab. A background
+    monitor watches for history changes (via PostgreSQL LISTEN/NOTIFY,
+    or audit-table polling as a fallback for SQLite); entry-point
+    changes are dispatched directly from the code paths that mutate
+    them; in-app notifications and broadcasts are pushed when
+    ``enable_notification_system`` is also true. When disabled,
+    polling remains the source of updates for all three. See the admin
+    guide "Server-Sent Events for real-time updates" for the full
+    architecture, monitoring guidance and proxy configuration.
+:Default: ``false``
+:Type: bool
+
+
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+``history_audit_monitor_poll_interval``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+:Description:
+    The interval in seconds between history audit table polls when
+    using the polling fallback (SQLite or when PostgreSQL
+    LISTEN/NOTIFY is unavailable). Only used when enable_sse_updates
+    is true. Lower values mean faster updates but more database
+    queries. Recommended range: 1-5 seconds.
+:Default: ``2``
+:Type: int
+
+
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 ``enable_notification_system``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -5815,6 +5879,9 @@
     finished, etc.
     The system allows notification scheduling and expiration, and
     users can opt-out of specific notification categories or channels.
+    Delivery is push-based via Server-Sent Events when
+    ``enable_sse_updates`` is also true, and falls back to 30-second
+    polling against ``/api/notifications/status`` otherwise.
     Admins can schedule and broadcast notifications that will be
     visible to all users, including special server-wide announcements
     such as scheduled maintenance, high load warnings, and event
diff --git a/doc/source/admin/index.rst b/doc/source/admin/index.rst
index 30dd639d8003..7b065f21d0fc 100644
--- a/doc/source/admin/index.rst
+++ b/doc/source/admin/index.rst
@@ -14,6 +14,7 @@ Galaxy Deployment & Administration
    nginx
    apache
    scaling
+   sse_updates
    cluster
    jobs
    job_metrics
diff --git a/doc/source/admin/nginx.md b/doc/source/admin/nginx.md
index eb523ee4cc2c..d140809ba195 100644
--- a/doc/source/admin/nginx.md
+++ b/doc/source/admin/nginx.md
@@ -233,6 +233,17 @@ previous section:
 
 ## Advanced Configuration Topics
 
+### Server-Sent Events (real-time updates)
+
+Galaxy can push history, entry-point and notification updates to the browser
+via a long-lived Server-Sent Events stream at ``/api/events/stream``. nginx
+will buffer that response by default, which breaks the stream — either rely
+on the ``X-Accel-Buffering: no`` header Galaxy already sets, or add an
+explicit ``location /api/events/stream`` block that disables buffering and
+raises the read/send timeouts. The full configuration block, monitoring
+guidance, and the architecture overview live in
+[Server-Sent Events for real-time updates](sse_updates.md#configuring-nginx).
+
 ### Sending Files With Nginx
 
 Galaxy sends files (e.g. dataset downloads) by opening the file and streaming it in chunks through the proxy server.
diff --git a/doc/source/admin/scaling.md b/doc/source/admin/scaling.md
index cba75870576a..40036a1ad4bf 100644
--- a/doc/source/admin/scaling.md
+++ b/doc/source/admin/scaling.md
@@ -190,6 +190,14 @@ Some of these options deserve explanation:
 4-12 workers should be able to handle hundreds if not thousands of requests per second.
 * `extra_args`: You can specify additional arguments to pass to gunicorn here.
 
+```{note}
+If you have enabled [Server-Sent Events for real-time updates](sse_updates.md), every connected browser tab holds an
+open connection to one Gunicorn worker for the lifetime of the tab. Gravity configures Gunicorn with the
+``uvicorn.workers.UvicornWorker`` async worker by default, so a single worker process can hold thousands of idle SSE
+connections without blocking — the limit you actually need to watch is memory headroom and the per-process
+file-descriptor ulimit, not the worker count.
+```
+
 Note that the performance option values given above are just examples and should be tuned per your specific needs.
 However, as given, they are a good place to start.
 
diff --git a/doc/source/admin/sse_updates.md b/doc/source/admin/sse_updates.md
new file mode 100644
index 000000000000..b845217bf69d
--- /dev/null
+++ b/doc/source/admin/sse_updates.md
@@ -0,0 +1,280 @@
+# Server-Sent Events (SSE) for real-time updates
+
+Galaxy can push history changes, interactive-tool entry-point changes, and
+in-app notifications to connected browsers via [Server-Sent Events][sse-mdn]
+instead of polling. This replaces the legacy 3-second history poll and
+10-second entry-point poll with a single long-lived HTTP connection per
+browser tab, dramatically reducing API load on busy servers and giving users
+sub-second update latency.
+
+This document describes the architecture, the configuration knobs, the
+metrics admins should watch, and how to configure NGINX so the long-lived
+event connection is not buffered or prematurely terminated.
+
+[sse-mdn]: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events
+
+## How it works
+
+A high-level view of the moving parts on the server:
+
+```
+   ┌────────────────┐  LISTEN/NOTIFY     ┌──────────────────────┐
+   │  Postgres      │ ─────────────────► │ HistoryAuditMonitor  │
+   │  history_audit │  (or audit poll    │ (one elected process) │
+   │                │   on SQLite)       └──────────┬───────────┘
+   └────────────────┘                               │ Kombu
+                                                    ▼
+                                     ┌─────────────────────────────┐
+                                     │  galaxy_queue_worker        │
+                                     │  control task fan-out:      │
+                                     │   • history_update          │
+                                     │   • entry_point_update      │
+                                     │   • notification_update     │
+                                     │   • broadcast_update        │
+                                     └──────────────┬──────────────┘
+                                                    │ in-process
+                                                    ▼
+                              ┌─────────────────────────────────────┐
+                              │  Each gunicorn worker:              │
+                              │   SSEConnectionManager → asyncio    │
+                              │   queue per connected browser tab   │
+                              └──────────────┬──────────────────────┘
+                                             │ HTTP chunked
+                                             ▼
+                              ┌─────────────────────────────────────┐
+                              │  Browser EventSource                │
+                              │  (single stream, multiplexed events)│
+                              └─────────────────────────────────────┘
+```
+
+Concretely:
+
+1. **One stream per browser, many event types.** The browser opens a single
+   `EventSource` against `/api/events/stream`. The same connection carries
+   `history_update`, `entry_point_update`, `notification_update`,
+   `broadcast_update`, and `notification_status` events.
+2. **Per-process registries.** Each Gunicorn worker keeps an
+   `SSEConnectionManager` that holds an `asyncio.Queue` per connected tab,
+   indexed by user id (and Galaxy session id, so anonymous users still
+   receive their own history's updates).
+3. **Producers.**
+   - History updates come from a `HistoryAuditMonitor` that watches
+     `history_audit` via PostgreSQL `LISTEN/NOTIFY` (instant) or by polling
+     the audit table on SQLite. Only one process in the cluster is the
+     producer, picked by `DatabaseHeartbeat` leader election. If a
+     standalone `galaxy-sse-monitor` process is running it always wins;
+     otherwise one webapp picks it up.
+   - Entry-point updates are dispatched directly from the code paths that
+     mutate interactive-tool entry points — there is no separate watcher.
+   - Notifications dispatch SSE events from `NotificationManager` whenever
+     a notification or broadcast is created.
+4. **Cross-process fan-out.** Producers don't know which worker holds the
+   recipient's connection, so all events go through a Kombu control task
+   broadcast on the internal AMQP bus. Every worker receives every event
+   and locally drops the ones for users it doesn't currently hold a
+   connection for.
+5. **Reconnect catch-up.** When a browser reconnects after a network blip,
+   it sends `Last-Event-ID`. The server replays an aggregated
+   `notification_status` covering everything since that timestamp. History
+   updates come with an `update_time` cursor in the payload, so the client
+   can request the delta itself.
+
+### Standalone monitor (recommended for production)
+
+The `galaxy-sse-monitor` console script (installed by the `galaxy-app`
+package) runs the `HistoryAuditMonitor` outside the webapp processes. This
+is the recommended layout for production because:
+
+- The webapp processes never compete with each other for the audit-monitor
+  role on cold starts.
+- Restarting the webapp tier doesn't briefly stall history updates while
+  another worker is elected.
+- The monitor needs only DB + AMQP access, so it can be sized
+  independently and runs with a much smaller resident-set than a webapp.
+
+A typical Gravity supervisor entry looks like the existing
+`galaxy-celery-worker` block — point at the same `galaxy.yml` and run
+`galaxy-sse-monitor` on its own. With the daemon present, the
+`HistoryAuditMonitor` registered on the webapp side stays idle (the
+heartbeat election picks the daemon) but is still wired up so it can take
+over if the daemon goes away.
+
+If `enable_sse_updates` is `false`, `galaxy-sse-monitor` will start, log a
+warning, and idle — it does no work and produces no events.
+
+## Configuration
+
+There is a single admin-facing flag for SSE-driven updates:
+
+```yaml
+galaxy:
+  enable_sse_updates: true
+```
+
+This controls **all three** SSE-driven paths (history, entry-point,
+notifications). When `false`:
+
+- `HistoryAuditMonitor` is not registered, so the cluster does no
+  `LISTEN/NOTIFY` or audit-table polling for history changes.
+- The browser falls back to its existing 3-second history poll and
+  10-second entry-point poll.
+- Notifications fall back to the existing 30-second polling against
+  `/api/notifications/status`.
+
+`enable_notification_system` is independent: it gates whether the
+notification system is available at all (notification creation, delivery,
+preferences, broadcasts). With `enable_notification_system: true`:
+
+- `enable_sse_updates: true` → notifications arrive via SSE.
+- `enable_sse_updates: false` → notifications are polled.
+
+With `enable_notification_system: false` the entire notification system is
+off — there is nothing to push or poll.
+
+The polling-fallback knob for the history audit monitor stays available:
+
+```yaml
+galaxy:
+  history_audit_monitor_poll_interval: 2  # seconds, SQLite / no-LISTEN only
+```
+
+This only matters when running on SQLite or in setups where PostgreSQL
+`LISTEN/NOTIFY` is unavailable.
+
+## What to monitor
+
+When statsd is configured (via `statsd_host` and friends), the SSE
+plumbing emits the following metrics. Capture these on the same dashboard
+you use for Gunicorn worker health:
+
+| Metric                                     | Type    | Source                | Meaning                                                      |
+| ------------------------------------------ | ------- | --------------------- | ------------------------------------------------------------ |
+| `galaxy.sse.connections.dropped`           | counter | `SSEConnectionManager` | A per-connection asyncio queue filled up; an event was lost. |
+| `galaxy.sse.dispatch.count` (tag: `task`) | counter | `SSEEventDispatcher`  | Control-task fan-outs by event kind.                         |
+| `galaxy.sse.dispatch.latency_ms` (tag: `task`) | timing | `SSEEventDispatcher`  | Wall time spent enqueueing the control task.                 |
+| `galaxy.sse.dispatch.skipped_no_qw`        | counter | `SSEEventDispatcher`  | Producer tried to dispatch with no queue worker bound — events would have been dropped. |
+
+You can also expose a "currently connected SSE clients" gauge if you wire
+one up: each `SSEConnectionManager` instance publishes
+`total_broadcast_connections` (all connections, including anonymous) and
+`total_per_user_connections` (connections bound to a specific user). These
+are per-worker numbers; sum across workers for a cluster total.
+
+Alerting recommendations:
+
+- **`galaxy.sse.connections.dropped` > 0 sustained** indicates a slow or
+  stuck client whose queue filled up. Occasional drops on a network blip
+  are normal; a steady rate is a bug or a misconfigured proxy holding
+  events back too long.
+- **`galaxy.sse.dispatch.skipped_no_qw` > 0** means events are being lost
+  because the producer process can't reach the AMQP bus. Check the
+  `amqp_internal_connection` config and the AMQP broker health.
+- **`galaxy.sse.dispatch.latency_ms` p95 climbing** points at an AMQP
+  bottleneck (broker load, network) — events will land late.
+
+In addition, watch the standard worker metrics. SSE connections are
+long-lived (tens of minutes is common), but Galaxy runs Gunicorn with
+`uvicorn.workers.UvicornWorker` (configured by Gravity by default), so
+each worker process is async and can hold thousands of idle SSE
+connections without blocking other requests. The practical concern on
+busy servers is therefore memory and file-descriptor headroom, not
+worker exhaustion: budget a few KB per connection plus one fd per
+connection per worker, and raise `ulimit -n` accordingly.
+
+## Configuring NGINX
+
+The SSE endpoint is served at `/api/events/stream`. It is a normal
+HTTP/1.1 chunked response, so it works through NGINX without any special
+modules — but you must turn off response buffering and raise the
+read/send timeouts, otherwise events will arrive in batched bursts (or
+not at all until the connection times out).
+
+Galaxy already sets `X-Accel-Buffering: no` on the response, which
+disables NGINX's response buffering for that one endpoint without
+affecting buffering on the rest of Galaxy. That alone is enough on most
+setups. The block below adds the read/send timeouts and HTTP/1.1
+upgrade-friendly headers explicitly so the connection survives long
+idle periods between events:
+
+```nginx
+        # Long-lived Server-Sent Events stream.
+        # Galaxy sends ``X-Accel-Buffering: no`` on the response, which
+        # disables nginx response buffering just for this endpoint.
+        # The keepalive comment fires every 30s so the read timeout
+        # only needs to be a comfortable margin above that.
+        location /api/events/stream {
+            proxy_pass http://unix:/srv/galaxy/var/gunicorn.sock;
+            proxy_set_header Host $http_host;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $scheme;
+
+            proxy_http_version 1.1;
+            proxy_set_header Connection "";
+
+            # Disable buffering and gzip explicitly. ``X-Accel-Buffering``
+            # already does this, but pinning it here also covers setups
+            # where a sub-filter strips upstream headers.
+            proxy_buffering off;
+            proxy_cache off;
+            gzip off;
+
+            # Keepalives fire every 30s; allow generous slack on top.
+            proxy_read_timeout 1h;
+            proxy_send_timeout 1h;
+        }
+```
+
+Place this `location` block **above** the catch-all `location /` block in
+your existing Galaxy `server {}` (see the [NGINX proxy guide](nginx.md)).
+NGINX matches longest prefix first, so the order doesn't matter for
+correctness, but keeping all the override blocks together at the top of
+the server block makes the special-case handling easy to find.
+
+If you serve Galaxy at a URL prefix (`/galaxy`), prefix the location too:
+
+```nginx
+        location /galaxy/api/events/stream {
+            proxy_pass http://unix:/srv/galaxy/var/gunicorn.sock:/galaxy;
+            # ...same body as above
+        }
+```
+
+### Other proxies
+
+If you front Galaxy with something other than NGINX, the same rules
+apply: disable response buffering for `/api/events/stream`, allow the
+connection to stay open for at least the `keepalive` interval (30 s by
+default) plus a healthy margin, and pass the request through HTTP/1.1
+without forcing `Connection: close`.
+
+- **Apache `mod_proxy_http`**: add `ProxyPass` with
+  `flushpackets=on flushwait=5` and bump `ProxyTimeout` to at least a
+  few minutes. Avoid `mod_deflate` on this endpoint.
+- **HAProxy**: the connection is a plain HTTP/1.1 chunked response and
+  needs no special handling beyond `timeout server` and
+  `timeout tunnel` raised above 30 s.
+- **Cloudflare and other CDNs**: many CDNs buffer HTTP/1.1 chunked
+  responses by default. Either bypass the CDN for `/api/events/stream`
+  or follow the CDN's documented pattern for SSE streaming.
+
+## Verifying the deployment
+
+After enabling `enable_sse_updates`, three quick checks confirm the
+stream is healthy end-to-end:
+
+1. From the browser DevTools Network tab, open Galaxy and look for a
+   `GET /api/events/stream` request that stays in the **pending** state
+   with a constantly incrementing transferred-bytes count. The response
+   `Content-Type` is `text/event-stream`.
+2. Trigger a history change (run a tool, rename a dataset). The Network
+   tab should show a `history_update` event in the EventStream view of
+   that connection within a second or two, and the history panel
+   refreshes without a polling round-trip.
+3. On the server side, `galaxy.sse.dispatch.count` should be ticking up
+   for each event kind your users exercise. If you wired up the
+   connection gauges, they should reflect roughly one connection per
+   open browser tab.
+
+If the connection opens but no events arrive, the most common causes
+are: a proxy buffering responses (revisit the NGINX section), or a
+producer that can't reach AMQP (see `galaxy.sse.dispatch.skipped_no_qw`).
diff --git a/lib/galaxy/app/__init__.py b/lib/galaxy/app/__init__.py
index e8b8ef1b9410..02c1b47c4236 100644
--- a/lib/galaxy/app/__init__.py
+++ b/lib/galaxy/app/__init__.py
@@ -65,6 +65,7 @@
 from galaxy.managers.folders import FolderManager
 from galaxy.managers.hdas import HDAManager
 from galaxy.managers.histories import HistoryManager
+from galaxy.managers.history_audit_monitor import HistoryAuditMonitor
 from galaxy.managers.interactivetool import InteractiveToolManager
 from galaxy.managers.jobs import (
     JobManager as JobQueryManager,
@@ -77,6 +78,8 @@
 from galaxy.managers.object_store_instances import UserObjectStoreResolverImpl
 from galaxy.managers.roles import RoleManager
 from galaxy.managers.session import GalaxySessionManager
+from galaxy.managers.sse import SSEConnectionManager
+from galaxy.managers.sse_dispatch import SSEEventDispatcher
 from galaxy.managers.tasks import (
     AsyncTasksManager,
     CeleryAsyncTasksManager,
@@ -150,8 +153,12 @@
 )
 from galaxy.tool_shed.cache import ToolShedRepositoryCache
 from galaxy.tool_shed.galaxy_install.client import InstallationTarget
-from galaxy.tool_shed.galaxy_install.installed_repository_manager import InstalledRepositoryManager
-from galaxy.tool_shed.galaxy_install.update_repository_manager import UpdateRepositoryManager
+from galaxy.tool_shed.galaxy_install.installed_repository_manager import (
+    InstalledRepositoryManager,
+)
+from galaxy.tool_shed.galaxy_install.update_repository_manager import (
+    UpdateRepositoryManager,
+)
 from galaxy.tool_util.data import ToolDataTableManager as BaseToolDataTableManager
 from galaxy.tool_util.deps import containers
 from galaxy.tool_util.deps.dependencies import AppInfo
@@ -250,7 +257,11 @@ def before_send(event, hint):
                         #  "cannot find 'file_name' while searching for 'species_chromosomes.file_name'"]
                         # If we don't do this issues are never properly grouped since by default the calling stack is inspected,
                         # and that is always unique in cheetah as it is dynamically generated.
-                        event["fingerprint"] = [str(exc_value), str(exc_value.tool_version), str(exc_value.__cause__)]
+                        event["fingerprint"] = [
+                            str(exc_value),
+                            str(exc_value.tool_version),
+                            str(exc_value.__cause__),
+                        ]
                         event.setdefault("tags", {}).update(
                             {
                                 "tool_is_latest": exc_value.is_latest,
@@ -307,7 +318,10 @@ def __init__(self, fsmon=False, **kwargs) -> None:
         config_file = kwargs.get("global_conf", {}).get("__file__", None)
         if config_file:
             log.debug('Using "galaxy.ini" config file: %s', config_file)
-        self._configure_models(check_migrate_databases=self.config.check_migrate_databases, config_file=config_file)
+        self._configure_models(
+            check_migrate_databases=self.config.check_migrate_databases,
+            config_file=config_file,
+        )
         # Security helper
         self._configure_security()
         self._register_singleton(IdEncodingHelper, self.security)
@@ -411,7 +425,11 @@ def _configure_toolbox(self):
         index_help = getattr(self.config, "index_tool_help", True)
         self.toolbox_search = self._register_singleton(
             ToolBoxSearch,
-            ToolBoxSearch(self.toolbox, index_dir=self.config.tool_search_index_dir, index_help=index_help),
+            ToolBoxSearch(
+                self.toolbox,
+                index_dir=self.config.tool_search_index_dir,
+                index_help=index_help,
+            ),
         )
 
     @property
@@ -496,8 +514,9 @@ def _configure_object_store(self, **kwds):
         templates = ConfiguredObjectStoreTemplates.from_app_config(self.config, vault_configured=vault_configured)
         self.object_store_templates = self._register_singleton(ConfiguredObjectStoreTemplates, templates)
         user_object_store_resolver = self._register_abstract_singleton(
-            UserObjectStoreResolver, UserObjectStoreResolverImpl  # type: ignore[type-abstract]
-        )  # Ignored because of https://github.com/python/mypy/issues/4717
+            UserObjectStoreResolver,  # type: ignore[type-abstract]  # https://github.com/python/mypy/issues/4717
+            UserObjectStoreResolverImpl,
+        )
         kwds["user_object_store_resolver"] = user_object_store_resolver
         self.object_store = build_object_store_from_config(self.config, **kwds)
 
@@ -600,7 +619,13 @@ class GalaxyManagerApplication(MinimalManagerApp, MinimalGalaxyApplication):
 
     model: GalaxyModelMapping
 
-    def __init__(self, configure_logging=True, use_converters=True, use_display_applications=True, **kwargs):
+    def __init__(
+        self,
+        configure_logging=True,
+        use_converters=True,
+        use_display_applications=True,
+        **kwargs,
+    ):
         super().__init__(**kwargs)
         self._register_singleton(MinimalManagerApp, self)  # type: ignore[type-abstract]
         self.execution_timer_factory = self._register_singleton(
@@ -617,7 +642,8 @@ def __init__(self, configure_logging=True, use_converters=True, use_display_appl
         # Initialize job metrics manager, needs to be in place before
         # config so per-destination modifications can be made.
         self.job_metrics = self._register_singleton(
-            JobMetrics, JobMetrics(self.config.job_metrics_config_file, self.config.job_metrics, app=self)
+            JobMetrics,
+            JobMetrics(self.config.job_metrics_config_file, self.config.job_metrics, app=self),
         )
         # Initialize the job management configuration
         self.job_config = self._register_singleton(jobs.JobConfiguration)
@@ -655,11 +681,30 @@ def __init__(self, configure_logging=True, use_converters=True, use_display_appl
         self.role_manager = self._register_singleton(RoleManager)
         self.job_manager = self._register_singleton(JobManager)
 
+        # AMQP connection + a publisher-ready queue worker. Celery worker processes
+        # inherit this via GalaxyManagerApplication so they can fan out SSE events
+        # to web workers (no consumer thread is started here — see bind_publisher).
+        self.amqp_internal_connection_obj = galaxy.queues.connection_from_config(self.config)
+        if self.amqp_internal_connection_obj is not None:
+            self.queue_worker = self._register_singleton(GalaxyQueueWorker, GalaxyQueueWorker(self))
+            self.queue_worker.bind_publisher()
+
+        # SSE dispatcher must be registered before NotificationManager so Lagom
+        # can auto-inject the Optional[SSEEventDispatcher] constructor arg.
+        self._register_singleton(
+            SSEEventDispatcher,
+            SSEEventDispatcher(
+                queue_worker=getattr(self, "queue_worker", None),
+                application_stack=self.application_stack,
+                statsd_client=self.execution_timer_factory.galaxy_statsd_client,
+            ),
+        )
         self.notification_manager = self._register_singleton(NotificationManager)
         self.interactivetool_manager = InteractiveToolManager(self)
 
         self.task_manager = self._register_abstract_singleton(
-            AsyncTasksManager, CeleryAsyncTasksManager  # type: ignore[type-abstract]  # https://github.com/python/mypy/issues/4717
+            AsyncTasksManager,  # type: ignore[type-abstract]  # https://github.com/python/mypy/issues/4717
+            CeleryAsyncTasksManager,
         )
 
         # ConfiguredFileSources
@@ -671,10 +716,12 @@ def __init__(self, configure_logging=True, use_converters=True, use_display_appl
         self._register_singleton(FileSourcePluginLoader, file_source_plugin_loader)
         self.file_source_templates = self._register_singleton(ConfiguredFileSourceTemplates, templates)
         self._register_singleton(
-            UserDefinedFileSourcesConfig, UserDefinedFileSourcesConfig.from_app_config(self.config)
+            UserDefinedFileSourcesConfig,
+            UserDefinedFileSourcesConfig.from_app_config(self.config),
         )
         user_defined_file_sources = self._register_abstract_singleton(
-            UserDefinedFileSources, UserDefinedFileSourcesImpl  # type: ignore[type-abstract]  # https://github.com/python/mypy/issues/4717
+            UserDefinedFileSources,  # type: ignore[type-abstract]  # https://github.com/python/mypy/issues/4717
+            UserDefinedFileSourcesImpl,
         )
         configured_file_source_conf: ConfiguredFileSourcesConf = ConfiguredFileSourcesConf.from_app_config(self.config)
         file_sources = ConfiguredFileSources(
@@ -690,7 +737,8 @@ def __init__(self, configure_logging=True, use_converters=True, use_display_appl
         # Load security policy.
         self.security_agent = self.model.security_agent
         self.host_security_agent = galaxy.model.security.HostAgent(
-            self.security_agent.sa_session, permitted_actions=self.security_agent.permitted_actions
+            self.security_agent.sa_session,
+            permitted_actions=self.security_agent.permitted_actions,
         )
 
         # We need the datatype registry for running certain tasks that modify HDAs, and to build the registry we need
@@ -791,6 +839,7 @@ def __init__(self, **kwargs) -> None:
             ("queue worker", self._shutdown_queue_worker),
             ("file watcher", self._shutdown_watcher),
             ("database heartbeat", self._shutdown_database_heartbeat),
+            ("history audit monitor", self._shutdown_history_audit_monitor),
             ("workflow scheduler", self._shutdown_scheduling_manager),
             ("object store", self._shutdown_object_store),
             ("job manager", self._shutdown_job_manager),
@@ -805,21 +854,33 @@ def __init__(self, **kwargs) -> None:
         # A lot of postfork initialization depends on the server name, ensure it is set immediately after forking before other postfork functions
         self.application_stack.register_postfork_function(self.application_stack.set_postfork_server_name, self)
         self.config.reload_sanitize_allowlist(explicit="sanitize_allowlist_file" in kwargs)
-        self.amqp_internal_connection_obj = galaxy.queues.connection_from_config(self.config)
-        # queue_worker *can* be initialized with a queue, but here we don't
-        # want to and we'll allow postfork to bind and start it.
-        self.queue_worker = self._register_singleton(GalaxyQueueWorker, GalaxyQueueWorker(self))
+        # amqp_internal_connection_obj and queue_worker are built in GalaxyManagerApplication
+        # (so Celery workers also get a publisher); here we only register the consumer path,
+        # which is started later via the application_stack postfork hook.
+        # SSE connection manager for real-time notification push.
+        # Consumed via ``depends(SSEConnectionManager)`` / ``app[SSEConnectionManager]``,
+        # so no module-level attribute is needed — keep the container wiring only.
+        self._register_singleton(
+            SSEConnectionManager,
+            SSEConnectionManager(
+                statsd_client=self.execution_timer_factory.galaxy_statsd_client,
+            ),
+        )
 
         # AI agent registry and service
         agent_registry = build_agent_registry(self.config)
         self._register_singleton(AgentRegistry, agent_registry)
-        self._register_singleton(AgentService, AgentService(self.config, JobQueryManager(self), agent_registry))
+        self._register_singleton(
+            AgentService,
+            AgentService(self.config, JobQueryManager(self), agent_registry),
+        )
 
         self.dependency_resolvers_view = self._register_singleton(
             DependencyResolversView, DependencyResolversView(self)
         )
         self.test_data_resolver = self._register_singleton(
-            TestDataResolver, TestDataResolver(file_dirs=self.config.tool_test_data_directories)
+            TestDataResolver,
+            TestDataResolver(file_dirs=self.config.tool_test_data_directories),
         )
         self.api_keys_manager = self._register_singleton(ApiKeyManager)
 
@@ -876,7 +937,9 @@ def __init__(self, **kwargs) -> None:
         # Start the heartbeat process if configured and available
         if self.config.use_heartbeat:
             self.heartbeat = heartbeat.Heartbeat(
-                self.config, period=self.config.heartbeat_interval, fname=self.config.heartbeat_log
+                self.config,
+                period=self.config.heartbeat_interval,
+                fname=self.config.heartbeat_log,
             )
             self.heartbeat.daemon = True
             self.application_stack.register_postfork_function(self.heartbeat.start)
@@ -886,7 +949,9 @@ def __init__(self, **kwargs) -> None:
             from galaxy.authnz import managers
 
             self.authnz_manager = managers.AuthnzManager(
-                self, self.config.oidc_config_file, self.config.oidc_backends_config_file
+                self,
+                self.config.oidc_config_file,
+                self.config.oidc_backends_config_file,
             )
 
             # If there is only a single external authentication provider in use
@@ -923,7 +988,12 @@ def __init__(self, **kwargs) -> None:
                 self.workflow_completion_hook_registry,
             )
             self.application_stack.register_postfork_function(self.workflow_completion_monitor.start)
-            self.haltables.append(("WorkflowCompletionMonitor", self.workflow_completion_monitor.shutdown_monitor))
+            self.haltables.append(
+                (
+                    "WorkflowCompletionMonitor",
+                    self.workflow_completion_monitor.shutdown_monitor,
+                )
+            )
 
         # Start the job manager
         self.application_stack.register_postfork_function(self.job_manager.start)
@@ -942,12 +1012,23 @@ def __init__(self, **kwargs) -> None:
         self.database_heartbeat.add_change_callback(self.watchers.change_state)
         self.application_stack.register_postfork_function(self.database_heartbeat.start)
 
+        # History audit monitor for SSE-based history updates. The monitor only
+        # runs on the single process elected via DatabaseHeartbeat's
+        # is_history_audit_monitor role — a standalone ``galaxy-sse-monitor``
+        # daemon wins that election when present, otherwise one webapp picks it
+        # up. start/stop are driven by heartbeat role transitions rather than
+        # postfork, so the monitor cleanly migrates when the leader dies.
+        if self.config.enable_sse_updates:
+            monitor = self._register_singleton(HistoryAuditMonitor)
+            self.database_heartbeat.add_audit_monitor_change_callback(monitor.on_role_change)
+
         # Start web stack message handling
         self.application_stack.register_postfork_function(self.application_stack.start)
         self.application_stack.register_postfork_function(self.queue_worker.bind_and_start)
         # Reload toolbox to pick up changes to toolbox made after master was ready
         self.application_stack.register_postfork_function(
-            lambda: reload_toolbox(self, save_integrated_tool_panel=False), post_fork_only=True
+            lambda: reload_toolbox(self, save_integrated_tool_panel=False),
+            post_fork_only=True,
         )
         # Delay toolbox index until after startup
         self.application_stack.register_postfork_function(
@@ -975,6 +1056,11 @@ def _shutdown_watcher(self):
     def _shutdown_database_heartbeat(self):
         self.database_heartbeat.shutdown()
 
+    def _shutdown_history_audit_monitor(self):
+        if not self.config.enable_sse_updates:
+            return
+        self[HistoryAuditMonitor].shutdown()
+
     def _shutdown_scheduling_manager(self):
         self.workflow_scheduling_manager.shutdown()
 
diff --git a/lib/galaxy/app_unittest_utils/galaxy_mock.py b/lib/galaxy/app_unittest_utils/galaxy_mock.py
index 82a0b7886655..9f4fe40139e4 100644
--- a/lib/galaxy/app_unittest_utils/galaxy_mock.py
+++ b/lib/galaxy/app_unittest_utils/galaxy_mock.py
@@ -119,6 +119,7 @@ class MockApp(di.Container, GalaxyDataTestApp):
     history_manager: HistoryManager
     job_metrics: JobMetrics
     vault: Optional[Vault] = None
+    execution_timer_factory: Any
     stop: bool
     is_webapp: bool = True
 
@@ -159,7 +160,7 @@ def __init__(self, config=None, **kwargs) -> None:
         self.application_stack = ApplicationStack()
         self.auth_manager = AuthManager(self.config)
         self.user_manager = UserManager(cast(BasicSharedApp, self))
-        self.execution_timer_factory = Bunch(get_timer=StructuredExecutionTimer)
+        self.execution_timer_factory = Bunch(get_timer=StructuredExecutionTimer, galaxy_statsd_client=None)
         self.interactivetool_manager = Bunch(create_interactivetool=lambda *args, **kwargs: None)
         self.is_job_handler = False
         self.biotools_metadata_source = None
diff --git a/lib/galaxy/authnz/psa_authnz.py b/lib/galaxy/authnz/psa_authnz.py
index 5d1189b38370..022272e3d9de 100644
--- a/lib/galaxy/authnz/psa_authnz.py
+++ b/lib/galaxy/authnz/psa_authnz.py
@@ -798,7 +798,6 @@ def _send_oidc_profile_update_notification(trans, user, updates: list[str]) -> N
             NotificationVariant,
             PersonalNotificationCategory,
         )
-        from galaxy.webapps.galaxy.services.notifications import NotificationService
 
         labels: dict[str, str] = {
             "email": "email address",
@@ -819,7 +818,7 @@ def _send_oidc_profile_update_notification(trans, user, updates: list[str]) -> N
             ),
             galaxy_url=None,
         )
-        NotificationService(trans.app.notification_manager).send_notification_internal(request, force_sync=True)
+        trans.app.notification_manager.send_notification_internal(request, force_sync=True)
     except Exception as exc:
         log.warning("OIDC profile update notification failed for user %s: %s", user.id, exc)
 
diff --git a/lib/galaxy/celery/__init__.py b/lib/galaxy/celery/__init__.py
index 67917cca6580..c9e8cb96daab 100644
--- a/lib/galaxy/celery/__init__.py
+++ b/lib/galaxy/celery/__init__.py
@@ -251,6 +251,10 @@ def schedule_task(task, interval):
     beat_schedule: dict[str, dict[str, Any]] = {}
     schedule_task("prune_history_audit_table", config.history_audit_table_prune_interval)
     schedule_task("cleanup_short_term_storage", config.short_term_storage_cleanup_interval)
+    schedule_task("prune_kombu_sqla_transport", config.kombu_sqla_transport_cleanup_interval)
+
+    if config.statsd_host:
+        schedule_task("emit_queue_metrics_task", config.queue_metrics_interval)
 
     if config.enable_notification_system:
         schedule_task("cleanup_expired_notifications", config.expired_notifications_cleanup_interval)
diff --git a/lib/galaxy/celery/helpers.py b/lib/galaxy/celery/helpers.py
new file mode 100644
index 000000000000..e771e340cb1d
--- /dev/null
+++ b/lib/galaxy/celery/helpers.py
@@ -0,0 +1,25 @@
+from celery.result import AsyncResult
+
+from galaxy.schema.schema import AsyncTaskResultSummary
+
+
+def async_task_summary(async_result: AsyncResult) -> AsyncTaskResultSummary:
+    name = None
+    try:
+        name = async_result.name
+    except AttributeError:
+        # if backend is disabled, we won't have this
+        pass
+    queue = None
+    try:
+        queue = async_result.queue
+    except AttributeError:
+        # if backend is disabled, we won't have this
+        pass
+
+    return AsyncTaskResultSummary(
+        id=str(async_result.id),
+        ignored=async_result.ignored,
+        name=name,
+        queue=queue,
+    )
diff --git a/lib/galaxy/celery/tasks.py b/lib/galaxy/celery/tasks.py
index b630fdce26a7..c5f5f3e064ff 100644
--- a/lib/galaxy/celery/tasks.py
+++ b/lib/galaxy/celery/tasks.py
@@ -9,13 +9,17 @@
     Any,
     Optional,
 )
+from urllib.parse import urlparse
 
+from lagom.exceptions import UnresolvableType
 from sqlalchemy import (
     and_,
+    create_engine,
     delete,
     exists,
     false,
     select,
+    text,
     update,
 )
 
@@ -40,6 +44,8 @@
 from galaxy.managers.markdown_util import generate_branded_pdf
 from galaxy.managers.model_stores import ModelStoreManager
 from galaxy.managers.notification import NotificationManager
+from galaxy.managers.queue_metrics import emit_queue_metrics
+from galaxy.managers.sse import SSEConnectionManager
 from galaxy.managers.tool_data import ToolDataImportManager
 from galaxy.managers.workflow_completion import WorkflowCompletionManager
 from galaxy.metadata.set_metadata import set_metadata_portable
@@ -76,7 +82,10 @@
     Vault,
 )
 from galaxy.short_term_storage import ShortTermStorageMonitor
-from galaxy.structured_app import MinimalManagerApp
+from galaxy.structured_app import (
+    MinimalManagerApp,
+    StructuredApp,
+)
 from galaxy.tools import create_tool_from_representation
 from galaxy.tools.data_fetch import do_fetch
 from galaxy.util import galaxy_directory
@@ -591,6 +600,38 @@ def prune_history_audit_table(sa_session: galaxy_scoped_session):
     model.HistoryAudit.prune(sa_session)
 
 
+@galaxy_task(action="cleaning up Kombu SQLAlchemy transport")
+def prune_kombu_sqla_transport(config: GalaxyAppConfiguration):
+    """Delete fully-consumed rows from the Kombu SQLAlchemy transport tables.
+
+    Kombu's SQLAlchemy transport marks consumed messages with ``visible=0`` but
+    never deletes them — without this task the control-queue tables grow
+    without bound on the default on-disk sqlite broker. On AMQP / Redis the
+    broker has native TTL, so this task is a no-op.
+    """
+    broker_url = config.amqp_internal_connection
+    if not broker_url:
+        log.debug("kombu cleanup: no broker URL configured, skipping")
+        return
+    scheme = urlparse(broker_url).scheme
+    if not scheme.startswith("sqlalchemy"):
+        log.debug("kombu cleanup: broker scheme %s is not sqlalchemy, skipping", scheme)
+        return
+
+    # Kombu's SQLA transport URL is ``sqlalchemy+<dialect>://...``. Strip the
+    # ``sqlalchemy+`` prefix to get an engine URL we can hand to SQLAlchemy.
+    sa_url = broker_url[len("sqlalchemy+") :] if broker_url.startswith("sqlalchemy+") else broker_url
+    engine = create_engine(sa_url)
+    try:
+        with engine.begin() as conn:
+            result = conn.execute(text("DELETE FROM kombu_message WHERE visible = 0"))
+            log.info("kombu cleanup: deleted %s consumed messages", result.rowcount)
+    except Exception:
+        log.exception("kombu cleanup: failed to prune kombu_message")
+    finally:
+        engine.dispose()
+
+
 @galaxy_task(action="clean up short term storage")
 def cleanup_short_term_storage(storage_monitor: ShortTermStorageMonitor):
     """Cleanup short term storage."""
@@ -628,6 +669,28 @@ def dispatch_pending_notifications(notification_manager: NotificationManager):
         log.info(f"Successfully dispatched {count} notifications.")
 
 
+@galaxy_task(action="emit queue and SSE observability metrics")
+def emit_queue_metrics_task(app: StructuredApp):
+    """Sample control-queue depth, SSE connection count, and worker rows → statsd.
+
+    Resolves the narrow collaborators ``emit_queue_metrics`` needs from the app
+    container and passes them in — keeps the emitter module free of
+    ``StructuredApp`` service-locator lookups.
+    """
+    try:
+        sse_manager: Optional[SSEConnectionManager] = app[SSEConnectionManager]
+    except UnresolvableType:
+        sse_manager = None
+
+    emit_queue_metrics(
+        statsd_client=app.execution_timer_factory.galaxy_statsd_client,
+        connection=app.amqp_internal_connection_obj,
+        application_stack=app.application_stack,
+        model=app.model,
+        sse_manager=sse_manager,
+    )
+
+
 @galaxy_task(action="clean up job working directories")
 def cleanup_jwds(sa_session: galaxy_scoped_session, object_store: BaseObjectStore, config: GalaxyAppConfiguration):
     """Cleanup job working directories for failed jobs that are older than X days"""
diff --git a/lib/galaxy/config/sample/galaxy.yml.sample b/lib/galaxy/config/sample/galaxy.yml.sample
index 524ad8dcf774..0f08881b5c74 100644
--- a/lib/galaxy/config/sample/galaxy.yml.sample
+++ b/lib/galaxy/config/sample/galaxy.yml.sample
@@ -512,6 +512,16 @@ galaxy:
   # history_audit database table. Set to 0 to disable pruning.
   #history_audit_table_prune_interval: 3600
 
+  # Time (in seconds) between attempts to delete fully-consumed rows
+  # from the Kombu SQLAlchemy transport tables (``kombu_message``). Only
+  # relevant when ``amqp_internal_connection`` uses a ``sqlalchemy+*``
+  # scheme (the default with an on-disk control.sqlite); the SQLAlchemy
+  # transport has no built-in TTL, so without this task the tables grow
+  # unbounded. The task no-ops on non-SQLAlchemy brokers (RabbitMQ/Redis
+  # honor per-message expiration natively). Set to 0 to disable the
+  # cleanup task.
+  #kombu_sqla_transport_cleanup_interval: 900
+
   # Where dataset files are stored. It must be accessible at the same
   # path on any cluster nodes that will run Galaxy jobs, unless using
   # Pulsar. The default value has been changed from 'files' to 'objects'
@@ -1962,6 +1972,11 @@ galaxy:
   # really. Do not set this in production environments.
   #statsd_mock_calls: false
 
+  # How often (in seconds) the Celery beat task emits queue-depth,
+  # SSE-connection, and WorkerProcess gauges. Only active when
+  # statsd_host is set. Set to 0 to disable.
+  #queue_metrics_interval: 15
+
   # Add an option to the library upload form which allows administrators
   # to upload a directory of files.
   #library_import_dir: null
@@ -3129,12 +3144,36 @@ galaxy:
   # Message to display on the export citations tool page
   #citations_export_message_html: When writing up your analysis, remember to include all references that should be cited in order to completely describe your work. Also, please remember to <a href="https://galaxyproject.org/citing-galaxy">cite Galaxy</a>.
 
+  # Enables real-time updates via Server-Sent Events (SSE), replacing
+  # the history (3 s), entry-point (10 s) and notification (30 s)
+  # polling loops with push events delivered over a single
+  # ``/api/events/stream`` connection per browser tab. A background
+  # monitor watches for history changes (via PostgreSQL LISTEN/NOTIFY,
+  # or audit-table polling as a fallback for SQLite); entry-point
+  # changes are dispatched directly from the code paths that mutate
+  # them; in-app notifications and broadcasts are pushed when
+  # ``enable_notification_system`` is also true. When disabled, polling
+  # remains the source of updates for all three. See the admin guide
+  # "Server-Sent Events for real-time updates" for the full
+  # architecture, monitoring guidance and proxy configuration.
+  #enable_sse_updates: false
+
+  # The interval in seconds between history audit table polls when using
+  # the polling fallback (SQLite or when PostgreSQL LISTEN/NOTIFY is
+  # unavailable). Only used when enable_sse_updates is true. Lower
+  # values mean faster updates but more database queries. Recommended
+  # range: 1-5 seconds.
+  #history_audit_monitor_poll_interval: 2
+
   # Enables the Notification System integrated in Galaxy.
   # Users can receive automatic notifications when a certain resource is
   # shared with them or when some long running operations have finished,
   # etc.
   # The system allows notification scheduling and expiration, and users
   # can opt-out of specific notification categories or channels.
+  # Delivery is push-based via Server-Sent Events when
+  # ``enable_sse_updates`` is also true, and falls back to 30-second
+  # polling against ``/api/notifications/status`` otherwise.
   # Admins can schedule and broadcast notifications that will be visible
   # to all users, including special server-wide announcements such as
   # scheduled maintenance, high load warnings, and event announcements,
diff --git a/lib/galaxy/config/schemas/config_schema.yml b/lib/galaxy/config/schemas/config_schema.yml
index 05c50b6eb390..7b5b21505f88 100644
--- a/lib/galaxy/config/schemas/config_schema.yml
+++ b/lib/galaxy/config/schemas/config_schema.yml
@@ -222,6 +222,18 @@ mapping:
           Time (in seconds) between attempts to remove old rows from the history_audit database table.
           Set to 0 to disable pruning.
 
+      kombu_sqla_transport_cleanup_interval:
+        type: int
+        default: 900
+        required: false
+        desc: |
+          Time (in seconds) between attempts to delete fully-consumed rows from the Kombu SQLAlchemy
+          transport tables (``kombu_message``). Only relevant when ``amqp_internal_connection`` uses a
+          ``sqlalchemy+*`` scheme (the default with an on-disk control.sqlite); the SQLAlchemy transport
+          has no built-in TTL, so without this task the tables grow unbounded. The task no-ops on
+          non-SQLAlchemy brokers (RabbitMQ/Redis honor per-message expiration natively). Set to 0 to
+          disable the cleanup task.
+
       file_path:
         type: str
         default: objects
@@ -2512,6 +2524,14 @@ mapping:
           Mock out statsd client calls - only used by testing infrastructure really.
           Do not set this in production environments.
 
+      queue_metrics_interval:
+        type: int
+        default: 15
+        required: false
+        desc: |
+          How often (in seconds) the Celery beat task emits queue-depth, SSE-connection,
+          and WorkerProcess gauges. Only active when statsd_host is set. Set to 0 to disable.
+
       library_import_dir:
         type: str
         required: false
@@ -4293,6 +4313,33 @@ mapping:
         desc: |
           Message to display on the export citations tool page
 
+      enable_sse_updates:
+        type: bool
+        default: false
+        required: false
+        desc: |
+          Enables real-time updates via Server-Sent Events (SSE), replacing the
+          history (3 s), entry-point (10 s) and notification (30 s) polling loops
+          with push events delivered over a single ``/api/events/stream``
+          connection per browser tab. A background monitor watches for history
+          changes (via PostgreSQL LISTEN/NOTIFY, or audit-table polling as a
+          fallback for SQLite); entry-point changes are dispatched directly from
+          the code paths that mutate them; in-app notifications and broadcasts
+          are pushed when ``enable_notification_system`` is also true.
+          When disabled, polling remains the source of updates for all three.
+          See the admin guide "Server-Sent Events for real-time updates" for the
+          full architecture, monitoring guidance and proxy configuration.
+
+      history_audit_monitor_poll_interval:
+        type: int
+        default: 2
+        required: false
+        desc: |
+          The interval in seconds between history audit table polls when using the polling
+          fallback (SQLite or when PostgreSQL LISTEN/NOTIFY is unavailable). Only used when
+          enable_sse_updates is true. Lower values mean faster updates but more
+          database queries. Recommended range: 1-5 seconds.
+
       enable_notification_system:
         type: bool
         default: false
@@ -4304,6 +4351,10 @@ mapping:
 
           The system allows notification scheduling and expiration, and users can opt-out of specific notification categories or channels.
 
+          Delivery is push-based via Server-Sent Events when ``enable_sse_updates``
+          is also true, and falls back to 30-second polling against
+          ``/api/notifications/status`` otherwise.
+
           Admins can schedule and broadcast notifications that will be visible to all users, including special server-wide announcements such as scheduled maintenance, high load warnings, and event announcements, to name a few examples.
 
       enable_mcp_server:
diff --git a/lib/galaxy/dependencies/pinned-typecheck-requirements.txt b/lib/galaxy/dependencies/pinned-typecheck-requirements.txt
index f0cb63fba34f..cced7deee47c 100644
--- a/lib/galaxy/dependencies/pinned-typecheck-requirements.txt
+++ b/lib/galaxy/dependencies/pinned-typecheck-requirements.txt
@@ -18,6 +18,7 @@ tomli==2.4.1 ; python_full_version < '3.11'
 types-awscrt==0.31.3
 types-bleach==6.3.0.20260408
 types-boto==2.49.18.20241019
+types-cachetools==6.2.0.20260408
 types-contextvars==2.4.7.3
 types-dataclasses==0.6.6
 types-docutils==0.22.3.20260408
diff --git a/lib/galaxy/managers/configuration.py b/lib/galaxy/managers/configuration.py
index bf49cce94d59..b85d474c02ba 100644
--- a/lib/galaxy/managers/configuration.py
+++ b/lib/galaxy/managers/configuration.py
@@ -229,6 +229,7 @@ def _config_is_truthy(item, key, **context):
             "tool_training_recommendations_link": _use_config,
             "tool_training_recommendations_api_url": _use_config,
             "enable_notification_system": _use_config,
+            "enable_sse_updates": _use_config,
             "instance_resource_url": _use_config,
             "instance_access_url": _use_config,
             "organization_name": _use_config,
diff --git a/lib/galaxy/managers/history_audit_monitor.py b/lib/galaxy/managers/history_audit_monitor.py
new file mode 100644
index 000000000000..c98b8e59cc15
--- /dev/null
+++ b/lib/galaxy/managers/history_audit_monitor.py
@@ -0,0 +1,318 @@
+"""Monitor for history audit table changes.
+
+Detects history changes via PostgreSQL LISTEN/NOTIFY (instant) or by polling
+the history_audit table (SQLite fallback). Dispatches SSE events to connected
+users via Kombu control tasks.
+
+Only active when ``enable_sse_updates`` is True in the Galaxy config.
+"""
+
+import logging
+import select
+import threading
+import time
+from collections import (
+    defaultdict,
+    OrderedDict,
+)
+from collections.abc import Iterator
+from datetime import (
+    datetime,
+    timedelta,
+)
+from typing import (
+    Any,
+    Optional,
+)
+
+from sqlalchemy import select as sa_select
+from sqlalchemy.engine import Engine
+
+from galaxy.config import GalaxyAppConfiguration
+from galaxy.managers.sse_dispatch import SSEEventDispatcher
+from galaxy.model import (
+    GalaxySessionToHistoryAssociation,
+    History,
+    HistoryAudit,
+)
+from galaxy.model.mapping import GalaxyModelMapping
+
+log = logging.getLogger(__name__)
+
+CHANNEL_NAME = "galaxy_history_update"
+OWNER_CACHE_MAX = 10_000
+DEBOUNCE_SECONDS = 0.2
+
+
+class _PgListenAdapter:
+    """Thin DBAPI-level adapter for PostgreSQL LISTEN/NOTIFY.
+
+    Hides the receiving-API differences between psycopg2 (``conn.poll()`` +
+    ``conn.notifies`` list, driven by ``select.select``) and psycopg3
+    (``conn.notifies(timeout=...)`` generator). The SA URL is used to inherit
+    DSN / SSL / auth config, but the connection itself is opened directly with
+    the DBAPI driver so it stays outside the SA pool — LISTEN connections must
+    live for the lifetime of the monitor and never be returned to the pool.
+    """
+
+    # Typed once at the class level so both driver branches can assign without
+    # re-annotating the same attribute.
+    _conn: Any
+    driver: str
+
+    def __init__(self, engine: Engine) -> None:
+        # Strip the SA ``+driver`` suffix so the raw DBAPI libraries accept the URL.
+        dsn = engine.url.set(drivername="postgresql").render_as_string(hide_password=False)
+        driver = engine.dialect.driver
+        if driver == "psycopg":
+            import psycopg  # conditional: psycopg3 driver
+
+            self._conn = psycopg.connect(dsn, autocommit=True)
+            self.driver = "psycopg3"
+        else:
+            import psycopg2  # conditional: psycopg2 driver
+
+            self._conn = psycopg2.connect(dsn)
+            self._conn.autocommit = True  # same effect as set_isolation_level(AUTOCOMMIT)
+            self.driver = "psycopg2"
+
+    def listen(self, channel: str) -> None:
+        with self._conn.cursor() as cursor:
+            cursor.execute(f"LISTEN {channel};")
+
+    def close(self) -> None:
+        try:
+            self._conn.close()
+        except Exception:
+            log.debug("Error closing LISTEN connection", exc_info=True)
+
+    def poll(self, timeout: float) -> Iterator[str]:
+        """Block up to ``timeout`` seconds and yield notification payloads.
+
+        Returns an empty iterator on timeout so callers can uniformly treat
+        "nothing received in this tick" regardless of driver.
+        """
+        if self.driver == "psycopg3":
+            # psycopg3: notifies() is a blocking generator bounded by ``timeout``.
+            yield from (n.payload for n in self._conn.notifies(timeout=timeout))
+            return
+        # psycopg2: block on the socket via select(), then drain notifies list.
+        if select.select([self._conn], [], [], timeout) == ([], [], []):
+            return
+        self._conn.poll()
+        while self._conn.notifies:
+            yield self._conn.notifies.pop(0).payload
+
+
+class HistoryAuditMonitor:
+    """Background thread that monitors history_audit for changes and dispatches SSE events.
+
+    On PostgreSQL: uses LISTEN/NOTIFY for instant notification.
+    On SQLite: polls history_audit table at a configurable interval.
+    """
+
+    def __init__(
+        self,
+        config: GalaxyAppConfiguration,
+        model: GalaxyModelMapping,
+        sse_dispatcher: SSEEventDispatcher,
+    ) -> None:
+        self._config = config
+        self._model = model
+        self._dispatcher = sse_dispatcher
+        self.poll_interval: int = config.history_audit_monitor_poll_interval
+        self._is_postgres: bool = "postgres" in model.engine.name
+        self._exit = threading.Event()
+        self._thread: Optional[threading.Thread] = None
+        self._active = False
+        # Bounded LRU cache: history_id -> (user_id, session_ids), refreshed on miss.
+        # For registered-owned histories: (user_id, ()); for anonymous histories:
+        # (None, (session_id, ...)) — a history can be associated with multiple
+        # sessions via GalaxySessionToHistoryAssociation.
+        self._history_owner_cache: OrderedDict[int, tuple[Optional[int], tuple[int, ...]]] = OrderedDict()
+
+    def start(self) -> None:
+        if self._active:
+            return
+        self._active = True
+        self._exit.clear()  # allow restart after a previous shutdown
+        target = self._listen_postgres if self._is_postgres else self._poll_audit_table
+        self._thread = threading.Thread(
+            target=target,
+            name="history_audit_monitor",
+            daemon=True,
+        )
+        self._thread.start()
+        log.info(
+            "HistoryAuditMonitor started (mode=%s, interval=%ds)",
+            "pg_listen" if self._is_postgres else "poll",
+            self.poll_interval,
+        )
+
+    def shutdown(self) -> None:
+        if not self._active:
+            return
+        self._active = False
+        self._exit.set()
+        if self._thread:
+            self._thread.join(timeout=5)
+        self._thread = None
+        log.info("HistoryAuditMonitor stopped")
+
+    def on_role_change(self, is_leader: bool) -> None:
+        """Heartbeat callback: start/stop the monitor as this process's election state changes."""
+        if is_leader:
+            self.start()
+        else:
+            self.shutdown()
+
+    # --- PostgreSQL LISTEN/NOTIFY mode ---
+
+    def _listen_postgres(self) -> None:
+        """LISTEN for history update notifications.
+
+        Works against both psycopg2 and psycopg3 — whichever driver the SA
+        engine was built with. Falls back to the SQLite polling path if the
+        DBAPI driver can't be imported or the initial LISTEN fails.
+        """
+        try:
+            adapter = _PgListenAdapter(self._model.engine)
+            adapter.listen(CHANNEL_NAME)
+        except Exception:
+            log.warning(
+                "Failed to establish PostgreSQL LISTEN connection, falling back to polling",
+                exc_info=True,
+            )
+            self._poll_audit_table()
+            return
+
+        log.debug("LISTEN %s established (driver=%s)", CHANNEL_NAME, adapter.driver)
+        pending: dict[int, float] = {}  # history_id -> first_seen_time
+
+        try:
+            while not self._exit.is_set():
+                received_any = False
+                for payload in adapter.poll(self.poll_interval):
+                    received_any = True
+                    try:
+                        history_id = int(payload)
+                    except (ValueError, TypeError):
+                        continue
+                    pending.setdefault(history_id, time.monotonic())
+
+                if not received_any:
+                    # Timeout — flush anything that's been pending since last tick
+                    if pending:
+                        self._dispatch_history_updates(set(pending.keys()))
+                        pending.clear()
+                    continue
+
+                # Debounce: dispatch events that have been pending long enough
+                now = time.monotonic()
+                ready = {hid for hid, ts in pending.items() if now - ts >= DEBOUNCE_SECONDS}
+                if ready:
+                    self._dispatch_history_updates(ready)
+                    for hid in ready:
+                        del pending[hid]
+        except Exception:
+            log.exception("HistoryAuditMonitor LISTEN loop error")
+        finally:
+            adapter.close()
+
+    # --- SQLite polling fallback ---
+
+    def _poll_audit_table(self) -> None:
+        """Poll history_audit for recent changes."""
+        last_check = datetime.utcnow() - timedelta(seconds=self.poll_interval)
+
+        while not self._exit.is_set():
+            try:
+                check_time = datetime.utcnow()
+                stmt = (
+                    sa_select(HistoryAudit.history_id)
+                    .where(HistoryAudit.update_time > last_check)
+                    .group_by(HistoryAudit.history_id)
+                )
+                with self._model.new_session() as session:
+                    changed_ids = set(session.scalars(stmt).all())
+
+                if changed_ids:
+                    self._dispatch_history_updates(changed_ids)
+
+                last_check = check_time
+            except Exception:
+                log.exception("HistoryAuditMonitor poll error")
+
+            self._exit.wait(self.poll_interval)
+
+    # --- Common dispatch logic ---
+
+    def _dispatch_history_updates(self, history_ids: set[int]) -> None:
+        """Map history_ids to user_ids / session_ids and send Kombu control task.
+
+        Raw integer history IDs are sent across the control queue; encoding is
+        deferred to the ``history_update`` task handler on the receiving side,
+        keeping this manager free of presentation concerns.
+        """
+        # Resolve owners for unknown history_ids
+        unknown = history_ids - self._history_owner_cache.keys()
+        if unknown:
+            self._refresh_owner_cache(unknown)
+
+        user_updates: dict[str, list[int]] = defaultdict(list)
+        session_updates: dict[str, list[int]] = defaultdict(list)
+        for history_id in history_ids:
+            entry = self._history_owner_cache.get(history_id)
+            if entry is None:
+                continue
+            user_id, session_ids = entry
+            if user_id is not None:
+                user_updates[str(user_id)].append(history_id)
+            else:
+                for session_id in session_ids:
+                    session_updates[str(session_id)].append(history_id)
+
+        if not user_updates and not session_updates:
+            return
+
+        self._dispatcher.history_update(
+            user_updates=dict(user_updates),
+            session_updates=dict(session_updates) if session_updates else None,
+        )
+
+    def _refresh_owner_cache(self, history_ids: set[int]) -> None:
+        """Look up ownership for given history_ids and update the bounded cache.
+
+        Registered-owned histories resolve with just ``History.user_id``. For
+        histories where ``user_id IS NULL`` we additionally fetch associated
+        ``galaxy_session.id`` values from ``GalaxySessionToHistoryAssociation``
+        so the anonymous SSE dispatch path can target the right browser.
+        """
+        try:
+            with self._model.new_session() as session:
+                stmt = sa_select(History.id, History.user_id).where(History.id.in_(history_ids))
+                anon_history_ids: set[int] = set()
+                for row in session.execute(stmt):
+                    hid, uid = row[0], row[1]
+                    self._history_owner_cache[hid] = (uid, ())
+                    self._history_owner_cache.move_to_end(hid)
+                    if uid is None:
+                        anon_history_ids.add(hid)
+
+                if anon_history_ids:
+                    assoc_stmt = sa_select(
+                        GalaxySessionToHistoryAssociation.history_id,
+                        GalaxySessionToHistoryAssociation.session_id,
+                    ).where(GalaxySessionToHistoryAssociation.history_id.in_(anon_history_ids))
+                    sessions_by_history: dict[int, list[int]] = defaultdict(list)
+                    for row in session.execute(assoc_stmt):
+                        hid, sid = row[0], row[1]
+                        if sid is not None:
+                            sessions_by_history[hid].append(sid)
+                    for hid, sids in sessions_by_history.items():
+                        self._history_owner_cache[hid] = (None, tuple(sids))
+
+            while len(self._history_owner_cache) > OWNER_CACHE_MAX:
+                self._history_owner_cache.popitem(last=False)
+        except Exception:
+            log.debug("Failed to refresh history owner cache", exc_info=True)
diff --git a/lib/galaxy/managers/interactivetool.py b/lib/galaxy/managers/interactivetool.py
index 009d8f8dc36f..21b872f146ef 100644
--- a/lib/galaxy/managers/interactivetool.py
+++ b/lib/galaxy/managers/interactivetool.py
@@ -6,6 +6,7 @@
 )
 from typing import (
     Any,
+    Optional,
     TYPE_CHECKING,
     Union,
 )
@@ -28,6 +29,7 @@
 )
 
 from galaxy import exceptions
+from galaxy.managers.sse_dispatch import SSEEventDispatcher
 from galaxy.model import (
     InteractiveToolEntryPoint,
     Job,
@@ -147,7 +149,11 @@ class InteractiveToolManager:
     Manager for dealing with InteractiveTools
     """
 
-    def __init__(self, app: "MinimalManagerApp") -> None:
+    def __init__(
+        self,
+        app: "MinimalManagerApp",
+        dispatcher: Optional[SSEEventDispatcher] = None,
+    ) -> None:
         self.app = app
         self.security = app.security
         self.sa_session = app.model.context
@@ -157,6 +163,12 @@ def __init__(self, app: "MinimalManagerApp") -> None:
             app.config.interactivetoolsproxy_map or app.config.interactivetools_map,
             self.encoder.encode_id,
         )
+        # Lagom can't auto-inject ``SSEEventDispatcher`` here because the
+        # ``app: "MinimalManagerApp"`` hint is only a forward reference
+        # (TYPE_CHECKING import), so ``get_type_hints`` on this signature
+        # fails. Resolve through the container explicitly — ``resolve_or_none``
+        # returns ``None`` for mocks/test apps that never registered one.
+        self.dispatcher = dispatcher if dispatcher is not None else app.resolve_or_none(SSEEventDispatcher)
 
     def create_entry_points(
         self, job: Job, tool: "Tool", entry_points=Union[Iterable[dict[str, Any]], None], flush: bool = True
@@ -198,6 +210,17 @@ def configure_entry_points(
                 configured.append(ep)
         if configured:
             self.sa_session.commit()
+            # Fan out an SSE push so the user's browser can refresh the entry
+            # point list immediately instead of waiting for the 10 s poll.
+            # Anonymous jobs fall back to polling — ``push_to_user`` keys on
+            # user_id, and anonymous clients sit in the broadcast-only set.
+            if self.dispatcher is not None and job.user_id is not None:
+                try:
+                    self.dispatcher.entry_point_update(user_id=job.user_id)
+                except Exception:
+                    # The DB commit is authoritative; the SSE event is best
+                    # effort. Never let a dispatch failure poison the caller.
+                    log.exception("Failed to dispatch entry_point_update SSE event for job %s", job.id)
         return dict(not_configured=not_configured, configured=configured)
 
     def save_entry_point(self, entry_point: InteractiveToolEntryPoint) -> None:
diff --git a/lib/galaxy/managers/notification.py b/lib/galaxy/managers/notification.py
index f2c654ca6f97..18de7e795a59 100644
--- a/lib/galaxy/managers/notification.py
+++ b/lib/galaxy/managers/notification.py
@@ -5,6 +5,7 @@
     cast,
     NamedTuple,
     Optional,
+    Union,
 )
 from urllib.parse import urlparse
 
@@ -28,6 +29,7 @@
 from typing_extensions import Protocol
 
 from galaxy import util
+from galaxy.celery.helpers import async_task_summary
 from galaxy.config import (
     GalaxyAppConfiguration,
     templates,
@@ -37,6 +39,7 @@
     ObjectNotFound,
 )
 from galaxy.managers.markdown_util import to_html
+from galaxy.managers.sse_dispatch import SSEEventDispatcher
 from galaxy.model import (
     GroupRoleAssociation,
     Notification,
@@ -49,6 +52,7 @@
 from galaxy.schema.notifications import (
     AnyNotificationContent,
     BroadcastNotificationCreateRequest,
+    BroadcastNotificationResponse,
     MandatoryNotificationCategory,
     MessageNotificationContent,
     NewSharedItemNotificationContent,
@@ -56,14 +60,17 @@
     NotificationCategorySettings,
     NotificationChannelSettings,
     NotificationCreateData,
+    NotificationCreatedResponse,
     NotificationCreateRequest,
     NotificationRecipients,
+    NotificationResponse,
     NotificationVariant,
     PersonalNotificationCategory,
     UpdateUserNotificationPreferencesRequest,
     UserNotificationPreferences,
     UserNotificationUpdateRequest,
 )
+from galaxy.schema.schema import AsyncTaskResultSummary
 
 log = logging.getLogger(__name__)
 
@@ -94,9 +101,15 @@ def send(self, notification: Notification, user: User):
 class NotificationManager:
     """Manager class to interact with the database models related with Notifications."""
 
-    def __init__(self, sa_session: galaxy_scoped_session, config: GalaxyAppConfiguration):
+    def __init__(
+        self,
+        sa_session: galaxy_scoped_session,
+        config: GalaxyAppConfiguration,
+        sse_dispatcher: Optional[SSEEventDispatcher] = None,
+    ):
         self.sa_session = sa_session
         self.config = config
+        self.sse_dispatcher = sse_dispatcher
         self.recipient_resolver = NotificationRecipientResolver(strategy=DefaultStrategy(sa_session))
         self.user_notification_columns: list[InstrumentedAttribute] = [
             Notification.id,
@@ -164,8 +177,36 @@ def send_notification_to_recipients(self, request: NotificationCreateRequest) ->
         notifications_sent = self._create_associations(notification, recipient_users)
         self.sa_session.commit()
 
+        # Push SSE events to connected users via control queue
+        user_ids = [user.id for user in recipient_users]
+        self._notify_users_via_sse(user_ids, notification)
+
         return notification, notifications_sent
 
+    def send_notification_internal(
+        self, request: NotificationCreateRequest, force_sync: bool = False
+    ) -> Union[NotificationCreatedResponse, AsyncTaskResultSummary]:
+        """Sends a notification to a list of recipients (users, groups or roles).
+
+        If `force_sync` is set to `True`, the notification recipients will be processed synchronously instead of
+        in a background task.
+
+        Note: This function is meant for internal use from other callers that don't need to check sender permissions.
+        """
+        if self.can_send_notifications_async and not force_sync:
+            # Local import: galaxy.celery.tasks imports NotificationManager at module load,
+            # so importing it at module level here would be a circular dependency.
+            from galaxy.celery.tasks import send_notification_to_recipients_async
+
+            result = send_notification_to_recipients_async.delay(request)
+            return async_task_summary(result)
+
+        notification, recipient_user_count = self.send_notification_to_recipients(request)
+        return NotificationCreatedResponse(
+            total_notifications_sent=recipient_user_count,
+            notification=NotificationResponse.model_validate(notification),
+        )
+
     def _create_associations(self, notification: Notification, users: list[User]) -> int:
         success_count = 0
         for user in users:
@@ -179,6 +220,26 @@ def _create_associations(self, notification: Notification, users: list[User]) ->
                 continue
         return success_count
 
+    def _notify_users_via_sse(self, user_ids: list[int], notification: Notification) -> None:
+        """Broadcast a control task to all workers to push SSE events to connected users."""
+        if not self.sse_dispatcher or not user_ids:
+            return
+        try:
+            payload = NotificationResponse.model_validate(notification).model_dump_json()
+            self.sse_dispatcher.notify_users(user_ids, payload)
+        except Exception:
+            log.warning("Failed to send SSE notification event", exc_info=True)
+
+    def _notify_broadcast_via_sse(self, notification: Notification) -> None:
+        """Broadcast a control task to all workers to push SSE broadcast events."""
+        if not self.sse_dispatcher:
+            return
+        try:
+            payload = BroadcastNotificationResponse.model_validate(notification).model_dump_json()
+            self.sse_dispatcher.notify_broadcast(payload)
+        except Exception:
+            log.warning("Failed to send SSE broadcast event", exc_info=True)
+
     def dispatch_pending_notifications_via_channels(self) -> int:
         """
         Dispatches all pending notifications to the users depending on the configured channels.
@@ -273,6 +334,7 @@ def create_broadcast_notification(self, request: BroadcastNotificationCreateRequ
         notification = self._create_notification_model(request)
         self.sa_session.add(notification)
         self.sa_session.commit()
+        self._notify_broadcast_via_sse(notification)
         return notification
 
     def get_user_notification(self, user: User, notification_id: int, active_only: Optional[bool] = True):
@@ -353,7 +415,10 @@ def get_all_broadcasted_notifications(self, since: Optional[datetime] = None, ac
         return result
 
     def update_user_notifications(
-        self, user: User, notification_ids: set[int], request: UserNotificationUpdateRequest
+        self,
+        user: User,
+        notification_ids: set[int],
+        request: UserNotificationUpdateRequest,
     ) -> int:
         """Updates a batch of notifications associated with the user using the requested values."""
         updated_row_count = 0
@@ -447,7 +512,8 @@ def cleanup_expired_notifications(self) -> CleanupResultSummary:
             UserNotificationAssociation.notification_id.in_(expired_notifications_stmt)
         )
         result = cast(
-            CursorResult, self.sa_session.execute(delete_stmt, execution_options={"synchronize_session": False})
+            CursorResult,
+            self.sa_session.execute(delete_stmt, execution_options={"synchronize_session": False}),
         )
         deleted_associations_count = result.rowcount
 
@@ -474,7 +540,10 @@ def _create_notification_model(
         return notification
 
     def _user_notifications_query(
-        self, user: User, since: Optional[datetime] = None, active_only: Optional[bool] = True
+        self,
+        user: User,
+        since: Optional[datetime] = None,
+        active_only: Optional[bool] = True,
     ):
         stmt = (
             select(*self.user_notification_columns)
@@ -552,7 +621,7 @@ def resolve_users(self, recipients: NotificationRecipients) -> list[User]:
         user_ids_from_roles_stmt = self._get_all_user_ids_from_roles_query(all_role_ids)
 
         union_stmt = union(user_ids_from_groups_stmt, user_ids_from_roles_stmt)
-        user_ids_from_groups_and_roles = {id for id, in self.sa_session.execute(union_stmt)}
+        user_ids_from_groups_and_roles = {id for (id,) in self.sa_session.execute(union_stmt)}
         unique_user_ids.update(user_ids_from_groups_and_roles)
 
         stmt = select(User).where(User.id.in_(unique_user_ids))
@@ -591,7 +660,7 @@ def _expand_group_and_roles_ids(self, group_ids: set[int], role_ids: set[int]) -
                 .where(GroupRoleAssociation.role_id.in_(role_ids))
                 .distinct()
             )
-            group_ids_from_roles = {id for id, in self.sa_session.execute(stmt) if id is not None}
+            group_ids_from_roles = {id for (id,) in self.sa_session.execute(stmt) if id is not None}
             new_group_ids = group_ids_from_roles - processed_group_ids
 
             # Get role IDs associated with any of the given group IDs
@@ -601,7 +670,7 @@ def _expand_group_and_roles_ids(self, group_ids: set[int], role_ids: set[int]) -
                 .where(GroupRoleAssociation.group_id.in_(group_ids))
                 .distinct()
             )
-            role_ids_from_groups = {id for id, in self.sa_session.execute(stmt) if id is not None}
+            role_ids_from_groups = {id for (id,) in self.sa_session.execute(stmt) if id is not None}
             new_role_ids = role_ids_from_groups - processed_role_ids
 
             # Stop if there are no new group or role IDs to process
@@ -713,7 +782,6 @@ def get_body(self, template_format: TemplateFormats) -> str:
 
 
 class MessageEmailNotificationTemplateBuilder(EmailNotificationTemplateBuilder):
-
     markdown_to = {
         TemplateFormats.HTML: to_html,
         TemplateFormats.TXT: lambda x: x,  # TODO: strip markdown?
@@ -730,9 +798,10 @@ def get_subject(self) -> str:
 
 
 class NewSharedItemEmailNotificationTemplateBuilder(EmailNotificationTemplateBuilder):
-
     def get_content(self, template_format: TemplateFormats) -> AnyNotificationContent:
-        content = NewSharedItemNotificationContent.model_construct(**self.notification.content)  # type: ignore[arg-type]
+        content = NewSharedItemNotificationContent.model_construct(
+            **self.notification.content
+        )  # type: ignore[arg-type]
         return content
 
     def get_subject(self) -> str:
@@ -741,7 +810,6 @@ def get_subject(self) -> str:
 
 
 class EmailNotificationChannelPlugin(NotificationChannelPlugin):
-
     # Register the supported email templates here
     email_templates_by_category: dict[PersonalNotificationCategory, type[EmailNotificationTemplateBuilder]] = {
         PersonalNotificationCategory.message: MessageEmailNotificationTemplateBuilder,
diff --git a/lib/galaxy/managers/queue_metrics.py b/lib/galaxy/managers/queue_metrics.py
new file mode 100644
index 000000000000..4bd15878bf83
--- /dev/null
+++ b/lib/galaxy/managers/queue_metrics.py
@@ -0,0 +1,164 @@
+"""Periodic gauge emitter for control-queue depth and SSE-connection counts.
+
+Scheduled by Celery beat (see ``galaxy.celery.__init__.setup_periodic_tasks``)
+at a fixed cadence. Opens a short-lived kombu connection, iterates the control
+queues returned by ``all_control_queues_for_declare`` and samples each queue's
+message-count via a passive declare. Also samples in-memory connection counts
+from ``SSEConnectionManager`` and the active-``WorkerProcess`` count from the
+database.
+
+All instrumentation no-ops when ``statsd_client`` is ``None`` — i.e. statsd
+isn't configured.
+
+The sub-emitters take narrow, typed collaborators (a kombu connection, an
+application stack, a model mapping, the statsd client, an SSE manager) rather
+than the whole ``StructuredApp``. The Celery task in
+``galaxy.celery.tasks.emit_queue_metrics_task`` is the composition root that
+resolves those narrow deps from the app and passes them in.
+"""
+
+import datetime
+import logging
+from collections import defaultdict
+from collections.abc import Callable
+from typing import (
+    Optional,
+    TYPE_CHECKING,
+)
+
+from sqlalchemy import (
+    func,
+    select,
+)
+
+from galaxy.managers.sse import SSEConnectionManager
+from galaxy.model import WorkerProcess
+from galaxy.model.mapping import GalaxyModelMapping
+from galaxy.model.orm.now import now
+from galaxy.queues import (
+    all_control_queues_for_declare,
+    DEFAULT_ACTIVE_PROCESS_WINDOW_SECONDS,
+)
+from galaxy.web_stack import ApplicationStack
+
+if TYPE_CHECKING:
+    from kombu import Connection
+
+    from galaxy.web.statsd_client import VanillaGalaxyStatsdClient
+
+log = logging.getLogger(__name__)
+
+
+def emit_sse_connection_gauges(
+    statsd_client: "VanillaGalaxyStatsdClient",
+    sse_manager: SSEConnectionManager,
+) -> None:
+    """Emit ``galaxy.sse.connections.active`` gauges by connection kind."""
+    statsd_client.timing(
+        "galaxy.sse.connections.active",
+        sse_manager.total_broadcast_connections,
+        tags={"kind": "broadcast"},
+    )
+    statsd_client.timing(
+        "galaxy.sse.connections.active",
+        sse_manager.total_per_user_connections,
+        tags={"kind": "per_user"},
+    )
+
+
+def emit_control_queue_depth(
+    statsd_client: "VanillaGalaxyStatsdClient",
+    connection: "Optional[Connection]",
+    application_stack: ApplicationStack,
+) -> None:
+    """Emit ``galaxy.control_queue.depth`` per active webapp/handler queue.
+
+    A per-queue passive declare can fail on transports that don't implement it
+    (e.g. the sqlalchemy kombu transport) or for queues that don't yet exist on
+    the broker. Those are expected and quiet — logged at DEBUG, no metric, move
+    on. Errors at the broker-connection layer propagate up so the caller can
+    surface them.
+    """
+    if connection is None:
+        return
+    queues = all_control_queues_for_declare(application_stack)
+    if not queues:
+        return
+    with connection.clone() as conn:
+        channel = conn.channel()
+        try:
+            for queue in queues:
+                try:
+                    declared = queue.bind(channel).queue_declare(passive=True)
+                except Exception:
+                    log.debug(
+                        "queue_metrics: passive declare failed for %s",
+                        queue.name,
+                        exc_info=True,
+                    )
+                    continue
+                statsd_client.timing(
+                    "galaxy.control_queue.depth",
+                    declared.message_count,
+                    tags={"queue_name": queue.name},
+                )
+        finally:
+            channel.close()
+
+
+def emit_worker_process_gauge(
+    statsd_client: "VanillaGalaxyStatsdClient",
+    model: GalaxyModelMapping,
+) -> None:
+    """Emit ``galaxy.worker_process.active`` gauge grouped by ``app_type``."""
+    cutoff = now() - datetime.timedelta(seconds=DEFAULT_ACTIVE_PROCESS_WINDOW_SECONDS)
+    stmt = (
+        select(WorkerProcess.app_type, func.count(WorkerProcess.id))
+        .where(WorkerProcess.update_time > cutoff)
+        .group_by(WorkerProcess.app_type)
+    )
+    counts: dict[str, int] = defaultdict(int)
+    with model.new_session() as session:
+        for app_type, count in session.execute(stmt):
+            counts[app_type or "unknown"] = int(count)
+    for app_type, count in counts.items():
+        statsd_client.timing(
+            "galaxy.worker_process.active",
+            count,
+            tags={"app_type": app_type},
+        )
+
+
+def _run(name: str, statsd_client: "VanillaGalaxyStatsdClient", fn: Callable[[], None]) -> None:
+    """Run a sub-emitter, isolating its failures.
+
+    A broken sub-emitter logs once at WARNING and increments
+    ``galaxy.queue_metrics.error`` (tagged by emitter name) so the failure is
+    observable in metrics without the Celery-beat wrapper logging on every
+    tick. The other sub-emitters continue to run on this tick.
+    """
+    try:
+        fn()
+    except Exception:
+        log.warning("queue_metrics: %s emitter failed", name, exc_info=True)
+        statsd_client.incr("galaxy.queue_metrics.error", tags={"emitter": name})
+
+
+def emit_queue_metrics(
+    statsd_client: "Optional[VanillaGalaxyStatsdClient]",
+    connection: "Optional[Connection]",
+    application_stack: ApplicationStack,
+    model: GalaxyModelMapping,
+    sse_manager: Optional[SSEConnectionManager],
+) -> None:
+    """Periodic entry-point — no-ops when statsd isn't configured."""
+    if statsd_client is None:
+        return
+    if sse_manager is not None:
+        _run("sse_connections", statsd_client, lambda: emit_sse_connection_gauges(statsd_client, sse_manager))
+    _run(
+        "control_queue_depth",
+        statsd_client,
+        lambda: emit_control_queue_depth(statsd_client, connection, application_stack),
+    )
+    _run("worker_process", statsd_client, lambda: emit_worker_process_gauge(statsd_client, model))
diff --git a/lib/galaxy/managers/sse.py b/lib/galaxy/managers/sse.py
new file mode 100644
index 000000000000..0284b1786b6b
--- /dev/null
+++ b/lib/galaxy/managers/sse.py
@@ -0,0 +1,242 @@
+"""Server-Sent Events (SSE) connection manager for real-time notifications.
+
+Manages per-worker in-memory mapping of user IDs to asyncio.Queue instances,
+enabling push of events from any thread (e.g. Kombu control queue worker)
+to async SSE endpoint handlers running in the uvicorn event loop.
+"""
+
+import asyncio
+import logging
+from collections import defaultdict
+from collections.abc import (
+    AsyncIterator,
+    Awaitable,
+    Callable,
+)
+from dataclasses import dataclass
+from datetime import datetime
+from typing import (
+    Optional,
+)
+
+from galaxy.model.orm.now import now
+from galaxy.web.statsd_client import VanillaGalaxyStatsdClient
+
+log = logging.getLogger(__name__)
+
+
+def make_event_id() -> str:
+    """Return an SSE ``id`` string for Last-Event-ID replay.
+
+    Uses ``galaxy.model.orm.now`` so the timestamp format matches the rest of
+    Galaxy's database-backed timestamps (timezone-naive UTC). Kept in one place
+    so producers and the parse path cannot drift.
+    """
+    return now().isoformat()
+
+
+def parse_event_id(event_id: str) -> Optional[datetime]:
+    """Inverse of :func:`make_event_id`. Returns ``None`` if unparseable."""
+    try:
+        return datetime.fromisoformat(event_id)
+    except (ValueError, TypeError):
+        return None
+
+
+#: Async callable returning True when the client has disconnected. The SSE
+#: stream loop polls this each iteration so managers don't depend on starlette.
+IsDisconnected = Callable[[], Awaitable[bool]]
+
+
+@dataclass
+class SSEEvent:
+    """An event to be sent to an SSE client."""
+
+    event: str  # e.g. "notification_update", "broadcast_update", "notification_status"
+    data: str  # JSON payload
+    id: Optional[str] = None  # ISO timestamp, used by EventSource as Last-Event-ID on reconnect
+
+    def to_wire(self) -> str:
+        """Serialize this event to the SSE wire format (``event:…\\ndata:…\\n[id:…\\n]\\n``)."""
+        frame = f"event: {self.event}\ndata: {self.data}\n"
+        if self.id:
+            frame += f"id: {self.id}\n"
+        return frame + "\n"
+
+
+class SSEConnectionManager:
+    """Per-worker manager for SSE connections.
+
+    Maps user_ids to sets of asyncio.Queue instances. Each SSE connection
+    gets its own queue. The manager is thread-safe for push operations
+    via ``loop.call_soon_threadsafe``.
+
+    Lifecycle:
+    - Instantiated once per Galaxy worker process (on app object).
+    - ``connect()`` is called from the SSE async endpoint (event loop thread).
+    - ``disconnect()`` is called from the SSE endpoint's ``finally`` block.
+    - ``push_to_user()`` / ``push_broadcast()`` are called from ANY thread
+      (typically the Kombu daemon thread via control task handlers).
+    """
+
+    def __init__(self, statsd_client: Optional[VanillaGalaxyStatsdClient] = None) -> None:
+        self._connections: dict[int, set[asyncio.Queue]] = defaultdict(set)
+        self._session_connections: dict[int, set[asyncio.Queue]] = defaultdict(set)
+        self._broadcast_connections: set[asyncio.Queue] = set()
+        self._loop: Optional[asyncio.AbstractEventLoop] = None
+        self._statsd_client = statsd_client
+
+    def _ensure_loop(self) -> None:
+        """Capture the running asyncio event loop. Must be called from async context."""
+        if self._loop is None or self._loop.is_closed():
+            self._loop = asyncio.get_running_loop()
+
+    # -- Called from ASYNC context (uvicorn event loop thread) --
+
+    def connect(self, user_id: Optional[int], galaxy_session_id: Optional[int] = None) -> asyncio.Queue:
+        """Register a new SSE connection. Returns a queue to await events from.
+
+        Called from the SSE endpoint handler (async context). A ``ready`` event is
+        enqueued immediately so that clients (and tests) can synchronize on the
+        server-side subscription rather than the underlying socket open event.
+
+        ``galaxy_session_id`` is the dispatch key for events that target a
+        specific browser session (e.g. history updates for anonymous users,
+        whose ``user_id`` is ``None``).
+        """
+        self._ensure_loop()
+        queue: asyncio.Queue = asyncio.Queue(maxsize=64)
+        if user_id is not None:
+            self._connections[user_id].add(queue)
+        if galaxy_session_id is not None:
+            self._session_connections[galaxy_session_id].add(queue)
+        self._broadcast_connections.add(queue)
+        queue.put_nowait(SSEEvent(event="ready", data=""))
+        log.debug(
+            "SSE connection opened for user_id=%s session_id=%s (total=%d)",
+            user_id,
+            galaxy_session_id,
+            len(self._broadcast_connections),
+        )
+        return queue
+
+    def disconnect(
+        self,
+        user_id: Optional[int],
+        queue: asyncio.Queue,
+        galaxy_session_id: Optional[int] = None,
+    ) -> None:
+        """Unregister an SSE connection.
+
+        Called from the SSE endpoint's ``finally`` block (async context).
+        """
+        if user_id is not None:
+            self._connections[user_id].discard(queue)
+            if not self._connections[user_id]:
+                del self._connections[user_id]
+        if galaxy_session_id is not None:
+            self._session_connections[galaxy_session_id].discard(queue)
+            if not self._session_connections[galaxy_session_id]:
+                del self._session_connections[galaxy_session_id]
+        self._broadcast_connections.discard(queue)
+        log.debug(
+            "SSE connection closed for user_id=%s session_id=%s (total=%d)",
+            user_id,
+            galaxy_session_id,
+            len(self._broadcast_connections),
+        )
+
+    # -- Called from ANY thread (Kombu thread or async) --
+
+    def push_to_user(self, user_id: int, event: SSEEvent) -> None:
+        """Thread-safe. Push an event to all SSE connections for a specific user."""
+        for queue in list(self._connections.get(user_id, [])):
+            self._safe_put(queue, event)
+
+    def push_to_session(self, galaxy_session_id: int, event: SSEEvent) -> None:
+        """Thread-safe. Push an event to all SSE connections for a specific galaxy_session.
+
+        Used to route per-browser events (e.g. history updates for anonymous
+        histories) when there is no registered ``user_id`` to key on.
+        """
+        for queue in list(self._session_connections.get(galaxy_session_id, [])):
+            self._safe_put(queue, event)
+
+    def push_broadcast(self, event: SSEEvent) -> None:
+        """Thread-safe. Push an event to ALL connected SSE clients."""
+        for queue in list(self._broadcast_connections):
+            self._safe_put(queue, event)
+
+    def _safe_put(self, queue: asyncio.Queue, event: SSEEvent) -> None:
+        """Cross the thread boundary safely using ``call_soon_threadsafe``."""
+        if self._loop is None or self._loop.is_closed():
+            return
+        try:
+            self._loop.call_soon_threadsafe(self._do_put, queue, event)
+        except RuntimeError:
+            # Event loop is closed or shutting down
+            pass
+
+    def _do_put(self, queue: asyncio.Queue, event: SSEEvent) -> None:
+        """Runs ON the event loop thread. Safe to touch asyncio.Queue here."""
+        try:
+            queue.put_nowait(event)
+        except asyncio.QueueFull:
+            log.warning("SSE queue full, dropping event: %s", event.event)
+            if self._statsd_client is not None:
+                self._statsd_client.incr("galaxy.sse.connections.dropped")
+
+    @property
+    def connected_user_ids(self) -> set[int]:
+        return set(self._connections.keys())
+
+    @property
+    def total_connections(self) -> int:
+        return len(self._broadcast_connections)
+
+    @property
+    def total_broadcast_connections(self) -> int:
+        """Number of all active SSE connections (includes anonymous).
+
+        Every connection is added to ``_broadcast_connections``; this is the
+        most accurate "SSE clients currently connected" gauge.
+        """
+        return len(self._broadcast_connections)
+
+    @property
+    def total_per_user_connections(self) -> int:
+        """Number of active SSE connections bound to a specific user_id."""
+        return sum(len(queues) for queues in self._connections.values())
+
+    # -- High-level streaming helper --
+
+    async def stream(
+        self,
+        is_disconnected: IsDisconnected,
+        user_id: Optional[int],
+        catch_up: Optional[SSEEvent] = None,
+        keepalive: float = 30.0,
+        galaxy_session_id: Optional[int] = None,
+    ) -> AsyncIterator[str]:
+        """Yield SSE-framed strings for one connected client.
+
+        Handles ``connect``, optional catch-up event priming, the main event
+        loop with a keepalive comment on timeout, disconnect detection, and
+        ``disconnect`` in ``finally``. The ``is_disconnected`` callable is
+        what the service passes in (typically ``request.is_disconnected`` from
+        starlette) so the manager stays framework-agnostic.
+        """
+        queue = self.connect(user_id, galaxy_session_id)
+        if catch_up is not None:
+            await queue.put(catch_up)
+        try:
+            while True:
+                if await is_disconnected():
+                    break
+                try:
+                    event: SSEEvent = await asyncio.wait_for(queue.get(), timeout=keepalive)
+                    yield event.to_wire()
+                except asyncio.TimeoutError:
+                    yield ": keepalive\n\n"
+        finally:
+            self.disconnect(user_id, queue, galaxy_session_id)
diff --git a/lib/galaxy/managers/sse_dispatch.py b/lib/galaxy/managers/sse_dispatch.py
new file mode 100644
index 000000000000..9fbfb8a16164
--- /dev/null
+++ b/lib/galaxy/managers/sse_dispatch.py
@@ -0,0 +1,166 @@
+"""Producer-side SSE helper: fans events out across Galaxy processes.
+
+Kept in its own module (separate from ``galaxy.managers.sse``) because the
+dispatcher depends on ``galaxy.queue_worker`` and ``galaxy.queues``, while
+``queue_worker`` in turn depends on the connection types in ``sse``. Splitting
+the producer (``SSEEventDispatcher``) from the connection state
+(``SSEConnectionManager`` / ``SSEEvent``) breaks that cycle without requiring
+inline imports in the hot path.
+"""
+
+import logging
+import threading
+import time
+from collections.abc import Callable
+from typing import (
+    Any,
+    Optional,
+)
+
+from cachetools import TTLCache
+from kombu import Queue
+
+from galaxy.managers.sse import make_event_id
+from galaxy.queue_worker import (
+    ControlTask,
+    GalaxyQueueWorker,
+)
+from galaxy.queues import all_control_queues_for_declare
+from galaxy.web.statsd_client import VanillaGalaxyStatsdClient
+from galaxy.web_stack import ApplicationStack
+
+log = logging.getLogger(__name__)
+
+
+class SSEEventDispatcher:
+    """Fans out SSE events across all Galaxy worker processes via the control queue.
+
+    Dependencies are injected individually so the dispatcher can be unit-tested
+    without a full ``app``. ``queue_worker`` is ``Optional`` because unit-test
+    mock apps and Galaxy configurations without AMQP don't construct one — the
+    dispatcher silently no-ops in that case.
+
+    ``statsd_client`` is optional — if ``None`` (statsd not configured), all
+    instrumentation becomes a cheap attribute-lookup no-op.
+    """
+
+    # TTL for the active-worker declare-queue cache. The WorkerProcess heartbeat
+    # writes every 60 s and ``all_control_queues_for_declare`` filters on a 120 s
+    # window, so a 30 s cache cannot produce a result that wasn't also valid in
+    # the non-cached call. Surfaced as a class constant so tests can monkey-patch.
+    _DECLARE_QUEUES_TTL_SECONDS = 30
+
+    def __init__(
+        self,
+        queue_worker: Optional[GalaxyQueueWorker],
+        application_stack: ApplicationStack,
+        statsd_client: Optional[VanillaGalaxyStatsdClient] = None,
+        clock: Callable[[], float] = time.monotonic,
+        # Factory return is typed ``Any`` so ``ControlTask`` itself and test-only
+        # duck-typed doubles (FakeControlTask/BoomControlTask/NoopControlTask)
+        # all satisfy the signature under mypy.
+        control_task_factory: Callable[[GalaxyQueueWorker], Any] = ControlTask,
+        queues_provider: Optional[Callable[[], list[Queue]]] = None,
+    ) -> None:
+        self._queue_worker = queue_worker
+        self._application_stack = application_stack
+        self._statsd_client = statsd_client
+        self._clock = clock
+        self._control_task_factory = control_task_factory
+        # Default provider closes over application_stack so tests can pass a
+        # plain ``lambda: [...]`` without needing a stack.
+        self._queues_provider = queues_provider or (
+            lambda: all_control_queues_for_declare(application_stack, webapp_only=True)
+        )
+        self._declare_queues_cache: TTLCache = TTLCache(maxsize=1, ttl=self._DECLARE_QUEUES_TTL_SECONDS, timer=clock)
+        self._declare_queues_lock = threading.RLock()
+
+    def _get_declare_queues(self) -> list[Queue]:
+        # Empty results (startup before DatabaseHeartbeat registers this process,
+        # or a transient DB error swallowed by the provider) must not be pinned
+        # for the full TTL — they'd silently drop every SSE event until the next
+        # expiry. Only cache non-empty results.
+        with self._declare_queues_lock:
+            try:
+                return self._declare_queues_cache["webapp"]
+            except KeyError:
+                queues = self._queues_provider()
+                if queues:
+                    self._declare_queues_cache["webapp"] = queues
+                return queues
+
+    def _send(self, task: str, kwargs: dict[str, Any]) -> None:
+        if self._queue_worker is None:
+            # AMQP not configured at all (e.g. unit-test mock app). Skip silently.
+            log.debug("SSE dispatch skipped: no queue_worker configured (task=%s)", task)
+            if self._statsd_client is not None:
+                self._statsd_client.incr("galaxy.sse.dispatch.skipped_no_qw")
+            return
+        if self._statsd_client is not None:
+            self._statsd_client.incr("galaxy.sse.dispatch.count", tags={"task": task})
+        # Only fan out to webapp processes — job handlers and workflow schedulers
+        # don't have browser SSE connections to push to.
+        declare_queues = self._get_declare_queues()
+        control_task = self._control_task_factory(self._queue_worker)
+        start_time = time.perf_counter() if self._statsd_client is not None else 0.0
+        try:
+            control_task.send_task(
+                payload={"task": task, "kwargs": kwargs},
+                routing_key="control.*",
+                expiration=10,
+                declare_queues=declare_queues,
+            )
+        finally:
+            if self._statsd_client is not None:
+                dt_ms = int((time.perf_counter() - start_time) * 1000)
+                self._statsd_client.timing("galaxy.sse.dispatch.latency_ms", dt_ms, tags={"task": task})
+
+    def notify_users(self, user_ids: list[int], payload: str, event_id: Optional[str] = None) -> None:
+        self._send(
+            "notify_users",
+            {
+                "user_ids": user_ids,
+                "payload": payload,
+                "event_id": event_id or make_event_id(),
+            },
+        )
+
+    def notify_broadcast(self, payload: str, event_id: Optional[str] = None) -> None:
+        self._send(
+            "notify_broadcast",
+            {
+                "payload": payload,
+                "event_id": event_id or make_event_id(),
+            },
+        )
+
+    def history_update(
+        self,
+        user_updates: dict[str, list[int]],
+        event_id: Optional[str] = None,
+        session_updates: Optional[dict[str, list[int]]] = None,
+    ) -> None:
+        kwargs: dict[str, Any] = {
+            "user_updates": user_updates,
+            "event_id": event_id or make_event_id(),
+        }
+        if session_updates:
+            # Only include when non-empty: anonymous histories are uncommon on
+            # most deployments, and an empty dict is wasted wire payload.
+            kwargs["session_updates"] = session_updates
+        self._send("history_update", kwargs)
+
+    def entry_point_update(self, user_id: int, event_id: Optional[str] = None) -> None:
+        """Fan out a wake-up ``entry_point_update`` event for one user.
+
+        The client always refetches the canonical entry-point list on receipt,
+        so no IDs are sent — keeping the payload small and the dispatch path
+        free of per-event encoding work.
+        """
+        self._send(
+            "entry_point_update",
+            {
+                "user_id": user_id,
+                "event_id": event_id or make_event_id(),
+            },
+        )
diff --git a/lib/galaxy/model/database_heartbeat.py b/lib/galaxy/model/database_heartbeat.py
index c16a8c64de03..143a94f1f116 100644
--- a/lib/galaxy/model/database_heartbeat.py
+++ b/lib/galaxy/model/database_heartbeat.py
@@ -17,6 +17,8 @@
 log = logging.getLogger(__name__)
 
 WEBAPP = "webapp"  # WorkerProcess.app_type for web apps.
+SSE_MONITOR = "sse_monitor"  # WorkerProcess.app_type for the standalone SSE monitor process.
+SSE_MONITOR_SERVER_PREFIX = "sse_monitor."  # server_name prefix used by the standalone SSE monitor process.
 
 
 class DatabaseHeartbeat:
@@ -27,7 +29,9 @@ def __init__(self, application_stack, heartbeat_interval=60):
         self.hostname = socket.gethostname()
         self._engine = application_stack.app.model.engine
         self._is_config_watcher = False
+        self._is_history_audit_monitor = False
         self._observers = []
+        self._audit_monitor_observers = []
         self.exit = threading.Event()
         self.thread = None
         self.active = False
@@ -72,6 +76,9 @@ def get_active_processes(self, last_seen_seconds=None):
     def add_change_callback(self, callback):
         self._observers.append(callback)
 
+    def add_audit_monitor_change_callback(self, callback):
+        self._audit_monitor_observers.append(callback)
+
     @property
     def is_config_watcher(self):
         return self._is_config_watcher
@@ -83,6 +90,28 @@ def is_config_watcher(self, value):
         for callback in self._observers:
             callback(self._is_config_watcher)
 
+    @property
+    def is_history_audit_monitor(self):
+        return self._is_history_audit_monitor
+
+    @is_history_audit_monitor.setter
+    def is_history_audit_monitor(self, value):
+        self._is_history_audit_monitor = value
+        log.debug(
+            "%s %s history audit monitor",
+            self.server_name,
+            "is" if self._is_history_audit_monitor else "is not",
+        )
+        for callback in self._audit_monitor_observers:
+            callback(self._is_history_audit_monitor)
+
+    def _app_type(self):
+        if self.application_stack.app.is_webapp:
+            return WEBAPP
+        if self.server_name.startswith(SSE_MONITOR_SERVER_PREFIX):
+            return SSE_MONITOR
+        return None
+
     def update_watcher_designation(self):
         expression = self._worker_process_identifying_clause()
         stmt = select(WorkerProcess).with_for_update(of=WorkerProcess).where(expression)
@@ -90,19 +119,37 @@ def update_watcher_designation(self):
             worker_process = session.scalars(stmt).first()
             if not worker_process:
                 worker_process = WorkerProcess(server_name=self.server_name, hostname=self.hostname)
-            if self.application_stack.app.is_webapp:
-                worker_process.app_type = WEBAPP
+            app_type = self._app_type()
+            if app_type is not None:
+                worker_process.app_type = app_type
             worker_process.update_time = now()
             worker_process.pid = self.pid
             session.add(worker_process)
+        active = list(self.get_active_processes(self.heartbeat_interval + 1))
         # We only want a single process watching the various config files on the file system.
         # We just pick the max server name for simplicity
-        webapp_servers = [
-            p.server_name for p in self.get_active_processes(self.heartbeat_interval + 1) if p.app_type == WEBAPP
-        ]
+        webapp_servers = [p.server_name for p in active if p.app_type == WEBAPP]
         is_config_watcher = bool(webapp_servers) and self.server_name == max(webapp_servers)
         if is_config_watcher != self.is_config_watcher:
             self.is_config_watcher = is_config_watcher
+        # The history-audit monitor is a single elected process too, but preference
+        # goes to a standalone sse_monitor daemon when one is running so the
+        # monitor's postgres LISTEN isn't blocked by webapp GIL pauses. If no
+        # dedicated process is registered we fall back to a webapp (same
+        # max-server_name tiebreaker as config_watcher).
+        audit_leader = self._elect_audit_leader(active, webapp_servers)
+        is_history_audit_monitor = audit_leader is not None and self.server_name == audit_leader
+        if is_history_audit_monitor != self.is_history_audit_monitor:
+            self.is_history_audit_monitor = is_history_audit_monitor
+
+    @staticmethod
+    def _elect_audit_leader(active, webapp_servers):
+        monitor_servers = [p.server_name for p in active if p.app_type == SSE_MONITOR]
+        if monitor_servers:
+            return min(monitor_servers)
+        if webapp_servers:
+            return max(webapp_servers)
+        return None
 
     def send_database_heartbeat(self):
         if self.active:
diff --git a/lib/galaxy/model/migrations/alembic/versions_gxy/b8d5e2f9a1c7_add_pg_notify_to_history_audit_triggers.py b/lib/galaxy/model/migrations/alembic/versions_gxy/b8d5e2f9a1c7_add_pg_notify_to_history_audit_triggers.py
new file mode 100644
index 000000000000..0fe49572c790
--- /dev/null
+++ b/lib/galaxy/model/migrations/alembic/versions_gxy/b8d5e2f9a1c7_add_pg_notify_to_history_audit_triggers.py
@@ -0,0 +1,62 @@
+"""Add pg_notify to history audit triggers
+
+Revision ID: b8d5e2f9a1c7
+Revises: f5e9e4bca542
+Create Date: 2026-04-17 14:30:00.000000
+
+The SSE-based history update pipeline (see `managers/history_audit_monitor.py`)
+depends on a PostgreSQL LISTEN on the `galaxy_history_update` channel. For
+existing installations, trigger functions installed by earlier migrations (most
+recently `c716ee82337b_replace_triggers`) do not emit the corresponding
+`pg_notify`, so the monitor wakes up only from the poll-timeout fallback and
+per-history events are never dispatched in real time. This revision replaces
+both audit trigger functions with versions that emit `pg_notify` for each
+affected history id, matching `model/triggers/update_audit_table.py` used on
+fresh installs.
+
+The STATEMENT-vs-ROW decision must match the trigger DEFINITION installed by
+`c716ee82337b` so the function body references the right context (`new_table`
+vs `NEW`); both use `version > 10` (and treat offline mode as STATEMENT).
+
+SQLite installations use the poll-only path and require no change.
+"""
+
+from alembic import op
+
+from galaxy.model.migrations.util import (
+    _is_sqlite,
+    transaction,
+)
+from galaxy.model.triggers.update_audit_table import (
+    build_trigger_fn,
+    fn_prefix,
+    use_statement_trigger,
+)
+
+revision = "b8d5e2f9a1c7"
+down_revision = "f5e9e4bca542"
+branch_labels = None
+depends_on = None
+
+
+def upgrade():
+    if _is_sqlite():
+        return
+    with transaction():
+        _install_functions(with_notify=True)
+
+
+def downgrade():
+    if _is_sqlite():
+        return
+    with transaction():
+        _install_functions(with_notify=False)
+
+
+def _install_functions(with_notify: bool) -> None:
+    version_info = op.get_bind().engine.dialect.server_version_info
+    # Offline mode (no live connection) matches c716ee82337b: assume STATEMENT.
+    statement = version_info is None or use_statement_trigger(version_info[0])
+    for id_field in ("history_id", "id"):
+        fn_name = f"{fn_prefix}_{id_field}"
+        op.execute(build_trigger_fn(fn_name, id_field, use_statement=statement, with_notify=with_notify))
diff --git a/lib/galaxy/model/triggers/update_audit_table.py b/lib/galaxy/model/triggers/update_audit_table.py
index 3c06c05830e0..be35d85da816 100644
--- a/lib/galaxy/model/triggers/update_audit_table.py
+++ b/lib/galaxy/model/triggers/update_audit_table.py
@@ -3,6 +3,9 @@
 # function name prefix
 fn_prefix = "fn_audit_history_by"
 
+# channel used by pg_notify so HistoryAuditMonitor can LISTEN for updates
+NOTIFY_CHANNEL = "galaxy_history_update"
+
 # map between source table and associated incoming id field
 trigger_config = {
     "history_dataset_association": "history_id",
@@ -11,6 +14,68 @@
 }
 
 
+def use_statement_trigger(version: int) -> bool:
+    """Return True when the postgres version supports the STATEMENT variant.
+
+    Fresh installs and the pg_notify migration share this predicate to ensure
+    the trigger function body (STATEMENT references new_table, ROW references NEW)
+    matches the trigger definition installed at that version.
+    """
+    return version > 10
+
+
+def build_trigger_fn(function_name: str, id_field: str, *, use_statement: bool, with_notify: bool = True) -> str:
+    """Build the plpgsql CREATE OR REPLACE FUNCTION body for an audit trigger.
+
+    Shared between runtime install (update_audit_table.install) and alembic
+    migrations so the two cannot drift.
+    """
+    if use_statement:
+        notify_block = (
+            f"""
+                    FOR _history_id IN SELECT DISTINCT {id_field} FROM new_table WHERE {id_field} IS NOT NULL
+                    LOOP
+                        PERFORM pg_notify('{NOTIFY_CHANNEL}', _history_id::text);
+                    END LOOP;
+            """
+            if with_notify
+            else ""
+        )
+        declare_block = "DECLARE _history_id integer;" if with_notify else ""
+        return f"""
+            CREATE OR REPLACE FUNCTION {function_name}()
+                RETURNS TRIGGER
+                LANGUAGE 'plpgsql'
+            AS $BODY$
+                {declare_block}
+                BEGIN
+                    INSERT INTO history_audit (history_id, update_time)
+                    SELECT DISTINCT {id_field}, clock_timestamp() AT TIME ZONE 'UTC'
+                    FROM new_table
+                    WHERE {id_field} IS NOT NULL
+                    ON CONFLICT DO NOTHING;
+                    {notify_block}
+                    RETURN NULL;
+                END;
+            $BODY$
+        """
+    notify_stmt = f"PERFORM pg_notify('{NOTIFY_CHANNEL}', NEW.{id_field}::text);" if with_notify else ""
+    return f"""
+        CREATE OR REPLACE FUNCTION {function_name}()
+            RETURNS TRIGGER
+            LANGUAGE 'plpgsql'
+        AS $BODY$
+            BEGIN
+                INSERT INTO history_audit (history_id, update_time)
+                VALUES (NEW.{id_field}, clock_timestamp() AT TIME ZONE 'UTC')
+                ON CONFLICT DO NOTHING;
+                {notify_stmt}
+                RETURN NULL;
+            END;
+        $BODY$
+    """
+
+
 def install(engine):
     """Install history audit table triggers"""
     sql = _postgres_install(engine) if "postgres" in engine.name else _sqlite_install()
@@ -41,47 +106,6 @@ def _postgres_install(engine):
 
     sql = []
 
-    # PostgreSQL trigger function template
-    # need to make separate functions purely because the incoming history_id field name will be
-    # different for different source tables. There may be a fancier way to dynamically choose
-    # between incoming fields, but having 2 triggers fns seems straightforward
-
-    def statement_trigger_fn(id_field):
-        fn = f"{fn_prefix}_{id_field}"
-
-        return f"""
-            CREATE OR REPLACE FUNCTION {fn}()
-                RETURNS TRIGGER
-                LANGUAGE 'plpgsql'
-            AS $BODY$
-                BEGIN
-                    INSERT INTO history_audit (history_id, update_time)
-                    SELECT DISTINCT {id_field}, clock_timestamp() AT TIME ZONE 'UTC'
-                    FROM new_table
-                    WHERE {id_field} IS NOT NULL
-                    ON CONFLICT DO NOTHING;
-                    RETURN NULL;
-                END;
-            $BODY$
-        """
-
-    def row_trigger_fn(id_field):
-        fn = f"{fn_prefix}_{id_field}"
-
-        return f"""
-            CREATE OR REPLACE FUNCTION {fn}()
-                RETURNS TRIGGER
-                LANGUAGE 'plpgsql'
-            AS $BODY$
-                BEGIN
-                    INSERT INTO history_audit (history_id, update_time)
-                    VALUES (NEW.{id_field}, clock_timestamp() AT TIME ZONE 'UTC')
-                    ON CONFLICT DO NOTHING;
-                    RETURN NULL;
-                END;
-            $BODY$
-        """
-
     def trigger_def(source_table: str, id_field: str, operation: str, version: int, when: str = "AFTER") -> str:
         fn = f"{fn_prefix}_{id_field}"
         # PostgreSQL supports many triggers per operation/table so the label can
@@ -93,7 +117,7 @@ def trigger_def(source_table: str, id_field: str, operation: str, version: int,
         # The use of the keyword PROCEDURE here is historical and deprecated (https://www.postgresql.org/docs/11/sql-createtrigger.html).
         function_keyword = "FUNCTION" if version >= 11 else "PROCEDURE"
         create_or_replace = "CREATE OR REPLACE" if version >= 14 else "CREATE"
-        if version >= 10 and when == "AFTER":
+        if use_statement_trigger(version) and when == "AFTER":
             return f"""
                 {create_or_replace} TRIGGER {trigger_name}
                 AFTER {operation}
@@ -114,10 +138,11 @@ def trigger_def(source_table: str, id_field: str, operation: str, version: int,
 
     # pick row or statement triggers depending on postgres version
     version = engine.dialect.server_version_info[0]
-    trigger_fn = statement_trigger_fn if version >= 10 else row_trigger_fn
+    statement = use_statement_trigger(version)
 
     for id_field in ["history_id", "id"]:
-        sql.append(trigger_fn(id_field))
+        fn_name = f"{fn_prefix}_{id_field}"
+        sql.append(build_trigger_fn(fn_name, id_field, use_statement=statement, with_notify=True))
 
     for source_table, id_field in trigger_config.items():
         for operation in ["UPDATE", "INSERT"]:
diff --git a/lib/galaxy/model/unittest_utils/data_app.py b/lib/galaxy/model/unittest_utils/data_app.py
index 0d424cca3623..2eb7e78bf3e7 100644
--- a/lib/galaxy/model/unittest_utils/data_app.py
+++ b/lib/galaxy/model/unittest_utils/data_app.py
@@ -9,6 +9,7 @@
 import os
 import shutil
 import tempfile
+from types import SimpleNamespace
 from typing import Optional
 
 from galaxy import (
@@ -105,6 +106,11 @@ def __init__(self, config: Optional[GalaxyDataTestConfig] = None, **kwd):
         model.setup_global_object_store_for_models(self.object_store)
         self.security_agent = self.model.security_agent
         self.tag_handler = GalaxyTagHandler(self.model.session)
+        # statsd/observability plumbing — stubbed out so paths that read
+        # ``app.execution_timer_factory.galaxy_statsd_client`` (e.g. the
+        # queue-worker instrumentation) degrade to a no-op if ever exercised
+        # against this mock instead of raising ``AttributeError``.
+        self.execution_timer_factory = SimpleNamespace(galaxy_statsd_client=None)
         self.init_datatypes()
 
     def init_datatypes(self):
diff --git a/lib/galaxy/queue_worker/__init__.py b/lib/galaxy/queue_worker/__init__.py
index 0e3d0f719b04..2bbff9e74e85 100644
--- a/lib/galaxy/queue_worker/__init__.py
+++ b/lib/galaxy/queue_worker/__init__.py
@@ -4,6 +4,7 @@
 """
 
 import importlib
+import json
 import logging
 import math
 import socket
@@ -12,8 +13,11 @@
 import time
 from inspect import ismodule
 from typing import (
+    Any,
+    cast,
     Optional,
     TYPE_CHECKING,
+    TypedDict,
 )
 
 from kombu import (
@@ -27,6 +31,10 @@
 import galaxy.queues
 from galaxy import util
 from galaxy.config import reload_config_options
+from galaxy.managers.sse import (
+    SSEConnectionManager,
+    SSEEvent,
+)
 from galaxy.model import User
 from galaxy.tools import ToolBox
 from galaxy.tools.data_manager.manager import DataManagers
@@ -43,7 +51,49 @@
     )
 
 
-def send_local_control_task(app: "StructuredApp", task: str, get_response: bool = False, kwargs: Optional[dict] = None):
+class NotifyUsersPayload(TypedDict, total=False):
+    """Wire contract for the ``notify_users`` control-task kwargs."""
+
+    user_ids: list[int]
+    payload: str
+    event_id: Optional[str]
+
+
+class NotifyBroadcastPayload(TypedDict, total=False):
+    """Wire contract for the ``notify_broadcast`` control-task kwargs."""
+
+    payload: str
+    event_id: Optional[str]
+
+
+class HistoryUpdatePayload(TypedDict, total=False):
+    """Wire contract for the ``history_update`` control-task kwargs.
+
+    ``user_updates`` maps stringified user IDs to lists of (unencoded) history IDs.
+    ``session_updates`` is the parallel route for anonymous-owned histories,
+    keyed by stringified ``galaxy_session.id`` (the dispatch key never leaves
+    the server — browsers never see it). Stringified because AMQP JSON
+    serialization coerces dict keys to strings.
+    """
+
+    user_updates: dict[str, list[int]]
+    session_updates: dict[str, list[int]]
+    event_id: Optional[str]
+
+
+class EntryPointUpdatePayload(TypedDict, total=False):
+    """Wire contract for the ``entry_point_update`` control-task kwargs."""
+
+    user_id: int
+    event_id: Optional[str]
+
+
+def send_local_control_task(
+    app: "StructuredApp",
+    task: str,
+    get_response: bool = False,
+    kwargs: Optional[dict] = None,
+) -> Any:
     """
     This sends a message to the process-local control worker, which is useful
     for one-time asynchronous tasks like recalculating user disk usage.
@@ -57,7 +107,16 @@ def send_local_control_task(app: "StructuredApp", task: str, get_response: bool
     return control_task.send_task(payload, routing_key, local=True, get_response=get_response)
 
 
-def send_control_task(app, task, noop_self=False, get_response=False, routing_key="control.*", kwargs=None):
+def send_control_task(
+    app: "StructuredApp",
+    task: str,
+    noop_self: bool = False,
+    get_response: bool = False,
+    routing_key: str = "control.*",
+    kwargs: Optional[dict] = None,
+    expiration: Optional[int] = None,
+    declare_queues: Optional[list[Queue]] = None,
+) -> Any:
     """
     This sends a control task out to all processes, useful for things like
     reloading a data table, which needs to happen individually in all
@@ -65,6 +124,9 @@ def send_control_task(app, task, noop_self=False, get_response=False, routing_ke
     Set noop_self to True to not run task for current process.
     Set get_response to True to wait for and return the task results
     as a list.
+    Set expiration to a number of seconds for message TTL.
+    Pass ``declare_queues`` to override the default active-processes list —
+    e.g. the SSE dispatcher uses this to restrict fan-out to webapp processes.
     """
     if kwargs is None:
         kwargs = {}
@@ -73,7 +135,13 @@ def send_control_task(app, task, noop_self=False, get_response=False, routing_ke
     if noop_self:
         payload["noop"] = app.config.server_name
     control_task = ControlTask(app.queue_worker)
-    return control_task.send_task(payload=payload, routing_key=routing_key, get_response=get_response)
+    return control_task.send_task(
+        payload=payload,
+        routing_key=routing_key,
+        get_response=get_response,
+        expiration=expiration,
+        declare_queues=declare_queues,
+    )
 
 
 class ControlTask:
@@ -107,10 +175,19 @@ def on_response(self, message):
         if message.properties["correlation_id"] == self.correlation_id:
             self.response = message.payload["result"]
 
-    def send_task(self, payload, routing_key, local=False, get_response=False, timeout=10):
+    def send_task(
+        self,
+        payload: dict,
+        routing_key: str,
+        local: bool = False,
+        get_response: bool = False,
+        timeout: int = 10,
+        expiration: Optional[int] = None,
+        declare_queues: Optional[list[Queue]] = None,
+    ):
         if local:
             declare_queues = self.control_queues
-        else:
+        elif declare_queues is None:
             declare_queues = self.declare_queues
         reply_to = None
         callback_queue = []
@@ -129,14 +206,24 @@ def send_task(self, payload, routing_key, local=False, get_response=False, timeo
                     correlation_id=self.correlation_id,
                     retry=True,
                     headers={"epoch": time.time()},
+                    expiration=expiration,
                 )
             if get_response:
-                with Consumer(self.connection, on_message=self.on_response, queues=callback_queue, no_ack=True):
+                with Consumer(
+                    self.connection,
+                    on_message=self.on_response,
+                    queues=callback_queue,
+                    no_ack=True,
+                ):
                     while self.response is self._response:
                         self.connection.drain_events(timeout=timeout)
                 return self.response
         except TimeoutError:
-            log.exception("Error waiting for task: '%s' sent with routing key '%s'", payload, routing_key)
+            log.exception(
+                "Error waiting for task: '%s' sent with routing key '%s'",
+                payload,
+                routing_key,
+            )
         except Exception:
             log.exception("Error queueing async task: '%s'. for %s", payload, routing_key)
 
@@ -189,7 +276,10 @@ def _get_new_toolbox(app: "UniverseApplication", save_integrated_tool_panel: boo
     tool_configs = app.config.tool_configs
 
     new_toolbox = ToolBox(
-        tool_configs, app.config.tool_path, app, save_integrated_tool_panel=save_integrated_tool_panel
+        tool_configs,
+        app.config.tool_path,
+        app,
+        save_integrated_tool_panel=save_integrated_tool_panel,
     )
     new_toolbox.data_manager_tools = app.toolbox.data_manager_tools
     app.datatypes_registry.load_datatype_converters(new_toolbox, use_cached=True)
@@ -309,6 +399,71 @@ def admin_job_lock(app, **kwargs):
     log.info(f"Administrative Job Lock is now set to {job_lock}. Jobs will {'not' if job_lock else 'now'} dispatch.")
 
 
+def notify_users(app: "MinimalManagerApp", **kwargs) -> None:
+    """Push SSE events to connected users on this worker process."""
+    payload = cast(NotifyUsersPayload, kwargs)
+    sse_manager = app[SSEConnectionManager]
+    event = SSEEvent(
+        event="notification_update",
+        data=payload.get("payload", "{}"),
+        id=payload.get("event_id"),
+    )
+    for user_id in payload.get("user_ids", []):
+        sse_manager.push_to_user(user_id, event)
+
+
+def notify_broadcast(app: "MinimalManagerApp", **kwargs) -> None:
+    """Push SSE broadcast events to all connected clients on this worker process."""
+    payload = cast(NotifyBroadcastPayload, kwargs)
+    sse_manager = app[SSEConnectionManager]
+    event = SSEEvent(
+        event="broadcast_update",
+        data=payload.get("payload", "{}"),
+        id=payload.get("event_id"),
+    )
+    sse_manager.push_broadcast(event)
+
+
+def history_update(app: "MinimalManagerApp", **kwargs) -> None:
+    """Push SSE history update events to connected users on this worker process.
+
+    Encodes integer history IDs here (not in the monitor) so the manager layer
+    stays free of presentation/security concerns. Handles both user-keyed
+    routing (registered users) and galaxy_session-keyed routing (anonymous
+    histories, which have ``user_id IS NULL``).
+    """
+    payload = cast(HistoryUpdatePayload, kwargs)
+    sse_manager = app[SSEConnectionManager]
+    event_id = payload.get("event_id")
+    encode = app.security.encode_id
+    for user_id_str, history_ids in payload.get("user_updates", {}).items():
+        user_id = int(user_id_str)
+        encoded_ids = [encode(hid) for hid in history_ids]
+        data = json.dumps({"history_ids": encoded_ids})
+        event = SSEEvent(event="history_update", data=data, id=event_id)
+        sse_manager.push_to_user(user_id, event)
+    for session_id_str, history_ids in payload.get("session_updates", {}).items():
+        session_id = int(session_id_str)
+        encoded_ids = [encode(hid) for hid in history_ids]
+        data = json.dumps({"history_ids": encoded_ids})
+        event = SSEEvent(event="history_update", data=data, id=event_id)
+        sse_manager.push_to_session(session_id, event)
+
+
+def entry_point_update(app: "MinimalManagerApp", **kwargs) -> None:
+    """Push a wake-up SSE event to a single connected user.
+
+    The payload is empty by design: the client refetches ``/api/entry_points``
+    (the canonical source) on receipt, so there's nothing to narrow or merge.
+    Dropping the IDs from the payload also avoids per-event ``encode_id`` work
+    at 1000+ events/s.
+    """
+    payload = cast(EntryPointUpdatePayload, kwargs)
+    sse_manager = app[SSEConnectionManager]
+    event = SSEEvent(event="entry_point_update", data="{}", id=payload.get("event_id"))
+    sse_manager.push_to_user(int(payload["user_id"]), event)
+
+
 control_message_to_task = {
     "create_panel_section": create_panel_section,
     "reload_tool": reload_tool,
@@ -324,6 +479,10 @@ def admin_job_lock(app, **kwargs):
     "reconfigure_watcher": reconfigure_watcher,
     "reload_tour": reload_tour,
     "reload_core_config": reload_core_config,
+    "notify_users": notify_users,
+    "notify_broadcast": notify_broadcast,
+    "history_update": history_update,
+    "entry_point_update": entry_point_update,
 }
 
 
@@ -354,7 +513,14 @@ def __init__(self, app, task_mapping=None):
         self.control_queues = []
         self.epoch = 0
 
-    def send_control_task(self, task, noop_self=False, get_response=False, routing_key="control.*", kwargs=None):
+    def send_control_task(
+        self,
+        task,
+        noop_self=False,
+        get_response=False,
+        routing_key="control.*",
+        kwargs=None,
+    ):
         return send_control_task(
             app=self.app,
             task=task,
@@ -369,14 +535,32 @@ def send_local_control_task(self, task, get_response=False, kwargs=None):
 
     @property
     def declare_queues(self):
-        # dynamically produce queues, allows addressing all known processes at a given time
+        # Dynamically produce queues, allows addressing all known processes at a given time.
         return galaxy.queues.all_control_queues_for_declare(self.app.application_stack)
 
-    def bind_and_start(self):
-        # This is post-forking, so we got the correct sever name
-        log.info("Binding and starting galaxy control worker for %s", self.app.config.server_name)
+    def bind_publisher(self):
+        """Set up the queues needed to PUBLISH control tasks (no consumer thread).
+
+        Safe to call from any process that needs to produce control messages — notably
+        Celery workers, which want to fan out SSE events to web workers but must not
+        start a consumer themselves.
+
+        Always (re)binds. A prefork call in ``GalaxyManagerApplication.__init__`` binds
+        using the parent's ``config.server_name``; under gunicorn with ``--preload``
+        the child's ``set_postfork_server_name`` mutates ``server_name`` to e.g.
+        ``main.1`` after fork. ``bind_and_start`` calls back into this so the
+        consumer's queues match what post-fork producers declare.
+        """
         self.exchange_queue, self.direct_queue = galaxy.queues.control_queues_from_config(self.app.config)
         self.control_queues = [self.exchange_queue, self.direct_queue]
+
+    def bind_and_start(self):
+        # This is post-forking, so we got the correct sever name
+        log.info(
+            "Binding and starting galaxy control worker for %s",
+            self.app.config.server_name,
+        )
+        self.bind_publisher()
         self.epoch = time.time()
         self.start()
 
@@ -388,8 +572,14 @@ def get_consumers(self, Consumer, channel):
 
     def process_task(self, body, message):
         result = "NO_RESULT"
+        task_name = body.get("task")
+        statsd_client = self.app.execution_timer_factory.galaxy_statsd_client
+        if statsd_client is not None and task_name is not None:
+            statsd_client.incr("galaxy.control_queue.task.count", tags={"task": task_name})
         if body["task"] in self.task_mapping:
             if body.get("noop", None) != self.app.config.server_name:
+                outcome = "ok"
+                handler_start = time.perf_counter() if statsd_client is not None else 0.0
                 try:
                     f = self.task_mapping[body["task"]]
                     if message.headers.get("epoch", math.inf) > self.epoch:
@@ -410,7 +600,16 @@ def process_task(self, body, message):
                         result = "NO_OP"
                 except Exception:
                     # this shouldn't ever throw an exception, but...
+                    outcome = "error"
                     log.exception("Error running control task type: %s", body["task"])
+                finally:
+                    if statsd_client is not None and task_name is not None:
+                        dt_ms = int((time.perf_counter() - handler_start) * 1000)
+                        statsd_client.timing(
+                            "galaxy.control_queue.task.latency_ms",
+                            dt_ms,
+                            tags={"task": task_name, "outcome": outcome},
+                        )
             else:
                 result = "NO_OP"
         else:
diff --git a/lib/galaxy/queues/__init__.py b/lib/galaxy/queues/__init__.py
index ab6edafcf5b0..a5fab0efc5fe 100644
--- a/lib/galaxy/queues/__init__.py
+++ b/lib/galaxy/queues/__init__.py
@@ -4,29 +4,67 @@
 
 """
 
+import datetime
+import logging
 import socket
-from typing import Optional
+from typing import (
+    Optional,
+    TYPE_CHECKING,
+)
 
 from kombu import (
     Connection,
     Exchange,
     Queue,
 )
+from sqlalchemy import select
+
+from galaxy.model import WorkerProcess
+from galaxy.model.orm.now import now
+
+if TYPE_CHECKING:
+    from galaxy.web_stack import ApplicationStack
+
+log = logging.getLogger(__name__)
 
 ALL_CONTROL = "control.*"
 galaxy_exchange = Exchange("galaxy_core_exchange", type="topic")
 
+DEFAULT_ACTIVE_PROCESS_WINDOW_SECONDS = 120
+# Matches WorkerProcess.app_type set by DatabaseHeartbeat for webapp processes.
+WEBAPP_APP_TYPE = "webapp"
 
-def all_control_queues_for_declare(application_stack):
+
+def all_control_queues_for_declare(application_stack: "ApplicationStack", webapp_only: bool = False) -> list[Queue]:
     """
     For in-memory routing (used by sqlalchemy-based transports), we need to be able to
     build the entire routing table in producers.
+
+    Queries ``WorkerProcess`` directly rather than going through
+    ``DatabaseHeartbeat`` so this works from Celery workers too — they have a
+    ``model`` but no heartbeat thread. Without this, a notification created in
+    a Celery task publishes a ``notify_users`` control task with an empty
+    ``declare`` list, so on the sqlalchemy+sqlite kombu transport the message
+    never lands in a web worker's queue.
+
+    When ``webapp_only`` is True, only returns queues for processes that have
+    registered themselves with ``app_type='webapp'``. This is what the SSE
+    dispatcher wants: job handlers and workflow schedulers have no browser
+    connections, so routing SSE events to them is wasted work.
     """
-    # Get all active processes and construct queues for each process
-    process_names = (
-        f"{p.server_name}@{p.hostname}" for p in application_stack.app.database_heartbeat.get_active_processes()
-    )
-    return [Queue(f"control.{server_name}", galaxy_exchange, routing_key="control.*") for server_name in process_names]
+    app = application_stack.app
+    try:
+        stmt = select(WorkerProcess).where(
+            WorkerProcess.update_time > now() - datetime.timedelta(seconds=DEFAULT_ACTIVE_PROCESS_WINDOW_SECONDS)
+        )
+        if webapp_only:
+            stmt = stmt.where(WorkerProcess.app_type == WEBAPP_APP_TYPE)
+        with app.model.new_session() as session:
+            processes = session.scalars(stmt).all()
+    except Exception:
+        log.debug("Failed to look up active processes for control-queue declare", exc_info=True)
+        return []
+    return [Queue(f"control.{p.server_name}@{p.hostname}", galaxy_exchange, routing_key="control.*") for p in processes]
 
 
 def control_queues_from_config(config):
diff --git a/lib/galaxy/selenium/has_playwright_driver.py b/lib/galaxy/selenium/has_playwright_driver.py
index d2cf3aa6aead..93235d763ec8 100644
--- a/lib/galaxy/selenium/has_playwright_driver.py
+++ b/lib/galaxy/selenium/has_playwright_driver.py
@@ -112,6 +112,7 @@ def timeout_for(self, **kwds):
 """
 
 import abc
+import logging
 from contextlib import contextmanager
 from typing import (
     Any,
@@ -157,6 +158,8 @@ def timeout_for(self, **kwds):
 from .wait_methods_mixin import WaitMethodsMixin
 from .web_element_protocol import WebElementProtocol
 
+logger = logging.getLogger(__name__)
+
 UNSPECIFIED_TIMEOUT = object()
 
 
@@ -1153,9 +1156,23 @@ def quit(self) -> None:
 
         This closes all windows/tabs and releases all system resources.
         The driver cannot be used after calling this method.
+
+        ``browser.close()`` can raise — in CI we've seen it time out or hit
+        target-detached errors — and if the exception escapes before
+        ``playwright.stop()`` runs, the per-instance asyncio loop is left
+        registered as "running" on the main thread. Every subsequent test's
+        ``sync_playwright().__enter__`` then refuses to start with "Playwright
+        Sync API inside the asyncio loop", cascading the whole shard into
+        errors. Always tear down the Playwright instance even if the browser
+        close failed.
         """
-        self.close()
-        self._playwright_resources.playwright.stop()
+        try:
+            self.close()
+        finally:
+            try:
+                self._playwright_resources.playwright.stop()
+            except Exception:
+                logger.exception("Error stopping Playwright instance during quit()")
 
 
 __all__ = (
diff --git a/lib/galaxy/sse_monitor/__init__.py b/lib/galaxy/sse_monitor/__init__.py
new file mode 100644
index 000000000000..740f55743242
--- /dev/null
+++ b/lib/galaxy/sse_monitor/__init__.py
@@ -0,0 +1,30 @@
+"""Standalone SSE monitor process.
+
+Runs the :class:`~galaxy.managers.history_audit_monitor.HistoryAuditMonitor`
+(and, in principle, any future SSE event producer) in its own OS process so a
+webapp's GIL pauses or fork-lifecycle events can never block history-update
+dispatch. Webapps still consume SSE events normally; only the *producer* moves.
+
+The process registers itself via :class:`~galaxy.model.database_heartbeat.DatabaseHeartbeat`
+with a server_name beginning with ``sse_monitor.``. The heartbeat's
+``is_history_audit_monitor`` election prefers any such process, so the monitor
+automatically migrates here when this daemon is running. If it stops, a webapp
+is re-elected within one heartbeat interval (~60s).
+
+Starting the daemon
+-------------------
+
+Installed::
+
+    GALAXY_CONFIG_FILE=/etc/galaxy/galaxy.yml galaxy-sse-monitor
+
+From a source checkout::
+
+    GALAXY_CONFIG_FILE=config/galaxy.yml python -m galaxy.sse_monitor
+
+For production deployments, run under systemd / supervisord alongside the
+webapp. A native ``gravity`` (``galaxyctl``) service type for this daemon is
+tracked as a follow-up against the ``gravity`` package — until that lands,
+either start the process via your process manager of choice, or omit it and
+let a webapp fall back to the monitor role.
+"""
diff --git a/lib/galaxy/sse_monitor/__main__.py b/lib/galaxy/sse_monitor/__main__.py
new file mode 100644
index 000000000000..9b1098de9fdc
--- /dev/null
+++ b/lib/galaxy/sse_monitor/__main__.py
@@ -0,0 +1,99 @@
+"""Entry point for the standalone ``galaxy-sse-monitor`` daemon.
+
+Loads a minimal :class:`GalaxyManagerApplication` (same shape the Celery
+workers use), forces the server_name to ``sse_monitor.<host>.<pid>`` so
+DatabaseHeartbeat's election picks this process as the history-audit monitor,
+and blocks on SIGINT/SIGTERM.
+"""
+
+import logging
+import os
+import signal
+import socket
+import sys
+import threading
+
+from galaxy.celery import get_app_properties
+from galaxy.model.database_heartbeat import DatabaseHeartbeat
+
+log = logging.getLogger("galaxy.sse_monitor")
+
+
+def _build_server_name() -> str:
+    return f"sse_monitor.{socket.gethostname()}.{os.getpid()}"
+
+
+def _build_app(server_name: str):
+    kwargs = get_app_properties() or {}
+    if not kwargs:
+        raise RuntimeError(
+            "GALAXY_CONFIG_FILE (or GALAXY_ROOT_DIR with an on-disk Galaxy config) is required "
+            "to start galaxy-sse-monitor"
+        )
+    kwargs = dict(kwargs)
+    kwargs["check_migrate_databases"] = False
+    kwargs["use_display_applications"] = False
+    kwargs["use_converters"] = False
+    kwargs["server_name"] = server_name
+
+    import galaxy.app
+
+    return galaxy.app.GalaxyManagerApplication(configure_logging=True, **kwargs)
+
+
+def main() -> int:
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s %(message)s")
+
+    server_name = _build_server_name()
+    log.info("Starting galaxy-sse-monitor as %s", server_name)
+
+    app = _build_app(server_name)
+
+    # GalaxyManagerApplication doesn't wire a DatabaseHeartbeat (that lives on
+    # UniverseApplication, which pulls in the webapp stack we don't need). We
+    # spin up our own and register the audit-monitor callback so election
+    # transitions start/stop the producer cleanly.
+    heartbeat = DatabaseHeartbeat(application_stack=app.application_stack)
+
+    monitor = None
+    if app.config.enable_sse_updates:
+        from galaxy.managers.history_audit_monitor import HistoryAuditMonitor
+
+        monitor = app[HistoryAuditMonitor]
+        heartbeat.add_audit_monitor_change_callback(monitor.on_role_change)
+    else:
+        log.warning("enable_sse_updates is False — galaxy-sse-monitor will idle with no producers")
+
+    heartbeat.start()
+
+    shutdown = threading.Event()
+
+    def _handle_signal(signum, _frame):
+        log.info("Received signal %s, shutting down galaxy-sse-monitor", signum)
+        shutdown.set()
+
+    signal.signal(signal.SIGINT, _handle_signal)
+    signal.signal(signal.SIGTERM, _handle_signal)
+
+    try:
+        shutdown.wait()
+    finally:
+        if monitor is not None:
+            try:
+                monitor.shutdown()
+            except Exception:
+                log.exception("Error shutting down HistoryAuditMonitor")
+        try:
+            heartbeat.shutdown()
+        except Exception:
+            log.exception("Error shutting down database heartbeat")
+        try:
+            app.shutdown()
+        except Exception:
+            log.exception("Error shutting down GalaxyManagerApplication")
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/lib/galaxy/structured_app/__init__.py b/lib/galaxy/structured_app/__init__.py
index f0fb04cb8945..2c8b7c84e526 100644
--- a/lib/galaxy/structured_app/__init__.py
+++ b/lib/galaxy/structured_app/__init__.py
@@ -41,6 +41,7 @@
 from galaxy.workflow.trs_proxy import TrsProxy
 
 if TYPE_CHECKING:
+    from galaxy.app import ExecutionTimerFactory
     from galaxy.config_watchers import ConfigWatchers
     from galaxy.jobs import JobConfiguration
     from galaxy.jobs.manager import JobManager
@@ -57,7 +58,9 @@
         WorkflowsManager,
     )
     from galaxy.tool_shed.galaxy_install.client import DataManagersInterface
-    from galaxy.tool_shed.galaxy_install.installed_repository_manager import InstalledRepositoryManager
+    from galaxy.tool_shed.galaxy_install.installed_repository_manager import (
+        InstalledRepositoryManager,
+    )
     from galaxy.tool_util.data import ToolDataTableManager
     from galaxy.tools import ToolBox
     from galaxy.tools.cache import ToolCache
@@ -174,6 +177,7 @@ class StructuredApp(MinimalManagerApp):
     vault: Vault
     webhooks_registry: WebhooksRegistry
     queue_worker: Any  # 'galaxy.queue_worker.GalaxyQueueWorker'
+    execution_timer_factory: "ExecutionTimerFactory"
     data_provider_registry: Any  # 'galaxy.visualization.data_providers.registry.DataProviderRegistry'
     tool_cache: "ToolCache"
     tool_shed_repository_cache: Optional[ToolShedRepositoryCache]
diff --git a/lib/galaxy/webapps/galaxy/api/events.py b/lib/galaxy/webapps/galaxy/api/events.py
new file mode 100644
index 000000000000..67120591227b
--- /dev/null
+++ b/lib/galaxy/webapps/galaxy/api/events.py
@@ -0,0 +1,62 @@
+"""
+API endpoint for Server-Sent Events (SSE) stream.
+
+Provides a unified event stream for all real-time push events (notifications,
+history updates, etc.) independent of the notification system configuration.
+"""
+
+import logging
+from typing import Optional
+
+from fastapi import (
+    Header,
+    Request,
+)
+from starlette.responses import StreamingResponse
+
+from galaxy.managers.context import ProvidesUserContext
+from galaxy.webapps.galaxy.services.events import EventsService
+from . import (
+    depends,
+    DependsOnTrans,
+    Router,
+)
+
+log = logging.getLogger(__name__)
+
+router = Router(tags=["events"])
+
+
+@router.cbv
+class FastAPIEvents:
+    service: EventsService = depends(EventsService)
+
+    @router.get(
+        "/api/events/stream",
+        summary="Server-Sent Events stream for real-time updates.",
+        response_class=StreamingResponse,
+    )
+    async def stream_events(
+        self,
+        request: Request,
+        trans: ProvidesUserContext = DependsOnTrans,
+        last_event_id: Optional[str] = Header(None, alias="Last-Event-ID"),
+    ) -> StreamingResponse:
+        """Opens a Server-Sent Events (SSE) connection that pushes real-time
+        updates for notifications, history changes, and other events.
+
+        On reconnect, the browser sends the ``Last-Event-ID`` header automatically.
+        If the notification system is enabled, any notifications created since that
+        timestamp are delivered as a catch-up ``notification_status`` event.
+
+        Anonymous users receive only broadcast events.
+        """
+        return StreamingResponse(
+            self.service.open_stream(trans, last_event_id, request.is_disconnected),
+            media_type="text/event-stream",
+            headers={
+                "Cache-Control": "no-cache",
+                "Connection": "keep-alive",
+                "X-Accel-Buffering": "no",
+            },
+        )
diff --git a/lib/galaxy/webapps/galaxy/api/tool_data.py b/lib/galaxy/webapps/galaxy/api/tool_data.py
index c7ae44f45aec..3d1f1a9abb8f 100644
--- a/lib/galaxy/webapps/galaxy/api/tool_data.py
+++ b/lib/galaxy/webapps/galaxy/api/tool_data.py
@@ -11,6 +11,7 @@
     Field,
 )
 
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import import_data_bundle
 from galaxy.managers.context import ProvidesUserContext
 from galaxy.managers.tool_data import ToolDataManager
@@ -25,7 +26,6 @@
     ToolDataItem,
 )
 from galaxy.webapps.base.api import GalaxyFileResponse
-from galaxy.webapps.galaxy.services.base import async_task_summary
 from . import (
     depends,
     DependsOnTrans,
diff --git a/lib/galaxy/webapps/galaxy/services/base.py b/lib/galaxy/webapps/galaxy/services/base.py
index 5e096c7a6191..e86924966186 100644
--- a/lib/galaxy/webapps/galaxy/services/base.py
+++ b/lib/galaxy/webapps/galaxy/services/base.py
@@ -7,8 +7,6 @@
     Optional,
 )
 
-from celery.result import AsyncResult
-
 from galaxy.exceptions import (
     AuthenticationRequired,
     ConfigDoesNotAllowException,
@@ -32,7 +30,6 @@
 )
 from galaxy.schema.fields import EncodedDatabaseIdField
 from galaxy.schema.schema import (
-    AsyncTaskResultSummary,
     ToolRequestDetailedModel,
     ToolRequestModel,
 )
@@ -185,28 +182,6 @@ def create_objects_from_store(
         )
 
 
-def async_task_summary(async_result: AsyncResult) -> AsyncTaskResultSummary:
-    name = None
-    try:
-        name = async_result.name
-    except AttributeError:
-        # if backend is disabled, we won't have this
-        pass
-    queue = None
-    try:
-        queue = async_result.queue
-    except AttributeError:
-        # if backend is disabled, we won't have this
-        pass
-
-    return AsyncTaskResultSummary(
-        id=str(async_result.id),
-        ignored=async_result.ignored,
-        name=name,
-        queue=queue,
-    )
-
-
 def _encode_tool_request(tool_request: ToolRequest, security: IdEncodingHelper) -> dict[str, Any]:
     """Encode request IDs using strongly-typed parameter walking."""
     tool_source_model = tool_request.tool_source
diff --git a/lib/galaxy/webapps/galaxy/services/datasets.py b/lib/galaxy/webapps/galaxy/services/datasets.py
index aae67a7ec74b..d3f7f8ee8b3a 100644
--- a/lib/galaxy/webapps/galaxy/services/datasets.py
+++ b/lib/galaxy/webapps/galaxy/services/datasets.py
@@ -24,6 +24,7 @@
     util,
     web,
 )
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import compute_dataset_hash
 from galaxy.datatypes.binary import Binary
 from galaxy.datatypes.dataproviders.exceptions import NoProviderAvailable
@@ -88,10 +89,7 @@
 )
 from galaxy.visualization.data_providers.registry import DataProviderRegistry
 from galaxy.webapps.base.controller import UsesVisualizationMixin
-from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
-    ServiceBase,
-)
+from galaxy.webapps.galaxy.services.base import ServiceBase
 
 log = logging.getLogger(__name__)
 
diff --git a/lib/galaxy/webapps/galaxy/services/events.py b/lib/galaxy/webapps/galaxy/services/events.py
new file mode 100644
index 000000000000..170c4c035efc
--- /dev/null
+++ b/lib/galaxy/webapps/galaxy/services/events.py
@@ -0,0 +1,44 @@
+"""Service layer for the unified SSE events endpoint.
+
+Unlike :class:`NotificationService.open_stream`, this service does **not**
+require the notification system to be enabled — ``/api/events/stream`` also
+serves history updates and other event types independent of the notification
+configuration. When notifications are disabled the catch-up event is simply
+skipped; the stream still delivers other push events.
+"""
+
+from collections.abc import AsyncIterator
+from typing import (
+    Optional,
+)
+
+from galaxy.managers.context import ProvidesUserContext
+from galaxy.managers.sse import (
+    IsDisconnected,
+    SSEConnectionManager,
+)
+from galaxy.webapps.galaxy.services.base import ServiceBase
+from galaxy.webapps.galaxy.services.notifications import NotificationService
+
+
+class EventsService(ServiceBase):
+    def __init__(self, sse_manager: SSEConnectionManager, notifications: NotificationService) -> None:
+        self.sse_manager = sse_manager
+        self.notifications = notifications
+
+    def open_stream(
+        self,
+        user_context: ProvidesUserContext,
+        last_event_id: Optional[str],
+        is_disconnected: IsDisconnected,
+    ) -> AsyncIterator[str]:
+        """Open an SSE events stream.
+
+        Anonymous users still register under their ``galaxy_session.id`` so the
+        server can route per-session events (e.g. ``history_update`` for
+        anonymous-owned histories) even when ``user_id`` is ``None``.
+        """
+        user_id = user_context.user.id if not user_context.anonymous else None
+        session_id = user_context.galaxy_session.id if user_context.galaxy_session else None
+        catch_up = self.notifications.build_status_catchup(user_context, last_event_id)
+        return self.sse_manager.stream(is_disconnected, user_id, catch_up=catch_up, galaxy_session_id=session_id)
diff --git a/lib/galaxy/webapps/galaxy/services/histories.py b/lib/galaxy/webapps/galaxy/services/histories.py
index 25e8982da5df..097fbc1c96f4 100644
--- a/lib/galaxy/webapps/galaxy/services/histories.py
+++ b/lib/galaxy/webapps/galaxy/services/histories.py
@@ -24,6 +24,7 @@
     exceptions as glx_exceptions,
     model,
 )
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import (
     import_model_store,
     prepare_history_download,
@@ -87,7 +88,6 @@
 from galaxy.short_term_storage import ShortTermStorageAllocator
 from galaxy.util import restore_text
 from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
     ConsumesModelStores,
     model_store_storage_target,
     ServesExportStores,
diff --git a/lib/galaxy/webapps/galaxy/services/history_contents.py b/lib/galaxy/webapps/galaxy/services/history_contents.py
index 581d1dc772e7..53926007bd31 100644
--- a/lib/galaxy/webapps/galaxy/services/history_contents.py
+++ b/lib/galaxy/webapps/galaxy/services/history_contents.py
@@ -21,6 +21,7 @@
 )
 
 from galaxy import exceptions
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import (
     change_datatype,
     materialize as materialize_task,
@@ -118,7 +119,6 @@
 from galaxy.short_term_storage import ShortTermStorageAllocator
 from galaxy.util.zipstream import ZipstreamWrapper
 from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
     ConsumesModelStores,
     ensure_celery_tasks_enabled,
     model_store_storage_target,
diff --git a/lib/galaxy/webapps/galaxy/services/invocations.py b/lib/galaxy/webapps/galaxy/services/invocations.py
index 29654d885a41..dc5cc27a5770 100644
--- a/lib/galaxy/webapps/galaxy/services/invocations.py
+++ b/lib/galaxy/webapps/galaxy/services/invocations.py
@@ -6,6 +6,7 @@
 
 from pydantic import Field
 
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import (
     prepare_invocation_download,
     write_invocation_to,
@@ -61,7 +62,6 @@
 from galaxy.security.idencoding import IdEncodingHelper
 from galaxy.short_term_storage import ShortTermStorageAllocator
 from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
     ConsumesModelStores,
     ensure_celery_tasks_enabled,
     model_store_storage_target,
diff --git a/lib/galaxy/webapps/galaxy/services/jobs.py b/lib/galaxy/webapps/galaxy/services/jobs.py
index fac6ea9ecf37..daadbe37b13a 100644
--- a/lib/galaxy/webapps/galaxy/services/jobs.py
+++ b/lib/galaxy/webapps/galaxy/services/jobs.py
@@ -16,6 +16,7 @@
     exceptions,
     model,
 )
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import queue_jobs
 from galaxy.managers import hdas
 from galaxy.managers.base import security_check
@@ -60,7 +61,6 @@
     ToolParameterBundleModel,
 )
 from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
     ServiceBase,
 )
 from .tools import validate_tool_for_running
diff --git a/lib/galaxy/webapps/galaxy/services/notifications.py b/lib/galaxy/webapps/galaxy/services/notifications.py
index ae3e44947197..7fd4cd696a59 100644
--- a/lib/galaxy/webapps/galaxy/services/notifications.py
+++ b/lib/galaxy/webapps/galaxy/services/notifications.py
@@ -5,7 +5,6 @@
     Union,
 )
 
-from galaxy.celery.tasks import send_notification_to_recipients_async
 from galaxy.exceptions import (
     AdminRequiredException,
     AuthenticationRequired,
@@ -14,6 +13,12 @@
 )
 from galaxy.managers.context import ProvidesUserContext
 from galaxy.managers.notification import NotificationManager
+from galaxy.managers.sse import (
+    make_event_id,
+    parse_event_id,
+    SSEConnectionManager,
+    SSEEvent,
+)
 from galaxy.model import User
 from galaxy.schema.fields import Security
 from galaxy.schema.notifications import (
@@ -35,15 +40,27 @@
     UserNotificationUpdateRequest,
 )
 from galaxy.schema.schema import AsyncTaskResultSummary
-from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
-    ServiceBase,
-)
+from galaxy.webapps.galaxy.services.base import ServiceBase
 
 
 class NotificationService(ServiceBase):
-    def __init__(self, notification_manager: NotificationManager):
+    def __init__(self, notification_manager: NotificationManager, sse_manager: SSEConnectionManager):
         self.notification_manager = notification_manager
+        self.sse_manager = sse_manager
+
+    @property
+    def notifications_enabled(self) -> bool:
+        return self.notification_manager.notifications_enabled
+
+    def send_internal_notification(
+        self, request: NotificationCreateRequest, force_sync: bool = False
+    ) -> Union[NotificationCreatedResponse, AsyncTaskResultSummary]:
+        """Send a system-emitted notification on behalf of internal callers (e.g. share flows).
+
+        Unlike :meth:`send_notification`, this skips admin/permission checks because the
+        caller has already resolved the recipient set and is not acting on user input.
+        """
+        return self.notification_manager.send_notification_internal(request, force_sync=force_sync)
 
     def send_notification(
         self, sender_context: ProvidesUserContext, payload: NotificationCreateRequestBody
@@ -62,28 +79,7 @@ def send_notification(
             recipients=payload.recipients,
             galaxy_url=galaxy_url,
         )
-        return self.send_notification_internal(request)
-
-    def send_notification_internal(
-        self, request: NotificationCreateRequest, force_sync: bool = False
-    ) -> Union[NotificationCreatedResponse, AsyncTaskResultSummary]:
-        """Sends a notification to a list of recipients (users, groups or roles).
-
-        If `force_sync` is set to `True`, the notification recipients will be processed synchronously instead of
-        in a background task.
-
-        Note: This function is meant for internal use from other services that don't need to check sender permissions.
-        """
-        if self.notification_manager.can_send_notifications_async and not force_sync:
-            result = send_notification_to_recipients_async.delay(request)
-            summary = async_task_summary(result)
-            return summary
-
-        notification, recipient_user_count = self.notification_manager.send_notification_to_recipients(request)
-        return NotificationCreatedResponse(
-            total_notifications_sent=recipient_user_count,
-            notification=NotificationResponse.model_validate(notification),
-        )
+        return self.notification_manager.send_notification_internal(request)
 
     def broadcast(
         self, sender_context: ProvidesUserContext, payload: BroadcastNotificationCreateRequest
@@ -99,6 +95,27 @@ def broadcast(
             total_notifications_sent=1, notification=NotificationResponse.model_validate(notification)
         )
 
+    def build_status_catchup(
+        self, user_context: ProvidesUserContext, last_event_id: Optional[str]
+    ) -> Optional[SSEEvent]:
+        """Build a ``notification_status`` SSE event covering everything since ``last_event_id``.
+
+        Returns ``None`` when catch-up isn't possible (no ``Last-Event-ID``,
+        unparseable timestamp, or notifications disabled) so callers can simply
+        pass the result to ``SSEConnectionManager.stream`` without extra guards.
+        """
+        if not last_event_id or not self.notification_manager.notifications_enabled:
+            return None
+        since = parse_event_id(last_event_id)
+        if since is None:
+            return None
+        catchup = self.get_notifications_status(user_context, since)
+        return SSEEvent(
+            event="notification_status",
+            data=catchup.model_dump_json(),
+            id=make_event_id(),
+        )
+
     def get_notifications_status(self, user_context: ProvidesUserContext, since: datetime) -> NotificationStatusSummary:
         """Returns the status of (unread or updated) notifications received by the user **since** a particular date and time.
 
diff --git a/lib/galaxy/webapps/galaxy/services/pages.py b/lib/galaxy/webapps/galaxy/services/pages.py
index d77560efd272..e95e6dd959de 100644
--- a/lib/galaxy/webapps/galaxy/services/pages.py
+++ b/lib/galaxy/webapps/galaxy/services/pages.py
@@ -4,6 +4,7 @@
 )
 
 from galaxy import exceptions
+from galaxy.celery.helpers import async_task_summary
 from galaxy.celery.tasks import prepare_pdf_download
 from galaxy.managers import base
 from galaxy.managers.markdown_util import (
@@ -34,7 +35,6 @@
 from galaxy.short_term_storage import ShortTermStorageAllocator
 from galaxy.webapps.galaxy.api.common import PageIdPathParam
 from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
     ensure_celery_tasks_enabled,
     ServiceBase,
 )
diff --git a/lib/galaxy/webapps/galaxy/services/sharable.py b/lib/galaxy/webapps/galaxy/services/sharable.py
index 33697adf51f2..eec9fb1e2579 100644
--- a/lib/galaxy/webapps/galaxy/services/sharable.py
+++ b/lib/galaxy/webapps/galaxy/services/sharable.py
@@ -177,17 +177,13 @@ def _get_users(self, trans, emails_or_ids: list[UserIdentifier]) -> tuple[set[Us
     def _send_notification_to_users(
         self, users_to_notify: set[User], item: SharableItem, status: ShareWithStatus, galaxy_url: Optional[str] = None
     ):
-        if (
-            self.notification_service.notification_manager.notifications_enabled
-            and not status.errors
-            and users_to_notify
-        ):
+        if self.notification_service.notifications_enabled and not status.errors and users_to_notify:
             request = SharedItemNotificationFactory.build_notification_request(
                 item, users_to_notify, status, galaxy_url
             )
             # We can set force_sync=True here because we already have the set of users to notify
             # and there is no need to resolve them asynchronously as no groups or roles are involved.
-            self.notification_service.send_notification_internal(request, force_sync=True)
+            self.notification_service.send_internal_notification(request, force_sync=True)
 
 
 class SharedItemNotificationFactory:
diff --git a/lib/galaxy/webapps/galaxy/services/users.py b/lib/galaxy/webapps/galaxy/services/users.py
index 39aa35c03a45..dd8fa7650c12 100644
--- a/lib/galaxy/webapps/galaxy/services/users.py
+++ b/lib/galaxy/webapps/galaxy/services/users.py
@@ -9,6 +9,7 @@
     exceptions as glx_exceptions,
     util,
 )
+from galaxy.celery.helpers import async_task_summary
 from galaxy.managers import api_keys
 from galaxy.managers.context import (
     ProvidesHistoryContext,
@@ -34,10 +35,7 @@
     UserModel,
 )
 from galaxy.security.idencoding import IdEncodingHelper
-from galaxy.webapps.galaxy.services.base import (
-    async_task_summary,
-    ServiceBase,
-)
+from galaxy.webapps.galaxy.services.base import ServiceBase
 from galaxy.webapps.galaxy.services.roles import role_to_model
 
 if TYPE_CHECKING:
diff --git a/lib/galaxy_test/base/sse.py b/lib/galaxy_test/base/sse.py
new file mode 100644
index 000000000000..d98ab7c6c528
--- /dev/null
+++ b/lib/galaxy_test/base/sse.py
@@ -0,0 +1,155 @@
+"""Shared helpers for SSE integration tests.
+
+The stream layer emits a ``ready`` event as the first frame on every connection
+so tests can synchronize on the server-side subscription rather than the
+underlying TCP socket. ``SSELineListener`` waits for that ``ready`` event before
+``start()`` returns, and propagates listener-thread exceptions back to the main
+thread instead of silently swallowing them.
+"""
+
+import queue
+import threading
+from collections.abc import Callable
+from typing import Optional
+
+import requests
+
+from galaxy.util.wait import wait_on
+
+CONNECT_TIMEOUT = 15
+DEFAULT_WAIT_TIMEOUT = 15
+
+
+def parse_sse_events(raw: str) -> list[dict]:
+    """Parse raw SSE text into a list of event dicts with ``event``, ``data``, and ``id`` keys."""
+    events: list[dict] = []
+    current: dict[str, str] = {}
+    for line in raw.split("\n"):
+        if line.startswith(":"):
+            continue  # comment / keepalive
+        if line == "":
+            if current:
+                events.append(current)
+                current = {}
+            continue
+        if ": " in line:
+            field, _, value = line.partition(": ")
+        else:
+            field, value = line.rstrip(":"), ""
+        if field in ("event", "data", "id"):
+            current[field] = value
+    if current:
+        events.append(current)
+    return events
+
+
+class SSEListenerError(Exception):
+    """Wraps an exception raised inside the listener thread."""
+
+
+class SSELineListener:
+    """Runs an SSE connection on a background thread and collects raw chunks.
+
+    ``start()`` blocks until the server-side ``ready`` event has been received,
+    guaranteeing that any event *posted after* ``start()`` returns will be seen
+    by this listener. Failures in the background thread are surfaced via
+    ``wait_for_event`` instead of silently timing out.
+    """
+
+    def __init__(
+        self,
+        url: str,
+        api_key: str,
+        headers: Optional[dict] = None,
+        timeout: int = 30,
+    ) -> None:
+        self.url = url
+        self.api_key = api_key
+        self.headers = headers or {}
+        self.timeout = timeout
+        self._collected: list[str] = []
+        self._stop = threading.Event()
+        self._ready = threading.Event()
+        self._errors: queue.Queue[BaseException] = queue.Queue()
+        self._thread = threading.Thread(target=self._listen, daemon=True)
+
+    def start(self) -> None:
+        self._thread.start()
+        wait_on(
+            lambda: True if self._ready.is_set() else None,
+            "SSE `ready` event",
+            timeout=CONNECT_TIMEOUT,
+        )
+        self._raise_if_errored()
+
+    def stop(self) -> None:
+        self._stop.set()
+        self._thread.join(timeout=5)
+
+    def wait_for_event(self, event_type: str, timeout: int = DEFAULT_WAIT_TIMEOUT) -> list[dict]:
+        """Block until at least one event of ``event_type`` has been observed, then return all such events."""
+
+        def _check():
+            self._raise_if_errored()
+            events = self.get_events(event_type)
+            return events if events else None
+
+        return wait_on(_check, f"SSE {event_type} event", timeout=timeout)
+
+    def wait_for_event_where(
+        self,
+        event_type: str,
+        predicate: Callable[[dict], bool],
+        timeout: int = DEFAULT_WAIT_TIMEOUT,
+    ) -> list[dict]:
+        """Block until at least one ``event_type`` event matches ``predicate``.
+
+        Returns every ``event_type`` event observed so far, not just matches, so
+        callers can still inspect the surrounding stream (e.g. assert what else
+        did or didn't appear) after the wait resolves.
+        """
+
+        def _check():
+            self._raise_if_errored()
+            events = self.get_events(event_type)
+            return events if any(predicate(e) for e in events) else None
+
+        return wait_on(_check, f"SSE {event_type} matching predicate", timeout=timeout)
+
+    def get_events(self, event_type: Optional[str] = None) -> list[dict]:
+        """Return all collected events so far, optionally filtered by type."""
+        all_events = parse_sse_events("".join(self._collected))
+        if event_type is None:
+            return all_events
+        return [e for e in all_events if e.get("event") == event_type]
+
+    def _raise_if_errored(self) -> None:
+        try:
+            err = self._errors.get_nowait()
+        except queue.Empty:
+            return
+        raise SSEListenerError(f"SSE listener thread failed: {err!r}") from err
+
+    def _listen(self) -> None:
+        try:
+            resp = requests.get(
+                self.url,
+                params={"key": self.api_key},
+                headers=self.headers,
+                stream=True,
+                timeout=self.timeout,
+            )
+            if resp.status_code != 200:
+                raise RuntimeError(f"SSE connect returned HTTP {resp.status_code}: {resp.text[:200]}")
+            for chunk in resp.iter_content(chunk_size=None, decode_unicode=True):
+                if chunk:
+                    self._collected.append(chunk)
+                    if not self._ready.is_set() and "event: ready" in "".join(self._collected):
+                        self._ready.set()
+                if self._stop.is_set():
+                    break
+            resp.close()
+        except Exception as exc:
+            self._errors.put(exc)
+            # Ensure start() doesn't hang forever on connection failure.
+            self._ready.set()
diff --git a/lib/galaxy_test/driver/integration_util.py b/lib/galaxy_test/driver/integration_util.py
index c7bc18c971cd..c15d3580ae80 100644
--- a/lib/galaxy_test/driver/integration_util.py
+++ b/lib/galaxy_test/driver/integration_util.py
@@ -22,8 +22,10 @@
     skip,
     SkipTest,
 )
+from urllib.parse import urljoin
 
 import pytest
+import requests
 
 from galaxy.app import UniverseApplication
 from galaxy.tool_util.verify.test_data import TestDataResolver
@@ -238,6 +240,16 @@ def _skip_unless_postgres(self):
         if not self._app.config.database_connection.startswith("post"):
             raise SkipTest("Test only valid for postgres")
 
+    def _decode_id(self, encoded_id: str) -> int:
+        """Decode an encoded API id to its raw int via the live app's security helper."""
+        return self._app.security.decode_id(encoded_id)
+
+    def _user_id_for_api_key(self, api_key: str) -> int:
+        """Return the raw integer ``User.id`` for the user owning ``api_key``."""
+        response = requests.get(urljoin(self.url, "api/users/current"), params={"key": api_key})
+        response.raise_for_status()
+        return self._decode_id(response.json()["id"])
+
     def _run_tool_test(self, *args, **kwargs):
         return self._test_driver.run_tool_test(*args, **kwargs)
 
diff --git a/lib/galaxy_test/selenium/framework.py b/lib/galaxy_test/selenium/framework.py
index 6b3b775705a1..d6dff43dd36d 100644
--- a/lib/galaxy_test/selenium/framework.py
+++ b/lib/galaxy_test/selenium/framework.py
@@ -3,6 +3,7 @@
 import datetime
 import errno
 import json
+import logging
 import os
 import traceback
 import unittest
@@ -76,6 +77,8 @@
 except ImportError:
     GalaxyTestDriver = None  # type: ignore[assignment, misc, unused-ignore]
 
+logger = logging.getLogger(__name__)
+
 
 def _load_config_file() -> None:
     """
@@ -433,10 +436,22 @@ def setup_selenium(self):
         self.target_url_from_selenium = self._target_url_from_selenium()
         self.snapshots = []
         self.setup_driver_and_session()
-        if self.run_as_admin and GALAXY_TEST_SELENIUM_ADMIN_USER_EMAIL == DEFAULT_ADMIN_USER:
-            self._setup_interactor()
-            self._setup_user(GALAXY_TEST_SELENIUM_ADMIN_USER_EMAIL)
-        self._try_setup_with_driver()
+        # Once the driver is allocated, any subsequent failure must still
+        # tear it down: pytest does not call tearDown when setUp raises, so
+        # without this the Playwright asyncio loop would stay registered as
+        # "running" on the main thread and cascade every subsequent test's
+        # setUp with "Sync API inside the asyncio loop".
+        try:
+            if self.run_as_admin and GALAXY_TEST_SELENIUM_ADMIN_USER_EMAIL == DEFAULT_ADMIN_USER:
+                self._setup_interactor()
+                self._setup_user(GALAXY_TEST_SELENIUM_ADMIN_USER_EMAIL)
+            self._try_setup_with_driver()
+        except Exception:
+            try:
+                self.tear_down_driver()
+            except Exception:
+                logger.exception("Error tearing down driver after setup_selenium failure")
+            raise
 
     def _try_setup_with_driver(self):
         try:
diff --git a/packages/app/setup.cfg b/packages/app/setup.cfg
index e1001c06c6a6..c455dc2cec44 100644
--- a/packages/app/setup.cfg
+++ b/packages/app/setup.cfg
@@ -101,6 +101,7 @@ test =
 console_scripts =
         galaxy-main = galaxy.main:main
         galaxy-dependencies = galaxy.dependencies.script:main
+        galaxy-sse-monitor = galaxy.sse_monitor.__main__:main
 
 [options.packages.find]
 where = src
diff --git a/packages/app/src/galaxy/sse_monitor b/packages/app/src/galaxy/sse_monitor
new file mode 120000
index 000000000000..e714ba834b24
--- /dev/null
+++ b/packages/app/src/galaxy/sse_monitor
@@ -0,0 +1 @@
+../../../../lib/galaxy/sse_monitor
\ No newline at end of file
diff --git a/pyproject.toml b/pyproject.toml
index ee68a0fc1c05..c3ec9ed16cf7 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -176,6 +176,7 @@ typecheck = [
     "pydantic>=2.7.4",  # for pydantic.mypy plugin
     "types-bleach",
     "types-boto",
+    "types-cachetools",
     "types-contextvars",
     "types-dataclasses",
     "types-docutils",
diff --git a/test/integration/test_entry_point_sse.py b/test/integration/test_entry_point_sse.py
new file mode 100644
index 000000000000..1929310a737b
--- /dev/null
+++ b/test/integration/test_entry_point_sse.py
@@ -0,0 +1,146 @@
+"""Integration tests for SSE-based interactive-tool entry-point update notifications.
+
+Mirrors ``test_history_sse.py``. Rather than spin up a real containerized
+interactive tool (which would require docker), these tests exercise the
+dispatch path by creating a ``Job`` and ``InteractiveToolEntryPoint`` rows
+directly via the live app's SQLAlchemy session and invoking
+``InteractiveToolManager.configure_entry_points`` with a stub ``ports_dict``.
+This is exactly the moment the event fires in production — the job runner's
+port-routing hook is the only upstream caller, and the SSE dispatch happens
+after the DB commit regardless of how the ports were obtained.
+"""
+
+import time
+from urllib.parse import urljoin
+from uuid import uuid4
+
+from galaxy.model import (
+    InteractiveToolEntryPoint,
+    Job,
+)
+from galaxy_test.base.populators import DatasetPopulator
+from galaxy_test.base.sse import SSELineListener
+from galaxy_test.driver.integration_util import IntegrationTestCase
+
+
+def _make_ports_dict(tool_port: int) -> dict:
+    """Stub the runner's port-routing payload — one tool_port, fake host/proto."""
+    return {
+        str(tool_port): {
+            "host": "host.invalid",
+            "port": 12345,
+            "protocol": "http",
+        }
+    }
+
+
+class TestEntryPointSSEIntegration(IntegrationTestCase):
+    dataset_populator: DatasetPopulator
+    framework_tool_and_types = True
+
+    @classmethod
+    def handle_galaxy_config_kwds(cls, config):
+        super().handle_galaxy_config_kwds(config)
+        config["enable_celery_tasks"] = False
+
+    def setUp(self):
+        super().setUp()
+        self.dataset_populator = DatasetPopulator(self.galaxy_interactor)
+
+    def _events_stream_url(self) -> str:
+        return urljoin(self.url, "api/events/stream")
+
+    def _create_it_job_with_entry_point(self, user_id: int, tool_port: int = 8888) -> tuple[int, int]:
+        """Create a minimal Job + unconfigured InteractiveToolEntryPoint row pair.
+
+        Returns ``(job_id, entry_point_id)``. The session used is the live
+        app's; the rows are real and survive the call.
+        """
+        sa_session = self._app.model.context
+        job = Job()
+        job.user_id = user_id
+        job.tool_id = "interactivetool_simple"
+        job.state = Job.states.RUNNING
+        sa_session.add(job)
+        sa_session.flush()
+        ep = InteractiveToolEntryPoint(
+            job=job,
+            tool_port=tool_port,
+            entry_url="/",
+            name="test entry point",
+            label="test",
+            requires_domain=True,
+            requires_path_in_url=False,
+            requires_path_in_header_named=None,
+        )
+        sa_session.add(ep)
+        sa_session.commit()
+        return job.id, ep.id
+
+    def test_entry_point_update_event_fires_on_configure(self):
+        """configure_entry_points should fire an ``entry_point_update`` wake-up event."""
+        api_key = self.galaxy_interactor.api_key
+        assert api_key is not None
+        user_id = self._user_id_for_api_key(api_key)
+        job_id, _ = self._create_it_job_with_entry_point(user_id)
+
+        listener = SSELineListener(self._events_stream_url(), api_key)
+        listener.start()
+        try:
+            sa_session = self._app.model.context
+            job = sa_session.get(Job, job_id)
+            assert job is not None
+            self._app.interactivetool_manager.configure_entry_points(job, _make_ports_dict(8888))
+
+            entry_point_events = listener.wait_for_event("entry_point_update")
+        finally:
+            listener.stop()
+
+        # The event carries no payload: the client refetches ``/api/entry_points``
+        # (the canonical source) on receipt, so the event just needs to arrive.
+        assert len(entry_point_events) >= 1, f"Expected entry_point_update wake-up, got: {entry_point_events}"
+        assert entry_point_events[0]["event"] == "entry_point_update"
+
+    def test_entry_point_update_is_scoped_to_owning_user(self):
+        """User A must not see entry_point_update events for user B's jobs.
+
+        The event has no payload to cross-check with, so we assert on event
+        count: user A's stream should receive exactly one event for its own
+        ``configure_entry_points`` call and none for user B's.
+        """
+        user_b = self._setup_user(f"{uuid4()}@galaxy.test")
+        _, user_b_api_key = self._setup_user_get_key(user_b["email"])
+        user_b_id = self._user_id_for_api_key(user_b_api_key)
+
+        user_a_api_key = self.galaxy_interactor.api_key
+        assert user_a_api_key is not None
+        user_a_id = self._user_id_for_api_key(user_a_api_key)
+
+        job_a_id, _ = self._create_it_job_with_entry_point(user_a_id, tool_port=7001)
+        job_b_id, _ = self._create_it_job_with_entry_point(user_b_id, tool_port=7002)
+
+        listener = SSELineListener(self._events_stream_url(), user_a_api_key)
+        listener.start()
+        try:
+            sa_session = self._app.model.context
+            job_b = sa_session.get(Job, job_b_id)
+            assert job_b is not None
+            # User B's job — user A must NOT see this.
+            self._app.interactivetool_manager.configure_entry_points(job_b, _make_ports_dict(7002))
+            # Give the broker a moment so a leaked event (if any) would arrive
+            # before we fire user A's event — the assertion would then catch
+            # more than one event on user A's stream.
+            time.sleep(0.5)
+
+            job_a = sa_session.get(Job, job_a_id)
+            assert job_a is not None
+            # User A's own job — this is what A's stream must observe.
+            self._app.interactivetool_manager.configure_entry_points(job_a, _make_ports_dict(7001))
+
+            entry_point_events = listener.wait_for_event("entry_point_update")
+        finally:
+            listener.stop()
+
+        assert (
+            len(entry_point_events) == 1
+        ), f"User A expected exactly one entry_point_update (own job); saw {len(entry_point_events)}: {entry_point_events}"
diff --git a/test/integration/test_history_sse.py b/test/integration/test_history_sse.py
new file mode 100644
index 000000000000..f60bfef1b15a
--- /dev/null
+++ b/test/integration/test_history_sse.py
@@ -0,0 +1,104 @@
+"""Integration tests for SSE-based history update notifications."""
+
+import json
+from urllib.parse import urljoin
+from uuid import uuid4
+
+import requests
+
+from galaxy_test.base.populators import DatasetPopulator
+from galaxy_test.base.sse import SSELineListener
+from galaxy_test.driver.integration_util import IntegrationTestCase
+
+
+class TestHistorySSEIntegration(IntegrationTestCase):
+    dataset_populator: DatasetPopulator
+    framework_tool_and_types = True
+
+    @classmethod
+    def handle_galaxy_config_kwds(cls, config):
+        super().handle_galaxy_config_kwds(config)
+        config["enable_celery_tasks"] = False
+        config["enable_sse_updates"] = True
+
+    def setUp(self):
+        super().setUp()
+        self.dataset_populator = DatasetPopulator(self.galaxy_interactor)
+
+    def _events_stream_url(self) -> str:
+        return urljoin(self.url, "api/events/stream")
+
+    def _create_history(self, name=None) -> str:
+        name = name or f"test_history_{uuid4()}"
+        response = self._post("histories", data={"name": name}, json=True)
+        self._assert_status_code_is_ok(response)
+        return response.json()["id"]
+
+    def test_history_update_contains_current_history_id(self):
+        """The history_update event should contain the history's encoded ID."""
+        history_id = self._create_history()
+
+        api_key = self.galaxy_interactor.api_key
+        assert api_key is not None
+        listener = SSELineListener(self._events_stream_url(), api_key)
+        listener.start()
+        try:
+            self.dataset_populator.new_dataset(history_id, wait=False)
+            history_events = listener.wait_for_event_where(
+                "history_update",
+                lambda e: history_id in json.loads(e["data"]).get("history_ids", []),
+            )
+            found = any(history_id in json.loads(e["data"]).get("history_ids", []) for e in history_events)
+            assert found, f"Expected history_id '{history_id}' in history_update events, got: {history_events}"
+        finally:
+            listener.stop()
+
+    def test_history_update_is_scoped_to_owning_user(self):
+        """User A must only see history_update events for their own histories.
+
+        Inverted positive assertion: after user B's upload, user A uploads to
+        their own history and we assert A's stream contains A's encoded id and
+        not B's. This avoids a sleep-based "no events" test that was prone to
+        flaking under slow CI.
+        """
+        user_b = self._setup_user(f"{uuid4()}@galaxy.test")
+        _, user_b_api_key = self._setup_user_get_key(user_b["email"])
+
+        user_a_history_id = self._create_history()
+
+        api_key = self.galaxy_interactor.api_key
+        assert api_key is not None
+        listener = SSELineListener(self._events_stream_url(), api_key)
+        listener.start()
+        try:
+            # User B creates a history and uploads to it. User A must NOT see this.
+            create_resp = requests.post(
+                urljoin(self.url, "api/histories"),
+                params={"key": user_b_api_key},
+                json={"name": "User B History"},
+            )
+            assert create_resp.status_code == 200
+            user_b_history_id = create_resp.json()["id"]
+
+            requests.post(
+                urljoin(self.url, f"api/histories/{user_b_history_id}/contents"),
+                params={"key": user_b_api_key},
+                json={"from_hda_id": None, "source": "pasted", "content": "user b content"},
+            )
+
+            # User A uploads to their own history — this is what A's stream must observe.
+            self.dataset_populator.new_dataset(user_a_history_id, wait=False)
+            history_events = listener.wait_for_event_where(
+                "history_update",
+                lambda e: user_a_history_id in json.loads(e["data"]).get("history_ids", []),
+            )
+        finally:
+            listener.stop()
+
+        seen_ids: set[str] = set()
+        for event in history_events:
+            seen_ids.update(json.loads(event["data"]).get("history_ids", []))
+        assert user_a_history_id in seen_ids, f"User A missed its own history_update: {history_events}"
+        assert (
+            user_b_history_id not in seen_ids
+        ), f"User A received history_update for user B's history ({user_b_history_id}): {history_events}"
diff --git a/test/integration/test_notification_sse.py b/test/integration/test_notification_sse.py
new file mode 100644
index 000000000000..7ec3bf7f2e53
--- /dev/null
+++ b/test/integration/test_notification_sse.py
@@ -0,0 +1,173 @@
+"""Integration tests for the notification SSE (Server-Sent Events) endpoint."""
+
+import json
+from typing import Optional
+from urllib.parse import urljoin
+from uuid import uuid4
+
+from galaxy_test.base.populators import DatasetPopulator
+from galaxy_test.base.sse import SSELineListener
+from galaxy_test.driver.integration_util import IntegrationTestCase
+
+
+def notification_test_data(subject: Optional[str] = None, message: Optional[str] = None) -> dict:
+    return {
+        "source": "integration_tests",
+        "variant": "info",
+        "category": "message",
+        "content": {
+            "category": "message",
+            "subject": subject or "Testing Subject",
+            "message": message or "Testing Message",
+        },
+    }
+
+
+def notification_broadcast_test_data(subject: Optional[str] = None, message: Optional[str] = None) -> dict:
+    return {
+        "source": "integration_tests",
+        "variant": "info",
+        "category": "broadcast",
+        "content": {
+            "category": "broadcast",
+            "subject": subject or "Testing Broadcast Subject",
+            "message": message or "Testing Broadcast Message",
+        },
+    }
+
+
+def _notification_subjects(events: list[dict]) -> list[str]:
+    """Extract ``content.subject`` from each SSE ``data`` payload.
+
+    Verifies JSON shape rather than substring-matching raw ``data`` strings — a
+    regression in the envelope (missing id, wrong serializer, content key
+    renamed) fails here instead of silently passing. Each ``data`` payload is
+    a ``NotificationResponse`` dump with a top-level ``content.subject``.
+    """
+    subjects = []
+    for event in events:
+        payload = json.loads(event["data"])
+        subjects.append(payload["content"]["subject"])
+    return subjects
+
+
+class TestNotificationSSEIntegration(IntegrationTestCase):
+    dataset_populator: DatasetPopulator
+    framework_tool_and_types = False
+
+    @classmethod
+    def handle_galaxy_config_kwds(cls, config):
+        super().handle_galaxy_config_kwds(config)
+        config["enable_celery_tasks"] = False
+        config["enable_notification_system"] = True
+
+    def setUp(self):
+        super().setUp()
+        self.dataset_populator = DatasetPopulator(self.galaxy_interactor)
+
+    def _stream_url(self) -> str:
+        return urljoin(self.url, "api/events/stream")
+
+    def test_sse_receives_notification_events(self):
+        """When a notification is created, the SSE stream should receive it."""
+        user = self._setup_user(f"{uuid4()}@galaxy.test")
+        _, user_api_key = self._setup_user_get_key(user["email"])
+
+        listener = SSELineListener(self._stream_url(), user_api_key)
+        listener.start()
+        try:
+            subject = f"sse_test_{uuid4()}"
+            request = {
+                "recipients": {"user_ids": [user["id"]]},
+                "notification": notification_test_data(subject=subject, message="SSE test notification"),
+            }
+            response = self._post("notifications", data=request, admin=True, json=True)
+            self._assert_status_code_is_ok(response)
+
+            notification_events = listener.wait_for_event("notification_update")
+        finally:
+            listener.stop()
+
+        assert subject in _notification_subjects(
+            notification_events
+        ), f"Expected subject '{subject}' in SSE events, got: {notification_events}"
+
+    def test_sse_receives_broadcast_events(self):
+        """When a broadcast is created, the SSE stream should receive it."""
+        api_key = self.galaxy_interactor.api_key
+        assert api_key is not None
+        listener = SSELineListener(self._stream_url(), api_key)
+        listener.start()
+        try:
+            subject = f"broadcast_sse_test_{uuid4()}"
+            payload = notification_broadcast_test_data(subject=subject)
+            response = self._post("notifications/broadcast", data=payload, admin=True, json=True)
+            self._assert_status_code_is_ok(response)
+
+            broadcast_events = listener.wait_for_event("broadcast_update")
+        finally:
+            listener.stop()
+
+        # Broadcast events carry a BroadcastNotificationResponse, which shares
+        # the top-level content.subject shape with per-user notifications.
+        broadcast_subjects = [json.loads(e["data"])["content"]["subject"] for e in broadcast_events]
+        assert subject in broadcast_subjects, f"Expected subject '{subject}' in broadcast events: {broadcast_events}"
+
+    def test_sse_catchup_on_reconnect(self):
+        """Reconnecting with Last-Event-ID should replay a catch-up notification_status event.
+
+        The ``Last-Event-ID`` value is the server-issued ID from a prior event,
+        not a client-side timestamp. This avoids clock-skew flake between the
+        test runner and the app in containerized CI.
+        """
+        user = self._setup_user(f"{uuid4()}@galaxy.test")
+        _, user_api_key = self._setup_user_get_key(user["email"])
+
+        # First connection: capture the server-issued event id of the first notification.
+        listener_1 = SSELineListener(self._stream_url(), user_api_key)
+        listener_1.start()
+        try:
+            subject_1 = f"first_{uuid4()}"
+            request = {
+                "recipients": {"user_ids": [user["id"]]},
+                "notification": notification_test_data(subject=subject_1),
+            }
+            response = self._post("notifications", data=request, admin=True, json=True)
+            self._assert_status_code_is_ok(response)
+            first_events = listener_1.wait_for_event("notification_update")
+        finally:
+            listener_1.stop()
+
+        last_event_id = next((e["id"] for e in first_events if e.get("id")), None)
+        assert last_event_id, f"No server-issued id on first notification event: {first_events}"
+
+        # Emit a second notification while disconnected; it should appear in the catch-up.
+        subject_2 = f"catchup_{uuid4()}"
+        request = {
+            "recipients": {"user_ids": [user["id"]]},
+            "notification": notification_test_data(subject=subject_2, message="Catch-up test"),
+        }
+        response = self._post("notifications", data=request, admin=True, json=True)
+        self._assert_status_code_is_ok(response)
+
+        # Reconnect with Last-Event-ID = the captured id. The catch-up must include
+        # the notification sent after that id but not the one that produced it.
+        listener_2 = SSELineListener(
+            self._stream_url(),
+            user_api_key,
+            headers={"Last-Event-ID": last_event_id},
+        )
+        listener_2.start()
+        try:
+            status_events = listener_2.wait_for_event("notification_status")
+        finally:
+            listener_2.stop()
+
+        replayed_subjects: list[str] = []
+        for event in status_events:
+            payload = json.loads(event["data"])
+            replayed_subjects.extend(n["content"]["subject"] for n in payload.get("notifications", []))
+        assert subject_2 in replayed_subjects, f"Missed catch-up of '{subject_2}': {status_events}"
+        assert (
+            subject_1 not in replayed_subjects
+        ), f"Last-Event-ID did not filter — '{subject_1}' replayed: {status_events}"
diff --git a/test/integration_selenium/test_entry_point_sse.py b/test/integration_selenium/test_entry_point_sse.py
new file mode 100644
index 000000000000..0011a7d12dfc
--- /dev/null
+++ b/test/integration_selenium/test_entry_point_sse.py
@@ -0,0 +1,113 @@
+"""Playwright E2E test for the interactive-tool entry-point SSE pipeline.
+
+Verifies that when an interactive-tool entry point transitions to ``configured``
+server-side (the runner's port-routing hook calls ``configure_entry_points``),
+a logged-in user's browser receives the ``entry_point_update`` SSE event and
+the entry-point store refetches, without the 10 s polling interval.
+
+This test stubs the server-side runner callback: it creates the Job and entry
+point directly and calls ``InteractiveToolManager.configure_entry_points`` on
+the live app. That invocation is the exact dispatch site in production.
+"""
+
+from uuid import uuid4
+
+from galaxy.model import (
+    InteractiveToolEntryPoint,
+    Job,
+)
+from galaxy.util.wait import wait_on
+from galaxy_test.selenium.framework import (
+    managed_history,
+    selenium_test,
+)
+from .framework import SeleniumIntegrationTestCase
+
+SSE_CONNECT_TIMEOUT_SECONDS = 15
+SSE_EVENT_TIMEOUT_SECONDS = 15
+
+
+class TestEntryPointSSESeleniumIntegration(SeleniumIntegrationTestCase):
+    ensure_registered = True
+
+    @classmethod
+    def handle_galaxy_config_kwds(cls, config):
+        super().handle_galaxy_config_kwds(config)
+        config["enable_celery_tasks"] = False
+        # App.vue only calls entryPointStore.startWatchingEntryPoints() when
+        # interactivetools_enable is True, and the store only opens an SSE
+        # connection when enable_sse_updates is True. Without both,
+        # __galaxy_sse_connected never becomes true and the gate below times out.
+        config["interactivetools_enable"] = True
+        config["enable_sse_updates"] = True
+
+    def _wait_for_sse_connected(self) -> None:
+        """Block until the frontend confirms the SSE pipeline is live."""
+        wait_on(
+            lambda: True if self.execute_script("return window.__galaxy_sse_connected === true") else None,
+            "window.__galaxy_sse_connected === true",
+            timeout=SSE_CONNECT_TIMEOUT_SECONDS,
+        )
+
+    def _last_sse_event_ts(self) -> int:
+        return self.execute_script("return window.__galaxy_sse_last_event_ts || 0") or 0
+
+    def _wait_for_sse_event_after(self, baseline_ts: int) -> None:
+        wait_on(
+            lambda: True if self._last_sse_event_ts() > baseline_ts else None,
+            "window.__galaxy_sse_last_event_ts advanced past baseline",
+            timeout=SSE_EVENT_TIMEOUT_SECONDS,
+        )
+
+    def _create_it_job_with_entry_point(self, tool_port: int = 8888) -> tuple[int, int]:
+        # Use the browser's cookie-authenticated user, not the API interactor's
+        # default: SSE connects under the Selenium-registered user, and the
+        # dispatch's user_id must match or push_to_user finds no queues.
+        user_info = self.api_get("users/current")
+        user_id = self._decode_id(user_info["id"])
+        sa_session = self._app.model.context
+        job = Job()
+        job.user_id = user_id
+        job.tool_id = "interactivetool_simple"
+        job.state = Job.states.RUNNING
+        sa_session.add(job)
+        sa_session.flush()
+        ep = InteractiveToolEntryPoint(
+            job=job,
+            tool_port=tool_port,
+            entry_url="/",
+            name=f"selenium entry {uuid4()}",
+            label="selenium",
+            requires_domain=True,
+            requires_path_in_url=False,
+            requires_path_in_header_named=None,
+        )
+        sa_session.add(ep)
+        sa_session.commit()
+        return job.id, ep.id
+
+    @selenium_test
+    @managed_history
+    def test_entry_point_update_pushed_via_sse(self):
+        """configure_entry_points should trigger an SSE push the client observes."""
+        # Navigate home so the entry-point store is mounted and subscribed.
+        self.home()
+        self._wait_for_sse_connected()
+        baseline_ts = self._last_sse_event_ts()
+        self.screenshot("entry_point_sse_before")
+
+        job_id, _ep_id = self._create_it_job_with_entry_point(tool_port=8888)
+
+        # Stub the runner hook: call configure_entry_points on the live app.
+        sa_session = self._app.model.context
+        job = sa_session.get(Job, job_id)
+        assert job is not None
+        self._app.interactivetool_manager.configure_entry_points(
+            job,
+            {"8888": {"host": "host.invalid", "port": 12345, "protocol": "http"}},
+        )
+
+        # Prove the update arrived via SSE (not polling): the composable's
+        # event-timestamp hook only advances when useSSE's listener fires.
+        self._wait_for_sse_event_after(baseline_ts)
+        self.screenshot("entry_point_sse_after")
diff --git a/test/integration_selenium/test_notification_sse.py b/test/integration_selenium/test_notification_sse.py
new file mode 100644
index 000000000000..34123952ce09
--- /dev/null
+++ b/test/integration_selenium/test_notification_sse.py
@@ -0,0 +1,145 @@
+"""Playwright E2E test for the notification SSE pipeline.
+
+Verifies that when an admin creates a notification via the API,
+a logged-in user sees it appear in the UI in real-time (within seconds)
+without a page refresh, proving the SSE push pipeline works end-to-end.
+"""
+
+from uuid import uuid4
+
+from galaxy.util.wait import wait_on
+from galaxy_test.selenium.framework import (
+    managed_history,
+    selenium_test,
+)
+from .framework import SeleniumIntegrationTestCase
+
+SSE_CONNECT_TIMEOUT_SECONDS = 15
+SSE_EVENT_TIMEOUT_SECONDS = 15
+
+
+class TestNotificationSSESeleniumIntegration(SeleniumIntegrationTestCase):
+    ensure_registered = True
+
+    @classmethod
+    def handle_galaxy_config_kwds(cls, config):
+        super().handle_galaxy_config_kwds(config)
+        config["enable_notification_system"] = True
+        config["enable_sse_updates"] = True
+        config["enable_celery_tasks"] = False
+
+    def _wait_for_sse_connected(self) -> None:
+        """Block until the frontend confirms the SSE pipeline is live.
+
+        Without this gate, the 30 s polling fallback silently masks a broken
+        SSE pipeline — the UI would still update, but via polling, and the
+        test would falsely pass.
+        """
+        wait_on(
+            lambda: True if self.execute_script("return window.__galaxy_sse_connected === true") else None,
+            "window.__galaxy_sse_connected === true",
+            timeout=SSE_CONNECT_TIMEOUT_SECONDS,
+        )
+
+    def _last_sse_event_ts(self) -> int:
+        """Return the last SSE event timestamp recorded by the composable, or 0."""
+        return self.execute_script("return window.__galaxy_sse_last_event_ts || 0") or 0
+
+    def _wait_for_sse_event_after(self, baseline_ts: int) -> None:
+        """Block until an SSE event arrives after ``baseline_ts``.
+
+        Guards against a silent regression where the UI update originates from
+        the polling fallback rather than the SSE push: ``__galaxy_sse_last_event_ts``
+        only advances when the composable's event listener fires.
+        """
+        wait_on(
+            lambda: True if self._last_sse_event_ts() > baseline_ts else None,
+            "window.__galaxy_sse_last_event_ts advanced past baseline",
+            timeout=SSE_EVENT_TIMEOUT_SECONDS,
+        )
+
+    @selenium_test
+    @managed_history
+    def test_notification_appears_via_sse(self):
+        """Send a notification via the API and verify it appears in the UI without refresh."""
+        # Get the browser-logged-in user's info via the browser's cookie. ``self._get``
+        # uses the API interactor's default-user key, which does not match the Selenium
+        # user created by ``ensure_registered``, so the SSE push would target a
+        # different user than the one watching the stream.
+        user_info = self.api_get("users/current")
+        user_id = user_info["id"]
+
+        # Navigate to notifications page so the store is watching.
+        # ``get()`` uses ``build_url()`` which handles trailing slashes on the
+        # base correctly; concatenating against ``target_url_from_selenium``
+        # can produce a double-slash that Galaxy routes differently.
+        self.get("user/notifications")
+        self._wait_for_sse_connected()
+        baseline_ts = self._last_sse_event_ts()
+        self.screenshot("notification_sse_before")
+
+        # Send a notification to this user via the admin API
+        subject = f"SSE E2E Test {uuid4()}"
+        notification_request = {
+            "recipients": {"user_ids": [user_id]},
+            "notification": {
+                "source": "integration_tests",
+                "variant": "info",
+                "category": "message",
+                "content": {
+                    "category": "message",
+                    "subject": subject,
+                    "message": "This notification was pushed via SSE",
+                },
+            },
+        }
+        response = self._post("notifications", data=notification_request, admin=True, json=True)
+        self._assert_status_code_is_ok(response)
+
+        # Prove the incoming update arrived via SSE: the event-timestamp hook
+        # only advances when useSSE's listener fires. If this times out while
+        # the UI still shows the notification, polling picked it up — a silent
+        # regression this assertion catches.
+        self._wait_for_sse_event_after(baseline_ts)
+        # Use xpath (works on both Selenium and Playwright backends); `text=` is a
+        # Playwright engine prefix that Selenium's css-selector path rejects.
+        self.wait_for_xpath_visible(f'//*[contains(text(), "{subject}")]', timeout=SSE_EVENT_TIMEOUT_SECONDS)
+        self.screenshot("notification_sse_after")
+
+    @selenium_test
+    @managed_history
+    def test_notification_bell_updates_via_sse(self):
+        """The notification bell indicator should update when a new notification arrives via SSE."""
+        # See ``test_notification_appears_via_sse`` — must use the browser's user,
+        # not the API interactor's default user.
+        user_info = self.api_get("users/current")
+        user_id = user_info["id"]
+
+        # Go to home page (bell is in masthead)
+        self.home()
+        self._wait_for_sse_connected()
+        baseline_ts = self._last_sse_event_ts()
+
+        # Send a notification
+        subject = f"Bell Test {uuid4()}"
+        notification_request = {
+            "recipients": {"user_ids": [user_id]},
+            "notification": {
+                "source": "integration_tests",
+                "variant": "info",
+                "category": "message",
+                "content": {
+                    "category": "message",
+                    "subject": subject,
+                    "message": "Testing bell indicator update via SSE",
+                },
+            },
+        }
+        response = self._post("notifications", data=notification_request, admin=True, json=True)
+        self._assert_status_code_is_ok(response)
+
+        self._wait_for_sse_event_after(baseline_ts)
+        # The activity-bar notifications item renders its unread-count badge as
+        # ``.nav-indicator`` (see ``ActivityItem.vue``) once ``totalUnreadCount > 0``.
+        self.wait_for_selector_visible("#activity-notifications .nav-indicator", timeout=SSE_EVENT_TIMEOUT_SECONDS)
+        self.screenshot("notification_bell_indicator")
diff --git a/test/unit/app/managers/test_queue_metrics.py b/test/unit/app/managers/test_queue_metrics.py
new file mode 100644
index 000000000000..8bacd21d4b0c
--- /dev/null
+++ b/test/unit/app/managers/test_queue_metrics.py
@@ -0,0 +1,227 @@
+"""Unit tests for :mod:`galaxy.managers.queue_metrics`.
+
+The ``SSEConnectionManager`` and the kombu connection are small hand-built
+fakes — no broker or database is required. Assertions are on the recorded
+state of the statsd client (counters, timings) rather than on mock call-lists
+so a regression that stops emitting a gauge fails the test for the right
+reason.
+
+The failure-isolation test drives real sub-emitters into their error paths by
+handing them genuinely broken collaborators (a connection whose ``clone()``
+raises, a model whose ``new_session()`` raises). That way the test exercises
+the real ``_run`` wrapper rather than asserting a monkey-patched side_effect.
+"""
+
+from dataclasses import (
+    dataclass,
+    field,
+)
+from types import SimpleNamespace
+from typing import (
+    cast,
+    Optional,
+)
+from unittest.mock import MagicMock
+
+import pytest
+
+from galaxy.managers import queue_metrics
+from galaxy.managers.sse import SSEConnectionManager
+from galaxy.model.mapping import GalaxyModelMapping
+from galaxy.web.statsd_client import VanillaGalaxyStatsdClient
+from galaxy.web_stack import ApplicationStack
+
+# ---------------------------------------------------------------------------
+# Fakes
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class FakeStatsdClient:
+    """In-memory statsd recorder — see docstring in ``test_sse_dispatch.py``."""
+
+    counters: dict[tuple[str, tuple[tuple[str, str], ...]], int] = field(default_factory=dict)
+    timings: list[tuple[str, float, tuple[tuple[str, str], ...]]] = field(default_factory=list)
+
+    def incr(self, metric: str, tags: Optional[dict[str, str]] = None) -> None:
+        key = (metric, tuple(sorted((tags or {}).items())))
+        self.counters[key] = self.counters.get(key, 0) + 1
+
+    def timing(self, metric: str, value: float, tags: Optional[dict[str, str]] = None) -> None:
+        self.timings.append((metric, value, tuple(sorted((tags or {}).items()))))
+
+    def counter(self, metric: str, tags: Optional[dict[str, str]] = None) -> int:
+        return self.counters.get((metric, tuple(sorted((tags or {}).items()))), 0)
+
+    def timings_for(self, metric: str) -> list[tuple[float, dict[str, str]]]:
+        return [(v, dict(t)) for m, v, t in self.timings if m == metric]
+
+
+def _fake_sse_manager(broadcast: int = 3, per_user: int = 5):
+    m = MagicMock(spec=SSEConnectionManager)
+    m.total_broadcast_connections = broadcast
+    m.total_per_user_connections = per_user
+    return m
+
+
+def _make_fake_queue(name: str, count: int):
+    """Fake kombu Queue exposing ``.bind(channel).queue_declare(passive=True).message_count``."""
+    declared = SimpleNamespace(message_count=count)
+    bound = SimpleNamespace(queue_declare=lambda passive: declared)
+    return SimpleNamespace(name=name, bind=lambda channel: bound)
+
+
+def _make_fake_connection(channel=None):
+    """Fake kombu Connection whose ``.clone()`` is usable as a context manager."""
+    channel = channel or MagicMock()
+    conn_cm = MagicMock()
+    conn_cm.channel.return_value = channel
+    cm = MagicMock()
+    cm.__enter__ = lambda self: conn_cm
+    cm.__exit__ = lambda self, *a: False
+    connection = MagicMock()
+    connection.clone.return_value = cm
+    return connection
+
+
+# ---------------------------------------------------------------------------
+# Sub-emitter tests — assert on recorded state
+# ---------------------------------------------------------------------------
+
+
+def test_emit_sse_connection_gauges_emits_both_kinds():
+    statsd = FakeStatsdClient()
+    queue_metrics.emit_sse_connection_gauges(
+        cast(VanillaGalaxyStatsdClient, statsd), _fake_sse_manager(broadcast=4, per_user=7)
+    )
+
+    assert statsd.timings_for("galaxy.sse.connections.active") == [
+        (4, {"kind": "broadcast"}),
+        (7, {"kind": "per_user"}),
+    ]
+
+
+def test_emit_control_queue_depth_emits_per_queue(monkeypatch):
+    statsd = FakeStatsdClient()
+    fake_queues = [
+        _make_fake_queue("control.main@h", 3),
+        _make_fake_queue("control.main.1@h", 0),
+    ]
+    monkeypatch.setattr(
+        queue_metrics,
+        "all_control_queues_for_declare",
+        lambda application_stack: fake_queues,
+    )
+
+    queue_metrics.emit_control_queue_depth(
+        cast(VanillaGalaxyStatsdClient, statsd),
+        _make_fake_connection(),
+        cast(ApplicationStack, MagicMock()),
+    )
+
+    assert statsd.timings_for("galaxy.control_queue.depth") == [
+        (3, {"queue_name": "control.main@h"}),
+        (0, {"queue_name": "control.main.1@h"}),
+    ]
+
+
+def test_emit_control_queue_depth_skips_failed_passive_declare(monkeypatch):
+    """One bad queue → we skip it and keep going for the rest."""
+    statsd = FakeStatsdClient()
+
+    def bad_declare(passive):
+        raise RuntimeError("queue does not exist yet")
+
+    good_queue = _make_fake_queue("control.good@h", 9)
+    bad_queue = SimpleNamespace(
+        name="control.bad@h",
+        bind=lambda channel: SimpleNamespace(queue_declare=bad_declare),
+    )
+    monkeypatch.setattr(queue_metrics, "all_control_queues_for_declare", lambda stack: [good_queue, bad_queue])
+
+    queue_metrics.emit_control_queue_depth(
+        cast(VanillaGalaxyStatsdClient, statsd),
+        _make_fake_connection(),
+        cast(ApplicationStack, MagicMock()),
+    )
+
+    assert statsd.timings_for("galaxy.control_queue.depth") == [
+        (9, {"queue_name": "control.good@h"}),
+    ]
+
+
+def test_emit_control_queue_depth_no_broker_connection_is_noop():
+    statsd = FakeStatsdClient()
+    queue_metrics.emit_control_queue_depth(
+        cast(VanillaGalaxyStatsdClient, statsd), None, cast(ApplicationStack, MagicMock())
+    )
+    assert statsd.timings == []
+
+
+# ---------------------------------------------------------------------------
+# emit_queue_metrics — aggregate entry-point
+# ---------------------------------------------------------------------------
+
+
+def test_emit_queue_metrics_is_silent_when_statsd_is_none():
+    """No statsd client → every sub-call is skipped, no DB or broker access."""
+    # Would raise AttributeError if the short-circuit didn't fire before any
+    # real collaborator was touched.
+    queue_metrics.emit_queue_metrics(
+        statsd_client=None,
+        connection=None,
+        application_stack=cast(ApplicationStack, MagicMock()),
+        model=cast(GalaxyModelMapping, MagicMock()),
+        sse_manager=None,
+    )
+
+
+def test_emit_queue_metrics_isolates_real_subemitter_failures(monkeypatch):
+    """When real sub-emitters raise, ``_run`` contains the failure and logs an error counter.
+
+    We drive the failures through the actual sub-emitter bodies — not monkey-
+    patched side_effects — by handing in a connection whose ``.clone()`` raises
+    (for ``emit_control_queue_depth``) and a model whose ``.new_session()``
+    raises (for ``emit_worker_process_gauge``). The SSE gauge is given healthy
+    collaborators and should still land.
+    """
+    statsd = FakeStatsdClient()
+    sse_manager = _fake_sse_manager(broadcast=2, per_user=1)
+
+    broken_connection = MagicMock()
+    broken_connection.clone.side_effect = RuntimeError("broker is gone")
+
+    broken_model = MagicMock()
+    broken_model.new_session.side_effect = RuntimeError("db is gone")
+
+    # Ensure the broker path reaches .clone() rather than short-circuiting on
+    # an empty queue list.
+    monkeypatch.setattr(
+        queue_metrics,
+        "all_control_queues_for_declare",
+        lambda stack: [_make_fake_queue("control.main@h", 0)],
+    )
+
+    # Must not raise — the SSE gauge still lands.
+    queue_metrics.emit_queue_metrics(
+        statsd_client=cast(VanillaGalaxyStatsdClient, statsd),
+        connection=broken_connection,
+        application_stack=cast(ApplicationStack, MagicMock()),
+        model=cast(GalaxyModelMapping, broken_model),
+        sse_manager=sse_manager,
+    )
+
+    # SSE gauge landed despite the other two failing.
+    sse_timings = statsd.timings_for("galaxy.sse.connections.active")
+    assert (2, {"kind": "broadcast"}) in sse_timings
+    assert (1, {"kind": "per_user"}) in sse_timings
+
+    # Each failing sub-emitter bumped its error counter tagged by name.
+    assert statsd.counter("galaxy.queue_metrics.error", {"emitter": "control_queue_depth"}) == 1
+    assert statsd.counter("galaxy.queue_metrics.error", {"emitter": "worker_process"}) == 1
+    # The healthy SSE sub-emitter did not.
+    assert statsd.counter("galaxy.queue_metrics.error", {"emitter": "sse_connections"}) == 0
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])
diff --git a/test/unit/app/managers/test_sse_dispatch.py b/test/unit/app/managers/test_sse_dispatch.py
new file mode 100644
index 000000000000..2118ccacdc0c
--- /dev/null
+++ b/test/unit/app/managers/test_sse_dispatch.py
@@ -0,0 +1,236 @@
+"""Unit tests for :class:`galaxy.managers.sse_dispatch.SSEEventDispatcher` observability.
+
+Focus is on the statsd instrumentation contract AND the effect of dispatch:
+counters/timers fire on the happy path and the ``_queue_worker is None``
+early-return, payloads reach the broker with the expected task+kwargs, and the
+dispatcher is a silent no-op when ``statsd_client`` is ``None``.
+
+These tests use lightweight fakes (``FakeStatsdClient``, ``FakeControlTask``)
+that record state we can assert against — rather than ``MagicMock`` call-lists —
+so a regression that silently drops dispatch or stops recording metrics fails
+the test for the right reason.
+"""
+
+from dataclasses import (
+    dataclass,
+    field,
+)
+from typing import (
+    Any,
+    Optional,
+)
+from unittest.mock import MagicMock
+
+import pytest
+
+from galaxy.managers.sse_dispatch import SSEEventDispatcher
+
+
+@dataclass
+class FakeStatsdClient:
+    """In-memory stand-in for ``VanillaGalaxyStatsdClient``.
+
+    Records ``incr`` and ``timing`` calls as plain data so tests assert on
+    observable state (counter totals, recorded timings) instead of mock
+    call-lists.
+    """
+
+    counters: dict[tuple[str, tuple[tuple[str, str], ...]], int] = field(default_factory=dict)
+    timings: list[tuple[str, float, tuple[tuple[str, str], ...]]] = field(default_factory=list)
+
+    def incr(self, metric: str, tags: Optional[dict[str, str]] = None) -> None:
+        key = (metric, tuple(sorted((tags or {}).items())))
+        self.counters[key] = self.counters.get(key, 0) + 1
+
+    def timing(self, metric: str, value: float, tags: Optional[dict[str, str]] = None) -> None:
+        self.timings.append((metric, value, tuple(sorted((tags or {}).items()))))
+
+    def counter(self, metric: str, tags: Optional[dict[str, str]] = None) -> int:
+        return self.counters.get((metric, tuple(sorted((tags or {}).items()))), 0)
+
+
+@dataclass
+class RecordedTask:
+    payload: dict[str, Any]
+    routing_key: str
+    expiration: Optional[int]
+    declare_queues: Any
+
+
+class FakeControlTask:
+    """Stand-in for ``ControlTask`` that records dispatches instead of touching AMQP."""
+
+    instances: list["FakeControlTask"] = []
+
+    def __init__(self, queue_worker) -> None:
+        self.queue_worker = queue_worker
+        self.sent: list[RecordedTask] = []
+        FakeControlTask.instances.append(self)
+
+    def send_task(
+        self,
+        payload: dict[str, Any],
+        routing_key: str,
+        expiration: Optional[int] = None,
+        declare_queues: Any = None,
+        **_: Any,
+    ) -> None:
+        self.sent.append(
+            RecordedTask(
+                payload=payload,
+                routing_key=routing_key,
+                expiration=expiration,
+                declare_queues=declare_queues,
+            )
+        )
+
+
+class BoomControlTask:
+    """``ControlTask`` fake whose ``send_task`` always raises — exercises the finally block."""
+
+    def __init__(self, queue_worker) -> None:
+        self.queue_worker = queue_worker
+
+    def send_task(self, **kwargs) -> None:
+        raise RuntimeError("broker down")
+
+
+@pytest.fixture
+def application_stack():
+    return MagicMock(name="ApplicationStack")
+
+
+@pytest.fixture
+def queue_worker():
+    return MagicMock(name="GalaxyQueueWorker")
+
+
+@pytest.fixture
+def statsd() -> FakeStatsdClient:
+    return FakeStatsdClient()
+
+
+@pytest.fixture(autouse=True)
+def _reset_fake_control_task_instances():
+    FakeControlTask.instances.clear()
+    yield
+    FakeControlTask.instances.clear()
+
+
+@pytest.fixture
+def fake_control_task() -> type[FakeControlTask]:
+    """Returns the fake class so tests can pass it as ``control_task_factory``."""
+    return FakeControlTask
+
+
+def _dispatcher_with_fakes(
+    queue_worker,
+    application_stack,
+    statsd,
+    control_task_factory=FakeControlTask,
+    queues=None,
+) -> SSEEventDispatcher:
+    """Build a dispatcher wired to injected fakes — no monkeypatching required."""
+    return SSEEventDispatcher(
+        queue_worker=queue_worker,
+        application_stack=application_stack,
+        statsd_client=statsd,
+        control_task_factory=control_task_factory,
+        queues_provider=lambda: queues if queues is not None else [],
+    )
+
+
+def test_dispatcher_no_op_when_queue_worker_is_none_and_no_statsd(application_stack):
+    """No statsd client set and no queue_worker → silent no-op, no AttributeError."""
+    dispatcher = SSEEventDispatcher(
+        queue_worker=None,
+        application_stack=application_stack,
+        statsd_client=None,
+    )
+    # Must not raise, must not attempt to declare queues.
+    dispatcher.notify_users([1, 2], "hello")
+    dispatcher.notify_broadcast("hi")
+    dispatcher.history_update({"1": [42]})
+
+
+def test_dispatcher_records_skipped_counter_when_queue_worker_is_none(application_stack, statsd):
+    """Two dispatches with no queue_worker → two skipped_no_qw increments, no timings."""
+    dispatcher = SSEEventDispatcher(
+        queue_worker=None,
+        application_stack=application_stack,
+        statsd_client=statsd,
+    )
+    dispatcher.notify_users([1], "hello")
+    dispatcher.notify_broadcast("world")
+
+    assert statsd.counter("galaxy.sse.dispatch.skipped_no_qw") == 2
+    # No latency timing — we never got as far as the broker call.
+    assert statsd.timings == []
+
+
+def test_dispatcher_enqueues_payload_and_records_metrics_on_send(
+    application_stack, queue_worker, statsd, fake_control_task
+):
+    """Happy path: payload reaches the broker AND counter+timer are recorded."""
+    dispatcher = _dispatcher_with_fakes(
+        queue_worker=queue_worker,
+        application_stack=application_stack,
+        statsd=statsd,
+        control_task_factory=fake_control_task,
+    )
+    dispatcher.notify_users([1, 2], "hello")
+
+    # Exactly one ControlTask constructed, one send_task recorded with the right
+    # payload — asserts the dispatch *effect*, not just that a mock was called.
+    assert len(fake_control_task.instances) == 1
+    sent = fake_control_task.instances[0].sent
+    assert len(sent) == 1
+    assert sent[0].payload["task"] == "notify_users"
+    assert sent[0].payload["kwargs"]["user_ids"] == [1, 2]
+    assert sent[0].payload["kwargs"]["payload"] == "hello"
+    assert "event_id" in sent[0].payload["kwargs"]
+    assert sent[0].routing_key == "control.*"
+    assert sent[0].expiration == 10
+
+    # Counter + timer both recorded with matching task tag.
+    assert statsd.counter("galaxy.sse.dispatch.count", {"task": "notify_users"}) == 1
+    assert len(statsd.timings) == 1
+    metric, _value, tags = statsd.timings[0]
+    assert metric == "galaxy.sse.dispatch.latency_ms"
+    assert dict(tags) == {"task": "notify_users"}
+
+
+def test_dispatcher_timer_still_fires_on_send_exception(application_stack, queue_worker, statsd):
+    """Timer lives in ``finally`` — broker errors don't mask the latency metric."""
+    dispatcher = _dispatcher_with_fakes(
+        queue_worker=queue_worker,
+        application_stack=application_stack,
+        statsd=statsd,
+        control_task_factory=BoomControlTask,
+    )
+    with pytest.raises(RuntimeError):
+        dispatcher.history_update({"7": [1]})
+
+    assert statsd.counter("galaxy.sse.dispatch.count", {"task": "history_update"}) == 1
+    assert len(statsd.timings) == 1
+    metric, _value, tags = statsd.timings[0]
+    assert metric == "galaxy.sse.dispatch.latency_ms"
+    assert dict(tags) == {"task": "history_update"}
+
+
+def test_dispatcher_no_statsd_means_no_instrumentation(application_stack, queue_worker, fake_control_task):
+    """When ``statsd_client`` is ``None`` instrumentation is bypassed entirely.
+
+    The dispatch still happens — we assert via the ControlTask fake — but there
+    is nothing to observe on the (absent) statsd side.
+    """
+    dispatcher = _dispatcher_with_fakes(
+        queue_worker=queue_worker,
+        application_stack=application_stack,
+        statsd=None,
+        control_task_factory=fake_control_task,
+    )
+    dispatcher.notify_broadcast("hi")
+    assert len(fake_control_task.instances) == 1
+    assert len(fake_control_task.instances[0].sent) == 1
+    assert fake_control_task.instances[0].sent[0].payload["task"] == "notify_broadcast"
diff --git a/test/unit/app/managers/test_sse_dispatch_cache.py b/test/unit/app/managers/test_sse_dispatch_cache.py
new file mode 100644
index 000000000000..3561610b36e9
--- /dev/null
+++ b/test/unit/app/managers/test_sse_dispatch_cache.py
@@ -0,0 +1,144 @@
+"""Tests for the TTL cache on ``SSEEventDispatcher._get_declare_queues``.
+
+The cache exists because ``_send`` is on a hot path (1000+ events/s at target
+load) and without it each dispatch fires a ``WorkerProcess`` DB query. The
+underlying data only changes on a 60 s heartbeat cadence, so a 30 s TTL is safe.
+"""
+
+from concurrent.futures import (
+    as_completed,
+    ThreadPoolExecutor,
+)
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from galaxy.managers.sse_dispatch import SSEEventDispatcher
+
+
+@pytest.fixture
+def fake_declare():
+    """Call-counting provider passed to ``SSEEventDispatcher`` via DI.
+
+    Returns a non-empty list by default so the cache stores a value — empty
+    results are intentionally not cached (see ``test_empty_result_not_cached``).
+    """
+    calls: dict[str, Any] = {"count": 0, "returns": [MagicMock(name="queue")]}
+
+    def _provider():
+        calls["count"] += 1
+        return calls["returns"]
+
+    calls["provider"] = _provider
+    return calls
+
+
+class FakeClock:
+    """Controllable time source passed to ``SSEEventDispatcher``.
+
+    Lets tests advance the dispatcher's TTL cache deterministically without
+    reaching into ``cachetools`` internals.
+    """
+
+    def __init__(self, start: float = 0.0) -> None:
+        self.now = start
+
+    def __call__(self) -> float:
+        return self.now
+
+    def advance(self, seconds: float) -> None:
+        self.now += seconds
+
+
+@pytest.fixture
+def clock() -> FakeClock:
+    return FakeClock()
+
+
+@pytest.fixture
+def dispatcher(clock, fake_declare):
+    """Build a dispatcher with a stub queue_worker, a counting queues_provider,
+    and a no-op ``ControlTask`` factory — no monkey-patching required.
+    """
+    queue_worker = MagicMock(name="queue_worker")
+    application_stack = MagicMock(name="application_stack")
+
+    class NoopControlTask:
+        def __init__(self, qw):
+            self.qw = qw
+
+        def send_task(self, **kwargs):
+            pass
+
+    return SSEEventDispatcher(
+        queue_worker=queue_worker,
+        application_stack=application_stack,
+        clock=clock,
+        control_task_factory=NoopControlTask,
+        queues_provider=fake_declare["provider"],
+    )
+
+
+def test_declare_queues_cached_within_ttl(dispatcher, fake_declare):
+    """Repeated dispatches inside the TTL window only hit the DB once."""
+    for _ in range(10):
+        dispatcher.notify_broadcast("payload")
+    assert fake_declare["count"] == 1
+
+
+def test_declare_queues_refetched_after_ttl(dispatcher, fake_declare, clock):
+    """Once the TTL expires, the next call refetches exactly once."""
+    dispatcher.notify_broadcast("payload")
+    assert fake_declare["count"] == 1
+
+    # Advance the injected clock past the TTL so the cache sees the entry as
+    # expired on the next read.
+    clock.advance(dispatcher._DECLARE_QUEUES_TTL_SECONDS + 1)
+
+    dispatcher.notify_broadcast("payload")
+    assert fake_declare["count"] == 2
+
+    # Further dispatches at the advanced time reuse the newly populated entry.
+    dispatcher.notify_broadcast("payload")
+    assert fake_declare["count"] == 2
+
+
+def test_empty_result_not_cached(dispatcher, fake_declare):
+    """An empty list must not be pinned in the cache for the full TTL.
+
+    Empty results arise during startup (before ``DatabaseHeartbeat`` writes the
+    row) and on swallowed DB errors. Caching them would silently drop every SSE
+    event until the next TTL expiry.
+    """
+    fake_declare["returns"] = []
+    dispatcher.notify_broadcast("payload")
+    dispatcher.notify_broadcast("payload")
+    assert fake_declare["count"] == 2
+
+    # Once the upstream starts returning a non-empty result, caching resumes.
+    fake_declare["returns"] = [MagicMock(name="queue")]
+    dispatcher.notify_broadcast("payload")
+    dispatcher.notify_broadcast("payload")
+    assert fake_declare["count"] == 3
+
+
+def test_declare_queues_thread_safe_single_query_under_load(dispatcher, fake_declare):
+    """Concurrent ``_send`` from many threads still only triggers one DB query.
+
+    With stampede protection (RLock around the miss) all 500 dispatches should
+    collapse to a single ``all_control_queues_for_declare`` call inside one TTL
+    window. The assertion is exact (== 1), not loose, because the lock
+    serializes the miss.
+    """
+    iterations = 500
+
+    def work():
+        dispatcher.notify_broadcast("payload")
+
+    with ThreadPoolExecutor(max_workers=16) as pool:
+        futures = [pool.submit(work) for _ in range(iterations)]
+        for future in as_completed(futures):
+            future.result()
+
+    assert fake_declare["count"] == 1
diff --git a/test/unit/app/queue_worker/test_queue_worker.py b/test/unit/app/queue_worker/test_queue_worker.py
index 09bc8a32fa54..857bdeb2cac2 100644
--- a/test/unit/app/queue_worker/test_queue_worker.py
+++ b/test/unit/app/queue_worker/test_queue_worker.py
@@ -1,6 +1,13 @@
 import datetime
 import time
+from dataclasses import (
+    dataclass,
+    field,
+)
 from math import inf
+from types import SimpleNamespace
+from typing import Optional
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -14,6 +21,27 @@
 from galaxy.web_stack import application_stack_instance
 
 
+@dataclass
+class FakeStatsdClient:
+    """In-memory statsd recorder — see docstring in ``test_sse_dispatch.py``."""
+
+    counters: dict[tuple[str, tuple[tuple[str, str], ...]], int] = field(default_factory=dict)
+    timings: list[tuple[str, float, tuple[tuple[str, str], ...]]] = field(default_factory=list)
+
+    def incr(self, metric: str, tags: Optional[dict[str, str]] = None) -> None:
+        key = (metric, tuple(sorted((tags or {}).items())))
+        self.counters[key] = self.counters.get(key, 0) + 1
+
+    def timing(self, metric: str, value: float, tags: Optional[dict[str, str]] = None) -> None:
+        self.timings.append((metric, value, tuple(sorted((tags or {}).items()))))
+
+    def counter(self, metric: str, tags: Optional[dict[str, str]] = None) -> int:
+        return self.counters.get((metric, tuple(sorted((tags or {}).items()))), 0)
+
+    def timings_for(self, metric: str) -> list[tuple[float, dict[str, str]]]:
+        return [(v, dict(t)) for m, v, t in self.timings if m == metric]
+
+
 def bar(app, **kwargs):
     app.some_var = "bar"
     app.tasks_executed.append("echo")
@@ -119,3 +147,86 @@ def wait_for_var(obj, var, value, tries=10, sleep=0.25):
         tries -= 1
         time.sleep(sleep)
     assert getattr(obj, var) == value
+
+
+# ---------------------------------------------------------------------------
+# process_task observability tests
+#
+# These don't need a real broker or DB — we just instantiate GalaxyQueueWorker
+# via ``__new__`` and drive ``process_task`` directly with fake ``body`` /
+# ``message`` objects.  The statsd client is pulled through
+# ``app.execution_timer_factory.galaxy_statsd_client``.
+# ---------------------------------------------------------------------------
+
+
+def _make_fake_worker(task_fn, statsd_client):
+    worker = GalaxyQueueWorker.__new__(GalaxyQueueWorker)
+    worker.app = SimpleNamespace(
+        config=SimpleNamespace(server_name="test.server"),
+        execution_timer_factory=SimpleNamespace(galaxy_statsd_client=statsd_client),
+    )
+    worker.task_mapping = {"echo": task_fn}
+    worker.epoch = 0
+    # ``producer`` is a read-only property on the mixin; we avoid the publisher
+    # path entirely by leaving ``reply_to`` out of the fake message properties.
+    return worker
+
+
+def _fake_message():
+    message = MagicMock()
+    message.headers = {"epoch": inf}  # always greater than worker.epoch
+    message.properties = {}  # no reply_to — skip publisher path
+    return message
+
+
+def test_process_task_emits_counter_and_ok_timer():
+    statsd = FakeStatsdClient()
+    calls: list[dict] = []
+
+    def handler(app, **kwargs):
+        calls.append(kwargs)
+        return "done"
+
+    worker = _make_fake_worker(handler, statsd)
+    worker.process_task({"task": "echo", "kwargs": {"x": 1}}, _fake_message())
+
+    # Handler actually ran — if the worker silently dropped the task, this list
+    # would be empty and the test would fail loudly.
+    assert calls == [{"x": 1}]
+    assert statsd.counter("galaxy.control_queue.task.count", {"task": "echo"}) == 1
+    assert len(statsd.timings_for("galaxy.control_queue.task.latency_ms")) == 1
+    _value, tags = statsd.timings_for("galaxy.control_queue.task.latency_ms")[0]
+    assert tags == {"task": "echo", "outcome": "ok"}
+
+
+def test_process_task_emits_error_timer_on_handler_exception():
+    statsd = FakeStatsdClient()
+    invocations: list[bool] = []
+
+    def boom(app, **kwargs):
+        invocations.append(True)
+        raise RuntimeError("handler failed")
+
+    worker = _make_fake_worker(boom, statsd)
+    # process_task swallows handler exceptions (logged, not raised).
+    worker.process_task({"task": "echo", "kwargs": {}}, _fake_message())
+
+    # Handler was actually invoked before raising.
+    assert invocations == [True]
+    assert statsd.counter("galaxy.control_queue.task.count", {"task": "echo"}) == 1
+    assert len(statsd.timings_for("galaxy.control_queue.task.latency_ms")) == 1
+    _value, tags = statsd.timings_for("galaxy.control_queue.task.latency_ms")[0]
+    assert tags == {"task": "echo", "outcome": "error"}
+
+
+def test_process_task_no_statsd_is_silent_no_op():
+    calls: list[dict] = []
+
+    def handler(app, **kwargs):
+        calls.append(kwargs)
+        return "done"
+
+    worker = _make_fake_worker(handler, statsd_client=None)
+    worker.process_task({"task": "echo", "kwargs": {"x": 2}}, _fake_message())
+    # Handler ran — the test guards both "no exception" AND "task not silently dropped".
+    assert calls == [{"x": 2}]
diff --git a/test/unit/authnz/test_psa_authnz.py b/test/unit/authnz/test_psa_authnz.py
index 6bd2429d0ebb..994d4aea12e1 100644
--- a/test/unit/authnz/test_psa_authnz.py
+++ b/test/unit/authnz/test_psa_authnz.py
@@ -9,10 +9,7 @@
 )
 from types import SimpleNamespace
 from typing import Optional
-from unittest.mock import (
-    MagicMock,
-    patch,
-)
+from unittest.mock import MagicMock
 
 import jwt
 import pytest
@@ -400,15 +397,16 @@ def test_oidc_config_custom_auth_pipeline_and_extra(mock_oidc_config_file, mock_
 def test_sync_user_profile_skips_when_account_interface_enabled():
     manager = MagicMock()
     session = MagicMock()
+    notify = MagicMock()
     app_config = SimpleNamespace(enable_account_interface=True, enable_notification_system=True)
-    app = SimpleNamespace(config=app_config, user_manager=manager, notification_manager=SimpleNamespace())
+    notification_manager = SimpleNamespace(send_notification_internal=notify)
+    app = SimpleNamespace(config=app_config, user_manager=manager, notification_manager=notification_manager)
     trans = SimpleNamespace(app=app, sa_session=session)
     strategy = SimpleNamespace(config={"GALAXY_TRANS": trans, "FIXED_DELEGATED_AUTH": True})
     user = SimpleNamespace(id=1, preferences={})
     details = {"email": "new@example.com", "username": "newname"}
 
-    with patch("galaxy.webapps.galaxy.services.notifications.NotificationService.send_notification_internal") as notify:
-        sync_user_profile(strategy=strategy, details=details, user=user)
+    sync_user_profile(strategy=strategy, details=details, user=user)
 
     manager.update_email.assert_not_called()
     manager.update_username.assert_not_called()
@@ -419,15 +417,16 @@ def test_sync_user_profile_skips_when_account_interface_enabled():
 def test_sync_user_profile_skips_when_fixed_delegated_auth_disabled():
     manager = MagicMock()
     session = MagicMock()
+    notify = MagicMock()
     app_config = SimpleNamespace(enable_account_interface=False, enable_notification_system=True)
-    app = SimpleNamespace(config=app_config, user_manager=manager, notification_manager=SimpleNamespace())
+    notification_manager = SimpleNamespace(send_notification_internal=notify)
+    app = SimpleNamespace(config=app_config, user_manager=manager, notification_manager=notification_manager)
     trans = SimpleNamespace(app=app, sa_session=session)
     strategy = SimpleNamespace(config={"GALAXY_TRANS": trans, "FIXED_DELEGATED_AUTH": False})
     user = SimpleNamespace(id=2, email="old@example.com", username="oldname", preferences={})
     details = {"email": "new@example.com", "username": "newname"}
 
-    with patch("galaxy.webapps.galaxy.services.notifications.NotificationService.send_notification_internal") as notify:
-        sync_user_profile(strategy=strategy, details=details, user=user)
+    sync_user_profile(strategy=strategy, details=details, user=user)
 
     manager.update_email.assert_not_called()
     manager.update_username.assert_not_called()
@@ -438,16 +437,16 @@ def test_sync_user_profile_skips_when_fixed_delegated_auth_disabled():
 def test_sync_user_profile_updates_when_account_interface_disabled():
     manager = MagicMock()
     session = MagicMock()
+    notify = MagicMock()
     app_config = SimpleNamespace(enable_account_interface=False, enable_notification_system=True)
-    notification_manager = SimpleNamespace(notifications_enabled=True)
+    notification_manager = SimpleNamespace(notifications_enabled=True, send_notification_internal=notify)
     app = SimpleNamespace(config=app_config, user_manager=manager, notification_manager=notification_manager)
     trans = SimpleNamespace(app=app, sa_session=session)
     strategy = SimpleNamespace(config={"GALAXY_TRANS": trans, "FIXED_DELEGATED_AUTH": True})
     user = SimpleNamespace(id=2, email="old@example.com", username="oldname", preferences={})
     details = {"email": "new@example.com", "username": "newname"}
 
-    with patch("galaxy.webapps.galaxy.services.notifications.NotificationService.send_notification_internal") as notify:
-        sync_user_profile(strategy=strategy, details=details, user=user)
+    sync_user_profile(strategy=strategy, details=details, user=user)
 
     manager.update_email.assert_called_once_with(
         trans, user, "new@example.com", commit=False, send_activation_email=False